fix(smoke): report sanitized operation failure diagnostics

.env.example

@@ -25,6 +25,12 @@ CRANK_OUTBOUND_ALLOWED_HOSTS=
 CRANK_OUTBOUND_DENIED_HOSTS=
 CRANK_OUTBOUND_MAX_REQUEST_BYTES=4194304
 CRANK_OUTBOUND_MAX_RESPONSE_BYTES=4194304
+CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES=
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH=8
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS=32
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES=262144
+CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS=10000
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES=10000
 CRANK_ENVIRONMENT=development
 CRANK_LOG_LEVEL=
 CRANK_SENTRY_DSN=

.gitea/workflows/ci.yml

@@ -232,7 +232,7 @@ jobs:
             --required-test 'operations page imports OpenAPI methods as drafts' \
             --required-test 'OpenAPI upload rejects invalid files locally and restores focus after Escape' \
             --required-test 'OpenAPI upload recovers from pagehide and a preview server error' \
-            --required-test 'OpenAPI upload invalidates active draft creation after language or workspace changes' \
+            --required-test 'OpenAPI apply preserves job authority after language or workspace changes' \
             --required-test 'OpenAPI upload only renders the latest selected file and clears reset or close races' \
             --required-test 'OpenAPI upload ignores a stale failure and renders only correlation identifiers'
           python3 ../../scripts/validate-capability-run.py \
@@ -286,17 +286,44 @@ jobs:
           CRANK_SESSION_SECRET=ci-session-secret
           CRANK_PASSWORD_PEPPER=ci-password-pepper
           CRANK_BOOTSTRAP_ADMIN_EMAIL=owner@crank.test
-          CRANK_BOOTSTRAP_ADMIN_PASSWORD=ci-admin-password
           CRANK_BOOTSTRAP_ADMIN_DISPLAY_NAME=CI Owner
           CRANK_BASE_URL=http://127.0.0.1
+          CRANK_ENVIRONMENT=ci
+          CRANK_OUTBOUND_ALLOWED_HOSTS=admin-api
           CRANK_PUBLISH_BIND=127.0.0.1
           CRANK_ADMIN_PUBLISH_PORT=0
           CRANK_MCP_PUBLISH_PORT=0
           CRANK_UI_PUBLISH_PORT=0
           CRANK_DEMO_SEED=true
           EOF
-          docker compose -f deploy/community/docker-compose.images.yml \
+          compose=(docker compose -f deploy/community/docker-compose.images.yml \
-            --env-file .tmp/community-smoke.env --profile local-db up -d --wait
+            --env-file .tmp/community-smoke.env --profile local-db)
+          "${compose[@]}" up -d --wait postgres
+          "${compose[@]}" run --rm migrate
+          bootstrap_json="$("${compose[@]}" run --rm --no-deps \
+            --entrypoint crank-migrate migrate admin-auth bootstrap-create \
+            --email owner@crank.test --display-name 'CI Owner')"
+          bootstrap_token="$(python3 -c \
+            'import json,sys; print(json.loads(sys.stdin.read())["bootstrap_token"])' \
+            <<<"$bootstrap_json")"
+          install -d -m 700 .tmp/community-bootstrap
+          printf '%s' "$bootstrap_token" > .tmp/community-bootstrap/token
+          printf '%s' 'ci-admin-password' > .tmp/community-bootstrap/password
+          printf '%s' 'ci-password-pepper' > .tmp/community-bootstrap/password-pepper
+          chmod 600 .tmp/community-bootstrap/token \
+            .tmp/community-bootstrap/password \
+            .tmp/community-bootstrap/password-pepper
+          "${compose[@]}" run --rm --no-deps \
+            -v "$PWD/.tmp/community-bootstrap:/run/bootstrap:ro" \
+            --entrypoint crank-migrate migrate admin-auth bootstrap-complete \
+            --token-file /run/bootstrap/token \
+            --password-file /run/bootstrap/password \
+            --password-pepper-file /run/bootstrap/password-pepper
+          rm -f .tmp/community-bootstrap/token \
+            .tmp/community-bootstrap/password \
+            .tmp/community-bootstrap/password-pepper
+          rmdir .tmp/community-bootstrap
+          "${compose[@]}" up -d --wait
 
       - name: Run authenticated Community image smoke
         env:

.gitea/workflows/release.yml

@@ -139,7 +139,7 @@ jobs:
             --required-test 'operations page imports OpenAPI methods as drafts' \
             --required-test 'OpenAPI upload rejects invalid files locally and restores focus after Escape' \
             --required-test 'OpenAPI upload recovers from pagehide and a preview server error' \
-            --required-test 'OpenAPI upload invalidates active draft creation after language or workspace changes' \
+            --required-test 'OpenAPI apply preserves job authority after language or workspace changes' \
             --required-test 'OpenAPI upload only renders the latest selected file and clears reset or close races' \
             --required-test 'OpenAPI upload ignores a stale failure and renders only correlation identifiers'
           python3 ../../scripts/validate-capability-run.py \
@@ -217,8 +217,8 @@ jobs:
           CRANK_SESSION_SECRET=release-smoke-session
           CRANK_PASSWORD_PEPPER=release-smoke-pepper
           CRANK_BOOTSTRAP_ADMIN_EMAIL=owner@crank.test
-          CRANK_BOOTSTRAP_ADMIN_PASSWORD=release-smoke-password
           CRANK_BASE_URL=http://127.0.0.1
+          CRANK_ENVIRONMENT=release-smoke
           CRANK_PUBLISH_BIND=127.0.0.1
           CRANK_ADMIN_PUBLISH_PORT=0
           CRANK_MCP_PUBLISH_PORT=0

apps/admin-api/Dockerfile

@@ -16,17 +16,17 @@ RUN --mount=type=cache,id=crank-admin-cargo-registry,target=/usr/local/cargo/reg
 
 FROM debian:bookworm-slim
 
-RUN apt-get update \
+COPY --from=builder /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
-    && apt-get install -y --no-install-recommends ca-certificates curl \
-    && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
 
 COPY --from=builder /tmp/admin-api /usr/local/bin/admin-api
 COPY --from=builder /tmp/crank-migrate /usr/local/bin/crank-migrate
 COPY apps/admin-api/docker-entrypoint.sh /usr/local/bin/crank-admin-entrypoint
+COPY scripts/docker-http-healthcheck.sh /usr/local/bin/crank-http-healthcheck
 
-RUN chmod 0755 /usr/local/bin/crank-admin-entrypoint
+RUN test -s /etc/ssl/certs/ca-certificates.crt \
+    && chmod 0755 /usr/local/bin/crank-admin-entrypoint /usr/local/bin/crank-http-healthcheck
 
 ENV CRANK_ADMIN_BIND=0.0.0.0:3001
 ENV CRANK_STORAGE_ROOT=/var/lib/crank/storage

apps/admin-api/src/main.rs

@@ -227,6 +227,7 @@ async fn run(
     .with_artifact_store(artifact_store.clone())
     .with_public_base_url(base_url)
     .with_outbound_http_policy(outbound_http_policy)
+    .with_external_reference_import(&config.external_references)?
     .with_identity_provider(std::sync::Arc::new(identity_provider))
     .build();
     if config.demo_seed {

apps/admin-api/src/request_context.rs

@@ -6,7 +6,7 @@ use axum::{
 };
 use crank_core::{CorrelationContext, RequestId, TraceContext};
 use crank_metrics::ExemplarTraceId;
-use crank_observability::{set_remote_trace_parent, with_request_correlation};
+use crank_observability::with_request_correlation;
 use tracing::{Instrument, info, info_span};
 
 pub const REQUEST_ID_HEADER: HeaderName = HeaderName::from_static("x-request-id");
@@ -124,11 +124,7 @@ fn one_auxiliary_header_within_budget(
 }
 
 fn set_canonical_parent(span: &tracing::Span, context: &TraceContext) {
-    let mut headers = axum::http::HeaderMap::new();
+    crank_trace::set_parent_from_trace_context(span, context);
-    if let Ok(value) = HeaderValue::from_str(context.traceparent()) {
-        headers.insert("traceparent", value);
-        set_remote_trace_parent(span, &headers);
-    }
 }
 
 #[cfg(test)]

apps/admin-api/src/service.rs

@@ -4,6 +4,7 @@ use std::sync::Arc;
 
 use base64::{Engine as _, engine::general_purpose::URL_SAFE_NO_PAD};
 use crank_artifacts::ArtifactStore;
+use crank_config::ExternalReferenceSettings;
 use crank_core::{
     AuditActor, AuditEvent, AuditEventId, AuditSink, AuditTarget, AuditTargetKind, AuthProfile,
     CapabilityProfile, CommunityCapabilityProfile, CorrelationContext, EditionCapabilities,
@@ -19,7 +20,8 @@ use crank_registry::{
     UsageBucket,
 };
 use crank_runtime::{
-    OutboundHttpPolicy, ResolvedAuth, RuntimeError, RuntimeExecutor, SecretCrypto,
+    ExternalReferenceFetcher, OutboundHttpPolicy, ResolvedAuth, RuntimeError, RuntimeExecutor,
+    SecretCrypto,
 };
 use crank_schema::{Schema, SchemaKind};
 use crank_trace::{DbOperation, ErrorCategory, Stage, StageOutcome, observe_db_query};
@@ -68,6 +70,9 @@ pub struct AdminService {
     audit_sink: Arc<dyn AuditSink>,
     capability_profile: Arc<dyn CapabilityProfile>,
     outbound_http_policy: OutboundHttpPolicy,
+    external_reference_fetcher: ExternalReferenceFetcher,
+    external_reference_normalization: crank_import::rest::NormalizationConfig,
+    external_reference_materialization_timeout: std::time::Duration,
     public_base_url: String,
 }
 
@@ -83,6 +88,9 @@ pub struct AdminServiceBuilder {
     audit_sink: Option<Arc<dyn AuditSink>>,
     capability_profile: Option<Arc<dyn CapabilityProfile>>,
     outbound_http_policy: OutboundHttpPolicy,
+    external_reference_fetcher: ExternalReferenceFetcher,
+    external_reference_normalization: crank_import::rest::NormalizationConfig,
+    external_reference_settings: ExternalReferenceSettings,
     public_base_url: String,
 }
 
@@ -220,6 +228,22 @@ impl AdminServiceBuilder {
         secret_crypto: SecretCrypto,
         runtime: RuntimeExecutor,
     ) -> Self {
+        let outbound_http_policy = OutboundHttpPolicy::default();
+        let external_reference_settings = ExternalReferenceSettings {
+            allowed_url_prefixes: Vec::new(),
+            max_depth: 8,
+            max_documents: 32,
+            max_fetch_bytes: 256 * 1024,
+            fetch_timeout_ms: 5_000,
+            max_expanded_nodes: 10_000,
+        };
+        let external_reference_fetcher = ExternalReferenceFetcher::try_new(
+            outbound_http_policy.clone(),
+            external_reference_settings.allowed_url_prefixes.clone(),
+            external_reference_settings.max_fetch_bytes,
+            std::time::Duration::from_millis(external_reference_settings.fetch_timeout_ms),
+        )
+        .expect("static external reference defaults are valid");
         Self {
             registry,
             storage_root,
@@ -231,7 +255,10 @@ impl AdminServiceBuilder {
             policy_engine: None,
             audit_sink: None,
             capability_profile: None,
-            outbound_http_policy: OutboundHttpPolicy::default(),
+            outbound_http_policy,
+            external_reference_fetcher,
+            external_reference_normalization: Default::default(),
+            external_reference_settings,
             public_base_url: "http://localhost:3000".to_owned(),
         }
     }
@@ -242,10 +269,41 @@ impl AdminServiceBuilder {
     }
 
     pub fn with_outbound_http_policy(mut self, policy: OutboundHttpPolicy) -> Self {
+        self.external_reference_fetcher = ExternalReferenceFetcher::try_new(
+            policy.clone(),
+            self.external_reference_settings
+                .allowed_url_prefixes
+                .clone(),
+            self.external_reference_settings.max_fetch_bytes,
+            std::time::Duration::from_millis(self.external_reference_settings.fetch_timeout_ms),
+        )
+        .expect("validated external reference settings remain valid");
         self.outbound_http_policy = policy;
         self
     }
 
+    pub fn with_external_reference_import(
+        mut self,
+        settings: &ExternalReferenceSettings,
+    ) -> Result<Self, crank_runtime::ExternalReferenceFetchError> {
+        self.external_reference_fetcher = ExternalReferenceFetcher::try_new(
+            self.outbound_http_policy.clone(),
+            settings.allowed_url_prefixes.clone(),
+            settings.max_fetch_bytes,
+            std::time::Duration::from_millis(settings.fetch_timeout_ms),
+        )?;
+        self.external_reference_normalization.max_reference_depth = settings.max_depth;
+        self.external_reference_normalization
+            .max_reference_documents = settings.max_documents;
+        self.external_reference_normalization
+            .max_external_document_bytes = settings.max_fetch_bytes;
+        self.external_reference_normalization.max_expanded_nodes = settings.max_expanded_nodes;
+        self.external_reference_normalization
+            .external_references_enabled = !settings.allowed_url_prefixes.is_empty();
+        self.external_reference_settings = settings.clone();
+        Ok(self)
+    }
+
     /// Reuses the process-wide immutable artifact authority for OpenAPI
     /// ingress and reconciliation. Tests may omit this and use their private
     /// storage root instead.
@@ -299,6 +357,11 @@ impl AdminServiceBuilder {
                 .capability_profile
                 .unwrap_or_else(|| Arc::new(CommunityCapabilityProfile)),
             outbound_http_policy: self.outbound_http_policy,
+            external_reference_fetcher: self.external_reference_fetcher,
+            external_reference_normalization: self.external_reference_normalization,
+            external_reference_materialization_timeout: std::time::Duration::from_millis(
+                self.external_reference_settings.fetch_timeout_ms,
+            ),
             public_base_url: self.public_base_url,
         }
     }

apps/admin-api/src/service/imports.rs

@@ -11,7 +11,7 @@ use crank_import::rest::{
 };
 use crank_registry::{
     ApplyImportJobRequest, ArtifactSourceId, ArtifactSourceSensitivity,
-    CreateArtifactSourceRequest, CreateImportJobRequest, DetachArtifactSourceRequest,
+    CreateArtifactSourceRequest, CreateImportJobRequest, FinishImportJobRequest,
     ImportConflictMode, ImportJobApplyResult, ImportJobId, ImportJobKind, ImportJobSourceEnvelope,
     ImportJobStatus, ImportOperationDraft, RegistryError,
 };
@@ -29,6 +29,13 @@ use crate::{
     },
 };
 
+mod external_references;
+mod job_contract;
+use external_references::ImportReplayContext;
+use job_contract::{
+    SourceDetachGuard, import_job_dependencies, import_job_normalization_config, import_job_source,
+};
+
 const IMPORT_JOB_TTL_HOURS: i64 = 24;
 const OPENAPI_PARSE_DEADLINE: std::time::Duration = std::time::Duration::from_secs(30);
 const OPENAPI_PARSE_CONCURRENCY: usize = 4;
@@ -95,9 +102,14 @@ impl AdminService {
             detach_guard.detach_now().await;
             return Err(ApiError::openapi_upload(locale, "source_integrity"));
         }
+        let mut dependencies = self
+            .materialize_external_reference_snapshots(workspace_id, &verified.bytes, now)
+            .await;
         let parsed = match parse_verified_preview(
             verified.bytes,
             artifact.artifact_ref().digest_hex().to_owned(),
+            dependencies.snapshots.clone(),
+            self.external_reference_normalization.clone(),
             locale,
             false,
         )
@@ -105,6 +117,7 @@ impl AdminService {
         {
             Ok(preview) => preview,
             Err(error) => {
+                dependencies.detach_all().await;
                 detach_guard.detach_now().await;
                 return Err(error);
             }
@@ -115,6 +128,7 @@ impl AdminService {
             .iter()
             .all(|group| group.operations.is_empty())
         {
+            dependencies.detach_all().await;
             detach_guard.detach_now().await;
             return Err(ApiError::openapi_upload(locale, "no_methods"));
         }
@@ -130,12 +144,15 @@ impl AdminService {
                 "source_id": source_envelope.source_id.as_str(),
                 "digest": source_envelope.digest.as_str(),
             },
+            "dependencies": dependencies.payload(),
+            "dependency_snapshots": dependencies.snapshot_payload(),
             "preview": preview_value,
             "preview_digest": preview_digest,
             "normalization": {
                 "normalizer_version": NORMALIZER_VERSION,
                 "projection_version": PROJECTION_VERSION,
                 "ir_fingerprint": parsed.ir_fingerprint,
+                "config": self.external_reference_normalization,
             },
         });
 
@@ -155,9 +172,11 @@ impl AdminService {
             })
             .await
         {
+            dependencies.detach_all().await;
             detach_guard.detach_now().await;
             return Err(ApiError::from(error));
         }
+        dependencies.disarm();
         detach_guard.disarm();
 
         Ok(OpenApiImportPreviewResponse {
@@ -234,6 +253,13 @@ impl AdminService {
         } else {
             ImportConflictMode::Rename
         };
+        let replay_context = ImportReplayContext {
+            workspace_id,
+            job_id,
+            application_key: &application_key,
+            conflict_mode,
+            finished_at: &finished_at,
+        };
         if job.status == ImportJobStatus::Completed {
             let applied = self
                 .registry
@@ -250,7 +276,18 @@ impl AdminService {
             return Ok(openapi_import_response(applied));
         }
 
-        let source = import_job_source(&job.preview_payload, locale)?;
+        let source = match import_job_source(&job.preview_payload, locale) {
+            Ok(source) => source,
+            Err(error) => {
+                return self
+                    .fail_openapi_import_or_replay(
+                        &replay_context,
+                        "import_source_verification_failed",
+                        error,
+                    )
+                    .await;
+            }
+        };
         let verified = match self
             .registry
             .read_artifact_source(
@@ -264,44 +301,123 @@ impl AdminService {
             Err(
                 error @ (RegistryError::SourceNotFound { .. } | RegistryError::SourceUnavailable),
             ) => {
-                // Another request may have finished and detached the source
+                return self
-                // after our unlocked Pending read. Re-read the job and ask
+                    .fail_openapi_import_or_replay(
-                // the registry for its canonical, locked replay instead of
+                        &replay_context,
-                // exposing a false `source_unavailable` outcome.
+                        "import_source_verification_failed",
-                let latest = self.registry.get_import_job(workspace_id, job_id).await?;
+                        openapi_source_error(locale, error),
-                if latest.is_some_and(|latest| latest.status == ImportJobStatus::Completed) {
+                    )
-                    let applied = self
+                    .await;
-                        .registry
+            }
-                        .apply_import_job(ApplyImportJobRequest {
+            Err(error @ RegistryError::SourceIntegrity) => {
-                            id: job_id,
+                return self
-                            workspace_id,
+                    .fail_openapi_import_or_replay(
-                            application_key: &application_key,
+                        &replay_context,
-                            conflict_mode,
+                        "import_source_verification_failed",
-                            operations: &[],
+                        openapi_source_error(locale, error),
-                            pre_skipped: &[],
+                    )
-                            finished_at: &finished_at,
+                    .await;
-                        })
-                        .await?;
-                    return Ok(openapi_import_response(applied));
-                }
-                return Err(openapi_source_error(locale, error));
             }
             Err(error) => return Err(openapi_source_error(locale, error)),
         };
         if verified.source.blob.artifact_ref != source.digest {
-            return Err(ApiError::openapi_upload(locale, "source_integrity"));
+            return self
+                .fail_openapi_import_or_replay(
+                    &replay_context,
+                    "import_source_verification_failed",
+                    ApiError::openapi_upload(locale, "source_integrity"),
+                )
+                .await;
         }
-        let legacy_v1 = job.preview_payload.get("normalization").is_none();
+        let replay_contract = match import_replay_contract(&job.preview_payload, locale) {
-        let parsed = parse_verified_preview(
+            Ok(contract) => contract,
-            verified.bytes,
+            Err(error) => {
-            source.digest.digest_hex().to_owned(),
+                return self
-            locale,
+                    .fail_openapi_import_or_replay(
-            legacy_v1,
+                        &replay_context,
-        )
+                        "import_replay_verification_failed",
-        .await?;
+                        error,
-        verify_preview_contract(&job.preview_payload, &parsed, locale)?;
+                    )
-        if preview_has_blocker(&parsed.preview) {
+                    .await;
-            return Err(ApiError::openapi_upload(locale, "invalid_document"));
+            }
+        };
+        let parsed = match replay_contract {
+            ImportReplayContract::PersistedV2(parsed) => parsed,
+            ImportReplayContract::LegacyV1 | ImportReplayContract::Current => {
+                let legacy_v1 = matches!(replay_contract, ImportReplayContract::LegacyV1);
+                let dependency_snapshots = match self
+                    .read_import_job_dependencies(workspace_id, &job.preview_payload, locale)
+                    .await
+                {
+                    Ok(snapshots) => snapshots,
+                    Err(error) => {
+                        return self
+                            .fail_openapi_import_or_replay(
+                                &replay_context,
+                                "import_dependency_verification_failed",
+                                error,
+                            )
+                            .await;
+                    }
+                };
+                let normalization_config = match import_job_normalization_config(
+                    &job.preview_payload,
+                    &self.external_reference_normalization,
+                    legacy_v1,
+                    locale,
+                ) {
+                    Ok(config) => config,
+                    Err(error) => {
+                        return self
+                            .fail_openapi_import_or_replay(
+                                &replay_context,
+                                "import_replay_verification_failed",
+                                error,
+                            )
+                            .await;
+                    }
+                };
+                let parsed = match parse_verified_preview(
+                    verified.bytes,
+                    source.digest.digest_hex().to_owned(),
+                    dependency_snapshots,
+                    normalization_config,
+                    locale,
+                    legacy_v1,
+                )
+                .await
+                {
+                    Ok(parsed) => parsed,
+                    Err(error) => {
+                        return self
+                            .fail_openapi_import_or_replay(
+                                &replay_context,
+                                "import_replay_verification_failed",
+                                error,
+                            )
+                            .await;
+                    }
+                };
+                if let Err(error) = verify_preview_contract(&job.preview_payload, &parsed, locale) {
+                    return self
+                        .fail_openapi_import_or_replay(
+                            &replay_context,
+                            "import_replay_verification_failed",
+                            error,
+                        )
+                        .await;
+                }
+                parsed
+            }
+        };
+        if preview_has_blocker(&parsed.preview, &selected) {
+            return self
+                .fail_openapi_import_or_replay(
+                    &replay_context,
+                    "reference_resolution_blocked",
+                    ApiError::openapi_upload(locale, "invalid_document"),
+                )
+                .await;
         }
         let mut candidates = BTreeMap::new();
         for group in &parsed.preview.groups {
@@ -372,6 +488,24 @@ impl AdminService {
     }
 }
 
+fn canonical_external_document_uri(base: Option<&str>, reference: &str) -> Option<String> {
+    if reference.starts_with('#') {
+        return None;
+    }
+    let mut url = match base {
+        Some(base) => url::Url::parse(base).ok()?.join(reference).ok()?,
+        None => url::Url::parse(reference).ok()?,
+    };
+    if !matches!(url.scheme(), "http" | "https")
+        || !url.username().is_empty()
+        || url.password().is_some()
+    {
+        return None;
+    }
+    url.set_fragment(None);
+    Some(url.to_string())
+}
+
 fn openapi_import_response(applied: ImportJobApplyResult) -> OpenApiImportCreateResponse {
     let created = applied
         .created
@@ -470,9 +604,72 @@ struct ParsedOpenApiPreview {
     ir_fingerprint: String,
 }
 
+enum ImportReplayContract {
+    LegacyV1,
+    PersistedV2(ParsedOpenApiPreview),
+    Current,
+}
+
+fn import_replay_contract(
+    payload: &serde_json::Value,
+    locale: OpenApiUploadLocale,
+) -> Result<ImportReplayContract, ApiError> {
+    let Some(normalization) = payload.get("normalization") else {
+        return Ok(ImportReplayContract::LegacyV1);
+    };
+    let normalizer = normalization
+        .get("normalizer_version")
+        .and_then(serde_json::Value::as_str)
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+    let projection = normalization
+        .get("projection_version")
+        .and_then(serde_json::Value::as_str)
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+    let ir_fingerprint = normalization
+        .get("ir_fingerprint")
+        .and_then(serde_json::Value::as_str)
+        .filter(|value| is_lower_sha256(value))
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+    let expected_preview_digest = payload
+        .get("preview_digest")
+        .and_then(serde_json::Value::as_str)
+        .filter(|value| is_lower_sha256(value))
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+
+    if normalizer == NORMALIZER_VERSION && projection == PROJECTION_VERSION {
+        return Ok(ImportReplayContract::Current);
+    }
+    if normalizer != "normalized-ir-v2" || projection != "preview-v2" {
+        return Err(ApiError::openapi_upload(locale, "source_integrity"));
+    }
+
+    let preview_value = payload
+        .get("preview")
+        .cloned()
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+    if preview_digest(&preview_value)? != expected_preview_digest {
+        return Err(ApiError::openapi_upload(locale, "source_integrity"));
+    }
+    let preview = serde_json::from_value(preview_value)
+        .map_err(|_| ApiError::openapi_upload(locale, "source_integrity"))?;
+    Ok(ImportReplayContract::PersistedV2(ParsedOpenApiPreview {
+        preview,
+        ir_fingerprint: ir_fingerprint.to_owned(),
+    }))
+}
+
+fn is_lower_sha256(value: &str) -> bool {
+    value.len() == 64
+        && value
+            .bytes()
+            .all(|byte| byte.is_ascii_digit() || (b'a'..=b'f').contains(&byte))
+}
+
 async fn parse_verified_preview(
     bytes: Vec<u8>,
     digest: String,
+    snapshots: Vec<crank_import::rest::ExternalDocumentSnapshot>,
+    normalization_config: crank_import::rest::NormalizationConfig,
     locale: OpenApiUploadLocale,
     legacy_v1: bool,
 ) -> Result<ParsedOpenApiPreview, ApiError> {
@@ -498,14 +695,18 @@ async fn parse_verified_preview(
         }
         let digest = crank_import::rest::SourceDigest::parse(digest)
             .map_err(|_| ApiError::openapi_upload(locale, "source_integrity"))?;
-        let ir =
+        let ir = crank_import::rest::normalize_verified_bundle(
-            crank_import::rest::normalize_verified_document(document, digest, &Default::default())
+            document,
-                .map_err(|error| match error {
+            digest,
-                    crank_import::rest::ImportParseError::NoMethods => {
+            &snapshots,
-                        ApiError::openapi_upload(locale, "no_methods")
+            &normalization_config,
-                    }
+        )
-                    _ => ApiError::openapi_upload(locale, "invalid_document"),
+        .map_err(|error| match error {
-                })?;
+            crank_import::rest::ImportParseError::NoMethods => {
+                ApiError::openapi_upload(locale, "no_methods")
+            }
+            _ => ApiError::openapi_upload(locale, "invalid_document"),
+        })?;
         let ir_fingerprint = preview_digest(
             &serde_json::to_value(&ir)
                 .map_err(|_| ApiError::openapi_upload(locale, "invalid_document"))?,
@@ -597,113 +798,21 @@ fn verify_preview_contract(
     }
 }
 
-fn preview_has_blocker(preview: &crank_import::rest::ImportPreview) -> bool {
+fn preview_has_blocker(
+    preview: &crank_import::rest::ImportPreview,
+    selected_operation_keys: &BTreeSet<String>,
+) -> bool {
     preview
         .findings
         .iter()
-        .chain(
-            preview
-                .groups
-                .iter()
-                .flat_map(|group| group.operations.iter())
-                .flat_map(|operation| operation.findings.iter()),
-        )
         .any(|finding| finding.severity == ImportFindingSeverity::Error)
-}
+        || preview
-
+            .groups
-fn import_job_source(
+            .iter()
-    payload: &serde_json::Value,
+            .flat_map(|group| group.operations.iter())
-    locale: OpenApiUploadLocale,
+            .filter(|operation| selected_operation_keys.contains(&operation.key))
-) -> Result<ImportJobSourceEnvelope, ApiError> {
+            .flat_map(|operation| operation.findings.iter())
-    let source = payload
+            .any(|finding| finding.severity == ImportFindingSeverity::Error)
-        .get("source")
-        .ok_or_else(|| ApiError::openapi_upload(locale, "source_unavailable"))?;
-    let source_id = source
-        .get("source_id")
-        .and_then(serde_json::Value::as_str)
-        .filter(|value| value.len() <= 132)
-        .ok_or_else(|| ApiError::openapi_upload(locale, "source_unavailable"))?;
-    let digest = source
-        .get("digest")
-        .and_then(serde_json::Value::as_str)
-        .and_then(|value| value.parse().ok())
-        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
-    Ok(ImportJobSourceEnvelope {
-        source_id: ArtifactSourceId::new(source_id),
-        digest,
-    })
-}
-
-struct SourceDetachGuard {
-    registry: crank_registry::PostgresRegistry,
-    workspace_id: WorkspaceId,
-    source_id: ArtifactSourceId,
-    expected_updated_at: OffsetDateTime,
-    armed: bool,
-}
-
-impl SourceDetachGuard {
-    fn new(
-        registry: crank_registry::PostgresRegistry,
-        workspace_id: WorkspaceId,
-        source_id: ArtifactSourceId,
-        expected_updated_at: OffsetDateTime,
-    ) -> Self {
-        Self {
-            registry,
-            workspace_id,
-            source_id,
-            expected_updated_at,
-            armed: true,
-        }
-    }
-
-    fn disarm(&mut self) {
-        self.armed = false;
-    }
-
-    async fn detach_now(&mut self) {
-        if !self.armed {
-            return;
-        }
-        self.armed = false;
-        // This is the normal error path: await the database transition so a
-        // caller receives a completed cleanup before it can retry. `Drop`
-        // remains only the cancellation/shutdown safety net.
-        let _ = self
-            .registry
-            .detach_artifact_source(DetachArtifactSourceRequest {
-                workspace_id: &self.workspace_id,
-                source_id: &self.source_id,
-                expected_updated_at: Some(self.expected_updated_at),
-                detached_at: OffsetDateTime::now_utc(),
-            })
-            .await;
-    }
-}
-
-impl Drop for SourceDetachGuard {
-    fn drop(&mut self) {
-        if !self.armed {
-            return;
-        }
-        let registry = self.registry.clone();
-        let workspace_id = self.workspace_id.clone();
-        let source_id = self.source_id.clone();
-        let expected_updated_at = Some(self.expected_updated_at);
-        if let Ok(handle) = tokio::runtime::Handle::try_current() {
-            handle.spawn(async move {
-                let _ = registry
-                    .detach_artifact_source(DetachArtifactSourceRequest {
-                        workspace_id: &workspace_id,
-                        source_id: &source_id,
-                        expected_updated_at,
-                        detached_at: OffsetDateTime::now_utc(),
-                    })
-                    .await;
-            });
-        }
-    }
 }
 
 fn openapi_application_key(payload: &OpenApiImportCreatePayload) -> Result<String, ApiError> {

apps/admin-api/src/service/imports/external_references.rs

@@ -0,0 +1,416 @@
+use super::*;
+use std::collections::VecDeque;
+use tracing::warn;
+
+pub(super) struct MaterializedDependencies {
+    pub(super) snapshots: Vec<crank_import::rest::ExternalDocumentSnapshot>,
+    envelopes: Vec<ImportJobSourceEnvelope>,
+    canonical_uris: Vec<String>,
+    guards: Vec<SourceDetachGuard>,
+}
+
+pub(super) struct ImportReplayContext<'a> {
+    pub(super) workspace_id: &'a WorkspaceId,
+    pub(super) job_id: &'a ImportJobId,
+    pub(super) application_key: &'a str,
+    pub(super) conflict_mode: ImportConflictMode,
+    pub(super) finished_at: &'a OffsetDateTime,
+}
+
+impl MaterializedDependencies {
+    fn empty() -> Self {
+        Self {
+            snapshots: Vec::new(),
+            envelopes: Vec::new(),
+            canonical_uris: Vec::new(),
+            guards: Vec::new(),
+        }
+    }
+
+    pub(super) fn payload(&self) -> Vec<serde_json::Value> {
+        self.envelopes
+            .iter()
+            .zip(&self.canonical_uris)
+            .map(|(dependency, canonical_uri)| {
+                json!({
+                    "source_id": dependency.source_id.as_str(),
+                    "digest": dependency.digest.as_str(),
+                    "canonical_uri": canonical_uri,
+                })
+            })
+            .collect()
+    }
+
+    pub(super) fn snapshot_payload(&self) -> Vec<serde_json::Value> {
+        self.snapshots
+            .iter()
+            .filter_map(|snapshot| {
+                self.envelopes
+                    .iter()
+                    .find(|dependency| dependency.digest.digest_hex() == snapshot.digest.as_str())
+                    .map(|dependency| {
+                        json!({
+                            "source_id": dependency.source_id.as_str(),
+                            "digest": dependency.digest.as_str(),
+                            "canonical_uri": snapshot.canonical_uri,
+                        })
+                    })
+            })
+            .collect()
+    }
+
+    pub(super) fn disarm(&mut self) {
+        for guard in &mut self.guards {
+            guard.disarm();
+        }
+    }
+
+    pub(super) async fn detach_all(&mut self) {
+        for guard in &mut self.guards {
+            guard.detach_now().await;
+        }
+    }
+}
+
+impl AdminService {
+    pub(super) async fn fail_openapi_import_or_replay(
+        &self,
+        context: &ImportReplayContext<'_>,
+        error_text: &'static str,
+        error: ApiError,
+    ) -> Result<OpenApiImportCreateResponse, ApiError> {
+        let empty = json!([]);
+        self.registry
+            .finish_import_job(FinishImportJobRequest {
+                id: context.job_id,
+                status: ImportJobStatus::Failed,
+                created_operation_ids: &empty,
+                error_text: Some(error_text),
+                finished_at: context.finished_at,
+            })
+            .await?;
+
+        // `finish_import_job` preserves an already Completed row under its
+        // lock. This read distinguishes that race from the terminal Failed
+        // transition and returns the immutable canonical application result.
+        let latest = self
+            .registry
+            .get_import_job(context.workspace_id, context.job_id)
+            .await?;
+        if latest.is_some_and(|job| job.status == ImportJobStatus::Completed) {
+            let applied = self
+                .registry
+                .apply_import_job(ApplyImportJobRequest {
+                    id: context.job_id,
+                    workspace_id: context.workspace_id,
+                    application_key: context.application_key,
+                    conflict_mode: context.conflict_mode,
+                    operations: &[],
+                    pre_skipped: &[],
+                    finished_at: context.finished_at,
+                })
+                .await?;
+            return Ok(openapi_import_response(applied));
+        }
+
+        Err(error)
+    }
+
+    pub(super) async fn materialize_external_reference_snapshots(
+        &self,
+        workspace_id: &WorkspaceId,
+        primary_bytes: &[u8],
+        created_at: OffsetDateTime,
+    ) -> MaterializedDependencies {
+        let materialized_count = std::sync::atomic::AtomicUsize::new(0);
+        match tokio::time::timeout(
+            self.external_reference_materialization_timeout,
+            self.materialize_external_reference_snapshots_inner(
+                workspace_id,
+                primary_bytes,
+                created_at,
+                &materialized_count,
+            ),
+        )
+        .await
+        {
+            Ok(dependencies) => dependencies,
+            Err(_) => {
+                materialization_failure(
+                    "chain",
+                    "timeout",
+                    materialized_count.load(std::sync::atomic::Ordering::Relaxed),
+                );
+                MaterializedDependencies::empty()
+            }
+        }
+    }
+
+    async fn materialize_external_reference_snapshots_inner(
+        &self,
+        workspace_id: &WorkspaceId,
+        primary_bytes: &[u8],
+        created_at: OffsetDateTime,
+        materialized_count: &std::sync::atomic::AtomicUsize,
+    ) -> MaterializedDependencies {
+        if !self
+            .external_reference_normalization
+            .external_references_enabled
+        {
+            return MaterializedDependencies::empty();
+        }
+        let Ok(primary) = std::str::from_utf8(primary_bytes) else {
+            materialization_failure("primary_decode", "invalid_utf8", 0);
+            return MaterializedDependencies::empty();
+        };
+        let Ok(primary_references) =
+            crank_import::rest::reference_uris(primary, &self.external_reference_normalization)
+        else {
+            materialization_failure("primary_scan", "invalid_document", 0);
+            return MaterializedDependencies::empty();
+        };
+        let mut queue = VecDeque::new();
+        for reference in primary_references {
+            if let Some(uri) = canonical_external_document_uri(None, &reference) {
+                queue.push_back((uri, 1usize));
+            }
+        }
+        let mut seen = BTreeSet::new();
+        let mut result = MaterializedDependencies::empty();
+        let mut dependency_by_digest = BTreeMap::<String, usize>::new();
+        while let Some((canonical_uri, depth)) = queue.pop_front() {
+            if !seen.insert(canonical_uri.clone())
+                || depth > self.external_reference_normalization.max_reference_depth
+                || seen.len()
+                    > self
+                        .external_reference_normalization
+                        .max_reference_documents
+            {
+                continue;
+            }
+            let bytes = match self.external_reference_fetcher.get(&canonical_uri).await {
+                Ok(bytes) => bytes,
+                Err(error) => {
+                    materialization_failure(
+                        "fetch",
+                        external_fetch_error_code(&error),
+                        result.snapshots.len(),
+                    );
+                    continue;
+                }
+            };
+            let Ok(document) = std::str::from_utf8(&bytes).map(str::to_owned) else {
+                materialization_failure(
+                    "dependency_decode",
+                    "invalid_utf8",
+                    result.snapshots.len(),
+                );
+                continue;
+            };
+            let Ok(references) = crank_import::rest::external_reference_uris(
+                &document,
+                &self.external_reference_normalization,
+            ) else {
+                materialization_failure(
+                    "dependency_scan",
+                    "invalid_document",
+                    result.snapshots.len(),
+                );
+                continue;
+            };
+            let store = self.artifact_store.clone();
+            let artifact_bytes = bytes.clone();
+            let Ok(Ok(artifact)) =
+                tokio::task::spawn_blocking(move || store.put_registered(&artifact_bytes)).await
+            else {
+                materialization_failure(
+                    "artifact_store",
+                    "storage_unavailable",
+                    result.snapshots.len(),
+                );
+                continue;
+            };
+            let digest_key = artifact.artifact_ref().as_str().to_owned();
+            let envelope_index = if let Some(index) = dependency_by_digest.get(&digest_key) {
+                *index
+            } else {
+                let source_id = ArtifactSourceId::new(new_prefixed_id("src_openapi_dep"));
+                let Ok(source) = self
+                    .registry
+                    .create_artifact_source(CreateArtifactSourceRequest {
+                        workspace_id,
+                        source_id: &source_id,
+                        artifact: &artifact,
+                        mime_type: "application/octet-stream",
+                        sensitivity: ArtifactSourceSensitivity::Internal,
+                        created_at,
+                    })
+                    .await
+                else {
+                    materialization_failure(
+                        "source_create",
+                        "registry_unavailable",
+                        result.snapshots.len(),
+                    );
+                    continue;
+                };
+                let mut guard = SourceDetachGuard::new(
+                    self.registry.clone(),
+                    workspace_id.clone(),
+                    source_id.clone(),
+                    source.updated_at,
+                );
+                let Ok(verified) = self
+                    .registry
+                    .read_artifact_source(
+                        std::sync::Arc::clone(&self.artifact_store),
+                        workspace_id,
+                        &source_id,
+                    )
+                    .await
+                else {
+                    materialization_failure(
+                        "source_verify",
+                        "source_unavailable",
+                        result.snapshots.len(),
+                    );
+                    guard.detach_now().await;
+                    continue;
+                };
+                if verified.source.blob.artifact_ref != *artifact.artifact_ref() {
+                    materialization_failure(
+                        "source_verify",
+                        "source_integrity",
+                        result.snapshots.len(),
+                    );
+                    guard.detach_now().await;
+                    continue;
+                }
+                let index = result.envelopes.len();
+                result.envelopes.push(ImportJobSourceEnvelope {
+                    source_id: source_id.clone(),
+                    digest: artifact.artifact_ref().clone(),
+                });
+                result.guards.push(guard);
+                dependency_by_digest.insert(digest_key, index);
+                index
+            };
+            let dependency = &result.envelopes[envelope_index];
+            let snapshot_digest = match crank_import::rest::SourceDigest::parse(
+                dependency.digest.digest_hex().to_owned(),
+            ) {
+                Ok(digest) => digest,
+                Err(_) => {
+                    materialization_failure("snapshot", "source_integrity", result.snapshots.len());
+                    continue;
+                }
+            };
+            result
+                .snapshots
+                .push(crank_import::rest::ExternalDocumentSnapshot {
+                    canonical_uri: canonical_uri.clone(),
+                    digest: snapshot_digest,
+                    document,
+                });
+            materialized_count.store(result.snapshots.len(), std::sync::atomic::Ordering::Relaxed);
+            result.canonical_uris.push(canonical_uri.clone());
+            for reference in references {
+                if let Some(uri) = canonical_external_document_uri(Some(&canonical_uri), &reference)
+                {
+                    queue.push_back((uri, depth.saturating_add(1)));
+                }
+            }
+        }
+        // Registry ownership is one envelope per immutable digest, while the
+        // resolver may map several canonical URIs to that same snapshot. Keep
+        // payload rows aligned with envelopes and add aliases separately.
+        if result.envelopes.len() != result.canonical_uris.len() {
+            let mut canonical_by_digest = BTreeMap::new();
+            for snapshot in &result.snapshots {
+                canonical_by_digest
+                    .entry(snapshot.digest.as_str().to_owned())
+                    .or_insert_with(|| snapshot.canonical_uri.clone());
+            }
+            result.canonical_uris = result
+                .envelopes
+                .iter()
+                .map(|envelope| {
+                    canonical_by_digest
+                        .get(envelope.digest.digest_hex())
+                        .cloned()
+                        .unwrap_or_default()
+                })
+                .collect();
+        }
+        result
+    }
+
+    pub(super) async fn read_import_job_dependencies(
+        &self,
+        workspace_id: &WorkspaceId,
+        payload: &serde_json::Value,
+        locale: OpenApiUploadLocale,
+    ) -> Result<Vec<crank_import::rest::ExternalDocumentSnapshot>, ApiError> {
+        let dependencies = import_job_dependencies(payload, locale)?;
+        let mut snapshots = Vec::with_capacity(dependencies.len());
+        for (canonical_uri, dependency) in dependencies {
+            let verified = self
+                .registry
+                .read_artifact_source(
+                    std::sync::Arc::clone(&self.artifact_store),
+                    workspace_id,
+                    &dependency.source_id,
+                )
+                .await
+                .map_err(|error| openapi_source_error(locale, error))?;
+            if verified.source.blob.artifact_ref != dependency.digest {
+                return Err(ApiError::openapi_upload(locale, "source_integrity"));
+            }
+            let document = String::from_utf8(verified.bytes)
+                .map_err(|_| ApiError::openapi_upload(locale, "source_integrity"))?;
+            let digest =
+                crank_import::rest::SourceDigest::parse(dependency.digest.digest_hex().to_owned())
+                    .map_err(|_| ApiError::openapi_upload(locale, "source_integrity"))?;
+            snapshots.push(crank_import::rest::ExternalDocumentSnapshot {
+                canonical_uri,
+                digest,
+                document,
+            });
+        }
+        Ok(snapshots)
+    }
+}
+fn external_fetch_error_code(error: &crank_runtime::ExternalReferenceFetchError) -> &'static str {
+    match error {
+        crank_runtime::ExternalReferenceFetchError::Disabled => "disabled",
+        crank_runtime::ExternalReferenceFetchError::InvalidUrl => "invalid_url",
+        crank_runtime::ExternalReferenceFetchError::TargetNotAllowed => "target_not_allowed",
+        crank_runtime::ExternalReferenceFetchError::RedirectNotAllowed => "redirect_not_allowed",
+        crank_runtime::ExternalReferenceFetchError::ResponseTooLarge { .. } => "response_too_large",
+        crank_runtime::ExternalReferenceFetchError::UnexpectedStatus { .. } => "unexpected_status",
+        crank_runtime::ExternalReferenceFetchError::Transport { timeout: true, .. } => "timeout",
+        crank_runtime::ExternalReferenceFetchError::Transport { .. } => "transport",
+        crank_runtime::ExternalReferenceFetchError::InvalidConfiguration => "invalid_configuration",
+    }
+}
+
+fn materialization_failure(stage: &'static str, error_code: &'static str, count: usize) {
+    warn!(
+        name: "admin.openapi_import.materialization_failed",
+        stage,
+        error_code,
+        count,
+        "external OpenAPI materialization failed"
+    );
+}
+
+#[cfg(test)]
+mod tests {
+    use super::external_fetch_error_code;
+
+    #[test]
+    fn unexpected_status_has_a_stable_safe_error_code() {
+        let error = crank_runtime::ExternalReferenceFetchError::UnexpectedStatus { status: 500 };
+        assert_eq!(external_fetch_error_code(&error), "unexpected_status");
+    }
+}

apps/admin-api/src/service/imports/job_contract.rs

@@ -0,0 +1,169 @@
+use std::collections::BTreeSet;
+
+use crank_core::WorkspaceId;
+use crank_registry::{ArtifactSourceId, DetachArtifactSourceRequest, ImportJobSourceEnvelope};
+use time::OffsetDateTime;
+
+use crate::{error::ApiError, service::OpenApiUploadLocale};
+
+use super::canonical_external_document_uri;
+
+const MAX_IMPORT_JOB_DOCUMENTS: usize = 32;
+
+pub(super) fn import_job_source(
+    payload: &serde_json::Value,
+    locale: OpenApiUploadLocale,
+) -> Result<ImportJobSourceEnvelope, ApiError> {
+    let source = payload
+        .get("source")
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_unavailable"))?;
+    let source_id = source
+        .get("source_id")
+        .and_then(serde_json::Value::as_str)
+        .filter(|value| value.len() <= 132)
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_unavailable"))?;
+    let digest = source
+        .get("digest")
+        .and_then(serde_json::Value::as_str)
+        .and_then(|value| value.parse().ok())
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+    Ok(ImportJobSourceEnvelope {
+        source_id: ArtifactSourceId::new(source_id),
+        digest,
+    })
+}
+
+pub(super) fn import_job_dependencies(
+    payload: &serde_json::Value,
+    locale: OpenApiUploadLocale,
+) -> Result<Vec<(String, ImportJobSourceEnvelope)>, ApiError> {
+    let empty = Vec::new();
+    let items = payload
+        .get("dependency_snapshots")
+        .and_then(serde_json::Value::as_array)
+        .unwrap_or(&empty);
+    if items.len() > MAX_IMPORT_JOB_DOCUMENTS {
+        return Err(ApiError::openapi_upload(locale, "source_integrity"));
+    }
+    let mut canonical_uris = BTreeSet::new();
+    items
+        .iter()
+        .map(|item| {
+            let canonical_uri = item
+                .get("canonical_uri")
+                .and_then(serde_json::Value::as_str)
+                .filter(|value| {
+                    canonical_external_document_uri(None, value).as_deref() == Some(*value)
+                })
+                .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+            if !canonical_uris.insert(canonical_uri.to_owned()) {
+                return Err(ApiError::openapi_upload(locale, "source_integrity"));
+            }
+            let source_id = item
+                .get("source_id")
+                .and_then(serde_json::Value::as_str)
+                .filter(|value| value.len() <= 132)
+                .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+            let digest = item
+                .get("digest")
+                .and_then(serde_json::Value::as_str)
+                .and_then(|value| value.parse().ok())
+                .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))?;
+            Ok((
+                canonical_uri.to_owned(),
+                ImportJobSourceEnvelope {
+                    source_id: ArtifactSourceId::new(source_id),
+                    digest,
+                },
+            ))
+        })
+        .collect()
+}
+
+pub(super) fn import_job_normalization_config(
+    payload: &serde_json::Value,
+    current: &crank_import::rest::NormalizationConfig,
+    legacy_v1: bool,
+    locale: OpenApiUploadLocale,
+) -> Result<crank_import::rest::NormalizationConfig, ApiError> {
+    if legacy_v1 {
+        return Ok(current.clone());
+    }
+    payload
+        .pointer("/normalization/config")
+        .cloned()
+        .ok_or_else(|| ApiError::openapi_upload(locale, "source_integrity"))
+        .and_then(|value| {
+            serde_json::from_value(value)
+                .map_err(|_| ApiError::openapi_upload(locale, "source_integrity"))
+        })
+}
+
+pub(super) struct SourceDetachGuard {
+    registry: crank_registry::PostgresRegistry,
+    workspace_id: WorkspaceId,
+    source_id: ArtifactSourceId,
+    expected_updated_at: OffsetDateTime,
+    armed: bool,
+}
+
+impl SourceDetachGuard {
+    pub(super) fn new(
+        registry: crank_registry::PostgresRegistry,
+        workspace_id: WorkspaceId,
+        source_id: ArtifactSourceId,
+        expected_updated_at: OffsetDateTime,
+    ) -> Self {
+        Self {
+            registry,
+            workspace_id,
+            source_id,
+            expected_updated_at,
+            armed: true,
+        }
+    }
+
+    pub(super) fn disarm(&mut self) {
+        self.armed = false;
+    }
+
+    pub(super) async fn detach_now(&mut self) {
+        if !self.armed {
+            return;
+        }
+        self.armed = false;
+        let _ = self
+            .registry
+            .detach_artifact_source(DetachArtifactSourceRequest {
+                workspace_id: &self.workspace_id,
+                source_id: &self.source_id,
+                expected_updated_at: Some(self.expected_updated_at),
+                detached_at: OffsetDateTime::now_utc(),
+            })
+            .await;
+    }
+}
+
+impl Drop for SourceDetachGuard {
+    fn drop(&mut self) {
+        if !self.armed {
+            return;
+        }
+        let registry = self.registry.clone();
+        let workspace_id = self.workspace_id.clone();
+        let source_id = self.source_id.clone();
+        let expected_updated_at = Some(self.expected_updated_at);
+        if let Ok(handle) = tokio::runtime::Handle::try_current() {
+            handle.spawn(async move {
+                let _ = registry
+                    .detach_artifact_source(DetachArtifactSourceRequest {
+                        workspace_id: &workspace_id,
+                        source_id: &source_id,
+                        expected_updated_at,
+                        detached_at: OffsetDateTime::now_utc(),
+                    })
+                    .await;
+            });
+        }
+    }
+}

apps/admin-api/tests/integration/common.rs

@@ -108,6 +108,28 @@ pub(super) fn build_test_app(
     })
 }
 
+pub(super) fn build_test_app_with_external_references(
+    registry: PostgresRegistry,
+    storage_root: std::path::PathBuf,
+    allowed_url_prefixes: Vec<String>,
+) -> Router {
+    build_app(AppState {
+        service: test_service_with_external_references_timeout(
+            registry,
+            storage_root,
+            allowed_url_prefixes,
+            // This HTTP fixture asserts classification of an immediate
+            // upstream status. Leave headroom for CI scheduling jitter so the
+            // unrelated chain-timeout branch cannot win first.
+            10_000,
+        ),
+        api_rate_limiter: crank_runtime::RequestRateLimiter::new(
+            crank_runtime::RequestRateLimitConfig::new(10_000, 10_000).unwrap(),
+        ),
+        trusted_proxy_ips: Vec::new(),
+    })
+}
+
 pub(super) fn build_test_app_with_audit_sink(
     registry: PostgresRegistry,
     storage_root: std::path::PathBuf,
@@ -152,6 +174,47 @@ pub(super) fn test_service(
     .build()
 }
 
+pub(super) fn test_service_with_external_references(
+    registry: PostgresRegistry,
+    storage_root: std::path::PathBuf,
+    allowed_url_prefixes: Vec<String>,
+) -> AdminService {
+    test_service_with_external_references_timeout(
+        registry,
+        storage_root,
+        allowed_url_prefixes,
+        2_000,
+    )
+}
+
+fn test_service_with_external_references_timeout(
+    registry: PostgresRegistry,
+    storage_root: std::path::PathBuf,
+    allowed_url_prefixes: Vec<String>,
+    fetch_timeout_ms: u64,
+) -> AdminService {
+    let outbound_policy = crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]);
+    let runtime = crank_runtime::community_with_outbound_policy(outbound_policy.clone()).build();
+    AdminServiceBuilder::new(
+        registry,
+        storage_root,
+        test_auth_settings(),
+        test_secret_crypto(),
+        runtime,
+    )
+    .with_external_reference_import(&crank_config::ExternalReferenceSettings {
+        allowed_url_prefixes,
+        max_depth: 8,
+        max_documents: 32,
+        max_fetch_bytes: 64 * 1024,
+        fetch_timeout_ms,
+        max_expanded_nodes: 10_000,
+    })
+    .unwrap()
+    .with_outbound_http_policy(outbound_policy)
+    .build()
+}
+
 pub(super) async fn spawn_upstream_server() -> String {
     let app = Router::new().route("/crm/leads", post(create_lead));
     let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();

apps/admin-api/tests/integration/openapi_import.rs

@@ -1,4 +1,6 @@
-use admin_api::service::{OpenApiImportCreatePayload, OpenApiUpload, OpenApiUploadLocale};
+use admin_api::service::{
+    AdminServiceBuilder, OpenApiImportCreatePayload, OpenApiUpload, OpenApiUploadLocale,
+};
 use crank_core::WorkspaceId;
 use crank_registry::ImportJobStatus;
 use serial_test::serial;
@@ -7,6 +9,8 @@ use super::common::{
     test_auth_settings, test_registry, test_secret_crypto, test_service, test_storage_root,
 };
 
+mod external_references;
+
 const OPENAPI3: &str = r#"
 openapi: 3.0.3
 info:
@@ -73,11 +77,11 @@ async fn previews_openapi_and_creates_draft_operations() {
     assert_eq!(preview_job.status, ImportJobStatus::Pending);
     assert_eq!(
         preview_job.preview_payload["normalization"]["normalizer_version"],
-        "normalized-ir-v2"
+        "normalized-ir-v3"
     );
     assert_eq!(
         preview_job.preview_payload["normalization"]["projection_version"],
-        "preview-v2"
+        "preview-v3"
     );
     assert!(preview_job.preview_payload["normalization"]["ir_fingerprint"].is_string());
 
@@ -236,7 +240,7 @@ async fn unknown_selected_operations_are_persisted_in_the_canonical_replay() {
 async fn concurrent_openapi_import_replays_the_same_atomic_result() {
     let registry = test_registry().await;
     let service = test_service(
-        registry,
+        registry.clone(),
         test_storage_root("openapi_import_replay"),
         test_auth_settings(),
         test_secret_crypto(),
@@ -271,6 +275,31 @@ async fn concurrent_openapi_import_replays_the_same_atomic_result() {
         service.list_operations(&workspace_id).await.unwrap().len(),
         1
     );
+    let completed_before = registry
+        .get_import_job(&workspace_id, &job_id)
+        .await
+        .unwrap()
+        .unwrap();
+    let empty = serde_json::json!([]);
+    let failed_at = time::OffsetDateTime::now_utc();
+    registry
+        .finish_import_job(crank_registry::FinishImportJobRequest {
+            id: &job_id,
+            status: ImportJobStatus::Failed,
+            created_operation_ids: &empty,
+            error_text: Some("late_verification_failure"),
+            finished_at: &failed_at,
+        })
+        .await
+        .unwrap();
+    let completed_after = registry
+        .get_import_job(&workspace_id, &job_id)
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(completed_after.status, ImportJobStatus::Completed);
+    assert_eq!(completed_after.error_text, completed_before.error_text);
+    assert_eq!(completed_after.finished_at, completed_before.finished_at);
 
     let conflicting_replay = service
         .create_openapi_import(
@@ -326,6 +355,13 @@ async fn apply_fails_closed_when_the_preview_parser_contract_drifts() {
         .await;
 
     assert!(result.is_err());
+    assert_failed_job_and_detached_sources(
+        &registry,
+        &workspace_id,
+        &job_id,
+        "import_replay_verification_failed",
+    )
+    .await;
     assert!(
         service
             .list_operations(&workspace_id)
@@ -395,6 +431,13 @@ async fn apply_fails_closed_for_each_normalization_contract_field_and_digest_sha
                 .await
                 .is_err()
         );
+        assert_failed_job_and_detached_sources(
+            &registry,
+            &workspace_id,
+            &job_id,
+            "import_replay_verification_failed",
+        )
+        .await;
         assert!(
             service
                 .list_operations(&workspace_id)
@@ -449,7 +492,7 @@ async fn legacy_job_without_normalization_contract_replays_until_expiry() {
 async fn blocker_finding_keeps_valid_preview_but_prevents_draft_mutation() {
     let registry = test_registry().await;
     let service = test_service(
-        registry,
+        registry.clone(),
         test_storage_root("openapi_import_blocker"),
         test_auth_settings(),
         test_secret_crypto(),
@@ -503,11 +546,124 @@ paths:
             .unwrap()
             .is_empty()
     );
+    assert_failed_job_and_detached_sources(
+        &registry,
+        &workspace_id,
+        &preview.job_id.as_str().into(),
+        "reference_resolution_blocked",
+    )
+    .await;
 }
 
 #[tokio::test]
 #[serial]
-async fn operation_level_reference_blocker_prevents_draft_mutation() {
+async fn operation_blockers_apply_only_to_selected_keys_and_full_selection_fails_atomically() {
+    let registry = test_registry().await;
+    let service = test_service(
+        registry.clone(),
+        test_storage_root("openapi_import_selected_blockers"),
+        test_auth_settings(),
+        test_secret_crypto(),
+    );
+    let workspace_id = WorkspaceId::new("ws_default");
+    let upload = OpenApiUpload {
+        bytes: br#"
+openapi: 3.1.0
+info: { title: Selected blockers }
+servers: [{ url: https://api.example.test }]
+paths:
+  /valid:
+    get:
+      operationId: validOperation
+      responses: { '200': { description: ok } }
+  /broken:
+    get:
+      operationId: brokenOperation
+      responses:
+        '200':
+          description: unresolved external schema
+          content:
+            application/json:
+              schema: { $ref: 'https://schemas.example.test/missing.yaml#/Result' }
+"#
+        .to_vec(),
+        mime_type: "application/yaml".to_owned(),
+        locale: OpenApiUploadLocale::En,
+    };
+
+    let valid_preview = service
+        .preview_openapi_import(&workspace_id, upload.clone())
+        .await
+        .unwrap();
+    assert!(
+        valid_preview.preview.findings.iter().all(|finding| {
+            finding.severity != crank_import::rest::ImportFindingSeverity::Error
+        })
+    );
+    let broken = valid_preview
+        .preview
+        .groups
+        .iter()
+        .flat_map(|group| &group.operations)
+        .find(|operation| operation.key == "GET /broken")
+        .unwrap();
+    assert!(
+        broken.findings.iter().any(|finding| {
+            finding.severity == crank_import::rest::ImportFindingSeverity::Error
+        })
+    );
+
+    let valid = service
+        .create_openapi_import(
+            &workspace_id,
+            &valid_preview.job_id.as_str().into(),
+            OpenApiImportCreatePayload {
+                selected_operation_keys: vec!["GET /valid".to_owned()],
+                server_url: None,
+                conflict_mode: "skip".to_owned(),
+            },
+        )
+        .await
+        .unwrap();
+    assert_eq!(valid.created.len(), 1);
+
+    let full_preview = service
+        .preview_openapi_import(&workspace_id, upload)
+        .await
+        .unwrap();
+    let result = service
+        .create_openapi_import(
+            &workspace_id,
+            &full_preview.job_id.as_str().into(),
+            OpenApiImportCreatePayload {
+                selected_operation_keys: vec!["GET /valid".to_owned(), "GET /broken".to_owned()],
+                server_url: None,
+                conflict_mode: "rename".to_owned(),
+            },
+        )
+        .await;
+    assert!(result.is_err());
+    assert_eq!(
+        service.list_operations(&workspace_id).await.unwrap().len(),
+        1,
+        "the full selection must not create either selected Draft"
+    );
+    let failed = registry
+        .get_import_job(&workspace_id, &full_preview.job_id.as_str().into())
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(failed.status, ImportJobStatus::Failed);
+    assert_eq!(
+        failed.error_text.as_deref(),
+        Some("reference_resolution_blocked")
+    );
+    assert!(failed.finished_at.is_some());
+}
+
+#[tokio::test]
+#[serial]
+async fn local_operation_reference_resolves_and_creates_draft() {
     let registry = test_registry().await;
     let service = test_service(
         registry,
@@ -520,6 +676,7 @@ async fn operation_level_reference_blocker_prevents_draft_mutation() {
         bytes: br#"
 openapi: 3.0.3
 info: { title: References }
+servers: [{ url: https://api.example.test }]
 paths:
   /referenced:
     get:
@@ -551,32 +708,26 @@ components:
         preview.preview.groups[0].operations[0]
             .findings
             .iter()
-            .any(|finding| {
+            .all(|finding| finding.code != "unresolved_reference")
-                finding.code == "unresolved_reference"
-                    && finding.severity == crank_import::rest::ImportFindingSeverity::Error
-            })
     );
+    assert_eq!(preview.preview.groups[0].operations[0].output_fields, 1);
 
-    assert!(
+    let created = service
-        service
+        .create_openapi_import(
-            .create_openapi_import(
+            &workspace_id,
-                &workspace_id,
+            &preview.job_id.as_str().into(),
-                &preview.job_id.as_str().into(),
+            OpenApiImportCreatePayload {
-                OpenApiImportCreatePayload {
+                selected_operation_keys: vec!["GET /referenced".to_owned()],
-                    selected_operation_keys: vec!["GET /referenced".to_owned()],
+                server_url: None,
-                    server_url: None,
+                conflict_mode: "skip".to_owned(),
-                    conflict_mode: "skip".to_owned(),
+            },
-                },
+        )
-            )
+        .await
-            .await
+        .unwrap();
-            .is_err()
+    assert_eq!(created.created.len(), 1);
-    );
+    assert_eq!(
-    assert!(
+        service.list_operations(&workspace_id).await.unwrap().len(),
-        service
+        1
-            .list_operations(&workspace_id)
-            .await
-            .unwrap()
-            .is_empty()
     );
 }
 
@@ -587,3 +738,44 @@ fn openapi_upload() -> OpenApiUpload {
         locale: OpenApiUploadLocale::En,
     }
 }
+
+async fn assert_failed_job_and_detached_sources(
+    registry: &crank_registry::PostgresRegistry,
+    workspace_id: &WorkspaceId,
+    job_id: &crank_registry::ImportJobId,
+    expected_error: &str,
+) {
+    let job = registry
+        .get_import_job(workspace_id, job_id)
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(job.status, ImportJobStatus::Failed);
+    assert!(job.finished_at.is_some());
+    assert_eq!(job.error_text.as_deref(), Some(expected_error));
+
+    let mut source_ids = vec![
+        job.preview_payload["source"]["source_id"]
+            .as_str()
+            .unwrap()
+            .to_owned(),
+    ];
+    source_ids.extend(
+        job.preview_payload["dependencies"]
+            .as_array()
+            .into_iter()
+            .flatten()
+            .map(|dependency| dependency["source_id"].as_str().unwrap().to_owned()),
+    );
+    for source_id in source_ids {
+        let lifecycle: String = sqlx::query_scalar(
+            "select lifecycle from artifact_sources where workspace_id = $1 and source_id = $2",
+        )
+        .bind(workspace_id.as_str())
+        .bind(source_id)
+        .fetch_one(registry.pool())
+        .await
+        .unwrap();
+        assert_eq!(lifecycle, "detached");
+    }
+}

apps/admin-api/tests/integration/openapi_import/external_references.rs

@@ -0,0 +1,538 @@
+use super::super::common::test_service_with_external_references;
+use super::*;
+use axum::{Router, routing::get};
+use std::sync::{
+    Arc,
+    atomic::{AtomicUsize, Ordering},
+};
+
+#[tokio::test]
+#[serial]
+async fn external_relative_chain_survives_reversed_builder_order_and_apply_never_refetches() {
+    let fetches = Arc::new(AtomicUsize::new(0));
+    let root_fetches = Arc::clone(&fetches);
+    let child_fetches = Arc::clone(&fetches);
+    let app = Router::new()
+        .route(
+            "/root.yaml",
+            get(move || {
+                let fetches = Arc::clone(&root_fetches);
+                async move {
+                    fetches.fetch_add(1, Ordering::SeqCst);
+                    "Item: { $ref: './child.yaml#/Item' }"
+                }
+            }),
+        )
+        .route(
+            "/child.yaml",
+            get(move || {
+                let fetches = Arc::clone(&child_fetches);
+                async move {
+                    fetches.fetch_add(1, Ordering::SeqCst);
+                    "Item: { type: object, required: [id], properties: { id: { type: string } } }"
+                }
+            }),
+        );
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    let origin = format!("http://{address}");
+
+    let registry = test_registry().await;
+    let service = test_service_with_external_references(
+        registry.clone(),
+        test_storage_root("openapi_import_external_snapshot"),
+        vec![format!("{origin}/")],
+    );
+    let workspace_id = WorkspaceId::new("ws_default");
+    let upload = OpenApiUpload {
+        bytes: format!(
+            r#"
+openapi: 3.1.0
+info: {{ title: External }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content:
+            application/json:
+              schema: {{ $ref: '{origin}/root.yaml#/Item' }}
+"#
+        )
+        .into_bytes(),
+        mime_type: "application/yaml".to_owned(),
+        locale: OpenApiUploadLocale::En,
+    };
+    let preview = service
+        .preview_openapi_import(&workspace_id, upload)
+        .await
+        .unwrap();
+    assert_eq!(fetches.load(Ordering::SeqCst), 2);
+    assert_eq!(preview.preview.groups[0].operations[0].output_fields, 1);
+    let job = registry
+        .get_import_job(&workspace_id, &preview.job_id.as_str().into())
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(
+        job.preview_payload["dependencies"]
+            .as_array()
+            .unwrap()
+            .len(),
+        2
+    );
+    assert_eq!(
+        job.preview_payload["dependency_snapshots"]
+            .as_array()
+            .unwrap()
+            .len(),
+        2
+    );
+
+    let applied = service
+        .create_openapi_import(
+            &workspace_id,
+            &preview.job_id.as_str().into(),
+            OpenApiImportCreatePayload {
+                selected_operation_keys: vec!["GET /items".to_owned()],
+                server_url: None,
+                conflict_mode: "skip".to_owned(),
+            },
+        )
+        .await
+        .unwrap();
+    assert_eq!(applied.created.len(), 1);
+    assert_eq!(fetches.load(Ordering::SeqCst), 2);
+    let active_sources: i64 = sqlx::query_scalar(
+        "select count(*) from artifact_sources
+          where source_id like 'src_openapi_%' and lifecycle = 'active'",
+    )
+    .fetch_one(registry.pool())
+    .await
+    .unwrap();
+    assert_eq!(active_sources, 0);
+}
+
+#[tokio::test]
+#[serial]
+async fn missing_external_snapshot_fails_closed_before_draft_mutation() {
+    let app = Router::new().route(
+        "/schemas.yaml",
+        get(|| async { "Item: { type: object, properties: { id: { type: string } } }" }),
+    );
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    let origin = format!("http://{address}");
+    let registry = test_registry().await;
+    let storage_root = test_storage_root("openapi_import_missing_external_snapshot");
+    let service = test_service_with_external_references(
+        registry.clone(),
+        storage_root.clone(),
+        vec![format!("{origin}/")],
+    );
+    let workspace_id = WorkspaceId::new("ws_default");
+    let preview = service
+        .preview_openapi_import(
+            &workspace_id,
+            OpenApiUpload {
+                bytes: format!(
+                    r#"
+openapi: 3.1.0
+info: {{ title: External integrity }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content: {{ application/json: {{ schema: {{ $ref: '{origin}/schemas.yaml#/Item' }} }} }}
+"#
+                )
+                .into_bytes(),
+                mime_type: "application/yaml".to_owned(),
+                locale: OpenApiUploadLocale::En,
+            },
+        )
+        .await
+        .unwrap();
+    let job = registry
+        .get_import_job(&workspace_id, &preview.job_id.as_str().into())
+        .await
+        .unwrap()
+        .unwrap();
+    let dependency_source_id = job.preview_payload["dependencies"][0]["source_id"]
+        .as_str()
+        .unwrap();
+    sqlx::query(
+        "update artifact_sources
+            set lifecycle = 'detached', updated_at = now(), detached_at = now()
+          where workspace_id = $1 and source_id = $2",
+    )
+    .bind(workspace_id.as_str())
+    .bind(dependency_source_id)
+    .execute(registry.pool())
+    .await
+    .unwrap();
+
+    assert!(
+        service
+            .create_openapi_import(
+                &workspace_id,
+                &preview.job_id.as_str().into(),
+                OpenApiImportCreatePayload {
+                    selected_operation_keys: vec!["GET /items".to_owned()],
+                    server_url: None,
+                    conflict_mode: "skip".to_owned(),
+                },
+            )
+            .await
+            .is_err()
+    );
+    assert!(
+        service
+            .list_operations(&workspace_id)
+            .await
+            .unwrap()
+            .is_empty()
+    );
+    assert_failed_job_and_detached_sources(
+        &registry,
+        &workspace_id,
+        &preview.job_id.as_str().into(),
+        "import_dependency_verification_failed",
+    )
+    .await;
+
+    let corrupt_preview = service
+        .preview_openapi_import(
+            &workspace_id,
+            OpenApiUpload {
+                bytes: format!(
+                    r#"
+openapi: 3.1.0
+info: {{ title: Corrupt external integrity }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: listItemsCorrupt
+      responses:
+        '200':
+          description: ok
+          content: {{ application/json: {{ schema: {{ $ref: '{origin}/schemas.yaml#/Item' }} }} }}
+"#
+                )
+                .into_bytes(),
+                mime_type: "application/yaml".to_owned(),
+                locale: OpenApiUploadLocale::En,
+            },
+        )
+        .await
+        .unwrap();
+    let corrupt_job = registry
+        .get_import_job(&workspace_id, &corrupt_preview.job_id.as_str().into())
+        .await
+        .unwrap()
+        .unwrap();
+    let dependency_ref: crank_artifacts::ArtifactRef = corrupt_job.preview_payload["dependencies"]
+        [0]["digest"]
+        .as_str()
+        .unwrap()
+        .parse()
+        .unwrap();
+    let dependency_path = storage_root
+        .join("sha256")
+        .join(&dependency_ref.digest_hex()[..2])
+        .join(dependency_ref.digest_hex());
+    std::fs::set_permissions(
+        &dependency_path,
+        std::os::unix::fs::PermissionsExt::from_mode(0o600),
+    )
+    .unwrap();
+    std::fs::write(&dependency_path, b"corrupt external dependency").unwrap();
+    std::fs::set_permissions(
+        &dependency_path,
+        std::os::unix::fs::PermissionsExt::from_mode(0o400),
+    )
+    .unwrap();
+
+    assert!(
+        service
+            .create_openapi_import(
+                &workspace_id,
+                &corrupt_preview.job_id.as_str().into(),
+                OpenApiImportCreatePayload {
+                    selected_operation_keys: vec!["GET /items".to_owned()],
+                    server_url: None,
+                    conflict_mode: "skip".to_owned(),
+                },
+            )
+            .await
+            .is_err()
+    );
+    assert_failed_job_and_detached_sources(
+        &registry,
+        &workspace_id,
+        &corrupt_preview.job_id.as_str().into(),
+        "import_dependency_verification_failed",
+    )
+    .await;
+}
+
+#[tokio::test]
+#[serial]
+async fn external_materialization_uses_one_wall_clock_timeout_and_cleans_cancelled_sources() {
+    let fetches = Arc::new(AtomicUsize::new(0));
+    let root_fetches = Arc::clone(&fetches);
+    let child_fetches = Arc::clone(&fetches);
+    let app = Router::new()
+        .route(
+            "/root.yaml",
+            get(move || {
+                let fetches = Arc::clone(&root_fetches);
+                async move {
+                    fetches.fetch_add(1, Ordering::SeqCst);
+                    "Item: { $ref: './slow-child.yaml#/Item' }"
+                }
+            }),
+        )
+        .route(
+            "/slow-child.yaml",
+            get(move || {
+                let fetches = Arc::clone(&child_fetches);
+                async move {
+                    fetches.fetch_add(1, Ordering::SeqCst);
+                    tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                    "Item: { type: object, properties: { id: { type: string } } }"
+                }
+            }),
+        );
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    let origin = format!("http://{address}");
+
+    let registry = test_registry().await;
+    let outbound_policy = crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]);
+    let runtime = crank_runtime::community_with_outbound_policy(outbound_policy.clone()).build();
+    let service = AdminServiceBuilder::new(
+        registry.clone(),
+        test_storage_root("openapi_import_chain_timeout"),
+        test_auth_settings(),
+        test_secret_crypto(),
+        runtime,
+    )
+    .with_external_reference_import(&crank_config::ExternalReferenceSettings {
+        allowed_url_prefixes: vec![format!("{origin}/")],
+        max_depth: 8,
+        max_documents: 32,
+        max_fetch_bytes: 64 * 1024,
+        fetch_timeout_ms: 200,
+        max_expanded_nodes: 10_000,
+    })
+    .unwrap()
+    .with_outbound_http_policy(outbound_policy)
+    .build();
+    let workspace_id = WorkspaceId::new("ws_default");
+    let started = tokio::time::Instant::now();
+    let preview = service
+        .preview_openapi_import(
+            &workspace_id,
+            OpenApiUpload {
+                bytes: format!(
+                    r#"
+openapi: 3.1.0
+info: {{ title: Timed chain }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: timedItems
+      responses:
+        '200':
+          description: ok
+          content: {{ application/json: {{ schema: {{ $ref: '{origin}/root.yaml#/Item' }} }} }}
+"#
+                )
+                .into_bytes(),
+                mime_type: "application/yaml".to_owned(),
+                locale: OpenApiUploadLocale::En,
+            },
+        )
+        .await
+        .unwrap();
+    assert!(started.elapsed() < std::time::Duration::from_millis(800));
+    assert_eq!(fetches.load(Ordering::SeqCst), 2);
+    let job = registry
+        .get_import_job(&workspace_id, &preview.job_id.as_str().into())
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(
+        job.preview_payload["dependencies"]
+            .as_array()
+            .unwrap()
+            .len(),
+        0
+    );
+
+    for _ in 0..100 {
+        let active_dependencies: i64 = sqlx::query_scalar(
+            "select count(*) from artifact_sources
+              where source_id like 'src_openapi_dep_%' and lifecycle = 'active'",
+        )
+        .fetch_one(registry.pool())
+        .await
+        .unwrap();
+        if active_dependencies == 0 {
+            break;
+        }
+        tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+    }
+    let detached_dependencies: i64 = sqlx::query_scalar(
+        "select count(*) from artifact_sources
+          where source_id like 'src_openapi_dep_%' and lifecycle = 'detached'",
+    )
+    .fetch_one(registry.pool())
+    .await
+    .unwrap();
+    assert_eq!(detached_dependencies, 1);
+}
+
+#[tokio::test]
+#[serial]
+async fn expired_job_detaches_primary_and_external_dependency_without_orphan() {
+    let app = Router::new().route(
+        "/schemas.yaml",
+        get(|| async { "Item: { type: object, properties: { id: { type: string } } }" }),
+    );
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    let origin = format!("http://{address}");
+    let registry = test_registry().await;
+    let service = test_service_with_external_references(
+        registry.clone(),
+        test_storage_root("openapi_import_external_expiry"),
+        vec![format!("{origin}/")],
+    );
+    let workspace_id = WorkspaceId::new("ws_default");
+    let preview = service
+        .preview_openapi_import(
+            &workspace_id,
+            OpenApiUpload {
+                bytes: format!(
+                    r#"
+openapi: 3.1.0
+info: {{ title: Expiring external }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content: {{ application/json: {{ schema: {{ $ref: '{origin}/schemas.yaml#/Item' }} }} }}
+"#
+                )
+                .into_bytes(),
+                mime_type: "application/yaml".to_owned(),
+                locale: OpenApiUploadLocale::En,
+            },
+        )
+        .await
+        .unwrap();
+    sqlx::query("update import_jobs set expires_at = now() - interval '1 second' where id = $1")
+        .bind(preview.job_id.as_str())
+        .execute(registry.pool())
+        .await
+        .unwrap();
+    let report = registry.cleanup_expired_import_jobs(16).await.unwrap();
+    assert_eq!(report.deleted_jobs, 1);
+    assert_eq!(report.detached_sources, 2);
+    let active_sources: i64 = sqlx::query_scalar(
+        "select count(*) from artifact_sources
+          where source_id like 'src_openapi_%' and lifecycle = 'active'",
+    )
+    .fetch_one(registry.pool())
+    .await
+    .unwrap();
+    assert_eq!(active_sources, 0);
+}
+#[tokio::test]
+#[serial]
+async fn exact_v2_contract_replays_persisted_preview_without_v3_or_dependency_reads() {
+    let registry = test_registry().await;
+    let service = test_service(
+        registry.clone(),
+        test_storage_root("openapi_import_v2_compatibility"),
+        test_auth_settings(),
+        test_secret_crypto(),
+    );
+    let workspace_id = WorkspaceId::new("ws_default");
+    let preview = service
+        .preview_openapi_import(&workspace_id, openapi_upload())
+        .await
+        .unwrap();
+    let job_id: crank_registry::ImportJobId = preview.job_id.as_str().into();
+    let missing_dependency = serde_json::json!([{
+        "source_id": "src_openapi_dep_v2_must_not_be_read",
+        "digest": format!("sha256:{}", "0".repeat(64)),
+        "canonical_uri": "https://schemas.example.test/v2.yaml"
+    }]);
+    sqlx::query(
+        "update import_jobs
+            set preview_payload = jsonb_set(
+                jsonb_set(
+                    jsonb_set(
+                        jsonb_set(
+                            preview_payload,
+                            '{normalization,normalizer_version}',
+                            to_jsonb('normalized-ir-v2'::text)
+                        ),
+                        '{normalization,projection_version}',
+                        to_jsonb('preview-v2'::text)
+                    ),
+                    '{normalization,ir_fingerprint}',
+                    to_jsonb($1::text)
+                ),
+                '{dependency_snapshots}',
+                $2::jsonb
+            )
+          where id = $3",
+    )
+    .bind("a".repeat(64))
+    .bind(missing_dependency)
+    .bind(job_id.as_str())
+    .execute(registry.pool())
+    .await
+    .unwrap();
+
+    let result = service
+        .create_openapi_import(
+            &workspace_id,
+            &job_id,
+            OpenApiImportCreatePayload {
+                selected_operation_keys: vec!["GET /v2/latest".to_owned()],
+                server_url: Some("https://api.frankfurter.dev".to_owned()),
+                conflict_mode: "skip".to_owned(),
+            },
+        )
+        .await
+        .unwrap();
+    assert_eq!(result.created.len(), 1);
+    let completed = registry
+        .get_import_job(&workspace_id, &job_id)
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(completed.status, ImportJobStatus::Completed);
+}

apps/admin-api/tests/integration/openapi_source.rs

@@ -1,30 +1,18 @@
-use std::{
-    io,
-    sync::{Arc, Mutex},
-};
-
-use admin_api::service::{OpenApiImportCreatePayload, OpenApiUpload, OpenApiUploadLocale};
-use crank_artifacts::MAX_ARTIFACT_BYTES;
-use crank_core::{Workspace, WorkspaceId, WorkspaceStatus};
-use crank_registry::{ArtifactSourceId, ArtifactSourceLifecycle, CreateWorkspaceRequest};
-use metrics_util::debugging::DebuggingRecorder;
-use opentelemetry::trace::TracerProvider as _;
-use opentelemetry_sdk::{
-    error::OTelSdkResult,
-    trace::{SdkTracerProvider, SpanData, SpanExporter},
-};
-use reqwest::multipart::{Form, Part};
-use serde_json::Value;
-use serial_test::serial;
-use time::{Duration, OffsetDateTime};
-use tracing_subscriber::{fmt::MakeWriter, layer::SubscriberExt};
-
 use super::common::{
     authorized_client, build_test_app, spawn_admin_api, test_auth_settings, test_registry,
     test_secret_crypto, test_service, test_storage_root,
 };
+use admin_api::service::{OpenApiImportCreatePayload, OpenApiUpload, OpenApiUploadLocale};
+use crank_artifacts::MAX_ARTIFACT_BYTES;
+use crank_core::{Workspace, WorkspaceId, WorkspaceStatus};
+use crank_registry::{ArtifactSourceId, ArtifactSourceLifecycle, CreateWorkspaceRequest};
+use reqwest::multipart::{Form, Part};
+use serde_json::Value;
+use serial_test::serial;
+use time::{Duration, OffsetDateTime};
 
 mod apply_failures;
+mod telemetry;
 
 const OPENAPI: &str = r#"
 openapi: 3.0.3
@@ -136,78 +124,6 @@ async fn multipart_requires_an_owner_membership_before_reading_the_file() {
     );
 }
 
-#[tokio::test(flavor = "multi_thread")]
-#[serial]
-async fn parser_canary_never_reaches_diagnostics_logs_traces_or_metrics() {
-    const CANARY: &str = "openapi-telemetry-secret-canary";
-
-    let recorder = DebuggingRecorder::new();
-    let snapshotter = recorder.snapshotter();
-    recorder
-        .install()
-        .expect("isolated integration test metrics recorder");
-
-    let writer = SharedLogWriter::default();
-    let exported = Arc::new(Mutex::new(Vec::new()));
-    let provider = SdkTracerProvider::builder()
-        .with_simple_exporter(CapturingExporter(Arc::clone(&exported)))
-        .build();
-    let tracer = provider.tracer("admin-openapi-source-test");
-    let subscriber = tracing_subscriber::registry()
-        .with(tracing_subscriber::fmt::layer().with_writer(writer.clone()))
-        .with(tracing_opentelemetry::layer().with_tracer(tracer));
-    let dispatch = tracing::Dispatch::new(subscriber);
-    tracing::dispatcher::set_global_default(dispatch)
-        .expect("isolated integration test tracing subscriber");
-
-    let registry = test_registry().await;
-    let app = build_test_app(registry, test_storage_root("openapi_telemetry_canary"));
-    let server = spawn_admin_api(app).await;
-    let client = authorized_client(&server).await;
-    let document = format!(
-        "openapi: 3.0.3\ninfo: {{ title: Canary }}\npaths:\n  /broken:\n    get:\n      description: {CANARY}\n      responses: ["
-    );
-    let response = client
-        .post(format!("{server}/imports/openapi/preview"))
-        .multipart(Form::new().part(
-            "file",
-            file_part(document.as_bytes(), "openapi.yaml", "application/yaml"),
-        ))
-        .send()
-        .await
-        .unwrap();
-    assert_eq!(response.status(), reqwest::StatusCode::BAD_REQUEST);
-    let trace_id = response
-        .headers()
-        .get("x-trace-id")
-        .unwrap()
-        .to_str()
-        .unwrap()
-        .to_owned();
-    let body = response.text().await.unwrap();
-    assert!(!body.contains(CANARY));
-    let diagnostics: Value = serde_json::from_str(&body).unwrap();
-    assert_eq!(diagnostics["error"]["trace_id"], trace_id);
-
-    provider.force_flush().unwrap();
-    let logs = writer.output();
-    assert!(!logs.contains(CANARY));
-    assert!(logs.contains(&trace_id));
-    let spans = exported.lock().unwrap();
-    let rendered_spans = format!("{spans:?}");
-    assert!(!rendered_spans.contains(CANARY));
-    drop(spans);
-    for (key, _, _, _) in snapshotter.snapshot().into_vec() {
-        assert!(!key.key().name().contains(CANARY));
-        assert!(
-            !key.key()
-                .labels()
-                .any(|label| { label.key().contains(CANARY) || label.value().contains(CANARY) })
-        );
-    }
-    provider.shutdown().unwrap();
-}
-
 #[tokio::test(flavor = "multi_thread")]
 #[serial]
 async fn multipart_boundary_rejections_are_localized_and_leave_no_entities() {
@@ -939,49 +855,3 @@ async fn wait_for_specific_source_lifecycle(
     }
     panic!("artifact source {source_id:?} did not reach {expected}");
 }
-
-#[derive(Clone, Default)]
-struct SharedLogWriter {
-    buffer: Arc<Mutex<Vec<u8>>>,
-}
-
-impl SharedLogWriter {
-    fn output(&self) -> String {
-        String::from_utf8(self.buffer.lock().unwrap().clone()).unwrap()
-    }
-}
-
-impl<'a> MakeWriter<'a> for SharedLogWriter {
-    type Writer = SharedLogGuard;
-
-    fn make_writer(&'a self) -> Self::Writer {
-        SharedLogGuard {
-            buffer: Arc::clone(&self.buffer),
-        }
-    }
-}
-
-struct SharedLogGuard {
-    buffer: Arc<Mutex<Vec<u8>>>,
-}
-
-impl io::Write for SharedLogGuard {
-    fn write(&mut self, bytes: &[u8]) -> io::Result<usize> {
-        self.buffer.lock().unwrap().extend_from_slice(bytes);
-        Ok(bytes.len())
-    }
-
-    fn flush(&mut self) -> io::Result<()> {
-        Ok(())
-    }
-}
-
-#[derive(Clone, Debug)]
-struct CapturingExporter(Arc<Mutex<Vec<SpanData>>>);
-
-impl SpanExporter for CapturingExporter {
-    async fn export(&self, batch: Vec<SpanData>) -> OTelSdkResult {
-        self.0.lock().unwrap().extend(batch);
-        Ok(())
-    }
-}

apps/admin-api/tests/integration/openapi_source/telemetry.rs

@@ -0,0 +1,237 @@
+use super::super::common::build_test_app_with_external_references;
+use super::*;
+use std::{
+    io,
+    sync::{Arc, Mutex},
+};
+
+use metrics_util::debugging::DebuggingRecorder;
+use opentelemetry::trace::TracerProvider as _;
+use opentelemetry_sdk::{
+    error::OTelSdkResult,
+    trace::{SdkTracerProvider, SpanData, SpanExporter},
+};
+use tracing_subscriber::{fmt::MakeWriter, layer::SubscriberExt};
+
+#[tokio::test(flavor = "multi_thread")]
+#[serial]
+async fn parser_canary_never_reaches_diagnostics_logs_traces_or_metrics() {
+    const CANARY: &str = "openapi-telemetry-secret-canary";
+
+    let recorder = DebuggingRecorder::new();
+    let snapshotter = recorder.snapshotter();
+    recorder
+        .install()
+        .expect("isolated integration test metrics recorder");
+
+    let writer = SharedLogWriter::default();
+    let exported = Arc::new(Mutex::new(Vec::new()));
+    let provider = SdkTracerProvider::builder()
+        .with_simple_exporter(CapturingExporter(Arc::clone(&exported)))
+        .build();
+    let tracer = provider.tracer("admin-openapi-source-test");
+    let subscriber = tracing_subscriber::registry()
+        .with(
+            tracing_subscriber::fmt::layer()
+                .with_ansi(false)
+                .with_writer(writer.clone()),
+        )
+        .with(tracing_opentelemetry::layer().with_tracer(tracer));
+    let dispatch = tracing::Dispatch::new(subscriber);
+    tracing::dispatcher::set_global_default(dispatch)
+        .expect("isolated integration test tracing subscriber");
+
+    let registry = test_registry().await;
+    let app = build_test_app(registry, test_storage_root("openapi_telemetry_canary"));
+    let server = spawn_admin_api(app).await;
+    let client = authorized_client(&server).await;
+    let document = format!(
+        "openapi: 3.0.3\ninfo: {{ title: Canary }}\npaths:\n  /broken:\n    get:\n      description: {CANARY}\n      responses: ["
+    );
+    let response = client
+        .post(format!("{server}/imports/openapi/preview"))
+        .multipart(Form::new().part(
+            "file",
+            file_part(document.as_bytes(), "openapi.yaml", "application/yaml"),
+        ))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(response.status(), reqwest::StatusCode::BAD_REQUEST);
+    let trace_id = response
+        .headers()
+        .get("x-trace-id")
+        .unwrap()
+        .to_str()
+        .unwrap()
+        .to_owned();
+    let body = response.text().await.unwrap();
+    assert!(!body.contains(CANARY));
+    let diagnostics: Value = serde_json::from_str(&body).unwrap();
+    assert_eq!(diagnostics["error"]["trace_id"], trace_id);
+
+    let route = format!("/{CANARY}.yaml");
+    let external = axum::Router::new().route(
+        &route,
+        axum::routing::get(|| async { (axum::http::StatusCode::INTERNAL_SERVER_ERROR, CANARY) }),
+    );
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, external).await.unwrap() });
+    let origin = format!("http://{address}");
+    let external_registry = test_registry().await;
+    let external_app = build_test_app_with_external_references(
+        external_registry,
+        test_storage_root("openapi_materialization_telemetry_canary"),
+        vec![format!("{origin}/")],
+    );
+    let external_server = spawn_admin_api(external_app).await;
+    let external_client = authorized_client(&external_server).await;
+    let external_document = format!(
+        r#"
+openapi: 3.1.0
+info: {{ title: Safe materialization }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      responses:
+        '200':
+          description: ok
+          content:
+            application/json:
+              schema: {{ $ref: '{origin}/{CANARY}.yaml#/Item' }}
+"#
+    );
+    let response = external_client
+        .post(format!("{external_server}/imports/openapi/preview"))
+        .multipart(Form::new().part(
+            "file",
+            file_part(
+                external_document.as_bytes(),
+                "external.yaml",
+                "application/yaml",
+            ),
+        ))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(response.status(), reqwest::StatusCode::OK);
+
+    provider.force_flush().unwrap();
+    let logs = writer.output();
+    assert!(!logs.contains(CANARY));
+    assert!(logs.contains(&trace_id));
+    let safe_materialization_failure = logs.lines().any(is_safe_materialization_failure);
+    assert!(
+        safe_materialization_failure,
+        "expected one bounded, sanitized materialization failure"
+    );
+    let spans = exported.lock().unwrap();
+    let rendered_spans = format!("{spans:?}");
+    assert!(!rendered_spans.contains(CANARY));
+    drop(spans);
+    for (key, _, _, _) in snapshotter.snapshot().into_vec() {
+        assert!(!key.key().name().contains(CANARY));
+        assert!(
+            !key.key()
+                .labels()
+                .any(|label| { label.key().contains(CANARY) || label.value().contains(CANARY) })
+        );
+    }
+    provider.shutdown().unwrap();
+}
+
+fn is_safe_materialization_failure(line: &str) -> bool {
+    if !line.contains("external OpenAPI materialization failed") || !line.contains("count=0") {
+        return false;
+    }
+    let immediate_status = contains_log_field(line, "stage", "fetch")
+        && contains_log_field(line, "error_code", "unexpected_status");
+    let bounded_timeout = contains_log_field(line, "stage", "chain")
+        && contains_log_field(line, "error_code", "timeout");
+    immediate_status || bounded_timeout
+}
+
+fn contains_log_field(line: &str, name: &str, value: &str) -> bool {
+    line.contains(&format!("{name}=\"{value}\"")) || line.contains(&format!("{name}={value}"))
+}
+
+#[test]
+fn materialization_log_accepts_only_coherent_safe_outcomes() {
+    for line in [
+        r#"external OpenAPI materialization failed stage="fetch" error_code="unexpected_status" count=0"#,
+        "external OpenAPI materialization failed stage=chain error_code=timeout count=0",
+    ] {
+        assert!(is_safe_materialization_failure(line));
+    }
+    assert!(!is_safe_materialization_failure(
+        r#"external OpenAPI materialization failed stage="fetch" error_code="timeout" count=0"#
+    ));
+}
+
+#[derive(Clone, Default)]
+struct SharedLogWriter {
+    buffer: Arc<Mutex<Vec<u8>>>,
+}
+
+impl SharedLogWriter {
+    fn output(&self) -> String {
+        String::from_utf8(self.buffer.lock().unwrap().clone()).unwrap()
+    }
+}
+
+impl<'a> MakeWriter<'a> for SharedLogWriter {
+    type Writer = SharedLogGuard;
+
+    fn make_writer(&'a self) -> Self::Writer {
+        SharedLogGuard {
+            buffer: Arc::clone(&self.buffer),
+            pending: Vec::new(),
+        }
+    }
+}
+
+struct SharedLogGuard {
+    buffer: Arc<Mutex<Vec<u8>>>,
+    pending: Vec<u8>,
+}
+
+impl io::Write for SharedLogGuard {
+    fn write(&mut self, bytes: &[u8]) -> io::Result<usize> {
+        self.pending.extend_from_slice(bytes);
+        Ok(bytes.len())
+    }
+
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
+    }
+}
+
+impl Drop for SharedLogGuard {
+    fn drop(&mut self) {
+        self.buffer.lock().unwrap().extend_from_slice(&self.pending);
+    }
+}
+
+#[test]
+fn shared_log_writer_publishes_complete_records_only() {
+    let writer = SharedLogWriter::default();
+    let mut guard = writer.make_writer();
+    io::Write::write_all(&mut guard, b"message").unwrap();
+    assert!(writer.output().is_empty());
+
+    io::Write::write_all(&mut guard, b" stage=\"fetch\"\n").unwrap();
+    drop(guard);
+    assert_eq!(writer.output(), "message stage=\"fetch\"\n");
+}
+
+#[derive(Clone, Debug)]
+struct CapturingExporter(Arc<Mutex<Vec<SpanData>>>);
+
+impl SpanExporter for CapturingExporter {
+    async fn export(&self, batch: Vec<SpanData>) -> OTelSdkResult {
+        self.0.lock().unwrap().extend(batch);
+        Ok(())
+    }
+}

apps/admin-api/tests/migration_command.rs

@@ -40,7 +40,7 @@ fn plan_is_deterministic_and_committed_contract_is_current() {
     );
     assert_eq!(first.stdout, second.stdout);
     let plan: serde_json::Value = serde_json::from_slice(&first.stdout).unwrap();
-    assert_eq!(plan["sequence"].as_array().unwrap().len(), 12);
+    assert_eq!(plan["sequence"].as_array().unwrap().len(), 13);
 
     let checked = command(&["plan", "--check"], None);
     assert!(
@@ -82,7 +82,7 @@ async fn database_only_config_can_apply_and_preflight_a_fresh_schema() {
     );
     let result: serde_json::Value = serde_json::from_slice(&preflight.stdout).unwrap();
     assert_eq!(result["status"], "current");
-    assert_eq!(result["version"], 12);
+    assert_eq!(result["version"], 13);
 }
 
 #[tokio::test]

apps/admin-api/tests/request_context_parent.rs

@@ -0,0 +1,39 @@
+use admin_api::request_context::{TRACE_ID_HEADER, apply_request_context};
+use axum::{Router, body::Body, http::Request, routing::get};
+use opentelemetry::trace::TracerProvider as _;
+use opentelemetry_sdk::trace::SdkTracerProvider;
+use tower::ServiceExt;
+use tracing_subscriber::layer::SubscriberExt;
+
+#[tokio::test(flavor = "current_thread")]
+async fn preserves_remote_trace_id_with_an_active_tracer_and_no_global_propagator() {
+    let provider = SdkTracerProvider::builder().build();
+    let tracer = provider.tracer("admin-request-context-parent-test");
+    let subscriber =
+        tracing_subscriber::registry().with(tracing_opentelemetry::layer().with_tracer(tracer));
+    let dispatch = tracing::Dispatch::new(subscriber);
+    let _dispatch_guard = tracing::dispatcher::set_default(&dispatch);
+    let app = Router::new()
+        .route("/probe", get(|| async { "ok" }))
+        .layer(axum::middleware::from_fn(apply_request_context));
+
+    let response = app
+        .oneshot(
+            Request::builder()
+                .uri("/probe")
+                .header(
+                    "traceparent",
+                    "00-0af7651916cd43dd8448eb211c80319c-b7ad6b7169203331-01",
+                )
+                .body(Body::empty())
+                .unwrap(),
+        )
+        .await
+        .unwrap();
+
+    assert_eq!(
+        response.headers()[TRACE_ID_HEADER],
+        "0af7651916cd43dd8448eb211c80319c"
+    );
+    provider.shutdown().unwrap();
+}

apps/mcp-server/Dockerfile

@@ -53,13 +53,15 @@ RUN --mount=type=cache,id=crank-mcp-cargo-registry,target=/usr/local/cargo/regis
 
 FROM debian:bookworm-slim
 
-RUN apt-get update \
+COPY --from=builder /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
-    && apt-get install -y --no-install-recommends ca-certificates curl \
-    && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
 
 COPY --from=builder /tmp/mcp-server /usr/local/bin/mcp-server
+COPY scripts/docker-http-healthcheck.sh /usr/local/bin/crank-http-healthcheck
+
+RUN test -s /etc/ssl/certs/ca-certificates.crt \
+    && chmod 0755 /usr/local/bin/crank-http-healthcheck
 
 ENV CRANK_MCP_BIND=0.0.0.0:3002
 

apps/ui/Dockerfile

@@ -10,7 +10,11 @@ COPY crank-community.png ./crank-community.png
 
 RUN npm run build
 
-FROM nginx:1.27-alpine
+FROM nginx:1.30.4-alpine3.24-slim
+
+RUN apk add --no-cache --upgrade \
+    'libcrypto3>=3.5.8-r0' \
+    'libssl3>=3.5.8-r0'
 
 COPY apps/ui/nginx.conf /etc/nginx/conf.d/default.conf
 COPY --from=build /app/dist /usr/share/nginx/html

apps/ui/css/onboarding.css

@@ -72,6 +72,17 @@
 body:has(.drawer.open) .onboarding-trigger,
 body:has(.drawer.open) .onboarding-panel { z-index: 149; }
 
+/* The wizard owns a fixed bottom action bar. Keep this optional helper above
+   it so the primary Continue action remains both visible and clickable. */
+body.wizard-page .onboarding-trigger {
+  bottom: calc(96px + env(safe-area-inset-bottom, 0px));
+}
+
+body.wizard-page .onboarding-panel {
+  bottom: calc(150px + env(safe-area-inset-bottom, 0px));
+  max-height: min(700px, calc(100vh - 178px));
+}
+
 @media (max-width: 640px) {
   .onboarding-trigger { right: 14px; bottom: 14px; }
   .onboarding-panel { right: 14px; bottom: 68px; }

apps/ui/scripts/playwright-stack.sh

@@ -17,6 +17,30 @@ POSTGRES_HOST="${CRANK_E2E_POSTGRES_HOST:-127.0.0.1}"
 USE_EXTERNAL_POSTGRES="${CRANK_E2E_USE_EXTERNAL_POSTGRES:-0}"
 ADMIN_EMAIL="${CRANK_E2E_ADMIN_EMAIL:-owner@crank.local}"
 ADMIN_PASSWORD="${CRANK_E2E_ADMIN_PASSWORD:-change-me-admin-password}"
+STARTUP_TIMEOUT_SECONDS="${CRANK_E2E_STARTUP_TIMEOUT_SECONDS:-120}"
+
+if [[ ! "$STARTUP_TIMEOUT_SECONDS" =~ ^[1-9][0-9]*$ ]]; then
+  echo "CRANK_E2E_STARTUP_TIMEOUT_SECONDS must be a positive integer" >&2
+  exit 1
+fi
+
+# These variables configure only this orchestration script. The application
+# rejects unknown variables in the owned CRANK_* namespace, so never leak the
+# E2E control plane into production process configuration.
+unset \
+  CRANK_E2E_ADMIN_EMAIL \
+  CRANK_E2E_ADMIN_PASSWORD \
+  CRANK_E2E_ADMIN_PORT \
+  CRANK_E2E_MCP_PORT \
+  CRANK_E2E_POSTGRES_DB \
+  CRANK_E2E_POSTGRES_HOST \
+  CRANK_E2E_POSTGRES_PASSWORD \
+  CRANK_E2E_POSTGRES_PORT \
+  CRANK_E2E_POSTGRES_USER \
+  CRANK_E2E_STARTUP_TIMEOUT_SECONDS \
+  CRANK_E2E_STREAM_FIXTURE_PORT \
+  CRANK_E2E_UI_PORT \
+  CRANK_E2E_USE_EXTERNAL_POSTGRES
 
 if [[ -f "$HOME/.cargo/env" ]]; then
   . "$HOME/.cargo/env"
@@ -32,6 +56,7 @@ if [[ -z "$PYTHON_BIN" ]]; then
 fi
 
 mkdir -p "$LOG_DIR"
+find "$LOG_DIR" -maxdepth 1 -type f -name '*.log' -delete
 
 (
   cd "$ROOT_DIR/apps/ui"
@@ -77,10 +102,23 @@ trap 'exit 130' INT TERM
 
 cleanup
 
+show_startup_log() {
+  local log_file="$1"
+  if [[ -f "$log_file" ]]; then
+    echo "---- $log_file ----" >&2
+    tail -n 120 "$log_file" >&2
+  fi
+}
+
 wait_for_port() {
-  local host="$1"
+  local label="$1"
-  local port="$2"
+  local host="$2"
-  until "$PYTHON_BIN" - "$host" "$port" <<'PY'
+  local port="$3"
+  local deadline=$((SECONDS + STARTUP_TIMEOUT_SECONDS))
+
+  echo "Waiting for $label on $host:$port"
+  while true; do
+    if "$PYTHON_BIN" - "$host" "$port" <<'PY'
 import socket, sys
 sock = socket.socket()
 sock.settimeout(0.5)
@@ -91,13 +129,75 @@ except OSError:
 finally:
     sock.close()
 PY
-  do
+    then
+      return 0
+    fi
+    if (( SECONDS >= deadline )); then
+      echo "Timed out waiting for $label on $host:$port after ${STARTUP_TIMEOUT_SECONDS}s" >&2
+      return 1
+    fi
+    sleep 1
+  done
+}
+
+wait_for_http() {
+  local label="$1"
+  local url="$2"
+  local pid_file="$3"
+  local log_file="$4"
+  local deadline=$((SECONDS + STARTUP_TIMEOUT_SECONDS))
+
+  echo "Waiting for $label at $url"
+  while true; do
+    if "$PYTHON_BIN" - "$url" <<'PY'
+import sys, urllib.request
+try:
+    with urllib.request.urlopen(sys.argv[1], timeout=0.5) as response:
+        sys.exit(0 if 200 <= response.status < 400 else 1)
+except Exception:
+    sys.exit(1)
+PY
+    then
+      return 0
+    fi
+
+    local pid=""
+    if [[ -f "$pid_file" ]]; then
+      pid="$(cat "$pid_file")"
+    fi
+    if [[ -z "$pid" ]] || ! kill -0 "$pid" >/dev/null 2>&1; then
+      echo "$label exited before becoming ready at $url" >&2
+      show_startup_log "$log_file"
+      return 1
+    fi
+    if (( SECONDS >= deadline )); then
+      echo "Timed out waiting for $label at $url after ${STARTUP_TIMEOUT_SECONDS}s" >&2
+      show_startup_log "$log_file"
+      return 1
+    fi
+    sleep 1
+  done
+}
+
+wait_for_postgres_container() {
+  local deadline=$((SECONDS + STARTUP_TIMEOUT_SECONDS))
+  echo "Waiting for PostgreSQL container $POSTGRES_CONTAINER"
+  while ! docker exec "$POSTGRES_CONTAINER" pg_isready -U "$POSTGRES_USER" -d "$POSTGRES_DB" >/dev/null 2>&1; do
+    if ! docker inspect "$POSTGRES_CONTAINER" >/dev/null 2>&1; then
+      echo "PostgreSQL container $POSTGRES_CONTAINER exited before becoming ready" >&2
+      return 1
+    fi
+    if (( SECONDS >= deadline )); then
+      echo "Timed out waiting for PostgreSQL container after ${STARTUP_TIMEOUT_SECONDS}s" >&2
+      docker logs "$POSTGRES_CONTAINER" >&2 || true
+      return 1
+    fi
     sleep 1
   done
 }
 
 if [[ "$USE_EXTERNAL_POSTGRES" = "1" ]]; then
-  wait_for_port "$POSTGRES_HOST" "$POSTGRES_PORT"
+  wait_for_port "PostgreSQL" "$POSTGRES_HOST" "$POSTGRES_PORT" || exit 1
 else
   docker run -d --rm \
     --name "$POSTGRES_CONTAINER" \
@@ -107,9 +207,7 @@ else
     -p "$POSTGRES_PORT:5432" \
     postgres:16-alpine >/dev/null
 
-  until docker exec "$POSTGRES_CONTAINER" pg_isready -U "$POSTGRES_USER" -d "$POSTGRES_DB" >/dev/null 2>&1; do
+  wait_for_postgres_container || exit 1
-    sleep 1
-  done
 fi
 
 export POSTGRES_HOST
@@ -141,40 +239,58 @@ chmod 700 "$CRANK_STORAGE_ROOT"
 ) &
 echo $! > "$TMP_DIR/http-fixture.pid"
 
-until curl -fsS "http://127.0.0.1:$STREAM_FIXTURE_PORT/health" >/dev/null 2>&1; do
+wait_for_http \
-  sleep 1
+  "Playwright HTTP fixture" \
-done
+  "http://127.0.0.1:$STREAM_FIXTURE_PORT/health" \
+  "$TMP_DIR/http-fixture.pid" \
+  "$LOG_DIR/http-fixture.log" || exit 1
 
-(
+if ! (
   cd "$ROOT_DIR"
   cargo run -p admin-api --bin crank-migrate -- apply >"$LOG_DIR/migrate.log" 2>&1
-)
+); then
+  echo "Failed to apply E2E database migrations" >&2
+  show_startup_log "$LOG_DIR/migrate.log"
+  exit 1
+fi
 
-BOOTSTRAP_JSON="$(
+if ! (
   cd "$ROOT_DIR"
   cargo run -p admin-api --bin crank-migrate -- admin-auth bootstrap-create \
     --email "$ADMIN_EMAIL" \
     --display-name "Crank E2E" \
     >"$LOG_DIR/bootstrap-create.log" 2>&1
-  tail -n 1 "$LOG_DIR/bootstrap-create.log"
+); then
-)"
+  echo "Failed to create the E2E bootstrap contract" >&2
-BOOTSTRAP_TOKEN="$("$PYTHON_BIN" - <<'PY' "$BOOTSTRAP_JSON"
+  show_startup_log "$LOG_DIR/bootstrap-create.log"
+  exit 1
+fi
+BOOTSTRAP_JSON="$(tail -n 1 "$LOG_DIR/bootstrap-create.log")"
+if ! BOOTSTRAP_TOKEN="$("$PYTHON_BIN" - <<'PY' "$BOOTSTRAP_JSON"
 import json, sys
 print(json.loads(sys.argv[1])["bootstrap_token"])
 PY
-)"
+)"; then
+  echo "Failed to parse the E2E bootstrap token" >&2
+  show_startup_log "$LOG_DIR/bootstrap-create.log"
+  exit 1
+fi
 printf '%s' "$BOOTSTRAP_TOKEN" >"$TMP_DIR/bootstrap-token.txt"
 printf '%s' "$ADMIN_PASSWORD" >"$TMP_DIR/admin-password.txt"
 printf '%s' "$CRANK_PASSWORD_PEPPER" >"$TMP_DIR/password-pepper.txt"
 
-(
+if ! (
   cd "$ROOT_DIR"
   cargo run -p admin-api --bin crank-migrate -- admin-auth bootstrap-complete \
     --token-file "$TMP_DIR/bootstrap-token.txt" \
     --password-file "$TMP_DIR/admin-password.txt" \
     --password-pepper-file "$TMP_DIR/password-pepper.txt" \
     >"$LOG_DIR/bootstrap-complete.log" 2>&1
-)
+); then
+  echo "Failed to complete the E2E admin bootstrap" >&2
+  show_startup_log "$LOG_DIR/bootstrap-complete.log"
+  exit 1
+fi
 
 (
   cd "$ROOT_DIR"
@@ -183,9 +299,11 @@ printf '%s' "$CRANK_PASSWORD_PEPPER" >"$TMP_DIR/password-pepper.txt"
 ) &
 echo $! > "$TMP_DIR/admin-api.pid"
 
-until curl -fsS "http://127.0.0.1:$ADMIN_PORT/health" >/dev/null 2>&1; do
+wait_for_http \
-  sleep 1
+  "admin-api" \
-done
+  "http://127.0.0.1:$ADMIN_PORT/health" \
+  "$TMP_DIR/admin-api.pid" \
+  "$LOG_DIR/admin-api.log" || exit 1
 
 (
   cd "$ROOT_DIR"
@@ -196,9 +314,11 @@ done
 ) &
 echo $! > "$TMP_DIR/mcp-server.pid"
 
-until curl -fsS "http://127.0.0.1:$MCP_PORT/health" >/dev/null 2>&1; do
+wait_for_http \
-  sleep 1
+  "mcp-server" \
-done
+  "http://127.0.0.1:$MCP_PORT/health" \
+  "$TMP_DIR/mcp-server.pid" \
+  "$LOG_DIR/mcp-server.log" || exit 1
 
 (
   cd "$ROOT_DIR/apps/ui"
@@ -206,9 +326,11 @@ done
 ) &
 echo $! > "$TMP_DIR/ui-server.pid"
 
-until curl -fsS "http://127.0.0.1:$UI_PORT/login" >/dev/null 2>&1; do
+wait_for_http \
-  sleep 1
+  "Playwright UI server" \
-done
+  "http://127.0.0.1:$UI_PORT/login" \
+  "$TMP_DIR/ui-server.pid" \
+  "$LOG_DIR/ui-server.log" || exit 1
 
 echo "Crank UI e2e stack is ready on http://127.0.0.1:$UI_PORT"
 

apps/ui/tests/e2e/wizard.spec.js

@@ -1,6 +1,25 @@
 const { test, expect } = require('@playwright/test');
 const { getCurrentWorkspace, login, localized } = require('./helpers');
 
+async function expectOnboardingOutsideWizardNavigation(page) {
+  await expect(page.getByTestId('onboarding-trigger')).toBeVisible();
+
+  const overlapsContinue = await page.locator('#btn-continue').evaluate((continueButton) => {
+    const trigger = document.querySelector('[data-testid="onboarding-trigger"]');
+    if (!trigger) return false;
+    const continueRect = continueButton.getBoundingClientRect();
+    const triggerRect = trigger.getBoundingClientRect();
+    return !(
+      triggerRect.right <= continueRect.left
+      || triggerRect.left >= continueRect.right
+      || triggerRect.bottom <= continueRect.top
+      || triggerRect.top >= continueRect.bottom
+    );
+  });
+
+  expect(overlapsContinue).toBe(false);
+}
+
 test('mobile wizard progress connector crosses the indicator centers', async ({ page }) => {
   await page.setViewportSize({ width: 720, height: 900 });
   await login(page);
@@ -24,6 +43,19 @@ test('mobile wizard progress connector crosses the indicator centers', async ({
   expect(geometry.topDelta).toBeLessThan(1);
 });
 
+test('onboarding helper does not cover wizard navigation', async ({ page }) => {
+  await login(page);
+  await page.goto('/wizard/');
+  await expectOnboardingOutsideWizardNavigation(page);
+});
+
+test('onboarding helper does not cover wizard navigation on mobile', async ({ page }) => {
+  await page.setViewportSize({ width: 390, height: 844 });
+  await login(page);
+  await page.goto('/wizard/');
+  await expectOnboardingOutsideWizardNavigation(page);
+});
+
 test('wizard loads and protocol selection updates flow', async ({ page }) => {
   await login(page);
   await page.goto('/wizard/');

crates/crank-adapter-rest/src/client.rs

@@ -25,7 +25,7 @@ use serde_json::Value;
 use tracing::{Instrument, Span};
 use tracing_opentelemetry::OpenTelemetrySpanExt;
 
-use crate::{RestAdapterError, RestRequest, RestResponse};
+use crate::{ExternalReferenceFetchError, RestAdapterError, RestRequest, RestResponse};
 
 #[derive(Clone, Debug)]
 pub struct RestAdapter {
@@ -33,6 +33,18 @@ pub struct RestAdapter {
     policy: OutboundHttpPolicy,
 }
 
+/// A deliberately separate, GET-only boundary for materializing external
+/// OpenAPI documents. It has no execution metrics, tracing propagation, or
+/// request construction semantics from [`RestAdapter`].
+#[derive(Clone, Debug)]
+pub struct ExternalReferenceFetcher {
+    client: Result<Client, Arc<str>>,
+    policy: OutboundHttpPolicy,
+    allowed_url_prefixes: Vec<String>,
+    max_response_bytes: usize,
+    timeout: Duration,
+}
+
 #[derive(Clone, Debug, PartialEq, Eq)]
 pub struct OutboundHttpPolicy {
     allowed_hosts: Vec<String>,
@@ -56,15 +68,7 @@ impl RestAdapter {
     }
 
     pub fn with_policy(policy: OutboundHttpPolicy) -> Self {
-        let resolver = Arc::new(PolicyDnsResolver {
+        let client = outbound_client(&policy);
-            policy: policy.clone(),
-        });
-        let client = Client::builder()
-            .redirect(redirect::Policy::none())
-            .no_proxy()
-            .dns_resolver(resolver)
-            .build()
-            .map_err(|error| Arc::<str>::from(error.to_string()));
 
         Self { client, policy }
     }
@@ -195,6 +199,153 @@ impl RestAdapter {
     }
 }
 
+impl ExternalReferenceFetcher {
+    pub fn try_new(
+        policy: OutboundHttpPolicy,
+        allowed_url_prefixes: Vec<String>,
+        max_response_bytes: usize,
+        timeout: Duration,
+    ) -> Result<Self, ExternalReferenceFetchError> {
+        if max_response_bytes == 0 || timeout.is_zero() {
+            return Err(ExternalReferenceFetchError::InvalidConfiguration);
+        }
+        let allowed_url_prefixes = allowed_url_prefixes
+            .into_iter()
+            .map(|prefix| canonical_external_reference_prefix(&prefix))
+            .collect::<Result<Vec<_>, _>>()?;
+        Ok(Self {
+            client: outbound_client(&policy),
+            policy,
+            allowed_url_prefixes,
+            max_response_bytes,
+            timeout,
+        })
+    }
+
+    /// Fetches one document with a bounded, headerless `GET`.
+    ///
+    /// URL fragments are stripped because they address JSON Pointer targets in
+    /// the fetched document rather than a network resource.
+    pub async fn get(&self, url: &str) -> Result<Vec<u8>, ExternalReferenceFetchError> {
+        if self.allowed_url_prefixes.is_empty() {
+            return Err(ExternalReferenceFetchError::Disabled);
+        }
+        let mut url =
+            reqwest::Url::parse(url).map_err(|_| ExternalReferenceFetchError::InvalidUrl)?;
+        url.set_fragment(None);
+        if !self
+            .allowed_url_prefixes
+            .iter()
+            .any(|prefix| matches_external_reference_prefix(&url, prefix))
+        {
+            return Err(ExternalReferenceFetchError::TargetNotAllowed);
+        }
+        self.policy
+            .validate_url(&url)
+            .map_err(external_policy_error)?;
+        let client = self
+            .client
+            .as_ref()
+            .map_err(|_| ExternalReferenceFetchError::InvalidConfiguration)?;
+        let response = client
+            .get(url)
+            .timeout(self.timeout)
+            .send()
+            .await
+            .map_err(external_transport_error)?;
+        let status = response.status();
+        if status.is_redirection() {
+            return Err(ExternalReferenceFetchError::RedirectNotAllowed);
+        }
+        if !status.is_success() {
+            return Err(ExternalReferenceFetchError::UnexpectedStatus {
+                status: status.as_u16(),
+            });
+        }
+        read_external_response_bytes(response, self.max_response_bytes).await
+    }
+}
+
+fn outbound_client(policy: &OutboundHttpPolicy) -> Result<Client, Arc<str>> {
+    let resolver = Arc::new(PolicyDnsResolver {
+        policy: policy.clone(),
+    });
+    Client::builder()
+        .redirect(redirect::Policy::none())
+        .no_proxy()
+        .dns_resolver(resolver)
+        .build()
+        .map_err(|error| Arc::<str>::from(error.to_string()))
+}
+
+fn canonical_external_reference_prefix(
+    prefix: &str,
+) -> Result<String, ExternalReferenceFetchError> {
+    let url = reqwest::Url::parse(prefix)
+        .map_err(|_| ExternalReferenceFetchError::InvalidConfiguration)?;
+    if !matches!(url.scheme(), "http" | "https")
+        || url.host_str().is_none()
+        || !url.username().is_empty()
+        || url.password().is_some()
+        || url.query().is_some()
+        || url.fragment().is_some()
+    {
+        return Err(ExternalReferenceFetchError::InvalidConfiguration);
+    }
+    Ok(url.to_string())
+}
+
+fn matches_external_reference_prefix(url: &reqwest::Url, prefix: &str) -> bool {
+    let url = url.as_str();
+    if !url.starts_with(prefix) {
+        return false;
+    }
+    let Some(next) = url.as_bytes().get(prefix.len()) else {
+        return true;
+    };
+    prefix.ends_with('/') || matches!(next, b'/' | b'?')
+}
+
+fn external_policy_error(error: RestAdapterError) -> ExternalReferenceFetchError {
+    match error {
+        RestAdapterError::TargetNotAllowed { .. } => ExternalReferenceFetchError::TargetNotAllowed,
+        _ => ExternalReferenceFetchError::InvalidConfiguration,
+    }
+}
+
+fn external_transport_error(error: reqwest::Error) -> ExternalReferenceFetchError {
+    ExternalReferenceFetchError::Transport {
+        timeout: error.is_timeout(),
+        connect: error.is_connect(),
+    }
+}
+
+async fn read_external_response_bytes(
+    response: reqwest::Response,
+    max_response_bytes: usize,
+) -> Result<Vec<u8>, ExternalReferenceFetchError> {
+    if response
+        .content_length()
+        .is_some_and(|length| length > max_response_bytes as u64)
+    {
+        return Err(ExternalReferenceFetchError::ResponseTooLarge {
+            limit_bytes: max_response_bytes,
+        });
+    }
+    let mut stream = response.bytes_stream();
+    let mut bytes = Vec::new();
+    while let Some(chunk) = stream.next().await {
+        let chunk = chunk.map_err(external_transport_error)?;
+        if bytes.len().saturating_add(chunk.len()) > max_response_bytes {
+            return Err(ExternalReferenceFetchError::ResponseTooLarge {
+                limit_bytes: max_response_bytes,
+            });
+        }
+        bytes.extend_from_slice(&chunk);
+    }
+    Ok(bytes)
+}
+
 fn upstream_outcome(error: &RestAdapterError) -> UpstreamOutcome {
     match error {
         RestAdapterError::UnexpectedStatus { status, .. } if (400..500).contains(status) => {

crates/crank-adapter-rest/src/error.rs

@@ -1,6 +1,26 @@
 use serde_json::Value;
 use thiserror::Error;
 
+#[derive(Debug, Error)]
+pub enum ExternalReferenceFetchError {
+    #[error("external references are disabled")]
+    Disabled,
+    #[error("external reference URL is invalid")]
+    InvalidUrl,
+    #[error("external reference target is not allowed")]
+    TargetNotAllowed,
+    #[error("external reference redirects are not allowed")]
+    RedirectNotAllowed,
+    #[error("external reference response exceeds the configured limit of {limit_bytes} bytes")]
+    ResponseTooLarge { limit_bytes: usize },
+    #[error("external reference endpoint returned status {status}")]
+    UnexpectedStatus { status: u16 },
+    #[error("external reference request failed")]
+    Transport { timeout: bool, connect: bool },
+    #[error("external reference fetch configuration is invalid")]
+    InvalidConfiguration,
+}
+
 #[derive(Debug, Error)]
 pub enum RestAdapterError {
     #[error("invalid base url: {url}")]

crates/crank-adapter-rest/src/lib.rs

@@ -8,8 +8,8 @@ use crank_core::{
     ProtocolAdapterError, RestTarget, RuntimeRequestContext, Target,
 };
 
-pub use client::{OutboundHttpPolicy, RestAdapter};
+pub use client::{ExternalReferenceFetcher, OutboundHttpPolicy, RestAdapter};
-pub use error::RestAdapterError;
+pub use error::{ExternalReferenceFetchError, RestAdapterError};
 pub use model::{RestRequest, RestResponse};
 
 #[async_trait]

crates/crank-adapter-rest/tests/integration.rs

@@ -1,4 +1,5 @@
 mod integration {
     mod client;
+    mod external_reference_fetcher;
     mod outbound_security;
 }

crates/crank-adapter-rest/tests/integration/external_reference_fetcher.rs

@@ -0,0 +1,151 @@
+use std::{
+    sync::{
+        Arc,
+        atomic::{AtomicUsize, Ordering},
+    },
+    time::Duration,
+};
+
+use axum::{Router, http::StatusCode, response::Redirect, routing::get};
+use crank_adapter_rest::{
+    ExternalReferenceFetchError, ExternalReferenceFetcher, OutboundHttpPolicy,
+};
+use tokio::net::TcpListener;
+
+#[tokio::test]
+async fn external_references_are_default_off_before_any_request() {
+    let requests = Arc::new(AtomicUsize::new(0));
+    let base_url = spawn_server(Arc::clone(&requests)).await;
+    let fetcher = ExternalReferenceFetcher::try_new(
+        OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
+        Vec::new(),
+        1024,
+        Duration::from_secs(1),
+    )
+    .unwrap();
+
+    let error = fetcher
+        .get(&format!("{base_url}/document"))
+        .await
+        .unwrap_err();
+
+    assert!(matches!(error, ExternalReferenceFetchError::Disabled));
+    assert_eq!(requests.load(Ordering::SeqCst), 0);
+}
+
+#[tokio::test]
+async fn fetcher_uses_prefix_and_actual_address_policy_then_returns_bounded_bytes() {
+    let base_url = spawn_server(Arc::new(AtomicUsize::new(0))).await;
+    let fetcher = ExternalReferenceFetcher::try_new(
+        OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
+        vec![base_url.clone()],
+        8,
+        Duration::from_secs(1),
+    )
+    .unwrap();
+
+    assert_eq!(
+        fetcher
+            .get(&format!("{base_url}/document#/components/schemas/A"))
+            .await
+            .unwrap(),
+        b"openapi".to_vec()
+    );
+    let error = fetcher.get(&format!("{base_url}/large")).await.unwrap_err();
+    assert!(matches!(
+        error,
+        ExternalReferenceFetchError::ResponseTooLarge { limit_bytes: 8 }
+    ));
+
+    let private_without_explicit_outbound_allow = ExternalReferenceFetcher::try_new(
+        OutboundHttpPolicy::default(),
+        vec![base_url.clone()],
+        1024,
+        Duration::from_secs(1),
+    )
+    .unwrap();
+    let error = private_without_explicit_outbound_allow
+        .get(&format!("{base_url}/document"))
+        .await
+        .unwrap_err();
+    assert!(matches!(
+        error,
+        ExternalReferenceFetchError::TargetNotAllowed
+    ));
+
+    let exact_path_fetcher = ExternalReferenceFetcher::try_new(
+        OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
+        vec![format!("{base_url}/document")],
+        1024,
+        Duration::from_secs(1),
+    )
+    .unwrap();
+    let error = exact_path_fetcher
+        .get(&format!("{base_url}/document-unrelated"))
+        .await
+        .unwrap_err();
+    assert!(matches!(
+        error,
+        ExternalReferenceFetchError::TargetNotAllowed
+    ));
+}
+
+#[tokio::test]
+async fn fetcher_rejects_redirects_and_userinfo_without_exposing_the_url() {
+    let requests = Arc::new(AtomicUsize::new(0));
+    let base_url = spawn_server(Arc::clone(&requests)).await;
+    let fetcher = ExternalReferenceFetcher::try_new(
+        OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
+        vec![base_url.clone()],
+        1024,
+        Duration::from_secs(1),
+    )
+    .unwrap();
+
+    let error = fetcher
+        .get(&format!("{base_url}/redirect"))
+        .await
+        .unwrap_err();
+    assert!(matches!(
+        error,
+        ExternalReferenceFetchError::RedirectNotAllowed
+    ));
+
+    let userinfo_url = base_url.replacen("http://", "http://user:credential@", 1);
+    let error = fetcher
+        .get(&format!("{userinfo_url}/document"))
+        .await
+        .unwrap_err();
+    assert!(matches!(
+        error,
+        ExternalReferenceFetchError::TargetNotAllowed
+    ));
+    let rendered = format!("{error:?} {error}");
+    assert!(!rendered.contains("credential"));
+    assert_eq!(requests.load(Ordering::SeqCst), 0);
+}
+
+async fn spawn_server(requests: Arc<AtomicUsize>) -> String {
+    let app = Router::new()
+        .route(
+            "/document",
+            get({
+                let requests = Arc::clone(&requests);
+                move || {
+                    let requests = Arc::clone(&requests);
+                    async move {
+                        requests.fetch_add(1, Ordering::SeqCst);
+                        (StatusCode::OK, "openapi")
+                    }
+                }
+            }),
+        )
+        .route("/large", get(|| async { "response too large" }))
+        .route("/redirect", get(|| async { Redirect::to("/document") }));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{address}")
+}

crates/crank-adapter-rest/tests/integration/outbound_security.rs

@@ -6,6 +6,7 @@ use std::{
         Arc,
         atomic::{AtomicUsize, Ordering},
     },
+    time::Duration,
 };
 
 use axum::{
@@ -14,7 +15,9 @@ use axum::{
     http::StatusCode,
     routing::{any, post},
 };
-use crank_adapter_rest::{OutboundHttpPolicy, RestAdapter, RestAdapterError, RestRequest};
+use crank_adapter_rest::{
+    ExternalReferenceFetcher, OutboundHttpPolicy, RestAdapter, RestAdapterError, RestRequest,
+};
 use crank_core::{HttpMethod, ProtocolAdapterError, RestTarget};
 use serde_json::{Value, json};
 use tokio::net::TcpListener;
@@ -94,7 +97,7 @@ async fn proxy_environment_is_ignored_by_default() {
         let adapter = RestAdapter::default();
         let request = json_request(json!({"payload": "proxy-env-canary"}));
         let rest_target = RestTarget {
-            base_url: target,
+            base_url: target.clone(),
             method: HttpMethod::Post,
             path_template: "/capture".to_owned(),
             static_headers: BTreeMap::new(),
@@ -102,6 +105,17 @@ async fn proxy_environment_is_ignored_by_default() {
         let _ = adapter.execute(&rest_target, &request).await.expect_err(
             "unresolvable target should fail locally instead of being sent through proxy env",
         );
+        let fetcher = ExternalReferenceFetcher::try_new(
+            OutboundHttpPolicy::default(),
+            vec!["http://public.example.test/".to_owned()],
+            1024,
+            Duration::from_secs(1),
+        )
+        .expect("valid external reference fetcher");
+        let _ = fetcher
+            .get(&target)
+            .await
+            .expect_err("external reference fetcher must not use proxy environment variables");
         return;
     }
 

crates/crank-artifacts/src/housekeeping.rs

@@ -1,10 +1,9 @@
 use std::{
     fmt,
     os::fd::{AsRawFd, FromRawFd, OwnedFd},
-    time::{Duration, SystemTime, UNIX_EPOCH},
 };
 
-use crate::temp_scan::{list_names_after, valid_temp_name};
+use crate::temp_scan::valid_temp_name;
 use crate::{
     ArtifactError, ArtifactStore, ReconciliationMutation, ReconciliationNamespace,
     ReconciliationScan, ReconciliationScanStop,
@@ -15,6 +14,10 @@ use crate::{
     },
 };
 
+mod stale_temp;
+
+pub use stale_temp::{StaleTemp, TempScan, TempScanCursor};
+
 /// Opaque bounded-scan continuation, valid only for an unchanged namespace.
 /// Discard it after any put, quarantine, delete, or external mutation.
 pub struct ReconciliationCursor {
@@ -80,49 +83,6 @@ pub struct ReconciliationCandidate {
     pub(crate) modified_nanoseconds: i64,
 }
 
-/// Opaque single-use evidence for a stale temporary inode under the pinned root.
-#[derive(Debug)]
-pub struct StaleTemp {
-    name: String,
-    shard: String,
-    root_dev: u64,
-    root_ino: u64,
-    dev: u64,
-    ino: u64,
-    modified_seconds: i64,
-    modified_nanoseconds: i64,
-    grace: Duration,
-}
-
-/// Opaque, inode-bound progress marker for the bounded stale-temp sweeper.
-///
-/// It deliberately advances by shard instead of retaining a directory offset:
-/// directory offsets are invalidated by a concurrent writer, while round-robin
-/// shard progress prevents a busy low-numbered shard from starving all others.
-#[derive(Clone, Debug)]
-pub struct TempScanCursor {
-    root_dev: u64,
-    root_ino: u64,
-    next_shard: u8,
-    shard_continuation: Option<TempShardContinuation>,
-}
-
-#[derive(Clone, Debug)]
-struct TempShardContinuation {
-    shard: u8,
-    dev: u64,
-    ino: u64,
-    cookie: i64,
-}
-
-#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
-pub struct TempScan {
-    pub scanned: usize,
-    pub omitted: usize,
-    /// True when this page completed a full round-robin traversal.
-    pub complete: bool,
-}
-
 impl ArtifactStore {
     pub(crate) fn scan_reconciliation_bounded(
         &self,
@@ -340,229 +300,6 @@ impl ArtifactStore {
             Err(_) => Ok(ReconciliationMutation::Retryable),
         }
     }
-    /// Scans at most `scan_budget` entries and returns `result_limit` stale capabilities.
-    pub fn scan_stale_temps(
-        &self,
-        grace: Duration,
-        scan_budget: usize,
-        result_limit: usize,
-    ) -> Result<(TempScan, Vec<StaleTemp>), ArtifactError> {
-        self.scan_stale_temps_after(grace, None, scan_budget, result_limit)
-            .map(|(report, candidates, _)| (report, candidates))
-    }
-
-    /// Resumes stale-temp cleanup from an opaque round-robin marker. A
-    /// continuation is valid only for this exact pinned root.
-    pub fn scan_stale_temps_after(
-        &self,
-        grace: Duration,
-        continuation: Option<TempScanCursor>,
-        scan_budget: usize,
-        result_limit: usize,
-    ) -> Result<(TempScan, Vec<StaleTemp>, TempScanCursor), ArtifactError> {
-        let root = self.root()?;
-        let _lock = RootLock::shared(root)?;
-        ensure_root_unchanged(root)?;
-        let sha = match open_existing_dir(root.fd.as_raw_fd(), b"sha256") {
-            Ok(fd) => fd,
-            Err(ArtifactError::NotFound) => {
-                return Ok((
-                    TempScan {
-                        complete: true,
-                        ..TempScan::default()
-                    },
-                    Vec::new(),
-                    TempScanCursor {
-                        root_dev: root.dev,
-                        root_ino: root.ino,
-                        next_shard: 0,
-                        shard_continuation: None,
-                    },
-                ));
-            }
-            Err(error) => return Err(error),
-        };
-        let (start_shard, mut shard_continuation) = match continuation {
-            Some(cursor) if cursor.root_dev == root.dev && cursor.root_ino == root.ino => {
-                (cursor.next_shard, cursor.shard_continuation)
-            }
-            Some(_) => return Err(ArtifactError::UnsafeRoot),
-            None => (0, None),
-        };
-        let mut report = TempScan::default();
-        let mut remaining = scan_budget;
-        let mut result = Vec::new();
-        if remaining == 0 {
-            return Ok((
-                report,
-                result,
-                TempScanCursor {
-                    root_dev: root.dev,
-                    root_ino: root.ino,
-                    next_shard: start_shard,
-                    shard_continuation,
-                },
-            ));
-        }
-        let mut next_shard = start_shard;
-        for offset in 0_u16..=255 {
-            if remaining == 0 {
-                break;
-            }
-            let shard_number = start_shard.wrapping_add(offset as u8);
-            next_shard = shard_number.wrapping_add(1);
-            let shard = format!("{shard_number:02x}");
-            let shard_fd = match open_existing_dir(sha.as_raw_fd(), shard.as_bytes()) {
-                Ok(fd) => fd,
-                Err(ArtifactError::NotFound) => {
-                    shard_continuation = None;
-                    continue;
-                }
-                Err(error) => return Err(error),
-            };
-            let stat = stat_fd(shard_fd.as_raw_fd())?;
-            let resume = shard_continuation.take().and_then(|continuation| {
-                if continuation.shard == shard_number
-                    && continuation.dev == stat.st_dev
-                    && continuation.ino == stat.st_ino
-                {
-                    Some(continuation.cookie)
-                } else {
-                    None
-                }
-            });
-            let page = list_names_after(shard_fd.as_raw_fd(), &mut remaining, &mut report, resume)?;
-            for entry in page.names {
-                let name = entry.name;
-                if !valid_temp_name(&name) {
-                    continue;
-                }
-                #[cfg(debug_assertions)]
-                let _ = crate::test_support::checkpoint("housekeeping_before_stat");
-                let stat = match nofollow_stat(shard_fd.as_raw_fd(), &name) {
-                    Ok(stat) => stat,
-                    Err(ArtifactError::NotFound) => continue,
-                    Err(error) => return Err(error),
-                };
-                if (stat.st_mode & libc::S_IFMT) != libc::S_IFREG || stat.st_nlink != 1 {
-                    continue;
-                }
-                let modified = UNIX_EPOCH
-                    .checked_add(Duration::new(
-                        stat.st_mtime.max(0) as u64,
-                        stat.st_mtime_nsec.max(0) as u32,
-                    ))
-                    .unwrap_or(SystemTime::UNIX_EPOCH);
-                if SystemTime::now()
-                    .duration_since(modified)
-                    .is_ok_and(|age| age >= grace)
-                {
-                    if result.len() == result_limit {
-                        report.omitted += 1;
-                        if result_limit > 0 {
-                            return Ok((
-                                report,
-                                result,
-                                TempScanCursor {
-                                    root_dev: root.dev,
-                                    root_ino: root.ino,
-                                    next_shard: shard_number,
-                                    shard_continuation: Some(TempShardContinuation {
-                                        shard: shard_number,
-                                        dev: stat.st_dev,
-                                        ino: stat.st_ino,
-                                        cookie: entry.cookie_before,
-                                    }),
-                                },
-                            ));
-                        }
-                        continue;
-                    }
-                    result.push(StaleTemp {
-                        name,
-                        shard: shard.clone(),
-                        root_dev: root.dev,
-                        root_ino: root.ino,
-                        dev: stat.st_dev,
-                        ino: stat.st_ino,
-                        modified_seconds: stat.st_mtime,
-                        modified_nanoseconds: stat.st_mtime_nsec,
-                        grace,
-                    });
-                }
-            }
-            if let Some(cookie) = page.continuation {
-                return Ok((
-                    report,
-                    result,
-                    TempScanCursor {
-                        root_dev: root.dev,
-                        root_ino: root.ino,
-                        next_shard: shard_number,
-                        shard_continuation: Some(TempShardContinuation {
-                            shard: shard_number,
-                            dev: stat.st_dev,
-                            ino: stat.st_ino,
-                            cookie,
-                        }),
-                    },
-                ));
-            }
-        }
-        report.complete = true;
-        Ok((
-            report,
-            result,
-            TempScanCursor {
-                root_dev: root.dev,
-                root_ino: root.ino,
-                next_shard,
-                shard_continuation: None,
-            },
-        ))
-    }
-
-    /// Deletes a revalidated stale inode under an exclusive lock, then fsyncs.
-    pub fn delete_stale_temp(&self, candidate: StaleTemp) -> Result<(), ArtifactError> {
-        let root = self.root()?;
-        let _lock = RootLock::exclusive(root)?;
-        ensure_root_unchanged(root)?;
-        if candidate.root_dev != root.dev || candidate.root_ino != root.ino {
-            return Err(ArtifactError::UnsafeRoot);
-        }
-        if !valid_temp_name(&candidate.name)
-            || candidate.shard.len() != 2
-            || !candidate.shard.bytes().all(is_lower_hex)
-        {
-            return Err(ArtifactError::UnsafeRoot);
-        }
-        let sha = open_existing_dir(root.fd.as_raw_fd(), b"sha256")?;
-        let shard = open_existing_dir(sha.as_raw_fd(), candidate.shard.as_bytes())?;
-        let stat = nofollow_stat(shard.as_raw_fd(), &candidate.name)?;
-        if stat.st_dev != candidate.dev
-            || stat.st_ino != candidate.ino
-            || stat.st_mtime != candidate.modified_seconds
-            || stat.st_mtime_nsec != candidate.modified_nanoseconds
-            || (stat.st_mode & libc::S_IFMT) != libc::S_IFREG
-            || stat.st_nlink != 1
-        {
-            return Err(ArtifactError::UnsafeRoot);
-        }
-        let modified = UNIX_EPOCH
-            .checked_add(Duration::new(
-                stat.st_mtime.max(0) as u64,
-                stat.st_mtime_nsec.max(0) as u32,
-            ))
-            .unwrap_or(SystemTime::UNIX_EPOCH);
-        if !SystemTime::now()
-            .duration_since(modified)
-            .is_ok_and(|age| age >= candidate.grace)
-        {
-            return Err(ArtifactError::UnsafeRoot);
-        }
-        unlinkat(shard.as_raw_fd(), candidate.name.as_bytes())?;
-        fsync_fd(shard.as_raw_fd())
-    }
 }
 
 fn reconciliation_cursor(

crates/crank-artifacts/src/housekeeping/stale_temp.rs

@@ -0,0 +1,280 @@
+use std::{
+    os::fd::AsRawFd,
+    time::{Duration, SystemTime, UNIX_EPOCH},
+};
+
+use super::{is_lower_hex, nofollow_stat};
+use crate::{
+    ArtifactError, ArtifactStore,
+    store::{RootLock, ensure_root_unchanged, fsync_fd, open_existing_dir, stat_fd, unlinkat},
+    temp_scan::{list_names_after, valid_temp_name},
+};
+
+/// Opaque single-use evidence for a stale temporary inode under the pinned root.
+#[derive(Debug)]
+pub struct StaleTemp {
+    name: String,
+    shard: String,
+    root_dev: u64,
+    root_ino: u64,
+    dev: u64,
+    ino: u64,
+    modified_seconds: i64,
+    modified_nanoseconds: i64,
+    grace: Duration,
+}
+
+/// Opaque, inode-bound progress marker for the bounded stale-temp sweeper.
+///
+/// It deliberately advances by shard instead of retaining a directory offset:
+/// directory offsets are invalidated by a concurrent writer, while round-robin
+/// shard progress prevents a busy low-numbered shard from starving all others.
+#[derive(Clone, Debug)]
+pub struct TempScanCursor {
+    root_dev: u64,
+    root_ino: u64,
+    next_shard: u8,
+    shard_continuation: Option<TempShardContinuation>,
+}
+
+#[derive(Clone, Debug)]
+struct TempShardContinuation {
+    shard: u8,
+    dev: u64,
+    ino: u64,
+    cookie: i64,
+}
+
+#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
+pub struct TempScan {
+    pub scanned: usize,
+    pub omitted: usize,
+    /// True when this page completed a full round-robin traversal.
+    pub complete: bool,
+}
+
+impl ArtifactStore {
+    /// Scans at most `scan_budget` entries and returns `result_limit` stale capabilities.
+    pub fn scan_stale_temps(
+        &self,
+        grace: Duration,
+        scan_budget: usize,
+        result_limit: usize,
+    ) -> Result<(TempScan, Vec<StaleTemp>), ArtifactError> {
+        self.scan_stale_temps_after(grace, None, scan_budget, result_limit)
+            .map(|(report, candidates, _)| (report, candidates))
+    }
+
+    /// Resumes stale-temp cleanup from an opaque round-robin marker. A
+    /// continuation is valid only for this exact pinned root.
+    pub fn scan_stale_temps_after(
+        &self,
+        grace: Duration,
+        continuation: Option<TempScanCursor>,
+        scan_budget: usize,
+        result_limit: usize,
+    ) -> Result<(TempScan, Vec<StaleTemp>, TempScanCursor), ArtifactError> {
+        let root = self.root()?;
+        let _lock = RootLock::shared(root)?;
+        ensure_root_unchanged(root)?;
+        let sha = match open_existing_dir(root.fd.as_raw_fd(), b"sha256") {
+            Ok(fd) => fd,
+            Err(ArtifactError::NotFound) => {
+                return Ok((
+                    TempScan {
+                        complete: true,
+                        ..TempScan::default()
+                    },
+                    Vec::new(),
+                    TempScanCursor {
+                        root_dev: root.dev,
+                        root_ino: root.ino,
+                        next_shard: 0,
+                        shard_continuation: None,
+                    },
+                ));
+            }
+            Err(error) => return Err(error),
+        };
+        let (start_shard, mut shard_continuation) = match continuation {
+            Some(cursor) if cursor.root_dev == root.dev && cursor.root_ino == root.ino => {
+                (cursor.next_shard, cursor.shard_continuation)
+            }
+            Some(_) => return Err(ArtifactError::UnsafeRoot),
+            None => (0, None),
+        };
+        let mut report = TempScan::default();
+        let mut remaining = scan_budget;
+        let mut result = Vec::new();
+        if remaining == 0 {
+            return Ok((
+                report,
+                result,
+                TempScanCursor {
+                    root_dev: root.dev,
+                    root_ino: root.ino,
+                    next_shard: start_shard,
+                    shard_continuation,
+                },
+            ));
+        }
+        let mut next_shard = start_shard;
+        for offset in 0_u16..=255 {
+            if remaining == 0 {
+                break;
+            }
+            let shard_number = start_shard.wrapping_add(offset as u8);
+            next_shard = shard_number.wrapping_add(1);
+            let shard = format!("{shard_number:02x}");
+            let shard_fd = match open_existing_dir(sha.as_raw_fd(), shard.as_bytes()) {
+                Ok(fd) => fd,
+                Err(ArtifactError::NotFound) => {
+                    shard_continuation = None;
+                    continue;
+                }
+                Err(error) => return Err(error),
+            };
+            let stat = stat_fd(shard_fd.as_raw_fd())?;
+            let resume = shard_continuation.take().and_then(|continuation| {
+                if continuation.shard == shard_number
+                    && continuation.dev == stat.st_dev
+                    && continuation.ino == stat.st_ino
+                {
+                    Some(continuation.cookie)
+                } else {
+                    None
+                }
+            });
+            let page = list_names_after(shard_fd.as_raw_fd(), &mut remaining, &mut report, resume)?;
+            for entry in page.names {
+                let name = entry.name;
+                if !valid_temp_name(&name) {
+                    continue;
+                }
+                #[cfg(debug_assertions)]
+                let _ = crate::test_support::checkpoint("housekeeping_before_stat");
+                let stat = match nofollow_stat(shard_fd.as_raw_fd(), &name) {
+                    Ok(stat) => stat,
+                    Err(ArtifactError::NotFound) => continue,
+                    Err(error) => return Err(error),
+                };
+                if (stat.st_mode & libc::S_IFMT) != libc::S_IFREG || stat.st_nlink != 1 {
+                    continue;
+                }
+                let modified = UNIX_EPOCH
+                    .checked_add(Duration::new(
+                        stat.st_mtime.max(0) as u64,
+                        stat.st_mtime_nsec.max(0) as u32,
+                    ))
+                    .unwrap_or(SystemTime::UNIX_EPOCH);
+                if SystemTime::now()
+                    .duration_since(modified)
+                    .is_ok_and(|age| age >= grace)
+                {
+                    if result.len() == result_limit {
+                        report.omitted += 1;
+                        if result_limit > 0 {
+                            return Ok((
+                                report,
+                                result,
+                                TempScanCursor {
+                                    root_dev: root.dev,
+                                    root_ino: root.ino,
+                                    next_shard: shard_number,
+                                    shard_continuation: Some(TempShardContinuation {
+                                        shard: shard_number,
+                                        dev: stat.st_dev,
+                                        ino: stat.st_ino,
+                                        cookie: entry.cookie_before,
+                                    }),
+                                },
+                            ));
+                        }
+                        continue;
+                    }
+                    result.push(StaleTemp {
+                        name,
+                        shard: shard.clone(),
+                        root_dev: root.dev,
+                        root_ino: root.ino,
+                        dev: stat.st_dev,
+                        ino: stat.st_ino,
+                        modified_seconds: stat.st_mtime,
+                        modified_nanoseconds: stat.st_mtime_nsec,
+                        grace,
+                    });
+                }
+            }
+            if let Some(cookie) = page.continuation {
+                return Ok((
+                    report,
+                    result,
+                    TempScanCursor {
+                        root_dev: root.dev,
+                        root_ino: root.ino,
+                        next_shard: shard_number,
+                        shard_continuation: Some(TempShardContinuation {
+                            shard: shard_number,
+                            dev: stat.st_dev,
+                            ino: stat.st_ino,
+                            cookie,
+                        }),
+                    },
+                ));
+            }
+        }
+        report.complete = true;
+        Ok((
+            report,
+            result,
+            TempScanCursor {
+                root_dev: root.dev,
+                root_ino: root.ino,
+                next_shard,
+                shard_continuation: None,
+            },
+        ))
+    }
+
+    /// Deletes a revalidated stale inode under an exclusive lock, then fsyncs.
+    pub fn delete_stale_temp(&self, candidate: StaleTemp) -> Result<(), ArtifactError> {
+        let root = self.root()?;
+        let _lock = RootLock::exclusive(root)?;
+        ensure_root_unchanged(root)?;
+        if candidate.root_dev != root.dev || candidate.root_ino != root.ino {
+            return Err(ArtifactError::UnsafeRoot);
+        }
+        if !valid_temp_name(&candidate.name)
+            || candidate.shard.len() != 2
+            || !candidate.shard.bytes().all(is_lower_hex)
+        {
+            return Err(ArtifactError::UnsafeRoot);
+        }
+        let sha = open_existing_dir(root.fd.as_raw_fd(), b"sha256")?;
+        let shard = open_existing_dir(sha.as_raw_fd(), candidate.shard.as_bytes())?;
+        let stat = nofollow_stat(shard.as_raw_fd(), &candidate.name)?;
+        if stat.st_dev != candidate.dev
+            || stat.st_ino != candidate.ino
+            || stat.st_mtime != candidate.modified_seconds
+            || stat.st_mtime_nsec != candidate.modified_nanoseconds
+            || (stat.st_mode & libc::S_IFMT) != libc::S_IFREG
+            || stat.st_nlink != 1
+        {
+            return Err(ArtifactError::UnsafeRoot);
+        }
+        let modified = UNIX_EPOCH
+            .checked_add(Duration::new(
+                stat.st_mtime.max(0) as u64,
+                stat.st_mtime_nsec.max(0) as u32,
+            ))
+            .unwrap_or(SystemTime::UNIX_EPOCH);
+        if !SystemTime::now()
+            .duration_since(modified)
+            .is_ok_and(|age| age >= candidate.grace)
+        {
+            return Err(ArtifactError::UnsafeRoot);
+        }
+        unlinkat(shard.as_raw_fd(), candidate.name.as_bytes())?;
+        fsync_fd(shard.as_raw_fd())
+    }
+}

crates/crank-artifacts/tests/artifact_store.rs

@@ -1,6 +1,5 @@
 use std::{
     fs,
-    io::{self, Read},
     os::unix::fs::PermissionsExt,
     path::PathBuf,
     process::{Child, Command, ExitStatus},
@@ -13,7 +12,7 @@ use std::{
 use crank_artifacts::test_support::{
     FaultAction, clear_checkpoint, set_checkpoint, wait_until_held,
 };
-use crank_artifacts::{ArtifactError, ArtifactRef, ArtifactStore, MAX_ARTIFACT_BYTES, StaleTemp};
+use crank_artifacts::{ArtifactError, ArtifactRef, ArtifactStore, MAX_ARTIFACT_BYTES};
 use sha2::{Digest, Sha256};
 #[cfg(debug_assertions)]
 use std::sync::{Mutex, OnceLock};
@@ -786,318 +785,3 @@ fn legacy_oversize_is_rejected() {
         ArtifactError::Integrity
     );
 }
-
-#[test]
-fn fifo_entries_fail_without_blocking_reads() {
-    let root = TestRoot::new("fifo");
-    let expected =
-        ArtifactRef::from_digest_hex(&format!("{:x}", Sha256::digest(b"fifo bytes"))).unwrap();
-    let sha = root.0.join("sha256");
-    let shard = sha.join(&expected.digest_hex()[..2]);
-    fs::create_dir(&sha).unwrap();
-    fs::create_dir(&shard).unwrap();
-    fs::set_permissions(&sha, fs::Permissions::from_mode(0o700)).unwrap();
-    fs::set_permissions(&shard, fs::Permissions::from_mode(0o700)).unwrap();
-    let blob_fifo = shard.join(expected.digest_hex());
-    let legacy = root.0.join("legacy");
-    fs::create_dir(&legacy).unwrap();
-    fs::set_permissions(&legacy, fs::Permissions::from_mode(0o700)).unwrap();
-    let legacy_fifo = legacy.join("source");
-    for path in [&blob_fifo, &legacy_fifo] {
-        let path = std::ffi::CString::new(path.as_os_str().as_encoded_bytes()).unwrap();
-        assert_eq!(unsafe { libc::mkfifo(path.as_ptr(), 0o400) }, 0);
-    }
-
-    let executable = std::env::current_exe().unwrap();
-    for kind in ["blob", "legacy"] {
-        let child = Command::new(&executable)
-            .args(["--exact", "fifo_read_child", "--nocapture"])
-            .env("CRANK_ARTIFACTS_FIFO_ROOT", &root.0)
-            .env("CRANK_ARTIFACTS_FIFO_KIND", kind)
-            .env("CRANK_ARTIFACTS_FIFO_REF", expected.as_str())
-            .spawn()
-            .unwrap();
-        assert!(wait_for_child(child, Duration::from_secs(2)).success());
-    }
-}
-
-#[test]
-fn fifo_read_child() {
-    let Some(root) = std::env::var_os("CRANK_ARTIFACTS_FIFO_ROOT") else {
-        return;
-    };
-    let root = PathBuf::from(root);
-    let kind = std::env::var("CRANK_ARTIFACTS_FIFO_KIND").unwrap();
-    let expected = ArtifactRef::parse(&std::env::var("CRANK_ARTIFACTS_FIFO_REF").unwrap()).unwrap();
-    let store = ArtifactStore::open(&root).unwrap();
-    let result = if kind == "blob" {
-        store.read(&expected)
-    } else {
-        store
-            .with_legacy_root(root.join("legacy"))
-            .read_legacy_file_url(
-                &format!("file://{}/legacy/source", root.display()),
-                &expected,
-            )
-    };
-    assert_eq!(result.unwrap_err(), ArtifactError::Integrity);
-}
-
-struct ShortReader {
-    bytes: Vec<u8>,
-    offset: usize,
-    interrupted: bool,
-}
-
-struct FailingReader;
-
-struct HousekeepingReader {
-    store: ArtifactStore,
-    candidate: Option<StaleTemp>,
-    bytes: io::Cursor<Vec<u8>>,
-}
-
-impl Read for FailingReader {
-    fn read(&mut self, _buf: &mut [u8]) -> io::Result<usize> {
-        Err(io::Error::other("private reader detail"))
-    }
-}
-impl Read for HousekeepingReader {
-    fn read(&mut self, buffer: &mut [u8]) -> io::Result<usize> {
-        if let Some(candidate) = self.candidate.take() {
-            self.store.delete_stale_temp(candidate).unwrap();
-        }
-        self.bytes.read(buffer)
-    }
-}
-impl Read for ShortReader {
-    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
-        if !self.interrupted {
-            self.interrupted = true;
-            return Err(io::Error::from(io::ErrorKind::Interrupted));
-        }
-        if self.offset == self.bytes.len() {
-            return Ok(0);
-        }
-        let amount = 1.min(buf.len()).min(self.bytes.len() - self.offset);
-        buf[..amount].copy_from_slice(&self.bytes[self.offset..self.offset + amount]);
-        self.offset += amount;
-        Ok(amount)
-    }
-}
-
-#[test]
-fn streaming_input_handles_short_reads_and_interruption() {
-    let root = TestRoot::new("stream");
-    let store = ArtifactStore::open(&root.0).unwrap();
-    let mut reader = ShortReader {
-        bytes: b"short chunk source".to_vec(),
-        offset: 0,
-        interrupted: false,
-    };
-    let stored = store.put_reader(&mut reader).unwrap();
-    assert_eq!(
-        store.read(&stored.artifact_ref).unwrap(),
-        b"short chunk source"
-    );
-
-    assert_eq!(
-        store.put_reader(&mut FailingReader).unwrap_err(),
-        ArtifactError::Storage
-    );
-    for error in [
-        ArtifactError::InvalidReference,
-        ArtifactError::Integrity,
-        ArtifactError::Storage,
-        ArtifactError::UnsafeRoot,
-    ] {
-        let diagnostic = format!("{error} {error:?}");
-        assert!(!diagnostic.contains("private reader detail"));
-        assert!(!diagnostic.contains(std::env::temp_dir().to_string_lossy().as_ref()));
-    }
-}
-
-#[test]
-fn user_reader_can_run_housekeeping_without_self_deadlock() {
-    use std::sync::mpsc;
-
-    let root = TestRoot::new("reader-housekeeping");
-    let store = ArtifactStore::open(&root.0).unwrap();
-    let stored = store.put(b"reader housekeeping shard").unwrap();
-    let temp = root
-        .0
-        .join("sha256")
-        .join(&stored.artifact_ref.digest_hex()[..2])
-        .join(".crank-artifact-tmp-v1-00000000000000000000000000000000-1-10");
-    fs::write(&temp, b"partial").unwrap();
-    let (_, candidates) = store.scan_stale_temps(Duration::ZERO, 128, 1).unwrap();
-    let mut reader = HousekeepingReader {
-        store: store.clone(),
-        candidate: candidates.into_iter().next(),
-        bytes: io::Cursor::new(b"reader bytes".to_vec()),
-    };
-    let writer_store = store.clone();
-    let (sender, receiver) = mpsc::channel();
-    thread::spawn(move || {
-        sender.send(writer_store.put_reader(&mut reader)).unwrap();
-    });
-
-    let stored = receiver
-        .recv_timeout(Duration::from_secs(2))
-        .expect("put_reader deadlocked while its reader ran housekeeping")
-        .unwrap();
-    assert_eq!(store.read(&stored.artifact_ref).unwrap(), b"reader bytes");
-    assert!(!temp.exists());
-}
-
-#[test]
-#[cfg(debug_assertions)]
-fn forked_crash_checkpoints_publish_only_complete_or_absent_blobs() {
-    let _guard = fault_guard();
-    if let Some(root) = std::env::var_os("CRANK_ARTIFACTS_CHILD_ROOT") {
-        set_checkpoint(
-            std::env::var("CRANK_ARTIFACTS_TEST_STAGE").unwrap(),
-            FaultAction::Exit,
-        );
-        let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
-        let _ = store.put(b"crash-consistent bytes");
-        panic!("checkpoint did not terminate the child");
-    }
-
-    for stage in ["write", "file_fsync", "publish", "directory_fsync"] {
-        let root = TestRoot::new(stage);
-        let status = Command::new(std::env::current_exe().unwrap())
-            .args([
-                "--exact",
-                "forked_crash_checkpoints_publish_only_complete_or_absent_blobs",
-                "--nocapture",
-            ])
-            .env("CRANK_ARTIFACTS_CHILD_ROOT", &root.0)
-            .env("CRANK_ARTIFACTS_TEST_STAGE", stage)
-            .status()
-            .unwrap();
-        assert_eq!(status.code(), Some(86), "stage={stage}");
-
-        let store = ArtifactStore::open(&root.0).unwrap();
-        let expected = ArtifactRef::from_digest_hex(&format!(
-            "{:x}",
-            Sha256::digest(b"crash-consistent bytes")
-        ))
-        .unwrap();
-        match store.read(&expected) {
-            Ok(bytes) => assert_eq!(bytes, b"crash-consistent bytes", "stage={stage}"),
-            Err(ArtifactError::NotFound) => {}
-            Err(error) => panic!("stage={stage} exposed an invalid final blob: {error}"),
-        }
-        let stored = store.put(b"crash-consistent bytes").unwrap();
-        assert_eq!(stored.artifact_ref, expected, "stage={stage}");
-        assert_eq!(
-            store.read(&expected).unwrap(),
-            b"crash-consistent bytes",
-            "stage={stage}"
-        );
-
-        let (_, candidates) = store.scan_stale_temps(Duration::ZERO, 128, 16).unwrap();
-        for candidate in candidates {
-            store.delete_stale_temp(candidate).unwrap();
-        }
-    }
-}
-
-#[test]
-#[cfg(debug_assertions)]
-fn disk_full_fault_seams_leave_a_retryable_store() {
-    let _guard = fault_guard();
-    if let Some(root) = std::env::var_os("CRANK_ARTIFACTS_CHILD_ROOT") {
-        set_checkpoint(
-            std::env::var("CRANK_ARTIFACTS_TEST_STAGE").unwrap(),
-            FaultAction::Fail,
-        );
-        let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
-        assert_eq!(
-            store.put(b"disk-full seam bytes").unwrap_err(),
-            ArtifactError::Storage
-        );
-        return;
-    }
-
-    for stage in ["write", "file_fsync", "publish", "directory_fsync"] {
-        let root = TestRoot::new(&format!("full-{stage}"));
-        let status = Command::new(std::env::current_exe().unwrap())
-            .args([
-                "--exact",
-                "disk_full_fault_seams_leave_a_retryable_store",
-                "--nocapture",
-            ])
-            .env("CRANK_ARTIFACTS_CHILD_ROOT", &root.0)
-            .env("CRANK_ARTIFACTS_TEST_STAGE", stage)
-            .status()
-            .unwrap();
-        assert!(status.success(), "stage={stage}");
-
-        let store = ArtifactStore::open(&root.0).unwrap();
-        let expected =
-            ArtifactRef::from_digest_hex(&format!("{:x}", Sha256::digest(b"disk-full seam bytes")))
-                .unwrap();
-        match store.read(&expected) {
-            Ok(bytes) => assert_eq!(bytes, b"disk-full seam bytes", "stage={stage}"),
-            Err(ArtifactError::NotFound) => {}
-            Err(error) => panic!("stage={stage} exposed an invalid final blob: {error}"),
-        }
-        assert_eq!(
-            store.put(b"disk-full seam bytes").unwrap().artifact_ref,
-            expected,
-            "stage={stage}"
-        );
-        assert_eq!(
-            store.read(&expected).unwrap(),
-            b"disk-full seam bytes",
-            "stage={stage}"
-        );
-    }
-}
-
-#[test]
-fn real_process_write_limit_leaves_no_partial_final_blob() {
-    if let Some(root) = std::env::var_os("CRANK_ARTIFACTS_CHILD_ROOT") {
-        // Keep the test's host filesystem untouched while making the actual
-        // write syscall fail as it would for a quota/disk-full condition.
-        unsafe {
-            assert_ne!(libc::signal(libc::SIGXFSZ, libc::SIG_IGN), libc::SIG_ERR);
-            let limit = libc::rlimit {
-                rlim_cur: 1,
-                rlim_max: 1,
-            };
-            assert_eq!(libc::setrlimit(libc::RLIMIT_FSIZE, &limit), 0);
-        }
-        let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
-        assert_eq!(
-            store.put(b"real write limit bytes").unwrap_err(),
-            ArtifactError::Storage
-        );
-        return;
-    }
-
-    let root = TestRoot::new("real-write-limit");
-    let status = Command::new(std::env::current_exe().unwrap())
-        .args([
-            "--exact",
-            "real_process_write_limit_leaves_no_partial_final_blob",
-            "--nocapture",
-        ])
-        .env("CRANK_ARTIFACTS_CHILD_ROOT", &root.0)
-        .status()
-        .unwrap();
-    assert!(status.success());
-
-    let store = ArtifactStore::open(&root.0).unwrap();
-    let expected =
-        ArtifactRef::from_digest_hex(&format!("{:x}", Sha256::digest(b"real write limit bytes")))
-            .unwrap();
-    assert_eq!(store.read(&expected).unwrap_err(), ArtifactError::NotFound);
-    assert_eq!(
-        store.put(b"real write limit bytes").unwrap().artifact_ref,
-        expected
-    );
-    assert_eq!(store.read(&expected).unwrap(), b"real write limit bytes");
-}

crates/crank-artifacts/tests/artifact_store_io.rs

@@ -0,0 +1,379 @@
+use std::{
+    fs,
+    io::{self, Read},
+    os::unix::fs::PermissionsExt,
+    path::PathBuf,
+    process::{Child, Command, ExitStatus},
+    sync::atomic::{AtomicU64, Ordering},
+    thread,
+    time::Duration,
+};
+
+#[cfg(debug_assertions)]
+use crank_artifacts::test_support::{FaultAction, set_checkpoint};
+use crank_artifacts::{ArtifactError, ArtifactRef, ArtifactStore, StaleTemp};
+use sha2::{Digest, Sha256};
+#[cfg(debug_assertions)]
+use std::sync::{Mutex, OnceLock};
+
+static NEXT_ROOT: AtomicU64 = AtomicU64::new(0);
+
+#[cfg(debug_assertions)]
+fn fault_guard() -> std::sync::MutexGuard<'static, ()> {
+    static GUARD: OnceLock<Mutex<()>> = OnceLock::new();
+    GUARD.get_or_init(|| Mutex::new(())).lock().unwrap()
+}
+
+struct TestRoot(PathBuf);
+
+impl TestRoot {
+    fn new(name: &str) -> Self {
+        let path = std::env::temp_dir().join(format!(
+            "crank-artifacts-{name}-{}-{}",
+            std::process::id(),
+            NEXT_ROOT.fetch_add(1, Ordering::Relaxed)
+        ));
+        fs::create_dir(&path).unwrap();
+        fs::set_permissions(&path, fs::Permissions::from_mode(0o700)).unwrap();
+        Self(path)
+    }
+}
+
+impl Drop for TestRoot {
+    fn drop(&mut self) {
+        let _ = fs::set_permissions(&self.0, fs::Permissions::from_mode(0o700));
+        let _ = fs::remove_dir_all(&self.0);
+    }
+}
+
+fn wait_for_child(mut child: Child, timeout: Duration) -> ExitStatus {
+    let deadline = std::time::Instant::now() + timeout;
+    loop {
+        if let Some(status) = child.try_wait().unwrap() {
+            return status;
+        }
+        if std::time::Instant::now() >= deadline {
+            let _ = child.kill();
+            let _ = child.wait();
+            panic!("child process did not finish within {timeout:?}");
+        }
+        thread::sleep(Duration::from_millis(10));
+    }
+}
+
+#[test]
+fn fifo_entries_fail_without_blocking_reads() {
+    let root = TestRoot::new("fifo");
+    let expected =
+        ArtifactRef::from_digest_hex(&format!("{:x}", Sha256::digest(b"fifo bytes"))).unwrap();
+    let sha = root.0.join("sha256");
+    let shard = sha.join(&expected.digest_hex()[..2]);
+    fs::create_dir(&sha).unwrap();
+    fs::create_dir(&shard).unwrap();
+    fs::set_permissions(&sha, fs::Permissions::from_mode(0o700)).unwrap();
+    fs::set_permissions(&shard, fs::Permissions::from_mode(0o700)).unwrap();
+    let blob_fifo = shard.join(expected.digest_hex());
+    let legacy = root.0.join("legacy");
+    fs::create_dir(&legacy).unwrap();
+    fs::set_permissions(&legacy, fs::Permissions::from_mode(0o700)).unwrap();
+    let legacy_fifo = legacy.join("source");
+    for path in [&blob_fifo, &legacy_fifo] {
+        let path = std::ffi::CString::new(path.as_os_str().as_encoded_bytes()).unwrap();
+        assert_eq!(unsafe { libc::mkfifo(path.as_ptr(), 0o400) }, 0);
+    }
+
+    let executable = std::env::current_exe().unwrap();
+    for kind in ["blob", "legacy"] {
+        let child = Command::new(&executable)
+            .args(["--exact", "fifo_read_child", "--nocapture"])
+            .env("CRANK_ARTIFACTS_FIFO_ROOT", &root.0)
+            .env("CRANK_ARTIFACTS_FIFO_KIND", kind)
+            .env("CRANK_ARTIFACTS_FIFO_REF", expected.as_str())
+            .spawn()
+            .unwrap();
+        assert!(wait_for_child(child, Duration::from_secs(2)).success());
+    }
+}
+
+#[test]
+fn fifo_read_child() {
+    let Some(root) = std::env::var_os("CRANK_ARTIFACTS_FIFO_ROOT") else {
+        return;
+    };
+    let root = PathBuf::from(root);
+    let kind = std::env::var("CRANK_ARTIFACTS_FIFO_KIND").unwrap();
+    let expected = ArtifactRef::parse(&std::env::var("CRANK_ARTIFACTS_FIFO_REF").unwrap()).unwrap();
+    let store = ArtifactStore::open(&root).unwrap();
+    let result = if kind == "blob" {
+        store.read(&expected)
+    } else {
+        store
+            .with_legacy_root(root.join("legacy"))
+            .read_legacy_file_url(
+                &format!("file://{}/legacy/source", root.display()),
+                &expected,
+            )
+    };
+    assert_eq!(result.unwrap_err(), ArtifactError::Integrity);
+}
+
+struct ShortReader {
+    bytes: Vec<u8>,
+    offset: usize,
+    interrupted: bool,
+}
+
+struct FailingReader;
+
+struct HousekeepingReader {
+    store: ArtifactStore,
+    candidate: Option<StaleTemp>,
+    bytes: io::Cursor<Vec<u8>>,
+}
+
+impl Read for FailingReader {
+    fn read(&mut self, _buf: &mut [u8]) -> io::Result<usize> {
+        Err(io::Error::other("private reader detail"))
+    }
+}
+
+impl Read for HousekeepingReader {
+    fn read(&mut self, buffer: &mut [u8]) -> io::Result<usize> {
+        if let Some(candidate) = self.candidate.take() {
+            self.store.delete_stale_temp(candidate).unwrap();
+        }
+        self.bytes.read(buffer)
+    }
+}
+
+impl Read for ShortReader {
+    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
+        if !self.interrupted {
+            self.interrupted = true;
+            return Err(io::Error::from(io::ErrorKind::Interrupted));
+        }
+        if self.offset == self.bytes.len() {
+            return Ok(0);
+        }
+        let amount = 1.min(buf.len()).min(self.bytes.len() - self.offset);
+        buf[..amount].copy_from_slice(&self.bytes[self.offset..self.offset + amount]);
+        self.offset += amount;
+        Ok(amount)
+    }
+}
+
+#[test]
+fn streaming_input_handles_short_reads_and_interruption() {
+    let root = TestRoot::new("stream");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let mut reader = ShortReader {
+        bytes: b"short chunk source".to_vec(),
+        offset: 0,
+        interrupted: false,
+    };
+    let stored = store.put_reader(&mut reader).unwrap();
+    assert_eq!(
+        store.read(&stored.artifact_ref).unwrap(),
+        b"short chunk source"
+    );
+
+    assert_eq!(
+        store.put_reader(&mut FailingReader).unwrap_err(),
+        ArtifactError::Storage
+    );
+    for error in [
+        ArtifactError::InvalidReference,
+        ArtifactError::Integrity,
+        ArtifactError::Storage,
+        ArtifactError::UnsafeRoot,
+    ] {
+        let diagnostic = format!("{error} {error:?}");
+        assert!(!diagnostic.contains("private reader detail"));
+        assert!(!diagnostic.contains(std::env::temp_dir().to_string_lossy().as_ref()));
+    }
+}
+
+#[test]
+fn user_reader_can_run_housekeeping_without_self_deadlock() {
+    use std::sync::mpsc;
+
+    let root = TestRoot::new("reader-housekeeping");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let stored = store.put(b"reader housekeeping shard").unwrap();
+    let temp = root
+        .0
+        .join("sha256")
+        .join(&stored.artifact_ref.digest_hex()[..2])
+        .join(".crank-artifact-tmp-v1-00000000000000000000000000000000-1-10");
+    fs::write(&temp, b"partial").unwrap();
+    let (_, candidates) = store.scan_stale_temps(Duration::ZERO, 128, 1).unwrap();
+    let mut reader = HousekeepingReader {
+        store: store.clone(),
+        candidate: candidates.into_iter().next(),
+        bytes: io::Cursor::new(b"reader bytes".to_vec()),
+    };
+    let writer_store = store.clone();
+    let (sender, receiver) = mpsc::channel();
+    thread::spawn(move || {
+        sender.send(writer_store.put_reader(&mut reader)).unwrap();
+    });
+
+    let stored = receiver
+        .recv_timeout(Duration::from_secs(2))
+        .expect("put_reader deadlocked while its reader ran housekeeping")
+        .unwrap();
+    assert_eq!(store.read(&stored.artifact_ref).unwrap(), b"reader bytes");
+    assert!(!temp.exists());
+}
+
+#[test]
+#[cfg(debug_assertions)]
+fn forked_crash_checkpoints_publish_only_complete_or_absent_blobs() {
+    let _guard = fault_guard();
+    if let Some(root) = std::env::var_os("CRANK_ARTIFACTS_CHILD_ROOT") {
+        set_checkpoint(
+            std::env::var("CRANK_ARTIFACTS_TEST_STAGE").unwrap(),
+            FaultAction::Exit,
+        );
+        let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
+        let _ = store.put(b"crash-consistent bytes");
+        panic!("checkpoint did not terminate the child");
+    }
+
+    for stage in ["write", "file_fsync", "publish", "directory_fsync"] {
+        let root = TestRoot::new(stage);
+        let status = Command::new(std::env::current_exe().unwrap())
+            .args([
+                "--exact",
+                "forked_crash_checkpoints_publish_only_complete_or_absent_blobs",
+                "--nocapture",
+            ])
+            .env("CRANK_ARTIFACTS_CHILD_ROOT", &root.0)
+            .env("CRANK_ARTIFACTS_TEST_STAGE", stage)
+            .status()
+            .unwrap();
+        assert_eq!(status.code(), Some(86), "stage={stage}");
+
+        let store = ArtifactStore::open(&root.0).unwrap();
+        let expected = ArtifactRef::from_digest_hex(&format!(
+            "{:x}",
+            Sha256::digest(b"crash-consistent bytes")
+        ))
+        .unwrap();
+        match store.read(&expected) {
+            Ok(bytes) => assert_eq!(bytes, b"crash-consistent bytes", "stage={stage}"),
+            Err(ArtifactError::NotFound) => {}
+            Err(error) => panic!("stage={stage} exposed an invalid final blob: {error}"),
+        }
+        let stored = store.put(b"crash-consistent bytes").unwrap();
+        assert_eq!(stored.artifact_ref, expected, "stage={stage}");
+        assert_eq!(
+            store.read(&expected).unwrap(),
+            b"crash-consistent bytes",
+            "stage={stage}"
+        );
+
+        let (_, candidates) = store.scan_stale_temps(Duration::ZERO, 128, 16).unwrap();
+        for candidate in candidates {
+            store.delete_stale_temp(candidate).unwrap();
+        }
+    }
+}
+
+#[test]
+#[cfg(debug_assertions)]
+fn disk_full_fault_seams_leave_a_retryable_store() {
+    let _guard = fault_guard();
+    if let Some(root) = std::env::var_os("CRANK_ARTIFACTS_CHILD_ROOT") {
+        set_checkpoint(
+            std::env::var("CRANK_ARTIFACTS_TEST_STAGE").unwrap(),
+            FaultAction::Fail,
+        );
+        let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
+        assert_eq!(
+            store.put(b"disk-full seam bytes").unwrap_err(),
+            ArtifactError::Storage
+        );
+        return;
+    }
+
+    for stage in ["write", "file_fsync", "publish", "directory_fsync"] {
+        let root = TestRoot::new(&format!("full-{stage}"));
+        let status = Command::new(std::env::current_exe().unwrap())
+            .args([
+                "--exact",
+                "disk_full_fault_seams_leave_a_retryable_store",
+                "--nocapture",
+            ])
+            .env("CRANK_ARTIFACTS_CHILD_ROOT", &root.0)
+            .env("CRANK_ARTIFACTS_TEST_STAGE", stage)
+            .status()
+            .unwrap();
+        assert!(status.success(), "stage={stage}");
+
+        let store = ArtifactStore::open(&root.0).unwrap();
+        let expected =
+            ArtifactRef::from_digest_hex(&format!("{:x}", Sha256::digest(b"disk-full seam bytes")))
+                .unwrap();
+        match store.read(&expected) {
+            Ok(bytes) => assert_eq!(bytes, b"disk-full seam bytes", "stage={stage}"),
+            Err(ArtifactError::NotFound) => {}
+            Err(error) => panic!("stage={stage} exposed an invalid final blob: {error}"),
+        }
+        assert_eq!(
+            store.put(b"disk-full seam bytes").unwrap().artifact_ref,
+            expected,
+            "stage={stage}"
+        );
+        assert_eq!(
+            store.read(&expected).unwrap(),
+            b"disk-full seam bytes",
+            "stage={stage}"
+        );
+    }
+}
+
+#[test]
+fn real_process_write_limit_leaves_no_partial_final_blob() {
+    if let Some(root) = std::env::var_os("CRANK_ARTIFACTS_CHILD_ROOT") {
+        // Keep the test's host filesystem untouched while making the actual
+        // write syscall fail as it would for a quota/disk-full condition.
+        unsafe {
+            assert_ne!(libc::signal(libc::SIGXFSZ, libc::SIG_IGN), libc::SIG_ERR);
+            let limit = libc::rlimit {
+                rlim_cur: 1,
+                rlim_max: 1,
+            };
+            assert_eq!(libc::setrlimit(libc::RLIMIT_FSIZE, &limit), 0);
+        }
+        let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
+        assert_eq!(
+            store.put(b"real write limit bytes").unwrap_err(),
+            ArtifactError::Storage
+        );
+        return;
+    }
+
+    let root = TestRoot::new("real-write-limit");
+    let status = Command::new(std::env::current_exe().unwrap())
+        .args([
+            "--exact",
+            "real_process_write_limit_leaves_no_partial_final_blob",
+            "--nocapture",
+        ])
+        .env("CRANK_ARTIFACTS_CHILD_ROOT", &root.0)
+        .status()
+        .unwrap();
+    assert!(status.success());
+
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let expected =
+        ArtifactRef::from_digest_hex(&format!("{:x}", Sha256::digest(b"real write limit bytes")))
+            .unwrap();
+    assert_eq!(store.read(&expected).unwrap_err(), ArtifactError::NotFound);
+    assert_eq!(
+        store.put(b"real write limit bytes").unwrap().artifact_ref,
+        expected
+    );
+    assert_eq!(store.read(&expected).unwrap(), b"real write limit bytes");
+}

crates/crank-config/src/debug.rs

@@ -1,8 +1,9 @@
 use std::fmt;
 
 use crate::{
-    AdminProcessConfig, CacheSettings, DatabaseSettings, McpProcessConfig, MetricsSettings,
+    AdminProcessConfig, CacheSettings, DatabaseSettings, ExternalReferenceSettings,
-    MigratorConfig, ObservabilitySettings, OtlpSettings, OutboundSettings, RuntimeSettings,
+    McpProcessConfig, MetricsSettings, MigratorConfig, ObservabilitySettings, OtlpSettings,
+    OutboundSettings, RuntimeSettings,
 };
 
 impl fmt::Debug for DatabaseSettings {
@@ -32,6 +33,18 @@ impl fmt::Debug for OutboundSettings {
             .finish()
     }
 }
+impl fmt::Debug for ExternalReferenceSettings {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("ExternalReferenceSettings")
+            .field("allowed_url_prefix_count", &self.allowed_url_prefixes.len())
+            .field("max_depth", &self.max_depth)
+            .field("max_documents", &self.max_documents)
+            .field("max_fetch_bytes", &self.max_fetch_bytes)
+            .field("fetch_timeout_ms", &self.fetch_timeout_ms)
+            .field("max_expanded_nodes", &self.max_expanded_nodes)
+            .finish()
+    }
+}
 impl fmt::Debug for RuntimeSettings {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         f.debug_struct("RuntimeSettings")
@@ -93,6 +106,7 @@ impl fmt::Debug for AdminProcessConfig {
         f.debug_struct("AdminProcessConfig")
             .field("database", &self.database)
             .field("runtime", &self.runtime)
+            .field("external_references", &self.external_references)
             .field("observability", &self.observability)
             .field("storage_root", &"configured")
             .field("session_secret", &self.session_secret)

crates/crank-config/src/lib.rs

@@ -15,8 +15,9 @@ pub use diagnostic::{ConfigError, Diagnostic, DiagnosticCode};
 pub use migrator::{MigratorConfig, parse_migrator};
 pub use process::{
     AdminProcessConfig, CacheBackend, CacheSettings, DatabaseSettings, DeprecationRecord,
-    EffectiveConfig, McpProcessConfig, MetricsSettings, ObservabilitySettings, OtlpSettings,
+    EffectiveConfig, ExternalReferenceSettings, McpProcessConfig, MetricsSettings,
-    OutboundSettings, PoolSettings, ProcessKind, RateLimitSettings, RuntimeSettings, parse_process,
+    ObservabilitySettings, OtlpSettings, OutboundSettings, PoolSettings, ProcessKind,
+    RateLimitSettings, RuntimeSettings, parse_process,
 };
 pub use schema::{
     FieldMode, FieldSpec, ProcessScope, Sensitivity, deployment_field_registry, field_registry,

crates/crank-config/src/process.rs

@@ -12,7 +12,9 @@ use std::{
     path::PathBuf,
 };
 use url::Url;
+mod external_references;
 mod list_parsers;
+pub use external_references::ExternalReferenceSettings;
 const MAX_ENV_VALUE_BYTES: usize = 8_192;
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub enum ProcessKind {
@@ -127,6 +129,7 @@ pub struct ObservabilitySettings {
 pub struct AdminProcessConfig {
     pub database: DatabaseSettings,
     pub runtime: RuntimeSettings,
+    pub external_references: ExternalReferenceSettings,
     pub observability: ObservabilitySettings,
     pub bind_addr: SocketAddr,
     pub storage_root: PathBuf,
@@ -154,8 +157,8 @@ pub struct McpProcessConfig {
 
 #[derive(Clone)]
 enum Projection {
-    Admin(AdminProcessConfig),
+    Admin(Box<AdminProcessConfig>),
-    Mcp(McpProcessConfig),
+    Mcp(Box<McpProcessConfig>),
 }
 
 #[derive(Clone)]
@@ -790,9 +793,10 @@ pub fn parse_process(
             {
                 parser.push(DiagnosticCode::UnsafeCombination, "admin.exposure.tls");
             }
-            Projection::Admin(AdminProcessConfig {
+            Projection::Admin(Box::new(AdminProcessConfig {
                 database,
                 runtime,
+                external_references: external_references::parse(&mut parser),
                 observability,
                 bind_addr,
                 storage_root: parser.absolute_path("CRANK_STORAGE_ROOT", "/var/lib/crank/storage"),
@@ -811,7 +815,7 @@ pub fn parse_process(
                 bootstrap_display_name: parser
                     .string("CRANK_BOOTSTRAP_ADMIN_DISPLAY_NAME", Some("Crank Owner")),
                 demo_seed: parser.boolean("CRANK_DEMO_SEED"),
-            })
+            }))
         }
         ProcessKind::McpServer => {
             let rps = parser.number("CRANK_MCP_RATE_LIMIT_RPS") as u32;
@@ -819,7 +823,7 @@ pub fn parse_process(
             if burst < rps {
                 parser.push(DiagnosticCode::UnsafeCombination, "mcp.rate_limit.burst");
             }
-            Projection::Mcp(McpProcessConfig {
+            Projection::Mcp(Box::new(McpProcessConfig {
                 database,
                 runtime,
                 observability,
@@ -829,7 +833,7 @@ pub fn parse_process(
                     requests_per_second: rps,
                     burst,
                 },
-            })
+            }))
         }
     };
 
@@ -867,6 +871,18 @@ fn fingerprint_parts(kind: ProcessKind, projection: &Projection) -> Vec<String>
                 format!("session_ttl={}", config.session_ttl_hours),
                 format!("trusted_proxies={:?}", config.trusted_proxy_ips),
                 format!("demo={}", config.demo_seed),
+                format!(
+                    "external_reference_prefixes={}",
+                    config.external_references.allowed_url_prefixes.join(",")
+                ),
+                format!(
+                    "external_reference_limits={}:{}:{}:{}:{}",
+                    config.external_references.max_depth,
+                    config.external_references.max_documents,
+                    config.external_references.max_fetch_bytes,
+                    config.external_references.fetch_timeout_ms,
+                    config.external_references.max_expanded_nodes,
+                ),
                 "storage=path-configured".to_owned(),
                 format!("session_secret={}", config.session_secret.is_configured()),
                 format!("pepper={}", config.password_pepper.is_configured()),

crates/crank-config/src/process/external_references.rs

@@ -0,0 +1,24 @@
+#[derive(Clone, Eq, PartialEq)]
+pub struct ExternalReferenceSettings {
+    /// Canonical HTTP(S) URL prefixes which opt an operator into remote `$ref` fetches.
+    /// An empty list is a deliberate default-deny switch.
+    pub allowed_url_prefixes: Vec<String>,
+    pub max_depth: usize,
+    pub max_documents: usize,
+    pub max_fetch_bytes: usize,
+    pub fetch_timeout_ms: u64,
+    pub max_expanded_nodes: usize,
+}
+
+pub(super) fn parse(parser: &mut super::Parser<'_>) -> ExternalReferenceSettings {
+    ExternalReferenceSettings {
+        allowed_url_prefixes: parser
+            .url_prefix_list("CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES"),
+        max_depth: parser.number("CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH") as usize,
+        max_documents: parser.number("CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS") as usize,
+        max_fetch_bytes: parser.number("CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES") as usize,
+        fetch_timeout_ms: parser.number("CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS"),
+        max_expanded_nodes: parser.number("CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES")
+            as usize,
+    }
+}

crates/crank-config/src/process/list_parsers.rs

@@ -4,6 +4,7 @@ use crate::{
     DiagnosticCode,
     validation::{parse_host_list, parse_ip_list},
 };
+use url::Url;
 
 impl super::Parser<'_> {
     pub(super) fn host_list(&mut self, name: &'static str) -> Vec<String> {
@@ -33,4 +34,41 @@ impl super::Parser<'_> {
         }
         parsed.items
     }
+
+    pub(super) fn url_prefix_list(&mut self, name: &'static str) -> Vec<String> {
+        let Some(raw) = self.optional(name) else {
+            return Vec::new();
+        };
+        let mut prefixes = Vec::new();
+        let mut invalid = false;
+        for item in raw.split(',') {
+            let item = item.trim();
+            let Ok(url) = Url::parse(item) else {
+                invalid = true;
+                continue;
+            };
+            if !matches!(url.scheme(), "http" | "https")
+                || url.host_str().is_none()
+                || !url.username().is_empty()
+                || url.password().is_some()
+                || url.query().is_some()
+                || url.fragment().is_some()
+            {
+                invalid = true;
+                continue;
+            }
+            let canonical = url.to_string();
+            if !prefixes.contains(&canonical) {
+                prefixes.push(canonical);
+            }
+        }
+        if invalid {
+            self.push(DiagnosticCode::InvalidType, name);
+        }
+        if prefixes.len() > 64 {
+            self.push(DiagnosticCode::OutOfRange, name);
+            prefixes.truncate(64);
+        }
+        prefixes
+    }
 }

crates/crank-config/src/schema.rs

@@ -78,7 +78,7 @@ macro_rules! f {
     };
 }
 
-static FIELDS: [FieldSpec; 59] = [
+static FIELDS: [FieldSpec; 65] = [
     FieldSpec {
         compatibility: Some("legacy URL form"),
         rules: &[
@@ -339,6 +339,78 @@ static FIELDS: [FieldSpec; 59] = [
         Some(67108864),
         Public
     ),
+    FieldSpec {
+        rules: &[
+            "empty list disables external OpenAPI reference fetching",
+            "each prefix must be canonical HTTP(S) without userinfo, query, or fragment",
+        ],
+        ..f!(
+            "import.external_references.allowed_url_prefixes",
+            "CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES",
+            AdminApi,
+            "url_prefix_list",
+            None,
+            Some(""),
+            None,
+            None,
+            Internal
+        )
+    },
+    f!(
+        "import.external_references.max_depth",
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH",
+        AdminApi,
+        "u32",
+        Some("edges"),
+        Some("8"),
+        Some(1),
+        Some(32),
+        Public
+    ),
+    f!(
+        "import.external_references.max_documents",
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS",
+        AdminApi,
+        "u32",
+        Some("documents"),
+        Some("32"),
+        Some(1),
+        Some(32),
+        Public
+    ),
+    f!(
+        "import.external_references.max_fetch_bytes",
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES",
+        AdminApi,
+        "u64",
+        Some("bytes"),
+        Some("262144"),
+        Some(1),
+        Some(4194304),
+        Public
+    ),
+    f!(
+        "import.external_references.fetch_timeout_ms",
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS",
+        AdminApi,
+        "u64",
+        Some("milliseconds"),
+        Some("10000"),
+        Some(1),
+        Some(300000),
+        Public
+    ),
+    f!(
+        "import.external_references.max_expanded_nodes",
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES",
+        AdminApi,
+        "u32",
+        Some("nodes"),
+        Some("10000"),
+        Some(1),
+        Some(100000),
+        Public
+    ),
     f!(
         "observability.environment",
         "CRANK_ENVIRONMENT",

crates/crank-config/tests/contract.rs

@@ -108,9 +108,64 @@ fn source_for(
 }
 
 #[test]
-fn registry_covers_exactly_the_59_observed_runtime_names() {
+fn external_reference_contract_is_default_off_and_validates_prefixes_and_limits() {
+    let config = parse_process(
+        ProcessKind::AdminApi,
+        ConfigSource::from_utf8(required_admin()),
+    )
+    .expect("the external reference contract has safe defaults");
+    let references = &config.admin().unwrap().external_references;
+    assert!(references.allowed_url_prefixes.is_empty());
+    assert_eq!(references.max_depth, 8);
+    assert_eq!(references.max_documents, 32);
+    assert_eq!(references.max_fetch_bytes, 262_144);
+    assert_eq!(references.fetch_timeout_ms, 10_000);
+    assert_eq!(references.max_expanded_nodes, 10_000);
+
+    let mut allowed = required_admin();
+    allowed.insert(
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES".to_owned(),
+        "https://schemas.example.test/openapi/,https://schemas.example.test/openapi/".to_owned(),
+    );
+    let config = parse_process(ProcessKind::AdminApi, ConfigSource::from_utf8(allowed)).unwrap();
+    assert_eq!(
+        config
+            .admin()
+            .unwrap()
+            .external_references
+            .allowed_url_prefixes,
+        vec!["https://schemas.example.test/openapi/".to_owned()]
+    );
+
+    let mut invalid = required_admin();
+    invalid.insert(
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES".to_owned(),
+        "https://user:secret@schemas.example.test/openapi/".to_owned(),
+    );
+    let error = parse_process(ProcessKind::AdminApi, ConfigSource::from_utf8(invalid))
+        .expect_err("userinfo must never become an operator allow rule");
+    assert!(error.diagnostics().iter().any(|diagnostic| {
+        diagnostic.code == DiagnosticCode::InvalidType
+            && diagnostic.field == "import.external_references.allowed_url_prefixes"
+    }));
+
+    let mut out_of_range = required_admin();
+    out_of_range.insert(
+        "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS".to_owned(),
+        "33".to_owned(),
+    );
+    let error = parse_process(ProcessKind::AdminApi, ConfigSource::from_utf8(out_of_range))
+        .expect_err("graph document limit must be bounded");
+    assert!(error.diagnostics().iter().any(|diagnostic| {
+        diagnostic.code == DiagnosticCode::OutOfRange
+            && diagnostic.field == "import.external_references.max_documents"
+    }));
+}
+
+#[test]
+fn registry_covers_exactly_the_65_observed_runtime_names() {
     let registry = field_registry();
-    assert_eq!(registry.len(), 59);
+    assert_eq!(registry.len(), 65);
     let unique = registry
         .iter()
         .map(|field| field.env_name)

crates/crank-config/tests/generation.rs

@@ -30,6 +30,9 @@ fn generated_reference_distinguishes_required_and_optional_fields() {
     assert!(reference.contains(
         "| `CRANK_LOG_LEVEL` | `observability.log_filter` | `Shared` | `string/-` | `blank` |"
     ));
+    assert!(reference.contains(
+        "| `CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES` | `import.external_references.allowed_url_prefixes` | `AdminApi` | `url_prefix_list/-` | `` |"
+    ));
 }
 
 #[test]

crates/crank-import/src/rest/mod.rs

@@ -8,19 +8,22 @@ mod normalize_schema;
 mod openapi3;
 mod payload;
 mod recommendations;
+mod reference;
 mod schema;
 mod swagger2;
 
 pub use model::{
-    ImportFinding, ImportFindingSeverity, ImportGroupPreview, ImportOperationCandidate,
+    ExternalDocumentSnapshot, ImportFinding, ImportFindingSeverity, ImportGroupPreview,
-    ImportPreview, ImportSourcePreview, NORMALIZER_VERSION, NormalizationConfig, NormalizedFinding,
+    ImportOperationCandidate, ImportPreview, ImportSourcePreview, NORMALIZER_VERSION,
-    NormalizedIr, NormalizedOperation, NormalizedParameter, NormalizedReference, NormalizedSchema,
+    NormalizationConfig, NormalizedFinding, NormalizedIr, NormalizedOperation, NormalizedParameter,
-    NormalizedSchemaKind, PROJECTION_VERSION, RestImportCandidate, RestImportDocument,
+    NormalizedReference, NormalizedSchema, NormalizedSchemaConstraints, NormalizedSchemaKind,
-    RestImportOperation, RestImportParameter, RestParameterLocation, SourceDigest, SourceIdentity,
+    PROJECTION_VERSION, ResolvedReferenceEdge, ResolvedReferenceGraph, ResolvedReferenceNode,
-    SourceLocation, UnresolvedReference,
+    RestImportCandidate, RestImportDocument, RestImportOperation, RestImportParameter,
+    RestParameterLocation, SourceDigest, SourceIdentity, SourceLocation, UnresolvedReference,
 };
 pub use normalize::{
-    ImportParseError, normalize_verified_document, preview_document, preview_document_legacy_v1,
+    ImportParseError, external_reference_uris, normalize_verified_bundle,
-    preview_from_ir, validate_normalized_ir,
+    normalize_verified_document, preview_document, preview_document_legacy_v1, preview_from_ir,
+    reference_uris, validate_normalized_ir,
 };
 pub use payload::operation_draft_from_candidate;

crates/crank-import/src/rest/model.rs

@@ -9,10 +9,10 @@ use serde_json::Value;
 /// The immutable contract used to normalize an OpenAPI source.  These names
 /// deliberately travel with an import job: changing either contract must not
 /// silently reinterpret a pending preview.
-pub const NORMALIZER_VERSION: &str = "normalized-ir-v2";
+pub const NORMALIZER_VERSION: &str = "normalized-ir-v3";
-pub const PROJECTION_VERSION: &str = "preview-v2";
+pub const PROJECTION_VERSION: &str = "preview-v3";
 
-#[derive(Clone, Debug, PartialEq, Eq, Serialize)]
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize)]
 pub struct SourceDigest(String);
 
 impl SourceDigest {
@@ -97,6 +97,19 @@ pub struct NormalizationConfig {
     pub max_collection_items: usize,
     pub max_aliases: usize,
     pub max_scalar_bytes: usize,
+    /// Maximum number of reference hops followed from one source location.
+    pub max_reference_depth: usize,
+    /// Maximum number of `$ref` occurrences inspected across the bundle.
+    pub max_references: usize,
+    /// Maximum number of immutable external documents supplied to the pure resolver.
+    pub max_reference_documents: usize,
+    /// Maximum number of nodes copied while expanding resolved references.
+    pub max_expanded_nodes: usize,
+    pub max_external_document_bytes: usize,
+    /// Signals that orchestration enabled external fetching. It never permits
+    /// I/O in this crate; it only distinguishes default-deny from a missing or
+    /// rejected supplied snapshot in exact findings.
+    pub external_references_enabled: bool,
 }
 
 impl Default for NormalizationConfig {
@@ -110,6 +123,12 @@ impl Default for NormalizationConfig {
             max_collection_items: 10_000,
             max_aliases: 128,
             max_scalar_bytes: 256 * 1024,
+            max_reference_depth: 32,
+            max_references: 4_096,
+            max_reference_documents: 32,
+            max_expanded_nodes: 100_000,
+            max_external_document_bytes: 256 * 1024,
+            external_references_enabled: false,
         }
     }
 }
@@ -139,6 +158,38 @@ pub struct UnresolvedReference {
     pub location: SourceLocation,
 }
 
+/// Immutable external input for the pure reference resolver. The caller owns
+/// URL policy and I/O; `crank-import` only consumes already verified bytes.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ExternalDocumentSnapshot {
+    pub canonical_uri: String,
+    pub digest: SourceDigest,
+    pub document: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ResolvedReferenceNode {
+    pub snapshot_digest: SourceDigest,
+    pub location: SourceLocation,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ResolvedReferenceEdge {
+    pub source: ResolvedReferenceNode,
+    pub target: ResolvedReferenceNode,
+    pub recursive: bool,
+}
+
+#[derive(Clone, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ResolvedReferenceGraph {
+    /// Sorted, deduplicated immutable dependency identities. Canonical URLs
+    /// deliberately do not enter the IR or public diagnostics.
+    #[serde(default)]
+    pub dependency_digests: Vec<SourceDigest>,
+    #[serde(default)]
+    pub edges: Vec<ResolvedReferenceEdge>,
+}
+
 /// An unresolved `$ref` preserved from the decoded source. This is deliberately
 /// broader than schema references: path items, reusable parameters and other
 /// object-level references remain available to a later resolution phase.
@@ -213,9 +264,34 @@ pub struct NormalizedSchema {
     pub location: SourceLocation,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub description: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub discriminator: Option<NormalizedDiscriminator>,
+    #[serde(default)]
+    pub constraints: NormalizedSchemaConstraints,
     pub kind: NormalizedSchemaKind,
 }
 
+#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
+pub struct NormalizedSchemaConstraints {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub minimum: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub maximum: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub min_length: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_length: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub pattern: Option<String>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct NormalizedDiscriminator {
+    pub property_name: String,
+    #[serde(default)]
+    pub mapping: BTreeMap<String, String>,
+}
+
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 #[serde(rename_all = "snake_case", tag = "type")]
 pub enum NormalizedSchemaKind {
@@ -280,6 +356,8 @@ pub struct NormalizedIr {
     pub paths: Vec<NormalizedPath>,
     #[serde(default)]
     pub unresolved_references: Vec<NormalizedReference>,
+    #[serde(default)]
+    pub reference_graph: ResolvedReferenceGraph,
     pub source: ImportSourcePreview,
     #[serde(default)]
     pub operations: Vec<NormalizedOperation>,
@@ -416,7 +494,7 @@ fn empty_source_location() -> SourceLocation {
     }
 }
 
-#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
 #[serde(rename_all = "snake_case")]
 pub enum RestParameterLocation {
     Path,

crates/crank-import/src/rest/normalize.rs

@@ -5,13 +5,13 @@ use thiserror::Error;
 
 use crate::rest::{
     model::{
-        CoverageDisposition, CoverageEntry, ImportFinding, ImportFindingSeverity,
+        CoverageDisposition, CoverageEntry, ExternalDocumentSnapshot, ImportFinding,
-        ImportGroupPreview, ImportPreview, ImportSourcePreview, NORMALIZER_VERSION,
+        ImportFindingSeverity, ImportGroupPreview, ImportPreview, ImportSourcePreview,
-        NormalizationConfig, NormalizedApiMetadata, NormalizedFinding, NormalizedIr,
+        NORMALIZER_VERSION, NormalizationConfig, NormalizedApiMetadata, NormalizedFinding,
-        NormalizedLiteral, NormalizedOperation, NormalizedParameter, NormalizedScalarKind,
+        NormalizedIr, NormalizedLiteral, NormalizedOperation, NormalizedParameter,
-        NormalizedSchema, NormalizedSchemaKind, PROJECTION_VERSION, RestImportDocument,
+        NormalizedScalarKind, NormalizedSchema, NormalizedSchemaKind, PROJECTION_VERSION,
-        RestImportOperation, RestImportParameter, SourceDigest, SourceIdentity, SourceLocation,
+        RestImportDocument, RestImportOperation, RestImportParameter, SourceDigest, SourceIdentity,
-        SourceSyntax,
+        SourceLocation, SourceSyntax,
     },
     openapi3,
     payload::candidate_from_operation,
@@ -39,6 +39,25 @@ pub fn preview_document(document: &str) -> Result<ImportPreview, ImportParseErro
     Ok(preview_from_ir(&ir))
 }
 
+/// Enumerates reference URIs without resolving or performing I/O. Intended
+/// for an orchestration layer that materializes a bounded immutable bundle.
+pub fn reference_uris(
+    document: &str,
+    config: &NormalizationConfig,
+) -> Result<Vec<String>, ImportParseError> {
+    super::reference::reference_uris(document, config)
+}
+
+/// Enumerates references in an already materialized external document. This
+/// keeps the historical primary-document API intact while applying the
+/// external-document byte limit at an explicit call site.
+pub fn external_reference_uris(
+    document: &str,
+    config: &NormalizationConfig,
+) -> Result<Vec<String>, ImportParseError> {
+    super::reference::external_reference_uris(document, config)
+}
+
 /// Compatibility projection for jobs created before `NormalizedIr`. It is
 /// deliberately isolated from the v2 pipeline and may be removed only after
 /// the import-job TTL has elapsed.
@@ -83,6 +102,18 @@ pub fn normalize_verified_document(
     document: &str,
     digest: SourceDigest,
     config: &NormalizationConfig,
+) -> Result<NormalizedIr, ImportParseError> {
+    normalize_verified_bundle(document, digest, &[], config)
+}
+
+/// Normalizes a verified primary document together with immutable external
+/// snapshots. This function is pure: callers must perform all URL policy,
+/// fetching, artifact persistence and digest verification beforehand.
+pub fn normalize_verified_bundle(
+    document: &str,
+    digest: SourceDigest,
+    snapshots: &[ExternalDocumentSnapshot],
+    config: &NormalizationConfig,
 ) -> Result<NormalizedIr, ImportParseError> {
     config
         .validate_versions()
@@ -99,6 +130,14 @@ pub fn normalize_verified_document(
     };
     let root = decode(document)?;
     validate_value_limits(&root, config)?;
+    for snapshot in snapshots {
+        if snapshot.document.len() > config.max_external_document_bytes {
+            return Err(ImportParseError::LimitExceeded);
+        }
+    }
+    let resolution = super::reference::resolve(root, digest.clone(), snapshots, config)?;
+    let root = resolution.root;
+    validate_value_limits(&root, config)?;
 
     let parsed = match root.get("openapi").and_then(Value::as_str) {
         Some(version) if supported_oas_version(version) => openapi3::parse_document(&root)?,
@@ -112,7 +151,15 @@ pub fn normalize_verified_document(
     if parsed.operations.is_empty() {
         return Err(ImportParseError::NoMethods);
     }
-    canonicalize(parsed, digest, config, root, source_syntax)
+    canonicalize(
+        parsed,
+        digest,
+        config,
+        root,
+        source_syntax,
+        resolution.graph,
+        resolution.findings,
+    )
 }
 
 pub fn preview_from_ir(ir: &NormalizedIr) -> ImportPreview {
@@ -202,6 +249,8 @@ fn canonicalize(
     config: &NormalizationConfig,
     root: Value,
     source_syntax: SourceSyntax,
+    reference_graph: crate::rest::model::ResolvedReferenceGraph,
+    resolution_findings: Vec<NormalizedFinding>,
 ) -> Result<NormalizedIr, ImportParseError> {
     let source = ImportSourcePreview {
         format: document.format,
@@ -260,6 +309,7 @@ fn canonicalize(
             }
         })
         .collect::<Vec<_>>();
+    findings.extend(resolution_findings);
     let mut operations = document
         .operations
         .into_iter()
@@ -466,6 +516,40 @@ fn canonicalize(
                 right.location.pointer.as_str(),
             ))
     });
+    let mut document_findings = Vec::with_capacity(findings.len());
+    for mut finding in findings.drain(..) {
+        let resolution_finding = matches!(
+            finding.code.as_str(),
+            "external_reference_disabled"
+                | "reference_target_missing"
+                | "reference_uri_malformed"
+                | "external_reference_unavailable"
+                | "reference_type_mismatch"
+                | "reference_graph_limit"
+                | "all_of_conflict"
+                | "unsupported_composition"
+                | "unsupported_discriminator"
+        );
+        let target = resolution_finding
+            .then(|| {
+                operations.iter_mut().find(|operation| {
+                    !finding.location.pointer.is_empty()
+                        && (finding.location.pointer == operation.location.pointer
+                            || finding
+                                .location
+                                .pointer
+                                .starts_with(&format!("{}/", operation.location.pointer)))
+                })
+            })
+            .flatten();
+        if let Some(operation) = target {
+            finding.operation_key = Some(operation.key.clone());
+            operation.findings.push(finding);
+        } else {
+            document_findings.push(finding);
+        }
+    }
+    findings = document_findings;
     for operation in &mut operations {
         sort_normalized_findings(&mut operation.findings);
     }
@@ -501,10 +585,9 @@ fn canonicalize(
         }
         sort_normalized_findings(&mut operation.findings);
     }
-    // Resolution is deliberately deferred. References already represented by
+    // References left in the finite projection are precisely failures which
-    // an operation schema use that operation's blocker; all other references
+    // the bounded resolver could not safely expand. Keep a blocker at the
-    // need a document-level blocker so omitted object-level semantics can
+    // closest affected operation; unrelated operations remain actionable.
-    // never reach Apply silently.
     findings.extend(
         unresolved_references
             .iter()
@@ -529,7 +612,11 @@ fn canonicalize(
             }),
     );
     sort_normalized_findings(&mut findings);
-    for finding in &findings {
+    for finding in findings.iter().chain(
+        operations
+            .iter()
+            .flat_map(|operation| operation.findings.iter()),
+    ) {
         if !coverage.iter().any(|entry| {
             entry.construct_id == finding.construct_id
                 && entry.location == finding.location
@@ -580,6 +667,7 @@ fn canonicalize(
             .unwrap_or_default(),
         paths: normalize_coverage::paths_from_operations(&operations, version),
         unresolved_references,
+        reference_graph,
         source,
         operations,
         findings,
@@ -593,9 +681,56 @@ pub fn validate_normalized_ir(ir: &NormalizedIr) -> Result<(), ImportParseError>
     if ir.normalizer_version != NORMALIZER_VERSION || ir.projection_version != PROJECTION_VERSION {
         return Err(ImportParseError::InvalidDocument);
     }
+    validate_reference_graph(ir)?;
     normalize_coverage::validate_coverage(ir)
 }
 
+fn validate_reference_graph(ir: &NormalizedIr) -> Result<(), ImportParseError> {
+    if ir
+        .reference_graph
+        .dependency_digests
+        .windows(2)
+        .any(|pair| pair[0] >= pair[1])
+    {
+        return Err(ImportParseError::InvalidDocument);
+    }
+    let allowed = ir
+        .reference_graph
+        .dependency_digests
+        .iter()
+        .chain(std::iter::once(&ir.source_identity.digest))
+        .map(SourceDigest::as_str)
+        .collect::<BTreeSet<_>>();
+    let edges = &ir.reference_graph.edges;
+    if edges
+        .windows(2)
+        .any(|pair| reference_edge_key(&pair[0]) >= reference_edge_key(&pair[1]))
+        || edges.iter().any(|edge| {
+            !allowed.contains(edge.source.snapshot_digest.as_str())
+                || !allowed.contains(edge.target.snapshot_digest.as_str())
+                || !(edge.source.location.pointer.is_empty()
+                    || edge.source.location.pointer.starts_with('/'))
+                || !(edge.target.location.pointer.is_empty()
+                    || edge.target.location.pointer.starts_with('/'))
+        })
+    {
+        return Err(ImportParseError::InvalidDocument);
+    }
+    Ok(())
+}
+
+fn reference_edge_key(
+    edge: &crate::rest::model::ResolvedReferenceEdge,
+) -> (&str, &str, &str, &str, bool) {
+    (
+        edge.source.snapshot_digest.as_str(),
+        edge.source.location.pointer.as_str(),
+        edge.target.snapshot_digest.as_str(),
+        edge.target.location.pointer.as_str(),
+        edge.recursive,
+    )
+}
+
 fn coverage_disposition_rank(disposition: &CoverageDisposition) -> u8 {
     match disposition {
         CoverageDisposition::Mapped => 0,
@@ -650,7 +785,10 @@ fn legacy_schema_value(schema: &NormalizedSchema) -> Value {
     let mut value = match &schema.kind {
         NormalizedSchemaKind::Reference { reference } => serde_json::json!({"$ref": reference.uri}),
         NormalizedSchemaKind::Composition { operator, variants } => {
-            serde_json::json!({operator: variants.iter().map(legacy_schema_value).collect::<Vec<_>>() })
+            serde_json::json!({
+                operator: variants.iter().map(legacy_schema_value).collect::<Vec<_>>(),
+                "x-crank-lossless-composition": true,
+            })
         }
         NormalizedSchemaKind::Object {
             properties,
@@ -718,7 +856,7 @@ fn schema_contains_reference(schema: &NormalizedSchema, location: &SourceLocatio
     }
 }
 
-fn validate_value_limits(
+pub(super) fn validate_value_limits(
     value: &Value,
     config: &NormalizationConfig,
 ) -> Result<(), ImportParseError> {

crates/crank-import/src/rest/normalize_schema.rs

@@ -1,8 +1,9 @@
 use serde_json::Value;
 
 use crate::rest::model::{
-    CoverageDisposition, CoverageEntry, NormalizedLiteral, NormalizedScalarKind, NormalizedSchema,
+    CoverageDisposition, CoverageEntry, NormalizedDiscriminator, NormalizedLiteral,
-    NormalizedSchemaKind, SourceLocation, UnresolvedReference,
+    NormalizedScalarKind, NormalizedSchema, NormalizedSchemaConstraints, NormalizedSchemaKind,
+    SourceLocation, UnresolvedReference,
 };
 
 use super::normalize_coverage::escape_pointer;
@@ -150,6 +151,35 @@ pub(super) fn typed_schema(
             .get("description")
             .and_then(Value::as_str)
             .map(ToOwned::to_owned),
+        discriminator: value.get("discriminator").and_then(|value| {
+            let property_name = value.get("propertyName")?.as_str()?.to_owned();
+            let mapping = value
+                .get("mapping")
+                .and_then(Value::as_object)
+                .map(|mapping| {
+                    mapping
+                        .iter()
+                        .filter_map(|(key, value)| {
+                            value.as_str().map(|value| (key.clone(), value.to_owned()))
+                        })
+                        .collect()
+                })
+                .unwrap_or_default();
+            Some(NormalizedDiscriminator {
+                property_name,
+                mapping,
+            })
+        }),
+        constraints: NormalizedSchemaConstraints {
+            minimum: value.get("minimum").and_then(Value::as_f64),
+            maximum: value.get("maximum").and_then(Value::as_f64),
+            min_length: value.get("minLength").and_then(Value::as_u64),
+            max_length: value.get("maxLength").and_then(Value::as_u64),
+            pattern: value
+                .get("pattern")
+                .and_then(Value::as_str)
+                .map(ToOwned::to_owned),
+        },
         kind,
     }
 }

crates/crank-import/src/rest/openapi3.rs

@@ -133,6 +133,7 @@ fn parse_document_v2(root: &Value) -> Result<RestImportDocument, ImportParseErro
                 parsed_parameters.findings,
             );
             operation_parameters.extend(parsed_parameters.parameters);
+            deduplicate_parameters(&mut operation_parameters);
             let operation_servers = parse_servers(
                 operation_value.get("servers"),
                 &format!("{operation_pointer}/servers"),
@@ -203,6 +204,13 @@ fn parse_document_v2(root: &Value) -> Result<RestImportDocument, ImportParseErro
     })
 }
 
+fn deduplicate_parameters(parameters: &mut Vec<RestImportParameter>) {
+    let mut seen = std::collections::BTreeSet::new();
+    parameters.reverse();
+    parameters.retain(|parameter| seen.insert((parameter.name.clone(), parameter.location)));
+    parameters.reverse();
+}
+
 pub fn parse_document_legacy_v1(root: &Value) -> Result<RestImportDocument, ImportParseError> {
     let version = root
         .get("openapi")

crates/crank-import/src/rest/reference.rs

@@ -0,0 +1,872 @@
+use std::collections::{BTreeMap, BTreeSet};
+
+use serde_json::{Map, Value};
+
+use crate::rest::model::{
+    ExternalDocumentSnapshot, ImportFindingSeverity, NormalizationConfig, NormalizedFinding,
+    ResolvedReferenceEdge, ResolvedReferenceGraph, ResolvedReferenceNode, SourceDigest,
+    SourceLocation,
+};
+
+use super::{ImportParseError, normalize_coverage::escape_pointer};
+
+pub(super) struct ResolutionResult {
+    pub root: Value,
+    pub graph: ResolvedReferenceGraph,
+    pub findings: Vec<NormalizedFinding>,
+}
+
+pub(super) fn reference_uris(
+    document: &str,
+    config: &NormalizationConfig,
+) -> Result<Vec<String>, ImportParseError> {
+    reference_uris_with_max_bytes(document, config, config.max_bytes)
+}
+
+pub(super) fn external_reference_uris(
+    document: &str,
+    config: &NormalizationConfig,
+) -> Result<Vec<String>, ImportParseError> {
+    reference_uris_with_max_bytes(document, config, config.max_external_document_bytes)
+}
+
+fn reference_uris_with_max_bytes(
+    document: &str,
+    config: &NormalizationConfig,
+    max_bytes: usize,
+) -> Result<Vec<String>, ImportParseError> {
+    fn collect(value: &Value, uris: &mut BTreeSet<String>) {
+        match value {
+            Value::Object(object) => {
+                if let Some(reference) = object.get("$ref").and_then(Value::as_str) {
+                    uris.insert(reference.to_owned());
+                }
+                for (key, child) in object {
+                    if is_literal_payload_key(key) {
+                        continue;
+                    }
+                    collect(child, uris);
+                }
+            }
+            Value::Array(items) => {
+                for child in items {
+                    collect(child, uris);
+                }
+            }
+            _ => {}
+        }
+    }
+    if document.len() > max_bytes
+        || super::normalize_limits::alias_count(document) > config.max_aliases
+    {
+        return Err(ImportParseError::LimitExceeded);
+    }
+    let root = decode_snapshot(document)?;
+    super::normalize::validate_value_limits(&root, config)?;
+    let mut uris = BTreeSet::new();
+    collect(&root, &mut uris);
+    Ok(uris.into_iter().collect())
+}
+
+struct Document {
+    uri: Option<String>,
+    digest: SourceDigest,
+    root: Value,
+    oas31: bool,
+}
+
+#[derive(Clone, Debug, Eq, Ord, PartialEq, PartialOrd)]
+struct NodeKey {
+    digest: String,
+    pointer: String,
+}
+
+#[derive(Clone, Copy)]
+struct TraversalLocation<'a> {
+    projection: &'a str,
+    origin: &'a str,
+}
+
+struct Resolver<'a> {
+    config: &'a NormalizationConfig,
+    documents: Vec<Document>,
+    by_uri: BTreeMap<String, usize>,
+    graph: ResolvedReferenceGraph,
+    findings: Vec<NormalizedFinding>,
+    references: usize,
+    expanded_nodes: usize,
+}
+
+pub(super) fn resolve(
+    root: Value,
+    primary_digest: SourceDigest,
+    snapshots: &[ExternalDocumentSnapshot],
+    config: &NormalizationConfig,
+) -> Result<ResolutionResult, ImportParseError> {
+    if snapshots.len() > config.max_reference_documents {
+        return Err(ImportParseError::LimitExceeded);
+    }
+    let mut documents = vec![Document {
+        uri: None,
+        digest: primary_digest,
+        oas31: is_oas31(&root),
+        root,
+    }];
+    let mut by_uri = BTreeMap::new();
+    let mut dependency_digests = BTreeSet::new();
+    let primary_oas31 = documents[0].oas31;
+    for snapshot in snapshots {
+        if snapshot.canonical_uri.is_empty()
+            || by_uri
+                .insert(snapshot.canonical_uri.clone(), documents.len())
+                .is_some()
+        {
+            return Err(ImportParseError::InvalidDocument);
+        }
+        if super::normalize_limits::alias_count(&snapshot.document) > config.max_aliases {
+            return Err(ImportParseError::LimitExceeded);
+        }
+        let root = decode_snapshot(&snapshot.document)?;
+        super::normalize::validate_value_limits(&root, config)?;
+        dependency_digests.insert(snapshot.digest.clone());
+        documents.push(Document {
+            uri: Some(snapshot.canonical_uri.clone()),
+            digest: snapshot.digest.clone(),
+            // External documents are fragments of the primary contract. In a
+            // 3.1 bundle they therefore use 3.1 `$ref` sibling semantics even
+            // when the fragment itself omits an `openapi` declaration.
+            oas31: primary_oas31 || is_oas31(&root),
+            root,
+        });
+    }
+    let mut resolver = Resolver {
+        config,
+        documents,
+        by_uri,
+        graph: ResolvedReferenceGraph {
+            dependency_digests: dependency_digests.into_iter().collect(),
+            edges: Vec::new(),
+        },
+        findings: Vec::new(),
+        references: 0,
+        expanded_nodes: 0,
+    };
+    let root = resolver.documents[0].root.clone();
+    let root = resolver.expand_value(
+        0,
+        root,
+        TraversalLocation {
+            projection: "",
+            origin: "",
+        },
+        0,
+        &mut Vec::new(),
+    )?;
+    resolver
+        .graph
+        .edges
+        .sort_by(|left, right| edge_key(left).cmp(&edge_key(right)));
+    resolver.graph.edges.dedup();
+    resolver.findings.sort_by(|left, right| {
+        (
+            &left.operation_key,
+            &left.construct_id,
+            &left.code,
+            &left.location.pointer,
+        )
+            .cmp(&(
+                &right.operation_key,
+                &right.construct_id,
+                &right.code,
+                &right.location.pointer,
+            ))
+    });
+    resolver.findings.dedup();
+    Ok(ResolutionResult {
+        root,
+        graph: resolver.graph,
+        findings: resolver.findings,
+    })
+}
+
+impl Resolver<'_> {
+    fn expand_value(
+        &mut self,
+        document_index: usize,
+        value: Value,
+        location: TraversalLocation<'_>,
+        depth: usize,
+        stack: &mut Vec<NodeKey>,
+    ) -> Result<Value, ImportParseError> {
+        if depth > 0 {
+            self.expanded_nodes = self.expanded_nodes.saturating_add(1);
+            if self.expanded_nodes > self.config.max_expanded_nodes {
+                self.push_finding("reference_graph_limit", location.projection);
+                return Ok(Value::Object(Map::new()));
+            }
+        }
+        match value {
+            Value::Object(mut object) => {
+                if let Some(reference) = object
+                    .get("$ref")
+                    .and_then(Value::as_str)
+                    .map(str::to_owned)
+                {
+                    return self.expand_reference(
+                        document_index,
+                        object,
+                        &reference,
+                        location,
+                        depth,
+                        stack,
+                    );
+                }
+                let keys = object.keys().cloned().collect::<Vec<_>>();
+                for key in keys {
+                    if let Some(child) = object.remove(&key) {
+                        if is_literal_payload_key(&key) {
+                            object.insert(key, child);
+                            continue;
+                        }
+                        let child_pointer =
+                            format!("{}/{}", location.projection, escape_pointer(&key));
+                        let child_origin_pointer =
+                            format!("{}/{}", location.origin, escape_pointer(&key));
+                        object.insert(
+                            key,
+                            self.expand_value(
+                                document_index,
+                                child,
+                                TraversalLocation {
+                                    projection: &child_pointer,
+                                    origin: &child_origin_pointer,
+                                },
+                                depth,
+                                stack,
+                            )?,
+                        );
+                    }
+                }
+                let composition_count = ["allOf", "oneOf", "anyOf"]
+                    .into_iter()
+                    .filter(|operator| object.contains_key(*operator))
+                    .count();
+                if composition_count > 1 {
+                    self.push_finding("unsupported_composition", location.projection);
+                }
+                if let Some(discriminator) = object.get("discriminator") {
+                    let valid = discriminator
+                        .get("propertyName")
+                        .and_then(Value::as_str)
+                        .is_some_and(|value| !value.is_empty())
+                        && discriminator.get("mapping").is_none_or(|mapping| {
+                            mapping
+                                .as_object()
+                                .is_some_and(|mapping| mapping.values().all(Value::is_string))
+                        })
+                        && (object.contains_key("oneOf") || object.contains_key("anyOf"));
+                    if !valid {
+                        self.push_finding("unsupported_discriminator", location.projection);
+                    }
+                }
+                self.merge_all_of(object, location.projection)
+            }
+            Value::Array(items) => Ok(Value::Array(
+                items
+                    .into_iter()
+                    .enumerate()
+                    .map(|(index, child)| {
+                        self.expand_value(
+                            document_index,
+                            child,
+                            TraversalLocation {
+                                projection: &format!("{}/{index}", location.projection),
+                                origin: &format!("{}/{index}", location.origin),
+                            },
+                            depth,
+                            stack,
+                        )
+                    })
+                    .collect::<Result<Vec<_>, _>>()?,
+            )),
+            other => Ok(other),
+        }
+    }
+
+    fn expand_reference(
+        &mut self,
+        document_index: usize,
+        mut source_object: Map<String, Value>,
+        reference: &str,
+        location: TraversalLocation<'_>,
+        depth: usize,
+        stack: &mut Vec<NodeKey>,
+    ) -> Result<Value, ImportParseError> {
+        self.references = self.references.saturating_add(1);
+        if self.references > self.config.max_references || depth >= self.config.max_reference_depth
+        {
+            self.push_finding(
+                "reference_graph_limit",
+                &format!("{}/$ref", location.projection),
+            );
+            return Ok(Value::Object(source_object));
+        }
+        let target = self.target(document_index, reference);
+        if matches!(target, Err(TargetError::MalformedFragment)) {
+            self.push_finding(
+                "reference_uri_malformed",
+                &format!("{}/$ref", location.projection),
+            );
+            return Ok(Value::Object(source_object));
+        }
+        let Some((target_document, target_pointer)) = target.ok().flatten() else {
+            let external = reference.starts_with("http://") || reference.starts_with("https://");
+            let code = if external {
+                if self.config.external_references_enabled {
+                    "external_reference_unavailable"
+                } else {
+                    "external_reference_disabled"
+                }
+            } else {
+                "reference_target_missing"
+            };
+            self.push_finding(code, &format!("{}/$ref", location.projection));
+            return Ok(Value::Object(source_object));
+        };
+        let source = ResolvedReferenceNode {
+            snapshot_digest: self.documents[document_index].digest.clone(),
+            location: SourceLocation {
+                pointer: format!("{}/$ref", location.origin),
+            },
+        };
+        let target = ResolvedReferenceNode {
+            snapshot_digest: self.documents[target_document].digest.clone(),
+            location: SourceLocation {
+                pointer: target_pointer.clone(),
+            },
+        };
+        let key = NodeKey {
+            digest: target.snapshot_digest.as_str().to_owned(),
+            pointer: target_pointer.clone(),
+        };
+        let recursive = stack.contains(&key);
+        self.graph.edges.push(ResolvedReferenceEdge {
+            source,
+            target,
+            recursive,
+        });
+        if recursive {
+            // The edge is the lossless representation. Expansion stops here,
+            // producing an opaque object for the finite preview projection.
+            return Ok(Value::Object(Map::new()));
+        }
+        let Some(target_value) = self.documents[target_document]
+            .root
+            .pointer(&target_pointer)
+            .cloned()
+        else {
+            self.push_finding(
+                "reference_target_missing",
+                &format!("{}/$ref", location.projection),
+            );
+            return Ok(Value::Object(source_object));
+        };
+        if !target_value.is_object() {
+            self.push_finding(
+                "reference_type_mismatch",
+                &format!("{}/$ref", location.projection),
+            );
+            return Ok(Value::Object(source_object));
+        }
+        stack.push(key);
+        let mut expanded = self.expand_value(
+            target_document,
+            target_value,
+            TraversalLocation {
+                projection: location.projection,
+                origin: &target_pointer,
+            },
+            depth + 1,
+            stack,
+        )?;
+        stack.pop();
+
+        // OAS 3.1 Schema Objects permit siblings next to `$ref`; OAS 3.0 and
+        // Swagger Reference Objects ignore them. The source document version
+        // controls the semantics at the reference site.
+        source_object.remove("$ref");
+        if self.documents[document_index].oas31 && !source_object.is_empty() {
+            let Value::Object(expanded_object) = &mut expanded else {
+                self.push_finding(
+                    "reference_type_mismatch",
+                    &format!("{}/$ref", location.projection),
+                );
+                return Ok(Value::Object(source_object));
+            };
+            for (key, sibling) in source_object {
+                let child_pointer = format!("{}/{}", location.projection, escape_pointer(&key));
+                let child_origin_pointer = format!("{}/{}", location.origin, escape_pointer(&key));
+                expanded_object.insert(
+                    key,
+                    self.expand_value(
+                        document_index,
+                        sibling,
+                        TraversalLocation {
+                            projection: &child_pointer,
+                            origin: &child_origin_pointer,
+                        },
+                        depth,
+                        stack,
+                    )?,
+                );
+            }
+        }
+        Ok(expanded)
+    }
+
+    fn target(
+        &self,
+        current: usize,
+        reference: &str,
+    ) -> Result<Option<(usize, String)>, TargetError> {
+        let (document, fragment) = reference.split_once('#').unwrap_or((reference, ""));
+        let fragment = percent_decode(fragment)?;
+        let pointer = if fragment.is_empty() {
+            String::new()
+        } else if fragment.starts_with('/') {
+            fragment
+        } else {
+            return Ok(None);
+        };
+        if document.is_empty() {
+            return Ok(Some((current, pointer)));
+        }
+        let canonical = if has_uri_scheme(document) {
+            canonical_absolute_uri(document)
+        } else {
+            let Some(base) = self.documents[current].uri.as_deref() else {
+                return Ok(None);
+            };
+            join_relative(base, document)
+        };
+        Ok(canonical.and_then(|canonical| {
+            self.by_uri
+                .get(&canonical)
+                .copied()
+                .map(|index| (index, pointer))
+        }))
+    }
+
+    fn merge_all_of(
+        &mut self,
+        mut object: Map<String, Value>,
+        pointer: &str,
+    ) -> Result<Value, ImportParseError> {
+        let branches = match object.remove("allOf") {
+            None => return Ok(Value::Object(object)),
+            Some(Value::Array(branches)) => branches,
+            Some(value) => {
+                object.insert("allOf".to_owned(), value);
+                self.push_finding("unsupported_composition", &format!("{pointer}/allOf"));
+                return Ok(Value::Object(object));
+            }
+        };
+        let original = branches.clone();
+        for branch in branches {
+            let Some(fields) = branch.as_object() else {
+                object.insert("allOf".to_owned(), Value::Array(original));
+                self.push_finding("all_of_conflict", &format!("{pointer}/allOf"));
+                return Ok(Value::Object(object));
+            };
+            if !merge_object(&mut object, fields) {
+                object.insert("allOf".to_owned(), Value::Array(original));
+                self.push_finding("all_of_conflict", &format!("{pointer}/allOf"));
+                return Ok(Value::Object(object));
+            }
+        }
+        Ok(Value::Object(object))
+    }
+
+    fn push_finding(&mut self, code: &str, pointer: &str) {
+        self.findings.push(NormalizedFinding {
+            code: code.to_owned(),
+            severity: ImportFindingSeverity::Error,
+            message: match code {
+                "external_reference_disabled" => "Внешняя ссылка отключена политикой импорта.",
+                "reference_target_missing" => {
+                    "Цель ссылки отсутствует или имеет неверный JSON Pointer."
+                }
+                "external_reference_unavailable" => {
+                    "Внешний snapshot недоступен или отклонён политикой импорта."
+                }
+                "reference_type_mismatch" => "Цель ссылки имеет неподдерживаемый тип.",
+                "reference_graph_limit" => "Граф ссылок превышает установленный предел.",
+                "all_of_conflict" => "Ветки allOf содержат несовместимые определения.",
+                "reference_uri_malformed" => "URI fragment ссылки содержит некорректное percent-кодирование.",
+                "unsupported_composition" => "Несколько операторов composition в одной schema не могут быть спроецированы без потерь.",
+                "unsupported_discriminator" => "Discriminator имеет неподдерживаемую или неполную структуру.",
+                _ => "Ссылка не может быть безопасно разрешена.",
+            }
+            .to_owned(),
+            construct_id: format!("source:{pointer}"),
+            location: SourceLocation {
+                pointer: pointer.to_owned(),
+            },
+            operation_key: None,
+        });
+    }
+}
+
+fn merge_object(target: &mut Map<String, Value>, source: &Map<String, Value>) -> bool {
+    for (key, value) in source {
+        match key.as_str() {
+            "properties" => {
+                let Some(source_properties) = value.as_object() else {
+                    return false;
+                };
+                let properties = target
+                    .entry(key.clone())
+                    .or_insert_with(|| Value::Object(Map::new()));
+                let Some(target_properties) = properties.as_object_mut() else {
+                    return false;
+                };
+                for (name, schema) in source_properties {
+                    if let Some(existing) = target_properties.get(name) {
+                        let (Some(existing), Some(schema)) =
+                            (existing.as_object(), schema.as_object())
+                        else {
+                            return false;
+                        };
+                        let mut merged = existing.clone();
+                        if !merge_object(&mut merged, schema) {
+                            return false;
+                        }
+                        target_properties.insert(name.clone(), Value::Object(merged));
+                    } else {
+                        target_properties.insert(name.clone(), schema.clone());
+                    }
+                }
+            }
+            "required" => {
+                let Some(source_required) = value.as_array() else {
+                    return false;
+                };
+                let required = target
+                    .entry(key.clone())
+                    .or_insert_with(|| Value::Array(Vec::new()));
+                let Some(target_required) = required.as_array_mut() else {
+                    return false;
+                };
+                target_required.extend(source_required.iter().cloned());
+                target_required.sort_by(|left, right| left.as_str().cmp(&right.as_str()));
+                target_required.dedup();
+            }
+            "minimum" | "maximum" => {
+                let Some(source_value) = value.as_f64() else {
+                    return false;
+                };
+                if target.contains_key(key) && target.get(key).and_then(Value::as_f64).is_none() {
+                    return false;
+                }
+                let merged = match (key.as_str(), target.get(key).and_then(Value::as_f64)) {
+                    ("minimum", Some(current)) => current.max(source_value),
+                    ("maximum", Some(current)) => current.min(source_value),
+                    _ => source_value,
+                };
+                let Some(number) = serde_json::Number::from_f64(merged) else {
+                    return false;
+                };
+                target.insert(key.clone(), Value::Number(number));
+                if constraint_contradiction(target, "minimum", "maximum") {
+                    return false;
+                }
+            }
+            "minLength" | "maxLength" => {
+                let Some(source_value) = value.as_u64() else {
+                    return false;
+                };
+                if target.contains_key(key) && target.get(key).and_then(Value::as_u64).is_none() {
+                    return false;
+                }
+                let merged = match (key.as_str(), target.get(key).and_then(Value::as_u64)) {
+                    ("minLength", Some(current)) => current.max(source_value),
+                    ("maxLength", Some(current)) => current.min(source_value),
+                    _ => source_value,
+                };
+                target.insert(key.clone(), Value::Number(merged.into()));
+                if integer_constraint_contradiction(target, "minLength", "maxLength") {
+                    return false;
+                }
+            }
+            _ => {
+                if target.get(key).is_some_and(|existing| existing != value) {
+                    return false;
+                }
+                target.insert(key.clone(), value.clone());
+            }
+        }
+    }
+    true
+}
+
+fn constraint_contradiction(target: &Map<String, Value>, minimum: &str, maximum: &str) -> bool {
+    match (
+        target.get(minimum).and_then(Value::as_f64),
+        target.get(maximum).and_then(Value::as_f64),
+    ) {
+        (Some(minimum), Some(maximum)) => minimum > maximum,
+        _ => false,
+    }
+}
+
+fn integer_constraint_contradiction(
+    target: &Map<String, Value>,
+    minimum: &str,
+    maximum: &str,
+) -> bool {
+    match (
+        target.get(minimum).and_then(Value::as_u64),
+        target.get(maximum).and_then(Value::as_u64),
+    ) {
+        (Some(minimum), Some(maximum)) => minimum > maximum,
+        _ => false,
+    }
+}
+
+fn edge_key(edge: &ResolvedReferenceEdge) -> (&str, &str, &str, &str, bool) {
+    (
+        edge.source.snapshot_digest.as_str(),
+        &edge.source.location.pointer,
+        edge.target.snapshot_digest.as_str(),
+        &edge.target.location.pointer,
+        edge.recursive,
+    )
+}
+
+fn decode_snapshot(document: &str) -> Result<Value, ImportParseError> {
+    if let Ok(value) = serde_json::from_str(document) {
+        return Ok(value);
+    }
+    let yaml: serde_yaml::Value =
+        serde_yaml::from_str(document).map_err(|_| ImportParseError::InvalidDocument)?;
+    serde_json::to_value(yaml).map_err(|_| ImportParseError::InvalidDocument)
+}
+
+fn is_oas31(root: &Value) -> bool {
+    root.get("openapi")
+        .and_then(Value::as_str)
+        .is_some_and(|version| version.starts_with("3.1."))
+}
+
+fn join_relative(base: &str, relative: &str) -> Option<String> {
+    let base = UriReference::parse(base)?;
+    let relative = UriReference::parse(relative)?;
+    let scheme = relative.scheme.or(base.scheme)?;
+    let authority = if relative.scheme.is_some() || relative.authority.is_some() {
+        relative.authority
+    } else {
+        base.authority
+    };
+    let (path, query) = if relative.scheme.is_some() || relative.authority.is_some() {
+        (remove_dot_segments(&relative.path), relative.query)
+    } else if relative.path.is_empty() {
+        (base.path, relative.query.or(base.query))
+    } else if relative.path.starts_with('/') {
+        (remove_dot_segments(&relative.path), relative.query)
+    } else {
+        (
+            remove_dot_segments(&merge_paths(
+                &base.path,
+                authority.is_some(),
+                &relative.path,
+            )),
+            relative.query,
+        )
+    };
+    UriReference {
+        scheme: Some(scheme),
+        authority,
+        path,
+        query,
+    }
+    .render()
+}
+
+fn canonical_absolute_uri(uri: &str) -> Option<String> {
+    let reference = UriReference::parse(uri)?;
+    let scheme = reference.scheme?;
+    UriReference {
+        scheme: Some(scheme),
+        authority: reference.authority,
+        path: remove_dot_segments(&reference.path),
+        query: reference.query,
+    }
+    .render()
+}
+
+#[derive(Clone, Debug)]
+struct UriReference<'a> {
+    scheme: Option<&'a str>,
+    authority: Option<&'a str>,
+    path: String,
+    query: Option<&'a str>,
+}
+
+impl<'a> UriReference<'a> {
+    fn parse(value: &'a str) -> Option<Self> {
+        if value.contains('\\') || value.contains('#') {
+            return None;
+        }
+        let (without_query, query) = value
+            .split_once('?')
+            .map_or((value, None), |(path, query)| (path, Some(query)));
+        let (scheme, rest) = if let Some(index) = without_query.find(':') {
+            let candidate = &without_query[..index];
+            if is_uri_scheme(candidate) {
+                (Some(candidate), &without_query[index + 1..])
+            } else {
+                (None, without_query)
+            }
+        } else {
+            (None, without_query)
+        };
+        let (authority, path) = if let Some(rest) = rest.strip_prefix("//") {
+            match rest.find('/') {
+                Some(index) => (Some(&rest[..index]), rest[index..].to_owned()),
+                None => (Some(rest), String::new()),
+            }
+        } else {
+            (None, rest.to_owned())
+        };
+        Some(Self {
+            scheme,
+            authority,
+            path,
+            query,
+        })
+    }
+
+    fn render(&self) -> Option<String> {
+        let scheme = self.scheme?;
+        let scheme = scheme.to_ascii_lowercase();
+        let mut value = format!("{scheme}:");
+        if let Some(authority) = self.authority {
+            value.push_str("//");
+            value.push_str(&canonical_authority(authority, &scheme));
+        }
+        if self.authority.is_some() && self.path.is_empty() {
+            value.push('/');
+        } else {
+            value.push_str(&self.path);
+        }
+        if let Some(query) = self.query {
+            value.push('?');
+            value.push_str(query);
+        }
+        Some(value)
+    }
+}
+
+fn canonical_authority(authority: &str, scheme: &str) -> String {
+    let authority = authority.to_ascii_lowercase();
+    let default_port = match scheme {
+        "http" => Some("80"),
+        "https" => Some("443"),
+        _ => None,
+    };
+    if let Some(default_port) = default_port
+        && let Some((host, port)) = authority.rsplit_once(':')
+        && port == default_port
+        && (!host.contains(':') || host.ends_with(']'))
+    {
+        return host.to_owned();
+    }
+    authority
+}
+
+fn has_uri_scheme(value: &str) -> bool {
+    value
+        .split_once(':')
+        .is_some_and(|(candidate, _)| is_uri_scheme(candidate))
+}
+
+fn is_uri_scheme(candidate: &str) -> bool {
+    let Some(first) = candidate.as_bytes().first() else {
+        return false;
+    };
+    first.is_ascii_alphabetic()
+        && candidate
+            .bytes()
+            .all(|byte| byte.is_ascii_alphanumeric() || matches!(byte, b'+' | b'-' | b'.'))
+}
+
+fn merge_paths(base_path: &str, has_authority: bool, relative_path: &str) -> String {
+    match base_path.rfind('/') {
+        Some(index) => format!("{}{}", &base_path[..=index], relative_path),
+        None if has_authority => format!("/{relative_path}"),
+        None => relative_path.to_owned(),
+    }
+}
+
+fn remove_dot_segments(path: &str) -> String {
+    let leading_slash = path.starts_with('/');
+    let trailing_slash = path.ends_with('/');
+    let mut output = Vec::new();
+    for segment in path.split('/') {
+        match segment {
+            "." => {}
+            ".." => {
+                output.pop();
+            }
+            _ => output.push(segment),
+        }
+    }
+    let mut result = output.join("/");
+    if leading_slash && !result.starts_with('/') {
+        result.insert(0, '/');
+    }
+    if trailing_slash && !result.ends_with('/') {
+        result.push('/');
+    }
+    result
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+enum TargetError {
+    MalformedFragment,
+}
+
+fn percent_decode(fragment: &str) -> Result<String, TargetError> {
+    let bytes = fragment.as_bytes();
+    let mut decoded = Vec::with_capacity(bytes.len());
+    let mut index = 0;
+    while index < bytes.len() {
+        if bytes[index] != b'%' {
+            decoded.push(bytes[index]);
+            index += 1;
+            continue;
+        }
+        let Some(high) = bytes.get(index + 1).and_then(|byte| hex_value(*byte)) else {
+            return Err(TargetError::MalformedFragment);
+        };
+        let Some(low) = bytes.get(index + 2).and_then(|byte| hex_value(*byte)) else {
+            return Err(TargetError::MalformedFragment);
+        };
+        decoded.push((high << 4) | low);
+        index += 3;
+    }
+    String::from_utf8(decoded).map_err(|_| TargetError::MalformedFragment)
+}
+
+fn hex_value(byte: u8) -> Option<u8> {
+    match byte {
+        b'0'..=b'9' => Some(byte - b'0'),
+        b'a'..=b'f' => Some(byte - b'a' + 10),
+        b'A'..=b'F' => Some(byte - b'A' + 10),
+        _ => None,
+    }
+}
+
+fn is_literal_payload_key(key: &str) -> bool {
+    matches!(key, "example" | "examples" | "default" | "enum" | "const") || key.starts_with("x-")
+}

crates/crank-import/src/rest/schema.rs

@@ -32,8 +32,33 @@ pub fn schema_from_openapi(
     let Some(value) = value else {
         return primitive(SchemaKind::String, required, description);
     };
-    // NormalizedIR keeps composition typed; the legacy preview adapter retains
+    // Only v3 NormalizedIR emits the private marker. Pending legacy-v1 jobs
-    // its historical first-branch projection for pending v1 jobs.
+    // retain their historical first-branch behavior, while modern oneOf/anyOf
+    // is projected losslessly through crank-schema's existing Oneof shape.
+    if value
+        .get("x-crank-lossless-composition")
+        .and_then(Value::as_bool)
+        == Some(true)
+        && let Some(items) = value
+            .get("oneOf")
+            .or_else(|| value.get("anyOf"))
+            .and_then(Value::as_array)
+    {
+        return Schema {
+            kind: SchemaKind::Oneof,
+            description: description.or_else(|| text(value, "description")),
+            required,
+            nullable: nullable(value),
+            default_value: value.get("default").cloned(),
+            fields: BTreeMap::new(),
+            items: None,
+            enum_values: Vec::new(),
+            variants: items
+                .iter()
+                .map(|item| schema_from_openapi(Some(item), true, None))
+                .collect(),
+        };
+    }
     let resolved = collapse_composition(value);
 
     if let Some(values) = resolved.get("enum").and_then(Value::as_array) {

crates/crank-import/src/rest/swagger2.rs

@@ -107,6 +107,7 @@ fn parse_document_v2(root: &Value) -> Result<RestImportDocument, ImportParseErro
                 parsed_parameters.findings,
             );
             operation_parameters.extend(parsed_parameters.parameters);
+            deduplicate_parameters(&mut operation_parameters);
             let tags = tags(
                 operation_value.get("tags"),
                 &format!("{path_pointer}/{method_name}/tags"),
@@ -183,6 +184,13 @@ fn parse_document_v2(root: &Value) -> Result<RestImportDocument, ImportParseErro
     })
 }
 
+fn deduplicate_parameters(parameters: &mut Vec<RestImportParameter>) {
+    let mut seen = std::collections::BTreeSet::new();
+    parameters.reverse();
+    parameters.retain(|parameter| seen.insert((parameter.name.clone(), parameter.location)));
+    parameters.reverse();
+}
+
 pub fn parse_document_legacy_v1(root: &Value) -> Result<RestImportDocument, ImportParseError> {
     let title = root
         .pointer("/info/title")

crates/crank-import/tests/normalization_details.rs

@@ -161,6 +161,58 @@ paths:
         assert!(ir.operations[0].request_body_schema.is_some());
     }
 
+    #[test]
+    fn operation_parameters_override_path_parameters_by_name_and_location() {
+        let openapi = r#"
+openapi: 3.1.0
+info: { title: OAS parameter overrides }
+servers: [{ url: https://example.test }]
+paths:
+  /items:
+    parameters:
+      - { name: page, in: query, description: path, schema: { type: integer } }
+    get:
+      operationId: listItems
+      parameters:
+        - { name: page, in: query, description: operation, schema: { type: string } }
+      responses: { '200': { description: ok } }
+"#;
+        let ir = normalize_document(openapi, &NormalizationConfig::default()).unwrap();
+        assert_eq!(ir.operations[0].parameters.len(), 1);
+        assert_eq!(
+            ir.operations[0].parameters[0].description.as_deref(),
+            Some("operation")
+        );
+        assert_eq!(
+            ir.operations[0].parameters[0].source_location.pointer,
+            "/paths/~1items/get/parameters/0"
+        );
+
+        let swagger = r#"
+swagger: '2.0'
+info: { title: Swagger parameter overrides }
+paths:
+  /items:
+    parameters:
+      - { name: page, in: query, description: path, type: integer }
+    get:
+      operationId: listItems
+      parameters:
+        - { name: page, in: query, description: operation, type: string }
+      responses: { '200': { description: ok } }
+"#;
+        let ir = normalize_document(swagger, &NormalizationConfig::default()).unwrap();
+        assert_eq!(ir.operations[0].parameters.len(), 1);
+        assert_eq!(
+            ir.operations[0].parameters[0].description.as_deref(),
+            Some("operation")
+        );
+        assert_eq!(
+            ir.operations[0].parameters[0].source_location.pointer,
+            "/paths/~1items/get/parameters/0"
+        );
+    }
+
     #[test]
     fn openapi_parameter_omissions_are_errors_at_the_dropped_item_pointer() {
         let document = r#"
@@ -220,10 +272,6 @@ components:
                     "/paths/~1items~1{id}/parameters/1",
                 ),
                 ("invalid_parameter", "/paths/~1items~1{id}/parameters/2",),
-                (
-                    "unresolved_parameter_reference",
-                    "/paths/~1items~1{id}/parameters/3",
-                ),
                 (
                     "missing_parameter_name",
                     "/paths/~1items~1{id}/get/parameters/0",
@@ -336,10 +384,6 @@ parameters:
                     "/paths/~1items~1{id}/parameters/1",
                 ),
                 ("invalid_parameter", "/paths/~1items~1{id}/parameters/2",),
-                (
-                    "unresolved_parameter_reference",
-                    "/paths/~1items~1{id}/parameters/3",
-                ),
                 (
                     "missing_parameter_name",
                     "/paths/~1items~1{id}/get/parameters/0",
@@ -688,7 +732,7 @@ paths:
     }
 
     #[test]
-    fn preserves_every_unresolved_reference_from_the_full_source_tree() {
+    fn resolves_local_reference_graph_and_preserves_external_blocker() {
         let document = r#"
 openapi: 3.1.0
 info: { title: References }
@@ -722,32 +766,24 @@ components:
             serde_json::to_vec(&first).unwrap(),
             serde_json::to_vec(&second).unwrap()
         );
-        assert_eq!(first.unresolved_references.len(), 10);
+        assert_eq!(first.unresolved_references.len(), 1);
         let references = first
             .unresolved_references
             .iter()
             .map(|reference| (reference.uri.as_str(), reference.location.pointer.as_str()))
             .collect::<std::collections::BTreeSet<_>>();
-        assert!(references.contains(&(
-            "#/components/parameters/Id",
-            "/paths/~1ok/parameters/0/$ref"
-        )));
-        assert!(references.contains(&(
-            "#/components/requestBodies/Body",
-            "/paths/~1ok/get/requestBody/$ref"
-        )));
-        assert!(references.contains(&(
-            "#/components/responses/Ok",
-            "/paths/~1ok/get/responses/200/$ref"
-        )));
-        assert!(references.contains(&("#/components/pathItems/Reusable", "/paths/~1reused/$ref")));
-        assert!(
-            references.contains(&("#/components/schemas/Loop", "/components/schemas/Loop/$ref"))
-        );
         assert!(references.contains(&(
             "https://example.test/schema.json#/Remote",
             "/components/schemas/Remote/$ref"
         )));
+        assert!(first.reference_graph.edges.len() >= 5);
+        assert!(
+            first
+                .reference_graph
+                .edges
+                .iter()
+                .any(|edge| edge.recursive)
+        );
         assert!(first.unresolved_references.iter().all(|reference| {
             reference.construct_id
                 == format!(

crates/crank-import/tests/preview.rs

@@ -94,7 +94,7 @@ definitions:
     }
 
     #[test]
-    fn previews_swagger2_and_preserves_unresolved_definitions() {
+    fn previews_swagger2_and_resolves_local_definitions() {
         let preview = preview_document(SWAGGER2).unwrap();
 
         assert_eq!(preview.source.format, "swagger");
@@ -105,7 +105,7 @@ definitions:
         let operation = &preview.groups[0].operations[0];
         assert_eq!(operation.suggested_name, "get_pet");
         assert_eq!(operation.input_fields, 1);
-        assert_eq!(operation.output_fields, 0);
+        assert_eq!(operation.output_fields, 2);
         assert_eq!(operation.draft.target.path_template, "/pets/{id}");
         assert_eq!(
             operation.draft.input_mapping.rules[0].target,

crates/crank-import/tests/reference_resolution.rs

@@ -0,0 +1,793 @@
+use crank_import::rest::{
+    ExternalDocumentSnapshot, ImportFindingSeverity, NormalizationConfig, NormalizedSchemaKind,
+    SourceDigest, external_reference_uris, normalize_verified_bundle, normalize_verified_document,
+    preview_from_ir, reference_uris,
+};
+
+fn digest(byte: char) -> SourceDigest {
+    SourceDigest::parse(byte.to_string().repeat(64)).unwrap()
+}
+
+fn normalize(document: &str) -> crank_import::rest::NormalizedIr {
+    normalize_verified_document(document, digest('a'), &NormalizationConfig::default()).unwrap()
+}
+
+#[test]
+fn resolves_local_schema_and_object_references_with_rfc6901_escaping() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Local refs }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items/{id}:
+    get:
+      operationId: getItem
+      parameters:
+        - { $ref: '#/components/parameters/Id' }
+      responses:
+        '200': { $ref: '#/components/responses/Ok' }
+components:
+  parameters:
+    Id: { name: id, in: path, required: true, schema: { type: string } }
+  responses:
+    Ok:
+      description: ok
+      content:
+        application/json:
+          schema: { $ref: '#/components/schemas/a~1b~0c' }
+  schemas:
+    a/b~c:
+      type: object
+      required: [id]
+      properties: { id: { type: string } }
+"#;
+    let ir = normalize(document);
+    let operation = &ir.operations[0];
+    assert_eq!(operation.parameters.len(), 1);
+    assert_eq!(operation.parameters[0].name, "id");
+    assert!(matches!(
+        operation.response_schema.as_ref().map(|schema| &schema.kind),
+        Some(NormalizedSchemaKind::Object { properties, .. }) if properties.contains_key("id")
+    ));
+    assert!(ir.unresolved_references.is_empty());
+    assert_eq!(ir.reference_graph.edges.len(), 3);
+    assert!(
+        ir.findings
+            .iter()
+            .chain(operation.findings.iter())
+            .all(|finding| finding.code != "unresolved_reference")
+    );
+}
+
+#[test]
+fn broken_reference_blocks_only_affected_candidate_and_full_preview_remains_visible() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Partial graph }
+servers: [{ url: https://api.example.test }]
+paths:
+  /broken:
+    get:
+      operationId: broken
+      responses:
+        '200':
+          description: nope
+          content: { application/json: { schema: { $ref: '#/components/schemas/Missing' } } }
+  /healthy:
+    get:
+      operationId: healthy
+      responses: { '204': { description: ok } }
+"#;
+    let ir = normalize(document);
+    let broken = ir
+        .operations
+        .iter()
+        .find(|operation| operation.path == "/broken")
+        .unwrap();
+    let healthy = ir
+        .operations
+        .iter()
+        .find(|operation| operation.path == "/healthy")
+        .unwrap();
+    assert!(broken.findings.iter().any(|finding| {
+        finding.code == "reference_target_missing"
+            && finding.severity == ImportFindingSeverity::Error
+    }));
+    assert!(healthy.findings.is_empty());
+    let preview = preview_from_ir(&ir);
+    assert_eq!(
+        preview
+            .groups
+            .iter()
+            .map(|group| group.operations.len())
+            .sum::<usize>(),
+        2
+    );
+}
+
+#[test]
+fn external_references_are_default_deny_without_network_or_snapshot() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: External deny }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content: { application/json: { schema: { $ref: 'https://schemas.example.test/root.yaml#/Item' } } }
+"#;
+    let ir = normalize(document);
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .any(|finding| finding.code == "external_reference_disabled")
+    );
+    assert!(ir.reference_graph.edges.is_empty());
+    assert!(ir.reference_graph.dependency_digests.is_empty());
+}
+
+#[test]
+fn resolves_supplied_external_snapshot_and_relative_chain_deterministically() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: External snapshots }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content: { application/json: { schema: { $ref: 'https://schemas.example.test/root.yaml#/Item' } } }
+"#;
+    let snapshots = vec![
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/root.yaml".to_owned(),
+            digest: digest('b'),
+            document: "Item: { $ref: 'child.yaml#/Child' }".to_owned(),
+        },
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/child.yaml".to_owned(),
+            digest: digest('c'),
+            document: "Child: { type: object, properties: { value: { type: integer } } }"
+                .to_owned(),
+        },
+    ];
+    let first = normalize_verified_bundle(
+        document,
+        digest('a'),
+        &snapshots,
+        &NormalizationConfig::default(),
+    )
+    .unwrap();
+    let second = normalize_verified_bundle(
+        document,
+        digest('a'),
+        &snapshots,
+        &NormalizationConfig::default(),
+    )
+    .unwrap();
+    assert_eq!(
+        serde_json::to_vec(&first).unwrap(),
+        serde_json::to_vec(&second).unwrap()
+    );
+    assert_eq!(
+        first.reference_graph.dependency_digests,
+        vec![digest('b'), digest('c')]
+    );
+    assert_eq!(first.reference_graph.edges.len(), 2);
+    assert_eq!(
+        first.reference_graph.edges[1].source.snapshot_digest,
+        digest('b')
+    );
+    assert!(first.unresolved_references.is_empty());
+}
+
+#[test]
+fn recursion_is_a_stable_graph_edge_without_unbounded_expansion() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Recursive }
+servers: [{ url: https://api.example.test }]
+paths:
+  /nodes:
+    get:
+      operationId: getNode
+      responses:
+        '200':
+          description: ok
+          content: { application/json: { schema: { $ref: '#/components/schemas/Node' } } }
+components:
+  schemas:
+    Node:
+      type: object
+      properties:
+        child: { $ref: '#/components/schemas/Node' }
+"#;
+    let ir = normalize(document);
+    assert!(ir.reference_graph.edges.iter().any(|edge| edge.recursive));
+    assert!(ir.unresolved_references.is_empty());
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .all(|finding| finding.code != "reference_graph_limit")
+    );
+}
+
+#[test]
+fn merges_compatible_all_of_and_blocks_conflicts_without_first_branch_loss() {
+    let compatible = r#"
+openapi: 3.1.0
+info: { title: AllOf }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    post:
+      operationId: createItem
+      requestBody:
+        content:
+          application/json:
+            schema:
+              allOf:
+                - { type: object, required: [id], properties: { id: { type: string } } }
+                - { type: object, required: [name], properties: { name: { type: string } } }
+      responses: { '204': { description: ok } }
+"#;
+    let ir = normalize(compatible);
+    assert!(matches!(
+        ir.operations[0].request_body_schema.as_ref().map(|schema| &schema.kind),
+        Some(NormalizedSchemaKind::Object { properties, required })
+            if properties.len() == 2 && required == &vec!["id".to_owned(), "name".to_owned()]
+    ));
+
+    let conflict = compatible.replace("{ name: { type: string } }", "{ id: { type: integer } }");
+    let ir = normalize(&conflict);
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .any(|finding| finding.code == "all_of_conflict")
+    );
+}
+
+#[test]
+fn preserves_one_of_discriminator_and_projects_all_alternatives() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Alternatives }
+servers: [{ url: https://api.example.test }]
+paths:
+  /events:
+    post:
+      operationId: createEvent
+      requestBody:
+        content:
+          application/json:
+            schema:
+              discriminator:
+                propertyName: kind
+                mapping: { text: '#/components/schemas/Text' }
+              oneOf:
+                - { type: object, properties: { text: { type: string } } }
+                - { type: object, properties: { count: { type: integer } } }
+      responses: { '204': { description: ok } }
+"#;
+    let ir = normalize(document);
+    let schema = ir.operations[0].request_body_schema.as_ref().unwrap();
+    assert_eq!(schema.discriminator.as_ref().unwrap().property_name, "kind");
+    assert!(matches!(
+        &schema.kind,
+        NormalizedSchemaKind::Composition { operator, variants }
+            if operator == "oneOf" && variants.len() == 2
+    ));
+    let preview = preview_from_ir(&ir);
+    let candidate = &preview.groups[0].operations[0];
+    assert_eq!(
+        candidate
+            .draft
+            .input_schema
+            .fields
+            .get("body")
+            .unwrap()
+            .variants
+            .len(),
+        2
+    );
+
+    let any_of = document
+        .replace("discriminator:\n                propertyName: kind\n                mapping: { text: '#/components/schemas/Text' }\n              oneOf:", "anyOf:");
+    let ir = normalize(&any_of);
+    assert!(matches!(
+        &ir.operations[0].request_body_schema.as_ref().unwrap().kind,
+        NormalizedSchemaKind::Composition { operator, variants }
+            if operator == "anyOf" && variants.len() == 2
+    ));
+    let preview = preview_from_ir(&ir);
+    assert_eq!(
+        preview.groups[0].operations[0].draft.input_schema.fields["body"]
+            .variants
+            .len(),
+        2
+    );
+}
+
+#[test]
+fn oas_31_applies_ref_siblings_while_oas_30_ignores_them() {
+    let template = |version: &str| {
+        format!(
+            r#"
+openapi: {version}
+info: {{ title: Siblings }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content:
+            application/json:
+              schema:
+                $ref: '#/components/schemas/Item'
+                description: sibling-description
+components:
+  schemas:
+    Item: {{ type: string, description: target-description }}
+"#
+        )
+    };
+    let v30 = normalize(&template("3.0.3"));
+    let v31 = normalize(&template("3.1.0"));
+    assert_eq!(
+        v30.operations[0]
+            .response_schema
+            .as_ref()
+            .unwrap()
+            .description
+            .as_deref(),
+        Some("target-description")
+    );
+    assert_eq!(
+        v31.operations[0]
+            .response_schema
+            .as_ref()
+            .unwrap()
+            .description
+            .as_deref(),
+        Some("sibling-description")
+    );
+}
+
+#[test]
+fn invalid_pointer_and_type_mismatch_are_exact_blockers_without_panic() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Invalid targets }
+servers: [{ url: https://api.example.test }]
+paths:
+  /missing:
+    get:
+      operationId: missing
+      responses:
+        '200': { description: ok, content: { application/json: { schema: { $ref: '#not-a-pointer' } } } }
+  /scalar:
+    get:
+      operationId: scalar
+      responses:
+        '200': { description: ok, content: { application/json: { schema: { $ref: '#/info/title' } } } }
+"#;
+    let ir = normalize(document);
+    let codes = ir
+        .operations
+        .iter()
+        .flat_map(|operation| {
+            operation
+                .findings
+                .iter()
+                .map(move |finding| (operation.path.as_str(), finding.code.as_str()))
+        })
+        .collect::<std::collections::BTreeSet<_>>();
+    assert!(codes.contains(&("/missing", "reference_target_missing")));
+    assert!(codes.contains(&("/scalar", "reference_type_mismatch")));
+}
+
+#[test]
+fn reference_depth_and_expanded_node_limits_fail_closed() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Bounded }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200': { description: ok, content: { application/json: { schema: { $ref: '#/components/schemas/A' } } } }
+components:
+  schemas:
+    A: { $ref: '#/components/schemas/B' }
+    B: { $ref: '#/components/schemas/C' }
+    C: { type: object, properties: { id: { type: string } } }
+"#;
+    let config = NormalizationConfig {
+        max_reference_depth: 1,
+        ..NormalizationConfig::default()
+    };
+    let ir = normalize_verified_document(document, digest('a'), &config).unwrap();
+    assert!(
+        ir.operations
+            .iter()
+            .flat_map(|operation| &operation.findings)
+            .any(|finding| finding.code == "reference_graph_limit")
+    );
+
+    let config = NormalizationConfig {
+        max_expanded_nodes: 8,
+        ..NormalizationConfig::default()
+    };
+    let ir = normalize_verified_document(document, digest('a'), &config).unwrap();
+    assert!(
+        ir.findings
+            .iter()
+            .chain(
+                ir.operations
+                    .iter()
+                    .flat_map(|operation| &operation.findings)
+            )
+            .any(|finding| finding.code == "reference_graph_limit")
+    );
+}
+
+#[test]
+fn invalid_discriminator_and_multiple_composition_operators_are_blockers() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Unsupported composition }
+servers: [{ url: https://api.example.test }]
+paths:
+  /events:
+    post:
+      operationId: createEvent
+      requestBody:
+        content:
+          application/json:
+            schema:
+              discriminator: { mapping: { bad: 42 } }
+              oneOf: [{ type: string }, { type: integer }]
+              anyOf: [{ type: boolean }, { type: string }]
+      responses: { '204': { description: ok } }
+"#;
+    let ir = normalize(document);
+    let codes = ir.operations[0]
+        .findings
+        .iter()
+        .map(|finding| finding.code.as_str())
+        .collect::<std::collections::BTreeSet<_>>();
+    assert!(codes.contains("unsupported_discriminator"));
+    assert!(codes.contains("unsupported_composition"));
+}
+
+#[test]
+fn reference_uri_scan_applies_external_size_depth_and_alias_limits_before_traversal() {
+    let config = NormalizationConfig {
+        max_bytes: 8,
+        max_external_document_bytes: 8,
+        ..NormalizationConfig::default()
+    };
+    assert_eq!(
+        reference_uris("external-document", &config),
+        Err(crank_import::rest::ImportParseError::LimitExceeded)
+    );
+
+    let config = NormalizationConfig {
+        max_bytes: 8,
+        max_external_document_bytes: 32,
+        ..NormalizationConfig::default()
+    };
+    assert_eq!(
+        external_reference_uris("external-document", &config),
+        Ok(Vec::new())
+    );
+
+    let config = NormalizationConfig {
+        max_depth: 1,
+        ..NormalizationConfig::default()
+    };
+    assert_eq!(
+        reference_uris("a: { b: { $ref: '#/x' } }", &config),
+        Err(crank_import::rest::ImportParseError::LimitExceeded)
+    );
+
+    let aliases = format!(
+        "items: [{}]",
+        std::iter::repeat_n("*a", 129)
+            .collect::<Vec<_>>()
+            .join(", ")
+    );
+    assert_eq!(
+        reference_uris(&aliases, &NormalizationConfig::default()),
+        Err(crank_import::rest::ImportParseError::LimitExceeded)
+    );
+}
+
+#[test]
+fn resolves_relative_references_using_rfc3986_paths_and_decoded_fragments() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Relative refs }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content: { application/json: { schema: { $ref: 'HTTPS://SCHEMAS.EXAMPLE.TEST:443/a/b/root.yaml#/Item' } } }
+"#;
+    let snapshots = vec![
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/a/b/root.yaml".to_owned(),
+            digest: digest('b'),
+            document: r#"
+Item:
+  type: object
+  properties:
+    dot: { $ref: './child.yaml#/Value' }
+    parent: { $ref: '../common.yaml#/Value' }
+    absolute: { $ref: '/shared.yaml#/Value' }
+"#
+            .to_owned(),
+        },
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/a/b/child.yaml".to_owned(),
+            digest: digest('c'),
+            document: "Value: { type: string }".to_owned(),
+        },
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/a/common.yaml".to_owned(),
+            digest: digest('d'),
+            document: "Value: { type: integer }".to_owned(),
+        },
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/shared.yaml".to_owned(),
+            digest: digest('e'),
+            document: "Value: { $ref: '#/a%7E1b' }\na/b: { type: boolean }".to_owned(),
+        },
+    ];
+    let ir = normalize_verified_bundle(
+        document,
+        digest('a'),
+        &snapshots,
+        &NormalizationConfig::default(),
+    )
+    .unwrap();
+    let Some(NormalizedSchemaKind::Object { properties, .. }) = ir.operations[0]
+        .response_schema
+        .as_ref()
+        .map(|schema| &schema.kind)
+    else {
+        panic!("response should be an expanded object schema");
+    };
+    assert_eq!(properties.len(), 3);
+    assert_eq!(ir.reference_graph.edges.len(), 5);
+    assert!(ir.findings.is_empty());
+}
+
+#[test]
+fn malformed_percent_encoded_fragment_is_a_blocker() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Malformed reference }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200': { description: ok, content: { application/json: { schema: { $ref: '#/components/%ZZ' } } } }
+"#;
+    let ir = normalize(document);
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .any(|finding| finding.code == "reference_uri_malformed")
+    );
+}
+
+#[test]
+fn does_not_follow_references_inside_literal_payloads() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Literal payloads }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200':
+          description: ok
+          content:
+            application/json:
+              schema:
+                type: object
+                example: { $ref: '#/components/schemas/Missing' }
+                examples: { sample: { value: { $ref: '#/components/schemas/Missing' } } }
+                default: { $ref: '#/components/schemas/Missing' }
+                enum: [{ $ref: '#/components/schemas/Missing' }]
+                const: { $ref: '#/components/schemas/Missing' }
+                x-fixture: { $ref: '#/components/schemas/Missing' }
+                properties: { known: { $ref: '#/components/schemas/Known' } }
+components:
+  schemas:
+    Known: { type: string }
+"#;
+    assert_eq!(
+        reference_uris(document, &NormalizationConfig::default()),
+        Ok(vec!["#/components/schemas/Known".to_owned()])
+    );
+    let ir = normalize(document);
+    assert_eq!(ir.reference_graph.edges.len(), 1);
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .all(|finding| finding.code != "reference_target_missing")
+    );
+}
+
+#[test]
+fn primary_oas31_applies_ref_siblings_in_external_snapshots_without_openapi_field() {
+    let source = |version: &str| {
+        format!(
+            r#"
+openapi: {version}
+info: {{ title: External siblings }}
+servers: [{{ url: https://api.example.test }}]
+paths:
+  /items:
+    get:
+      operationId: listItems
+      responses:
+        '200': {{ description: ok, content: {{ application/json: {{ schema: {{ $ref: 'https://schemas.example.test/root.yaml#/Item' }} }} }} }}
+"#
+        )
+    };
+    let snapshots = vec![
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/root.yaml".to_owned(),
+            digest: digest('b'),
+            document: "Item: { $ref: 'child.yaml#/Base', description: sibling }".to_owned(),
+        },
+        ExternalDocumentSnapshot {
+            canonical_uri: "https://schemas.example.test/child.yaml".to_owned(),
+            digest: digest('c'),
+            document: "Base: { type: string, description: target }".to_owned(),
+        },
+    ];
+    let v31 = normalize_verified_bundle(
+        &source("3.1.0"),
+        digest('a'),
+        &snapshots,
+        &NormalizationConfig::default(),
+    )
+    .unwrap();
+    let v30 = normalize_verified_bundle(
+        &source("3.0.3"),
+        digest('a'),
+        &snapshots,
+        &NormalizationConfig::default(),
+    )
+    .unwrap();
+    assert_eq!(
+        v31.operations[0]
+            .response_schema
+            .as_ref()
+            .unwrap()
+            .description
+            .as_deref(),
+        Some("sibling")
+    );
+    assert_eq!(
+        v30.operations[0]
+            .response_schema
+            .as_ref()
+            .unwrap()
+            .description
+            .as_deref(),
+        Some("target")
+    );
+}
+
+#[test]
+fn all_of_intersects_constraints_and_nested_properties_without_losing_conflicts() {
+    let source = r#"
+openapi: 3.1.0
+info: { title: allOf intersections }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    post:
+      operationId: createItem
+      requestBody:
+        content:
+          application/json:
+            schema:
+              allOf:
+                - type: object
+                  minimum: 1
+                  maximum: 10
+                  minLength: 2
+                  maxLength: 12
+                  properties: { nested: { type: object, properties: { left: { type: string } } } }
+                - type: object
+                  minimum: 4
+                  maximum: 8
+                  minLength: 5
+                  maxLength: 9
+                  properties: { nested: { type: object, properties: { right: { type: integer } } } }
+      responses: { '204': { description: ok } }
+"#;
+    let ir = normalize(source);
+    let schema = ir.operations[0].request_body_schema.as_ref().unwrap();
+    assert_eq!(schema.constraints.minimum, Some(4.0));
+    assert_eq!(schema.constraints.maximum, Some(8.0));
+    assert_eq!(schema.constraints.min_length, Some(5));
+    assert_eq!(schema.constraints.max_length, Some(9));
+    let NormalizedSchemaKind::Object { properties, .. } = &schema.kind else {
+        panic!("merged schema should remain an object");
+    };
+    let NormalizedSchemaKind::Object { properties, .. } = &properties["nested"].kind else {
+        panic!("nested property should remain an object");
+    };
+    assert!(properties.contains_key("left") && properties.contains_key("right"));
+
+    let conflicting = source.replace("maximum: 8", "maximum: 3");
+    let ir = normalize(&conflicting);
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .any(|finding| finding.code == "all_of_conflict")
+    );
+}
+
+#[test]
+fn non_array_all_of_is_preserved_and_reported() {
+    let document = r#"
+openapi: 3.1.0
+info: { title: Invalid allOf }
+servers: [{ url: https://api.example.test }]
+paths:
+  /items:
+    post:
+      operationId: createItem
+      requestBody:
+        content: { application/json: { schema: { allOf: { type: string } } } }
+      responses: { '204': { description: ok } }
+"#;
+    let ir = normalize(document);
+    assert!(
+        ir.operations[0]
+            .findings
+            .iter()
+            .any(|finding| finding.code == "unsupported_composition")
+    );
+    assert!(matches!(
+        ir.operations[0]
+            .request_body_schema
+            .as_ref()
+            .map(|schema| &schema.kind),
+        Some(NormalizedSchemaKind::Unknown)
+    ));
+}

crates/crank-import/tests/unit.rs

@@ -122,7 +122,7 @@ paths:
     }
 
     #[test]
-    fn keeps_references_typed_without_resolving_them() {
+    fn resolves_local_references_into_typed_graph() {
         let ir = normalize_document(SWAGGER2, &NormalizationConfig::default()).unwrap();
         let operation = &ir.operations[0];
         assert!(matches!(
@@ -130,13 +130,15 @@ paths:
                 .response_schema
                 .as_ref()
                 .map(|schema| &schema.kind),
-            Some(crank_import::rest::NormalizedSchemaKind::Reference { .. })
+            Some(crank_import::rest::NormalizedSchemaKind::Object { .. })
         ));
+        assert!(ir.unresolved_references.is_empty());
+        assert!(!ir.reference_graph.edges.is_empty());
         assert!(
             operation
                 .findings
                 .iter()
-                .any(|finding| finding.code == "unresolved_reference")
+                .all(|finding| finding.code != "unresolved_reference")
         );
     }
 

crates/crank-registry/src/migrations.rs

@@ -11,6 +11,7 @@ mod onboarding_product_events_v11;
 mod owned_relations;
 mod platform_key_name_reuse_v6;
 mod schema_guard;
+mod schema_guard_legacy_v1;
 mod schema_guard_v10;
 mod schema_guard_v11;
 mod schema_guard_v12;

crates/crank-registry/src/migrations/authority.rs

@@ -11,6 +11,9 @@ use super::platform_key_name_reuse_v6;
 use super::schema_guard::{
     OWNED_RELATIONS, relation_exists, validate_required_relations, validate_schema_fingerprint,
 };
+use super::schema_guard_legacy_v1::{
+    validate_ledgerless_baseline_fingerprint, validate_ledgerless_optional_fingerprints,
+};
 use super::{BASELINE_CHECKSUM, BASELINE_VERSION, apply_baseline};
 use crate::ext::ExtensionMigration;
 use sqlx::{PgConnection, PgPool, Row, Transaction, query};
@@ -389,6 +392,19 @@ impl MigrationAuthority {
         if from < 13 {
             artifact_cleanup_indexes_v13::apply(&mut transaction, &Self::sequence()[12]).await?;
         }
+        match inspect(&mut transaction).await? {
+            MigrationPreflight::Current {
+                version: CURRENT_VERSION,
+            } => {}
+            _ => {
+                return Err(MigrationError::new(
+                    "apply_failed",
+                    "apply.postflight",
+                    Some(CURRENT_VERSION),
+                    "restore_known_good_backup",
+                ));
+            }
+        }
         transaction
             .commit()
             .await
@@ -408,9 +424,15 @@ async fn inspect(connection: &mut PgConnection) -> Result<MigrationPreflight, Mi
             owned_exists |= relation_exists(connection, relation).await?;
         }
         if owned_exists {
+            if is_ledgerless_legacy_baseline(connection, canonical_exists).await? {
+                return Ok(MigrationPreflight::MigrationRequired {
+                    current: 0,
+                    target: CURRENT_VERSION,
+                });
+            }
             return Err(MigrationError::new(
                 "partial_sequence",
-                "preflight.core",
+                "preflight.core_missing",
                 None,
                 "restore_known_good_backup",
             ));
@@ -541,6 +563,47 @@ async fn inspect(connection: &mut PgConnection) -> Result<MigrationPreflight, Mi
         Ok(MigrationPreflight::Current { version: current })
     }
 }
+async fn is_ledgerless_legacy_baseline(
+    connection: &mut PgConnection,
+    canonical_exists: bool,
+) -> Result<bool, MigrationError> {
+    if canonical_exists {
+        return Ok(false);
+    }
+    for relation in [
+        "__crank_migration_legacy_audit",
+        "master_key_identities",
+        "master_key_rotations",
+        "admin_bootstrap_contracts",
+        "admin_login_backoff",
+        "admin_security_audit_events",
+        "product_events",
+        "product_event_daily_rollups",
+        "onboarding_selections",
+        "artifact_blobs",
+        "artifact_sources",
+    ] {
+        if relation_exists(connection, relation).await? {
+            return Ok(false);
+        }
+    }
+    for relation in owned_relations::BASELINE {
+        if !relation_exists(connection, relation).await? {
+            return Ok(false);
+        }
+    }
+    validate_required_relations(connection, owned_relations::BASELINE, 1).await?;
+    validate_ledgerless_baseline_fingerprint(connection).await?;
+    let mcp_ledger = relation_exists(connection, "__crank_mcp_migrations").await?;
+    if mcp_ledger {
+        return Ok(false);
+    }
+    let mcp_sessions = relation_exists(connection, "mcp_transport_sessions").await?;
+    let extension_ledger = relation_exists(connection, "__crank_ext_migrations").await?;
+    validate_ledgerless_optional_fingerprints(connection, mcp_sessions, extension_ledger).await?;
+    inspect_optional_legacy_for_ledgerless_baseline(connection).await?;
+    Ok(true)
+}
 async fn validate_core_ledger(connection: &mut PgConnection) -> Result<(), MigrationError> {
     let rows = query("select version, description, checksum from __crank_core_migrations order by version limit 2")
         .fetch_all(connection)
@@ -549,7 +612,7 @@ async fn validate_core_ledger(connection: &mut PgConnection) -> Result<(), Migra
     if rows.len() != 1 {
         return Err(MigrationError::new(
             "partial_sequence",
-            "preflight.core",
+            "preflight.core_cardinality",
             None,
             "restore_known_good_backup",
         ));
@@ -577,9 +640,21 @@ async fn validate_core_ledger(connection: &mut PgConnection) -> Result<(), Migra
     Ok(())
 }
 async fn inspect_optional_legacy(connection: &mut PgConnection) -> Result<(), MigrationError> {
+    inspect_optional_legacy_with_policy(connection, false).await
+}
+async fn inspect_optional_legacy_for_ledgerless_baseline(
+    connection: &mut PgConnection,
+) -> Result<(), MigrationError> {
+    inspect_optional_legacy_with_policy(connection, true).await
+}
+async fn inspect_optional_legacy_with_policy(
+    connection: &mut PgConnection,
+    allow_sessions_without_ledger: bool,
+) -> Result<(), MigrationError> {
     let mcp_ledger = relation_exists(connection, "__crank_mcp_migrations").await?;
     let mcp_sessions = relation_exists(connection, "mcp_transport_sessions").await?;
-    if mcp_ledger != mcp_sessions {
+    if mcp_ledger != mcp_sessions && !(allow_sessions_without_ledger && !mcp_ledger && mcp_sessions)
+    {
         return Err(MigrationError::new(
             "legacy_conflict",
             "preflight.legacy_mcp",

crates/crank-registry/src/migrations/authority_tests.rs

@@ -8,7 +8,7 @@ fn sequence_is_deterministic_and_append_only() {
     MigrationAuthority::validate_sequence().unwrap();
     assert_eq!(
         first.iter().map(|item| item.version).collect::<Vec<_>>(),
-        vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
+        vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
     );
     assert_eq!(first[0].checksum, "crank-community-baseline-v1");
     assert_eq!(first[0].source_digest, BASELINE_SOURCE_SHA256);

crates/crank-registry/src/migrations/schema_guard.rs

@@ -325,6 +325,15 @@ pub(super) async fn validate_schema_fingerprint(
             .iter()
             .filter_map(|row| row.try_get::<String, _>("column_name").ok())
             .collect::<Vec<_>>();
+        let required = required
+            .iter()
+            .copied()
+            .filter(|column| {
+                !(current_version == 1
+                    && *table == "__crank_ext_migrations"
+                    && *column == "checksum")
+            })
+            .collect::<Vec<_>>();
         if actual.len() != required.len()
             || required
                 .iter()
@@ -337,6 +346,9 @@ pub(super) async fn validate_schema_fingerprint(
         if !relation_exists(connection, table).await? {
             continue;
         }
+        if current_version == 1 && *table == "__crank_ext_migrations" && *column == "checksum" {
+            continue;
+        }
         let row = query(
             "select data_type, is_nullable from information_schema.columns
              where table_schema = current_schema() and table_name = $1 and column_name = $2",
@@ -424,23 +436,25 @@ pub(super) async fn validate_schema_fingerprint(
             return Err(schema_error(current_version));
         }
     }
-    let required_indexes = [
+    if relation_exists(connection, "mcp_transport_sessions").await? {
-        "mcp_transport_sessions_workspace_agent_idx",
+        let required_indexes = [
-        "mcp_transport_sessions_expires_at_idx",
+            "mcp_transport_sessions_workspace_agent_idx",
-    ];
+            "mcp_transport_sessions_expires_at_idx",
-    for index in required_indexes {
+        ];
-        let present = query(
+        for index in required_indexes {
-            "select exists (select 1 from pg_catalog.pg_indexes
+            let present = query(
-             where schemaname = current_schema() and indexname = $1) as present",
+                "select exists (select 1 from pg_catalog.pg_indexes
-        )
+                 where schemaname = current_schema() and indexname = $1) as present",
-        .bind(index)
+            )
-        .fetch_one(&mut *connection)
+            .bind(index)
-        .await
+            .fetch_one(&mut *connection)
-        .map_err(|_| MigrationError::storage("preflight.schema"))?
+            .await
-        .try_get::<bool, _>("present")
+            .map_err(|_| MigrationError::storage("preflight.schema"))?
-        .map_err(|_| MigrationError::storage("preflight.schema"))?;
+            .try_get::<bool, _>("present")
-        if !present {
+            .map_err(|_| MigrationError::storage("preflight.schema"))?;
-            return Err(schema_error(current_version));
+            if !present {
+                return Err(schema_error(current_version));
+            }
         }
     }
     if current_version >= 3 {

crates/crank-registry/src/migrations/schema_guard_legacy_v1.rs

@@ -0,0 +1,196 @@
+use sha2::{Digest, Sha256};
+use sqlx::PgConnection;
+
+use super::authority::MigrationError;
+
+// Exact PostgreSQL 16 catalog contract produced by the last published
+// pre-ledger Community schema (commit 8318e4b).
+const LEDGERLESS_BASELINE_FINGERPRINT_SHA256: &str =
+    "36624ca42a28c1388f9c50e6d6c489a8e5ca1f73441a894d9af89e5e115b542c";
+const LEDGERLESS_MCP_FINGERPRINTS_SHA256: &[&str] = &[
+    // Initial published session table.
+    "9240c3d85dbc9eeddb1cd99661ec8f7d8d6ece0e62ec486151dcc951a7f3c81c",
+    // Published session table after supports_elicitation was added.
+    "ac9f99a7e667552a07480d5d45379dd0a8b529b00ba98855ce6725535198249f",
+];
+const LEDGERLESS_EXTENSION_FINGERPRINT_SHA256: &str =
+    "0809a80c0bb6e80f68d0557006c4b62f2e63556f61f6fdedc5abc324950e2587";
+
+const BASELINE_RELATIONS: &[&str] = &[
+    "workspaces",
+    "users",
+    "memberships",
+    "user_sessions",
+    "invitation_tokens",
+    "platform_api_keys",
+    "operations",
+    "operation_versions",
+    "published_operations",
+    "operation_samples",
+    "descriptors",
+    "agents",
+    "agent_versions",
+    "published_agents",
+    "agent_operation_bindings",
+    "secrets",
+    "secret_versions",
+    "auth_profiles",
+    "workspace_upstreams",
+    "yaml_import_jobs",
+    "import_jobs",
+    "approval_requests",
+    "invocation_logs",
+    "usage_rollups",
+];
+
+pub(super) async fn validate_ledgerless_baseline_fingerprint(
+    connection: &mut PgConnection,
+) -> Result<(), MigrationError> {
+    let actual = catalog_fingerprint(connection, BASELINE_RELATIONS).await?;
+    if actual == LEDGERLESS_BASELINE_FINGERPRINT_SHA256 {
+        Ok(())
+    } else {
+        Err(fingerprint_error())
+    }
+}
+
+pub(super) async fn validate_ledgerless_optional_fingerprints(
+    connection: &mut PgConnection,
+    has_mcp_sessions: bool,
+    has_extension_ledger: bool,
+) -> Result<(), MigrationError> {
+    if has_mcp_sessions {
+        let actual = catalog_fingerprint(connection, &["mcp_transport_sessions"]).await?;
+        if !LEDGERLESS_MCP_FINGERPRINTS_SHA256.contains(&actual.as_str()) {
+            return Err(fingerprint_error());
+        }
+    }
+    if has_extension_ledger {
+        let actual = catalog_fingerprint(connection, &["__crank_ext_migrations"]).await?;
+        if actual != LEDGERLESS_EXTENSION_FINGERPRINT_SHA256 {
+            return Err(fingerprint_error());
+        }
+    }
+    Ok(())
+}
+
+async fn catalog_fingerprint(
+    connection: &mut PgConnection,
+    relations: &[&str],
+) -> Result<String, MigrationError> {
+    let relations = relations
+        .iter()
+        .map(|value| (*value).to_owned())
+        .collect::<Vec<_>>();
+    let unsafe_catalog_state = sqlx::query_scalar::<_, bool>(
+        "with selected(table_name) as (select unnest($1::text[]))
+         select
+             exists (
+                 select 1 from pg_catalog.pg_class c
+                 join pg_catalog.pg_namespace n on n.oid = c.relnamespace
+                 join selected s on s.table_name = c.relname
+                 where n.nspname = current_schema()
+                   and (c.relpersistence <> 'p' or c.relrowsecurity
+                        or c.relforcerowsecurity or c.relreplident <> 'd')
+             )
+             or exists (
+                 select 1 from pg_catalog.pg_index i
+                 join pg_catalog.pg_class t on t.oid = i.indrelid
+                 join pg_catalog.pg_namespace n on n.oid = t.relnamespace
+                 join selected s on s.table_name = t.relname
+                 where n.nspname = current_schema()
+                   and (not i.indisvalid or not i.indisready or not i.indislive)
+             )
+             or exists (
+                 select 1 from pg_catalog.pg_policy p
+                 join pg_catalog.pg_class t on t.oid = p.polrelid
+                 join pg_catalog.pg_namespace n on n.oid = t.relnamespace
+                 join selected s on s.table_name = t.relname
+                 where n.nspname = current_schema()
+             )
+             or exists (
+                 select 1 from pg_catalog.pg_trigger tg
+                 join pg_catalog.pg_class t on t.oid = tg.tgrelid
+                 join pg_catalog.pg_namespace n on n.oid = t.relnamespace
+                 join selected s on s.table_name = t.relname
+                 where n.nspname = current_schema() and tg.tgenabled <> 'O'
+             )",
+    )
+    .bind(relations.clone())
+    .fetch_one(&mut *connection)
+    .await
+    .map_err(|_| MigrationError::storage("preflight.legacy_fingerprint"))?;
+    if unsafe_catalog_state {
+        return Err(fingerprint_error());
+    }
+    let fingerprint = sqlx::query_scalar::<_, String>(
+        "with baseline(table_name) as (select unnest($1::text[])), relation_rows as (
+            select jsonb_build_array('relation', c.relname, c.relkind::text) item
+            from pg_catalog.pg_class c
+            join pg_catalog.pg_namespace n on n.oid = c.relnamespace
+            join baseline b on b.table_name = c.relname
+            where n.nspname = current_schema()
+         ), column_rows as (
+            select jsonb_build_array(
+                'column', c.table_name, c.column_name,
+                c.data_type, c.udt_name, c.is_nullable, coalesce(c.column_default, '')
+            ) item
+            from information_schema.columns c
+            join baseline b using (table_name)
+            where c.table_schema = current_schema()
+         ), constraint_rows as (
+            select jsonb_build_array(
+                'constraint', t.relname, c.conname, c.contype::text,
+                c.convalidated, pg_get_constraintdef(c.oid, true)
+            ) item
+            from pg_catalog.pg_constraint c
+            join pg_catalog.pg_class t on t.oid = c.conrelid
+            join pg_catalog.pg_namespace n on n.oid = t.relnamespace
+            join baseline b on b.table_name = t.relname
+            where n.nspname = current_schema()
+         ), index_rows as (
+            select jsonb_build_array(
+                'index', t.relname, idx.relname,
+                replace(pg_get_indexdef(i.indexrelid), format('%I.', current_schema()), '')
+            ) item
+            from pg_catalog.pg_index i
+            join pg_catalog.pg_class idx on idx.oid = i.indexrelid
+            join pg_catalog.pg_class t on t.oid = i.indrelid
+            join pg_catalog.pg_namespace n on n.oid = t.relnamespace
+            join baseline b on b.table_name = t.relname
+            where n.nspname = current_schema()
+         ), trigger_rows as (
+            select jsonb_build_array(
+                'trigger', t.relname, tg.tgname,
+                replace(pg_get_triggerdef(tg.oid, true), format('%I.', current_schema()), '')
+            ) item
+            from pg_catalog.pg_trigger tg
+            join pg_catalog.pg_class t on t.oid = tg.tgrelid
+            join pg_catalog.pg_namespace n on n.oid = t.relnamespace
+            join baseline b on b.table_name = t.relname
+            where n.nspname = current_schema() and not tg.tgisinternal
+         ), all_rows as (
+            select item from relation_rows
+            union all select item from column_rows
+            union all select item from constraint_rows
+            union all select item from index_rows
+            union all select item from trigger_rows
+         )
+         select coalesce(jsonb_agg(item order by item::text), '[]'::jsonb)::text
+         from all_rows",
+    )
+    .bind(relations)
+    .fetch_one(connection)
+    .await
+    .map_err(|_| MigrationError::storage("preflight.legacy_fingerprint"))?;
+    Ok(format!("{:x}", Sha256::digest(fingerprint.as_bytes())))
+}
+
+fn fingerprint_error() -> MigrationError {
+    MigrationError::new(
+        "partial_sequence",
+        "preflight.legacy_fingerprint",
+        Some(1),
+        "restore_known_good_backup",
+    )
+}

crates/crank-registry/src/postgres/import_job.rs

@@ -4,6 +4,7 @@ use crank_artifacts::ArtifactRef;
 
 const APPLICATION_RESULT_KEY: &str = "_crank_application_result";
 const IMPORT_JOB_CLEANUP_BATCH: u32 = 128;
+const MAX_IMPORT_JOB_DOCUMENTS: usize = 32;
 const DANGLING_OPENAPI_SOURCE_GRACE: time::Duration = time::Duration::minutes(5);
 
 impl PostgresRegistry {
@@ -12,6 +13,7 @@ impl PostgresRegistry {
         request: CreateImportJobRequest<'_>,
     ) -> Result<(), RegistryError> {
         validate_source_envelope(request.source, request.preview_payload)?;
+        validate_dependency_envelopes(request.preview_payload)?;
         let mut transaction = self.pool.begin().await?;
         sqlx::query(
             "insert into import_jobs (
@@ -81,6 +83,26 @@ impl PostgresRegistry {
         &self,
         request: FinishImportJobRequest<'_>,
     ) -> Result<(), RegistryError> {
+        let mut transaction = self.pool.begin().await?;
+        let row = sqlx::query(
+            "select workspace_id, status, preview_payload
+               from import_jobs
+              where id = $1
+              for update",
+        )
+        .bind(request.id.as_str())
+        .fetch_optional(&mut *transaction)
+        .await?
+        .ok_or_else(|| RegistryError::ImportJobNotFound {
+            job_id: request.id.as_str().to_owned(),
+        })?;
+        let current_status =
+            deserialize_enum_text::<ImportJobStatus>(row.try_get("status")?, "status")?;
+        if request.status == ImportJobStatus::Failed && current_status == ImportJobStatus::Completed
+        {
+            transaction.commit().await?;
+            return Ok(());
+        }
         let result = sqlx::query(
             "update import_jobs
                 set status = $2,
@@ -94,7 +116,7 @@ impl PostgresRegistry {
         .bind(request.created_operation_ids)
         .bind(request.error_text)
         .bind(request.finished_at)
-        .execute(&self.pool)
+        .execute(&mut *transaction)
         .await?;
 
         if result.rows_affected() == 0 {
@@ -103,6 +125,25 @@ impl PostgresRegistry {
             });
         }
 
+        if request.status == ImportJobStatus::Failed {
+            let workspace_id = WorkspaceId::new(row.try_get::<String, _>("workspace_id")?);
+            let payload = row.try_get::<Value, _>("preview_payload")?;
+            // Failure finalization must itself be fail-safe. The strict
+            // envelope parsers are used before Apply; cleanup only needs the
+            // bounded source identities and must not roll back the terminal
+            // state because some other payload field was corrupted.
+            for source_id in cleanup_source_ids(&payload) {
+                let _ = detach_source_in_transaction(
+                    &mut transaction,
+                    &workspace_id,
+                    &source_id,
+                    *request.finished_at,
+                )
+                .await?;
+            }
+        }
+        transaction.commit().await?;
+
         Ok(())
     }
 
@@ -120,24 +161,18 @@ impl PostgresRegistry {
             }
             Err(error) => {
                 tx.rollback().await?;
-                let error_text = error.to_string();
+                if !matches!(error, RegistryError::ImportJobAlreadyApplied { .. }) {
-                let _ = sqlx::query(
+                    let empty = serde_json::json!([]);
-                    "update import_jobs
+                    let _ = self
-                     set status = $3,
+                        .finish_import_job(FinishImportJobRequest {
-                         error_text = $4,
+                            id: request.id,
-                         finished_at = $5::timestamptz
+                            status: ImportJobStatus::Failed,
-                     where id = $1
+                            created_operation_ids: &empty,
-                       and workspace_id = $2
+                            error_text: Some("import_apply_failed"),
-                       and status <> $6",
+                            finished_at: request.finished_at,
-                )
+                        })
-                .bind(request.id.as_str())
+                        .await;
-                .bind(request.workspace_id.as_str())
+                }
-                .bind(serialize_enum_text(&ImportJobStatus::Failed, "status")?)
-                .bind(error_text)
-                .bind(request.finished_at)
-                .bind(serialize_enum_text(&ImportJobStatus::Completed, "status")?)
-                .execute(&self.pool)
-                .await;
                 Err(error)
             }
         }
@@ -176,17 +211,13 @@ impl PostgresRegistry {
             let payload = row.try_get::<Value, _>("preview_payload")?;
             // The job itself is expired regardless of whether a legacy or
             // corrupt payload can be decoded. Do not let one bad row roll
-            // back cleanup for every tenant.
+            // back cleanup for every workspace.
-            if let Ok(Some(source)) = source_from_payload(&payload)
+            for source_id in cleanup_source_ids(&payload) {
-                && detach_source_in_transaction(
+                if detach_source_in_transaction(&mut transaction, &workspace_id, &source_id, now)
-                    &mut transaction,
+                    .await?
-                    &workspace_id,
+                {
-                    &source.source_id,
+                    detached_sources += 1;
-                    now,
+                }
-                )
-                .await?
-            {
-                detached_sources += 1;
             }
         }
         let deleted = if expired_ids.is_empty() {
@@ -214,7 +245,14 @@ impl PostgresRegistry {
                     select 1
                       from import_jobs j
                      where j.workspace_id = s.workspace_id
-                       and j.preview_payload -> 'source' ->> 'source_id' = s.source_id
+                       and (
+                         j.preview_payload -> 'source' ->> 'source_id' = s.source_id
+                         or exists (
+                           select 1
+                             from jsonb_array_elements(coalesce(j.preview_payload -> 'dependencies', '[]'::jsonb)) d
+                            where d ->> 'source_id' = s.source_id
+                         )
+                       )
                 )
               order by s.created_at, s.workspace_id, s.source_id
               limit $2
@@ -283,6 +321,7 @@ async fn apply_import_job_transaction(
         source_from_payload(&preview_payload)?.ok_or(RegistryError::InvalidArtifactSource {
             field: "import_job_source",
         })?;
+    let dependencies = dependencies_from_payload(&preview_payload)?;
 
     if status == ImportJobStatus::Completed
         && let Some(result) = stored_application_result(&preview_payload)?
@@ -323,6 +362,31 @@ async fn apply_import_job_transaction(
     if !source_is_current {
         return Err(RegistryError::SourceUnavailable);
     }
+    for dependency in &dependencies {
+        let recorded = sqlx::query(
+            "select s.lifecycle, b.artifact_ref
+               from artifact_sources s
+               join artifact_blobs b on b.digest = s.blob_digest
+              where s.workspace_id = $1 and s.source_id = $2
+              for update",
+        )
+        .bind(request.workspace_id.as_str())
+        .bind(dependency.source_id.as_str())
+        .fetch_optional(&mut **tx)
+        .await?;
+        let current = recorded
+            .as_ref()
+            .map(|row| {
+                let lifecycle = row.try_get::<String, _>("lifecycle").ok();
+                let artifact_ref = row.try_get::<String, _>("artifact_ref").ok();
+                lifecycle.as_deref() == Some("active")
+                    && artifact_ref.as_deref() == Some(dependency.digest.as_str())
+            })
+            .unwrap_or(false);
+        if !current {
+            return Err(RegistryError::SourceUnavailable);
+        }
+    }
 
     let mut result = ImportJobApplyResult {
         application_key: request.application_key.to_owned(),
@@ -408,6 +472,15 @@ async fn apply_import_job_transaction(
         *request.finished_at,
     )
     .await?;
+    for dependency in dependencies {
+        let _ = detach_source_in_transaction(
+            tx,
+            request.workspace_id,
+            &dependency.source_id,
+            *request.finished_at,
+        )
+        .await?;
+    }
 
     Ok(result)
 }
@@ -433,6 +506,11 @@ fn validate_source_envelope(
     Ok(())
 }
 
+fn validate_dependency_envelopes(preview_payload: &Value) -> Result<(), RegistryError> {
+    dependencies_from_payload(preview_payload)?;
+    Ok(())
+}
+
 fn source_from_payload(payload: &Value) -> Result<Option<ImportJobSourceEnvelope>, RegistryError> {
     let Some(source) = payload.get("source") else {
         return Ok(None);
@@ -466,6 +544,73 @@ fn source_from_payload(payload: &Value) -> Result<Option<ImportJobSourceEnvelope
     }))
 }
 
+fn dependencies_from_payload(
+    payload: &Value,
+) -> Result<Vec<ImportJobSourceEnvelope>, RegistryError> {
+    let Some(value) = payload.get("dependencies") else {
+        return Ok(Vec::new());
+    };
+    let items = value
+        .as_array()
+        .filter(|items| items.len() <= MAX_IMPORT_JOB_DOCUMENTS)
+        .ok_or(RegistryError::InvalidArtifactSource {
+            field: "import_job_dependencies",
+        })?;
+    let mut dependencies = Vec::with_capacity(items.len());
+    let mut identities = std::collections::BTreeSet::new();
+    for item in items {
+        let source_id = item
+            .get("source_id")
+            .and_then(Value::as_str)
+            .filter(|value| value.len() <= 132)
+            .ok_or(RegistryError::InvalidArtifactSource {
+                field: "import_job_dependencies",
+            })?;
+        let digest = item.get("digest").and_then(Value::as_str).ok_or(
+            RegistryError::InvalidArtifactSource {
+                field: "import_job_dependencies",
+            },
+        )?;
+        let digest =
+            ArtifactRef::parse(digest).map_err(|_| RegistryError::InvalidArtifactSource {
+                field: "import_job_dependencies",
+            })?;
+        if !identities.insert((source_id.to_owned(), digest.as_str().to_owned())) {
+            return Err(RegistryError::InvalidArtifactSource {
+                field: "import_job_dependencies",
+            });
+        }
+        dependencies.push(ImportJobSourceEnvelope {
+            source_id: ArtifactSourceId::new(source_id),
+            digest,
+        });
+    }
+    Ok(dependencies)
+}
+
+fn cleanup_source_ids(payload: &Value) -> Vec<ArtifactSourceId> {
+    let mut identities = std::collections::BTreeSet::new();
+    if let Some(source_id) = payload
+        .pointer("/source/source_id")
+        .and_then(Value::as_str)
+        .filter(|value| value.len() <= 132)
+    {
+        identities.insert(source_id.to_owned());
+    }
+    if let Some(dependencies) = payload.get("dependencies").and_then(Value::as_array) {
+        for source_id in dependencies
+            .iter()
+            .take(MAX_IMPORT_JOB_DOCUMENTS)
+            .filter_map(|item| item.get("source_id"))
+            .filter_map(Value::as_str)
+            .filter(|value| value.len() <= 132)
+        {
+            identities.insert(source_id.to_owned());
+        }
+    }
+    identities.into_iter().map(ArtifactSourceId::new).collect()
+}
+
 async fn detach_source_in_transaction(
     transaction: &mut Transaction<'_, Postgres>,
     workspace_id: &WorkspaceId,

crates/crank-registry/tests/integration/migrations.rs

@@ -2,6 +2,7 @@ use crank_registry::{MigrationAuthority, MigrationPreflight, PostgresRegistry};
 use sqlx::Row;
 
 mod artifact_metadata;
+mod legacy_adoption;
 mod rollback;
 
 static EVENT_TRIGGER_TEST_LOCK: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());
@@ -153,7 +154,7 @@ async fn changed_checksum_fails_closed_without_repair() {
     );
 }
 #[tokio::test]
-async fn legacy_core_baseline_is_consolidated_without_data_loss() {
+async fn ledgerless_legacy_baseline_is_consolidated_without_data_loss() {
     let database_url = crank_test_support::postgres_schema_url("test_legacy_core").await;
     let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
     MigrationAuthority::apply(&pool).await.unwrap();
@@ -221,7 +222,8 @@ async fn legacy_core_baseline_is_consolidated_without_data_loss() {
     }
     sqlx::query(
         "drop table __crank_migrations, __crank_migration_legacy_audit,
-         __crank_mcp_migrations, mcp_transport_sessions, __crank_ext_migrations",
+         __crank_mcp_migrations, mcp_transport_sessions, __crank_ext_migrations,
+         __crank_core_migrations",
     )
     .execute(&pool)
     .await
@@ -229,7 +231,7 @@ async fn legacy_core_baseline_is_consolidated_without_data_loss() {
     assert_eq!(
         MigrationAuthority::preflight(&pool).await.unwrap(),
         MigrationPreflight::MigrationRequired {
-            current: 1,
+            current: 0,
             target: 13,
         }
     );
@@ -937,6 +939,20 @@ async fn any_owned_relation_without_core_ledger_is_partial() {
         .unwrap();
     let error = MigrationAuthority::preflight(&pool).await.unwrap_err();
     assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.core_missing");
+}
+#[tokio::test]
+async fn empty_core_ledger_is_reported_separately() {
+    let database_url = crank_test_support::postgres_schema_url("test_empty_core_ledger").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    MigrationAuthority::apply(&pool).await.unwrap();
+    sqlx::query("delete from __crank_core_migrations")
+        .execute(&pool)
+        .await
+        .unwrap();
+    let error = MigrationAuthority::preflight(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.core_cardinality");
 }
 #[tokio::test]
 async fn current_ledger_with_structural_drift_fails_closed() {

crates/crank-registry/tests/integration/migrations/legacy_adoption.rs

@@ -0,0 +1,439 @@
+use super::*;
+
+async fn ledgerless_v1(pool: &sqlx::PgPool) {
+    MigrationAuthority::apply(pool).await.unwrap();
+    remove_v4_schema(pool).await;
+    remove_v3_schema(pool).await;
+    sqlx::raw_sql(
+        "drop table __crank_migrations, __crank_migration_legacy_audit;
+         drop table __crank_mcp_migrations;
+         drop index mcp_transport_sessions_expires_at_idx;
+         drop table __crank_ext_migrations;
+         create table __crank_ext_migrations (
+             extension_name text not null,
+             version integer not null,
+             applied_at timestamptz not null default now(),
+             primary key (extension_name, version)
+         );
+         drop table __crank_core_migrations;",
+    )
+    .execute(pool)
+    .await
+    .unwrap();
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_optional_index_drift_is_rejected_without_writes() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_index").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::raw_sql(
+        "drop index mcp_transport_sessions_workspace_agent_idx;
+         create index mcp_transport_sessions_workspace_agent_idx
+             on mcp_transport_sessions(id);",
+    )
+    .execute(&pool)
+    .await
+    .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_disabled_integrity_triggers_is_rejected_without_writes() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_triggers").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query("alter table memberships disable trigger all")
+        .execute(&pool)
+        .await
+        .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}
+
+#[tokio::test]
+async fn published_ledgerless_baseline_upgrades_without_data_loss() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_baseline").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query(
+        "insert into operations
+            (id, workspace_id, name, display_name, protocol, status, created_at, updated_at)
+         values ('op_ledgerless', 'ws_default', 'ledgerless', 'Ledgerless', 'rest', 'draft', now(), now())",
+    )
+    .execute(&pool)
+    .await
+    .unwrap();
+
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::MigrationRequired {
+            current: 0,
+            target: 13,
+        }
+    );
+    MigrationAuthority::apply(&pool).await.unwrap();
+
+    let operation_count: i64 =
+        sqlx::query_scalar("select count(*) from operations where id = 'op_ledgerless'")
+            .fetch_one(&pool)
+            .await
+            .unwrap();
+    assert_eq!(operation_count, 1);
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::Current { version: 13 }
+    );
+}
+
+#[tokio::test]
+async fn published_ledgerless_historical_column_layout_upgrades_without_data_loss() {
+    let database_url =
+        crank_test_support::postgres_schema_url("test_ledgerless_historical_layout").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::raw_sql(
+        "alter table platform_api_keys
+             drop column key_kind,
+             drop column expires_at,
+             drop column allowed_origins_json;
+         alter table platform_api_keys
+             add column key_kind text not null default 'mcp_client',
+             add column expires_at timestamptz null,
+             add column allowed_origins_json jsonb not null default '[]'::jsonb;
+
+         drop index approval_requests_pending_fingerprint_idx;
+         alter table approval_requests
+             drop column execution_started_at,
+             drop column execution_attempts,
+             drop column request_fingerprint,
+             add column confirmation_title text not null default '',
+             add column confirmation_body text not null default '';
+         alter table approval_requests
+             alter column confirmation_title drop default,
+             alter column confirmation_body drop default,
+             drop column confirmation_title,
+             drop column confirmation_body,
+             add column execution_started_at timestamptz null,
+             add column execution_attempts integer not null default 0,
+             add column request_fingerprint text null;
+         create unique index approval_requests_pending_fingerprint_idx
+             on approval_requests(agent_id, operation_id, operation_version, request_fingerprint)
+             where status = 'pending' and request_fingerprint is not null;
+
+         insert into operations
+            (id, workspace_id, name, display_name, protocol, status, created_at, updated_at)
+         values ('op_historical_layout', 'ws_default', 'historical-layout', 'Historical layout', 'rest', 'draft', now(), now());
+         insert into agents
+            (id, workspace_id, slug, display_name, description, status, created_at, updated_at)
+         values ('agent_historical_layout', 'ws_default', 'historical-layout', 'Historical layout', '', 'draft', now(), now());
+         insert into platform_api_keys
+            (id, workspace_id, agent_id, name, prefix, secret_hash, key_kind, scopes_json, status,
+             created_at, expires_at, allowed_origins_json)
+         values ('key_historical_layout', 'ws_default', 'agent_historical_layout', 'Historical layout',
+                 'cp_', 'hash', 'admin', '[]'::jsonb, 'active', now(),
+                 '2030-01-02 03:04:05+00'::timestamptz, '[\"https://example.test\"]'::jsonb);
+         insert into approval_requests
+            (id, workspace_id, agent_id, operation_id, operation_version, status, risk_level,
+             request_payload_json, created_at, expires_at, execution_started_at,
+             execution_attempts, request_fingerprint)
+         values ('approval_historical_layout', 'ws_default', 'agent_historical_layout', 'op_historical_layout', 1,
+                 'pending', 'high', '{\"layout\":\"historical\"}'::jsonb, now(), now() + interval '1 hour',
+                 '2029-02-03 04:05:06+00'::timestamptz, 3, 'historical-fingerprint');",
+    )
+    .execute(&pool)
+    .await
+    .unwrap();
+
+    let historical_ordinals: Vec<(String, i32)> = sqlx::query_as(
+        "select table_name || '.' || column_name, ordinal_position
+         from information_schema.columns
+         where table_schema = current_schema()
+           and ((table_name = 'platform_api_keys'
+                 and column_name in ('key_kind', 'expires_at', 'allowed_origins_json'))
+             or (table_name = 'approval_requests'
+                 and column_name in ('execution_started_at', 'execution_attempts', 'request_fingerprint')))
+         order by table_name, ordinal_position",
+    )
+    .fetch_all(&pool)
+    .await
+    .unwrap();
+    assert_eq!(
+        historical_ordinals,
+        vec![
+            ("approval_requests.execution_started_at".to_owned(), 22),
+            ("approval_requests.execution_attempts".to_owned(), 23),
+            ("approval_requests.request_fingerprint".to_owned(), 24),
+            ("platform_api_keys.key_kind".to_owned(), 15),
+            ("platform_api_keys.expires_at".to_owned(), 16),
+            ("platform_api_keys.allowed_origins_json".to_owned(), 17),
+        ],
+        "fixture must reproduce the published in-place column layout",
+    );
+
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::MigrationRequired {
+            current: 0,
+            target: 13,
+        }
+    );
+    MigrationAuthority::apply(&pool).await.unwrap();
+
+    let key_count: i64 = sqlx::query_scalar(
+        "select count(*) from platform_api_keys
+         where id = 'key_historical_layout'
+           and key_kind = 'admin'
+           and expires_at = '2030-01-02 03:04:05+00'::timestamptz
+           and allowed_origins_json = '[\"https://example.test\"]'::jsonb",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert_eq!(
+        key_count, 1,
+        "platform key row must survive the layout upgrade"
+    );
+    let approval_count: i64 = sqlx::query_scalar(
+        "select count(*) from approval_requests
+         where id = 'approval_historical_layout'
+           and request_payload_json = '{\"layout\":\"historical\"}'::jsonb
+           and execution_started_at = '2029-02-03 04:05:06+00'::timestamptz
+           and execution_attempts = 3
+           and request_fingerprint = 'historical-fingerprint'",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert_eq!(
+        approval_count, 1,
+        "approval row must survive the layout upgrade"
+    );
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::Current { version: 13 }
+    );
+}
+
+#[tokio::test]
+async fn published_initial_mcp_layout_is_accepted() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_initial_mcp").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::raw_sql(
+        "drop table mcp_transport_sessions;
+         create table mcp_transport_sessions (
+             id text primary key,
+             protocol_version text not null,
+             initialized boolean not null default false,
+             workspace_slug text not null,
+             agent_slug text not null,
+             created_at timestamptz not null,
+             updated_at timestamptz not null,
+             expires_at timestamptz null
+         );
+         create index mcp_transport_sessions_workspace_agent_idx
+             on mcp_transport_sessions(workspace_slug, agent_slug, updated_at desc);",
+    )
+    .execute(&pool)
+    .await
+    .unwrap();
+
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::MigrationRequired {
+            current: 0,
+            target: 13,
+        }
+    );
+    MigrationAuthority::apply(&pool).await.unwrap();
+    let supports_elicitation: bool = sqlx::query_scalar(
+        "select exists (
+             select 1 from information_schema.columns
+             where table_schema = current_schema()
+               and table_name = 'mcp_transport_sessions'
+               and column_name = 'supports_elicitation'
+         )",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(supports_elicitation);
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::Current { version: 13 }
+    );
+}
+
+#[tokio::test]
+async fn published_in_place_mcp_upgrade_layout_is_accepted() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_mcp_upgrade").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::raw_sql(
+        "drop table mcp_transport_sessions;
+         create table mcp_transport_sessions (
+             id text primary key,
+             protocol_version text not null,
+             initialized boolean not null default false,
+             workspace_slug text not null,
+             agent_slug text not null,
+             created_at timestamptz not null,
+             updated_at timestamptz not null,
+             expires_at timestamptz null
+         );
+         create index mcp_transport_sessions_workspace_agent_idx
+             on mcp_transport_sessions(workspace_slug, agent_slug, updated_at desc);
+         alter table mcp_transport_sessions
+             add column supports_elicitation boolean not null default false;",
+    )
+    .execute(&pool)
+    .await
+    .unwrap();
+
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::MigrationRequired {
+            current: 0,
+            target: 13,
+        }
+    );
+    MigrationAuthority::apply(&pool).await.unwrap();
+    assert_eq!(
+        MigrationAuthority::preflight(&pool).await.unwrap(),
+        MigrationPreflight::Current { version: 13 }
+    );
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_future_drift_is_rejected_without_writes() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_drift").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query("alter table invocation_logs add column trace_id text")
+        .execute(&pool)
+        .await
+        .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_default_drift_is_rejected_without_writes() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_default").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query("alter table workspaces alter column status set default 'active'")
+        .execute(&pool)
+        .await
+        .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_constraint_drift_is_rejected_without_writes() {
+    let database_url =
+        crank_test_support::postgres_schema_url("test_ledgerless_constraint_definition").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query(
+        "alter table workspaces
+         add constraint workspaces_status_nonempty check (status <> '')",
+    )
+    .execute(&pool)
+    .await
+    .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_rls_drift_is_rejected_without_writes() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_rls").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query("alter table workspaces enable row level security")
+        .execute(&pool)
+        .await
+        .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}
+
+#[tokio::test]
+async fn ledgerless_baseline_with_nullability_drift_is_rejected_without_writes() {
+    let database_url = crank_test_support::postgres_schema_url("test_ledgerless_constraint").await;
+    let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
+    ledgerless_v1(&pool).await;
+    sqlx::query("alter table workspaces alter column status drop not null")
+        .execute(&pool)
+        .await
+        .unwrap();
+
+    let error = MigrationAuthority::apply(&pool).await.unwrap_err();
+    assert_eq!(error.code(), "partial_sequence");
+    assert_eq!(error.stage(), "preflight.legacy_fingerprint");
+    let core_exists: bool = sqlx::query_scalar(
+        "select to_regclass(format('%I.%I', current_schema(), '__crank_core_migrations')) is not null",
+    )
+    .fetch_one(&pool)
+    .await
+    .unwrap();
+    assert!(!core_exists, "rejected adoption must remain read-only");
+}

crates/crank-runtime/src/lib.rs

@@ -25,7 +25,9 @@ pub use cache::{
 pub use cache_factory::{
     BuiltinCacheBackendFactory, CacheBackendFactory, SharedCacheBackendFactory,
 };
-pub use crank_adapter_rest::OutboundHttpPolicy;
+pub use crank_adapter_rest::{
+    ExternalReferenceFetchError, ExternalReferenceFetcher, OutboundHttpPolicy,
+};
 pub use error::RuntimeError;
 pub use execution_failure::{normalize_runtime_error, normalize_runtime_error_for_operation};
 pub use execution_request::{

deploy/community/.env.example

@@ -25,6 +25,12 @@ CRANK_OUTBOUND_ALLOWED_HOSTS=
 CRANK_OUTBOUND_DENIED_HOSTS=
 CRANK_OUTBOUND_MAX_REQUEST_BYTES=4194304
 CRANK_OUTBOUND_MAX_RESPONSE_BYTES=4194304
+CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES=
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH=8
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS=32
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES=262144
+CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS=10000
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES=10000
 CRANK_ENVIRONMENT=production
 CRANK_LOG_LEVEL=
 CRANK_SENTRY_DSN=

deploy/community/.env.images.example

@@ -27,6 +27,12 @@ CRANK_OUTBOUND_ALLOWED_HOSTS=
 CRANK_OUTBOUND_DENIED_HOSTS=
 CRANK_OUTBOUND_MAX_REQUEST_BYTES=4194304
 CRANK_OUTBOUND_MAX_RESPONSE_BYTES=4194304
+CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES=
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH=8
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS=32
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES=262144
+CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS=10000
+CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES=10000
 CRANK_ENVIRONMENT=production
 CRANK_LOG_LEVEL=
 CRANK_SENTRY_DSN=

deploy/community/docker-compose.images.yml

@@ -127,12 +127,18 @@ services:
       CRANK_OUTBOUND_DENIED_HOSTS: ${CRANK_OUTBOUND_DENIED_HOSTS:-}
       CRANK_OUTBOUND_MAX_REQUEST_BYTES: ${CRANK_OUTBOUND_MAX_REQUEST_BYTES:-4194304}
       CRANK_OUTBOUND_MAX_RESPONSE_BYTES: ${CRANK_OUTBOUND_MAX_RESPONSE_BYTES:-4194304}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES:-}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH:-8}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS:-32}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES:-262144}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS: ${CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS:-10000}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES:-10000}
     volumes:
       - artifact_storage:${CRANK_STORAGE_ROOT:-/var/lib/crank/storage}
     ports:
       - "${CRANK_PUBLISH_BIND:-127.0.0.1}:${CRANK_ADMIN_PUBLISH_PORT:-3001}:3001"
     healthcheck:
-      test: ["CMD", "curl", "--fail", "http://127.0.0.1:3001/ready"]
+      test: ["CMD", "/usr/local/bin/crank-http-healthcheck", "3001", "/ready"]
       interval: 15s
       timeout: 5s
       retries: 5
@@ -193,7 +199,7 @@ services:
     ports:
       - "${CRANK_PUBLISH_BIND:-127.0.0.1}:${CRANK_MCP_PUBLISH_PORT:-3002}:3002"
     healthcheck:
-      test: ["CMD", "curl", "--fail", "http://127.0.0.1:3002/ready"]
+      test: ["CMD", "/usr/local/bin/crank-http-healthcheck", "3002", "/ready"]
       interval: 15s
       timeout: 5s
       retries: 5

deploy/community/docker-compose.yml

@@ -42,6 +42,7 @@ services:
     depends_on:
       postgres:
         condition: service_healthy
+        required: false
     environment:
       POSTGRES_HOST: ${POSTGRES_HOST:-postgres}
       POSTGRES_PORT: ${POSTGRES_PORT:-5432}
@@ -130,12 +131,18 @@ services:
       CRANK_OUTBOUND_DENIED_HOSTS: ${CRANK_OUTBOUND_DENIED_HOSTS:-}
       CRANK_OUTBOUND_MAX_REQUEST_BYTES: ${CRANK_OUTBOUND_MAX_REQUEST_BYTES:-4194304}
       CRANK_OUTBOUND_MAX_RESPONSE_BYTES: ${CRANK_OUTBOUND_MAX_RESPONSE_BYTES:-4194304}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES:-}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH:-8}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS:-32}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES:-262144}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS: ${CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS:-10000}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES:-10000}
     volumes:
       - artifact_storage:${CRANK_STORAGE_ROOT:-/var/lib/crank/storage}
     ports:
       - "${CRANK_PUBLISH_BIND:-127.0.0.1}:${CRANK_ADMIN_PUBLISH_PORT:-3001}:3001"
     healthcheck:
-      test: ["CMD", "curl", "--fail", "http://127.0.0.1:3001/ready"]
+      test: ["CMD", "/usr/local/bin/crank-http-healthcheck", "3001", "/ready"]
       interval: 15s
       timeout: 5s
       retries: 5
@@ -196,7 +203,7 @@ services:
     ports:
       - "${CRANK_PUBLISH_BIND:-127.0.0.1}:${CRANK_MCP_PUBLISH_PORT:-3002}:3002"
     healthcheck:
-      test: ["CMD", "curl", "--fail", "http://127.0.0.1:3002/ready"]
+      test: ["CMD", "/usr/local/bin/crank-http-healthcheck", "3002", "/ready"]
       interval: 15s
       timeout: 5s
       retries: 5

docker-compose.yml

@@ -109,6 +109,12 @@ services:
       CRANK_OUTBOUND_DENIED_HOSTS: ${CRANK_OUTBOUND_DENIED_HOSTS:-}
       CRANK_OUTBOUND_MAX_REQUEST_BYTES: ${CRANK_OUTBOUND_MAX_REQUEST_BYTES:-4194304}
       CRANK_OUTBOUND_MAX_RESPONSE_BYTES: ${CRANK_OUTBOUND_MAX_RESPONSE_BYTES:-4194304}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES:-}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH:-8}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS:-32}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES:-262144}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS: ${CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS:-10000}
+      CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES: ${CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES:-10000}
     depends_on:
       migrate:
         condition: service_completed_successfully
@@ -121,7 +127,7 @@ services:
     ports:
       - "${CRANK_PUBLISH_BIND:-127.0.0.1}:3001:3001"
     healthcheck:
-      test: ["CMD", "curl", "--fail", "http://127.0.0.1:3001/ready"]
+      test: ["CMD", "/usr/local/bin/crank-http-healthcheck", "3001", "/ready"]
       interval: 15s
       timeout: 5s
       retries: 5
@@ -184,7 +190,7 @@ services:
     ports:
       - "${CRANK_PUBLISH_BIND:-127.0.0.1}:3002:3002"
     healthcheck:
-      test: ["CMD", "curl", "--fail", "http://127.0.0.1:3002/health"]
+      test: ["CMD", "/usr/local/bin/crank-http-healthcheck", "3002", "/health"]
       interval: 15s
       timeout: 5s
       retries: 5

docs/migrations.md

@@ -16,7 +16,7 @@ docker compose -f deploy/community/docker-compose.yml --env-file deploy/communit
 2. Создайте и проверьте согласованный backup PostgreSQL и artifact storage. Для первой пустой установки зафиксируйте, что восстанавливать нечего.
 3. Проверьте immutable plan: `cargo run -p admin-api --bin crank-migrate -- plan --check` в source checkout либо `<compose> run --rm migrate crank-migrate plan` для образа.
 4. Примените sequence: `<compose> run --rm migrate crank-migrate apply`.
-5. Повторите preflight и убедитесь в `{"status":"current","version":12}`.
+5. Повторите preflight и убедитесь в `{"status":"current","version":13}`.
 6. Только теперь запускайте long-running services: `<compose> up -d`.
 
 Обычный `up` также содержит обязательный migration job, но при upgrade он не заменяет предварительные preflight и backup. Migrator делает до десяти bounded попыток подключения с секундной паузой и затем безопасно завершается ошибкой.
@@ -90,6 +90,9 @@ cargo run -p admin-api --bin crank-migrate -- plan --check
   sensitivity и source lifecycle принадлежат scoped relation. Claim token/expiry
   зарезервированы bounded all-or-none contract для последующего reconciliation,
   но эта версия не запускает cleanup и не удаляет physical blobs.
+- V13 — expand-only indexes, поддерживающие bounded artifact/import cleanup и
+  reconciliation ранее зарезервированных claim token/expiry; эта версия сама
+  не выполняет cleanup, не переписывает данные и не удаляет physical blobs.
 - Каждая версия имеет contiguous `i64` version, стабильное имя, lowercase SHA-256, owner, phase, explicit readable schema min/max и backfill policy.
 - `migrate` требует bounded cursor/batch policy; `contract` дополнительно требует tracked compatibility evidence и закрытого окна.
 - Добавление descriptor без executable implementation блокируется `invalid_contract` до DB I/O.

docs/runtime-config.md

@@ -27,6 +27,12 @@ Crank настраивается через переменные окружен
 | `CRANK_OUTBOUND_DENIED_HOSTS` | `outbound.denied_hosts` | `Shared` | `host_list/-` | `` | `-` | `Internal` | `Effective` |
 | `CRANK_OUTBOUND_MAX_REQUEST_BYTES` | `outbound.max_request_bytes` | `Shared` | `u64/bytes` | `4194304` | `1..=67108864` | `Public` | `Effective` |
 | `CRANK_OUTBOUND_MAX_RESPONSE_BYTES` | `outbound.max_response_bytes` | `Shared` | `u64/bytes` | `4194304` | `1..=67108864` | `Public` | `Effective` |
+| `CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES` | `import.external_references.allowed_url_prefixes` | `AdminApi` | `url_prefix_list/-` | `` | `-` | `Internal` | `Effective` |
+| `CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH` | `import.external_references.max_depth` | `AdminApi` | `u32/edges` | `8` | `1..=32` | `Public` | `Effective` |
+| `CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS` | `import.external_references.max_documents` | `AdminApi` | `u32/documents` | `32` | `1..=32` | `Public` | `Effective` |
+| `CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES` | `import.external_references.max_fetch_bytes` | `AdminApi` | `u64/bytes` | `262144` | `1..=4194304` | `Public` | `Effective` |
+| `CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS` | `import.external_references.fetch_timeout_ms` | `AdminApi` | `u64/milliseconds` | `10000` | `1..=300000` | `Public` | `Effective` |
+| `CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES` | `import.external_references.max_expanded_nodes` | `AdminApi` | `u32/nodes` | `10000` | `1..=100000` | `Public` | `Effective` |
 | `CRANK_ENVIRONMENT` | `observability.environment` | `Shared` | `label/-` | `development` | `-` | `Public` | `Effective` |
 | `CRANK_LOG_LEVEL` | `observability.log_filter` | `Shared` | `string/-` | `blank` | `-` | `Public` | `Effective` |
 | `CRANK_SENTRY_DSN` | `observability.sentry_dsn` | `Shared` | `url/-` | `blank` | `-` | `Secret` | `Effective` |

docs/schemas/migration-sequence.json

@@ -192,6 +192,22 @@
       "source_digest": "052058da53cd861b2a1af81243b34cc35204d665a9276f8ece563e061f8ebbfe",
       "transactional": true,
       "version": 12
+    },
+    {
+      "backfill": {
+        "kind": "none"
+      },
+      "checksum": "abe822c967e1b2cc3966ede988e05eb2d9b2a062b539c45aaa8cad677395ab05",
+      "compatibility": "n-minus-one-readable",
+      "contract_evidence": null,
+      "name": "artifact-cleanup-indexes-v13",
+      "owner": "crank-registry",
+      "phase": "expand",
+      "readable_schema_max": 13,
+      "readable_schema_min": 12,
+      "source_digest": "abe822c967e1b2cc3966ede988e05eb2d9b2a062b539c45aaa8cad677395ab05",
+      "transactional": true,
+      "version": 13
     }
   ]
 }

docs/schemas/runtime-config.schema.json

@@ -330,6 +330,99 @@
       "compatibility": null,
       "rules": []
     },
+    {
+      "semantic_path": "import.external_references.allowed_url_prefixes",
+      "env_name": "CRANK_IMPORT_EXTERNAL_REFERENCE_ALLOWED_URL_PREFIXES",
+      "process": "admin_api",
+      "value_type": "url_prefix_list",
+      "unit": null,
+      "default": "",
+      "required": false,
+      "minimum": null,
+      "maximum": null,
+      "sensitivity": "internal",
+      "mode": "effective",
+      "compatibility": null,
+      "rules": [
+        "empty list disables external OpenAPI reference fetching",
+        "each prefix must be canonical HTTP(S) without userinfo, query, or fragment"
+      ]
+    },
+    {
+      "semantic_path": "import.external_references.max_depth",
+      "env_name": "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DEPTH",
+      "process": "admin_api",
+      "value_type": "u32",
+      "unit": "edges",
+      "default": "8",
+      "required": false,
+      "minimum": 1,
+      "maximum": 32,
+      "sensitivity": "public",
+      "mode": "effective",
+      "compatibility": null,
+      "rules": []
+    },
+    {
+      "semantic_path": "import.external_references.max_documents",
+      "env_name": "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_DOCUMENTS",
+      "process": "admin_api",
+      "value_type": "u32",
+      "unit": "documents",
+      "default": "32",
+      "required": false,
+      "minimum": 1,
+      "maximum": 32,
+      "sensitivity": "public",
+      "mode": "effective",
+      "compatibility": null,
+      "rules": []
+    },
+    {
+      "semantic_path": "import.external_references.max_fetch_bytes",
+      "env_name": "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_FETCH_BYTES",
+      "process": "admin_api",
+      "value_type": "u64",
+      "unit": "bytes",
+      "default": "262144",
+      "required": false,
+      "minimum": 1,
+      "maximum": 4194304,
+      "sensitivity": "public",
+      "mode": "effective",
+      "compatibility": null,
+      "rules": []
+    },
+    {
+      "semantic_path": "import.external_references.fetch_timeout_ms",
+      "env_name": "CRANK_IMPORT_EXTERNAL_REFERENCE_FETCH_TIMEOUT_MS",
+      "process": "admin_api",
+      "value_type": "u64",
+      "unit": "milliseconds",
+      "default": "10000",
+      "required": false,
+      "minimum": 1,
+      "maximum": 300000,
+      "sensitivity": "public",
+      "mode": "effective",
+      "compatibility": null,
+      "rules": []
+    },
+    {
+      "semantic_path": "import.external_references.max_expanded_nodes",
+      "env_name": "CRANK_IMPORT_EXTERNAL_REFERENCE_MAX_EXPANDED_NODES",
+      "process": "admin_api",
+      "value_type": "u32",
+      "unit": "nodes",
+      "default": "10000",
+      "required": false,
+      "minimum": 1,
+      "maximum": 100000,
+      "sensitivity": "public",
+      "mode": "effective",
+      "compatibility": null,
+      "rules": []
+    },
     {
       "semantic_path": "observability.environment",
       "env_name": "CRANK_ENVIRONMENT",

scripts/authenticated-product-smoke.py

@@ -21,10 +21,17 @@ class SmokeError(RuntimeError):
     pass
 
 
-def safe_error(stage: str, code: str, status: int | None = None) -> SmokeError:
+def safe_error(
+    stage: str,
+    code: str,
+    status: int | None = None,
+    trace_id: str | None = None,
+) -> SmokeError:
     message = f"stage={stage[:64]} code={code[:96]}"
     if status is not None:
         message += f" status={status}"
+    if trace_id is not None:
+        message += f" trace_id={trace_id}"
     return SmokeError(message[:256])
 
 
@@ -134,6 +141,7 @@ class Client:
     def __init__(self, base_url: str, timeout_seconds: int) -> None:
         self.base_url = base_url.rstrip("/")
         self.timeout_seconds = timeout_seconds
+        self.csrf_token: str | None = None
         cookie_jar = http.cookiejar.CookieJar()
         self.opener = urllib.request.build_opener(
             urllib.request.HTTPCookieProcessor(cookie_jar)
@@ -156,6 +164,21 @@ class Client:
         request_headers = {"Accept": "application/json"}
         if headers:
             request_headers.update(headers)
+        if (
+            self.csrf_token
+            and method not in ("GET", "HEAD", "OPTIONS")
+            and (
+                path_or_url.startswith("/api/admin/")
+                or path_or_url.startswith("/api/auth/")
+            )
+            and path_or_url
+            not in (
+                "/api/auth/login",
+                "/api/auth/bootstrap/complete",
+                "/api/auth/session/csrf",
+            )
+        ):
+            request_headers.setdefault("x-csrf-token", self.csrf_token)
         if payload is not None:
             data = json.dumps(payload).encode("utf-8")
             request_headers["Content-Type"] = "application/json"
@@ -203,12 +226,21 @@ def admin_path(workspace_id: str, suffix: str) -> str:
 
 
 def login(client: Client, email: str, password: str) -> None:
-    client.request_json(
+    response = client.request_json(
         "POST",
         "/api/auth/login",
         {"email": email, "password": password},
         expected=(200,),
     )
+    session = require_object(response.body, "login")
+    csrf_token = session.get("csrf_token")
+    if (
+        not isinstance(csrf_token, str)
+        or not 32 <= len(csrf_token) <= 256
+        or not all(character.isalnum() or character in "-_." for character in csrf_token)
+    ):
+        raise safe_error("login", "invalid_csrf_token")
+    client.csrf_token = csrf_token
 
 
 def resolve_workspace(
@@ -261,6 +293,29 @@ def create_operation(
         raise safe_error("operation_create", "invalid_response") from error
 
 
+def is_safe_diagnostic_identifier(value: Any) -> bool:
+    return (
+        isinstance(value, str)
+        and 1 <= len(value) <= 64
+        and value[0].islower()
+        and value[0].isascii()
+        and all(
+            character.isascii()
+            and (character.islower() or character.isdigit() or character == "_")
+            for character in value
+        )
+    )
+
+
+def is_safe_trace_id(value: Any) -> bool:
+    return (
+        isinstance(value, str)
+        and len(value) == 32
+        and value != "0" * 32
+        and all(character in "0123456789abcdef" for character in value)
+    )
+
+
 def run_operation_test(
     client: Client,
     workspace_id: str,
@@ -272,9 +327,24 @@ def run_operation_test(
         admin_path(workspace_id, f"/operations/{operation_id}/test-runs"),
         {"version": operation_version, "input": {"probe": "ok"}},
     ).body
-    if not isinstance(result, dict) or result.get("ok") is not True:
+    if isinstance(result, dict) and result.get("ok") is True:
+        return
+    if not isinstance(result, dict) or result.get("ok") is not False:
         raise safe_error("operation_test", "outcome_not_ok")
 
+    errors = result.get("errors")
+    failure = errors[0] if isinstance(errors, list) and errors else None
+    code = failure.get("code") if isinstance(failure, dict) else None
+    stage = failure.get("stage") if isinstance(failure, dict) else None
+    trace_id = result.get("trace_id")
+    if (
+        is_safe_diagnostic_identifier(code)
+        and is_safe_diagnostic_identifier(stage)
+        and is_safe_trace_id(trace_id)
+    ):
+        raise safe_error(stage, code, trace_id=trace_id)
+    raise safe_error("operation_test", "outcome_not_ok")
+
 
 def operation_etag(client: Client, workspace_id: str, operation_id: str) -> str:
     response = client.request_json(
@@ -326,6 +396,17 @@ def create_agent(client: Client, workspace_id: str, agent_slug: str) -> tuple[st
         raise safe_error("agent_create", "invalid_response") from error
 
 
+def agent_etag(client: Client, workspace_id: str, agent_id: str) -> str:
+    response = client.request_json(
+        "GET",
+        admin_path(workspace_id, f"/agents/{agent_id}"),
+    )
+    etag = response.headers.get("ETag") if response.headers is not None else None
+    if not isinstance(etag, str) or len(etag) > 128 or not etag.startswith('"') or not etag.endswith('"'):
+        raise safe_error("agent_precondition", "invalid_response")
+    return etag
+
+
 def edit_and_archive_operation(
     client: Client,
     workspace_id: str,
@@ -377,11 +458,13 @@ def bind_and_publish_agent(
                 "enabled": True,
             }
         ],
+        headers={"If-Match": agent_etag(client, workspace_id, agent_id)},
     )
     published = client.request_json(
         "POST",
         admin_path(workspace_id, f"/agents/{agent_id}/publish"),
         {"version": agent_version},
+        headers={"If-Match": agent_etag(client, workspace_id, agent_id)},
     ).body
     try:
         published_version = int(published["published_version"])
@@ -436,8 +519,15 @@ def cleanup_smoke_assets(
     if agent_id:
         try:
             client.request_json(
-                "DELETE",
+                "POST",
-                admin_path(workspace_id, f"/agents/{agent_id}"),
+                admin_path(workspace_id, f"/agents/{agent_id}/unpublish"),
+                headers={"If-Match": agent_etag(client, workspace_id, agent_id)},
+                expected=(200, 404),
+            )
+            client.request_json(
+                "POST",
+                admin_path(workspace_id, f"/agents/{agent_id}/archive"),
+                headers={"If-Match": agent_etag(client, workspace_id, agent_id)},
                 expected=(200, 404),
             )
         except SmokeError as error:

scripts/collect-capability-baseline.py

@@ -90,26 +90,52 @@ def playwright_verdict(report: dict[str, Any], required_titles: list[str]) -> tu
         return "fail", counts
     tests = list(iter_tests(report))
     if not tests:
+        if required_titles:
+            counts["failed"] = len(set(required_titles))
+            return "fail", counts
         counts["not_run"] = 1
         return "not_run", counts
     required = {title: False for title in required_titles}
     for test, title in tests:
+        if required and title not in required:
+            continue
         status = test.get("status")
-        results = test.get("results") if isinstance(test.get("results"), list) else []
+        results = test.get("results")
-        result_statuses = [result.get("status") for result in results if isinstance(result, dict)]
+        if not isinstance(results, list):
-        retries = [result.get("retry", 0) for result in results if isinstance(result, dict)]
-        if len(result_statuses) != len(results):
             counts["not_run"] += 1
             continue
-        if status == "flaky" or any(isinstance(retry, int) and retry > 0 for retry in retries):
+        result_statuses: list[Any] = []
-            counts["flaky"] += 1
+        retries: list[int] = []
+        malformed = False
+        for result in results:
+            if not isinstance(result, dict):
+                malformed = True
+                break
+            result_status = result.get("status")
+            retry = result.get("retry", 0)
+            if result_status not in ("passed", "failed", "timedOut", "skipped", "interrupted") \
+                or type(retry) is not int or retry < 0:
+                malformed = True
+                break
+            result_statuses.append(result_status)
+            retries.append(retry)
+        if malformed:
+            counts["not_run"] += 1
+            continue
+        final_status = result_statuses[-1] if result_statuses else None
+        if status in ("unexpected", "failed", "timedOut", "interrupted") \
+            or final_status in ("failed", "timedOut", "interrupted"):
+            counts["failed"] += 1
+        elif status == "flaky":
+            if final_status == "passed" and any(retry > 0 for retry in retries):
+                counts["flaky"] += 1
+            else:
+                counts["not_run"] += 1
         elif status == "skipped" or (not results and status in ("skipped", "expected")):
             counts["skipped"] += 1
-        elif status in ("unexpected", "failed", "timedOut", "interrupted") or any(
+        elif any(retry > 0 for retry in retries):
-            result_status in ("failed", "timedOut", "interrupted") for result_status in result_statuses
+            counts["flaky"] += 1
-        ):
+        elif status == "expected" and results and all(result_status == "passed" for result_status in result_statuses):
-            counts["failed"] += 1
-        elif results and all(result_status == "passed" for result_status in result_statuses):
             counts["passed"] += 1
             if title in required:
                 required[title] = True

scripts/deploy-community.sh

@@ -26,13 +26,28 @@ compose_profiles=""
 if [ "$cache_backend" = "valkey" ] || [ "$cache_backend" = "redis" ]; then
     compose_profiles="--profile cache"
 fi
-
 compose() {
     # Intentional word splitting: compose_profiles is either empty or two arguments.
     # shellcheck disable=SC2086
     docker compose $compose_profiles "$@"
 }
 
+compose_up() {
+    postgres_scale=""
+    if [ "$(env_value POSTGRES_HOST postgres)" != "postgres" ] \
+        && compose config --services | grep -Fxq postgres; then
+        postgres_scale="--scale postgres=0"
+    fi
+    # Intentional word splitting: postgres_scale is either empty or two arguments.
+    # shellcheck disable=SC2086
+    compose up -d --remove-orphans $postgres_scale
+}
+
+show_failure_diagnostics() {
+    compose ps >&2 || true
+    compose logs --no-color migrate >&2 || true
+}
+
 wait_for_stack() {
     readiness_path="$1"
     attempt=1
@@ -48,6 +63,34 @@ wait_for_stack() {
     return 1
 }
 
+find_artifact_container() {
+    expected_root="$1"
+    expected_mount="$(printf 'volume\t%s' "$expected_root")"
+
+    for service in admin-api artifact-storage-init; do
+        candidates="$(compose ps -aq "$service" 2>/dev/null || true)"
+        # Intentional word splitting: Docker container IDs cannot contain whitespace.
+        # shellcheck disable=SC2086
+        set -- $candidates
+        if [ "$#" -gt 1 ]; then
+            echo "Artifact backup found multiple $service containers; refusing an ambiguous volume source" >&2
+            return 1
+        fi
+        if [ "$#" -eq 1 ]; then
+            candidate="$1"
+            oneoff="$(docker inspect --format '{{index .Config.Labels "com.docker.compose.oneoff"}}' "$candidate" 2>/dev/null || true)"
+            mounts="$(docker inspect --format '{{range .Mounts}}{{printf "%s\t%s\n" .Type .Destination}}{{end}}' "$candidate" 2>/dev/null || true)"
+            if [ "$oneoff" != "True" ] && [ "$oneoff" != "true" ] \
+                && printf '%s\n' "$mounts" | grep -Fxq "$expected_mount"; then
+                printf '%s' "$candidate"
+                return 0
+            fi
+        fi
+    done
+
+    return 1
+}
+
 create_backup() {
     timestamp="$(date -u +%Y%m%dT%H%M%SZ)"
     backup_dir="$(pwd)/backups/${timestamp}"
@@ -86,10 +129,9 @@ create_backup() {
             --username "$postgres_user" --dbname "$postgres_db" \
             --format custom --file /backup/postgres.dump
 
-    admin_container="$(compose ps -q admin-api 2>/dev/null || true)"
+    storage_root="$(env_value_from "$backup_env_file" CRANK_STORAGE_ROOT /var/lib/crank/storage)"
-    if [ -n "$admin_container" ]; then
+    if artifact_container="$(find_artifact_container "$storage_root")"; then
-        storage_root="$(env_value_from "$backup_env_file" CRANK_STORAGE_ROOT /var/lib/crank/storage)"
+        docker run --rm --volumes-from "$artifact_container:ro" \
-        docker run --rm --volumes-from "$admin_container" \
             -v "$backup_dir:/backup" alpine:3.21 \
             tar -C "$storage_root" -czf /backup/artifacts.tar.gz .
     elif [ "$previous_deployment" = true ]; then
@@ -104,6 +146,76 @@ create_backup() {
         | sort -nr | awk 'NR > 5 { print $2 }' | xargs -r rm -rf
 }
 
+validate_backup_migration() {
+    admin_image="$(env_value CRANK_ADMIN_API_IMAGE)"
+    if [ -z "$admin_image" ]; then
+        echo "CRANK_ADMIN_API_IMAGE is required for shadow migration validation" >&2
+        return 1
+    fi
+
+    shadow_id="crank-migration-check-$$"
+    shadow_network="${shadow_id}-network"
+    shadow_postgres="${shadow_id}-postgres"
+    shadow_password="crank-shadow-migration-password"
+    shadow_admin="crank_shadow_admin"
+    shadow_user="crank_shadow_owner"
+    shadow_cleanup() {
+        docker rm -fv "$shadow_postgres" >/dev/null 2>&1 || true
+        docker network rm "$shadow_network" >/dev/null 2>&1 || true
+    }
+    trap shadow_cleanup EXIT
+    trap 'exit 130' HUP INT TERM
+
+    docker network create "$shadow_network" >/dev/null
+    docker run -d --name "$shadow_postgres" --network "$shadow_network" \
+        -e POSTGRES_USER="$shadow_admin" \
+        -e POSTGRES_PASSWORD="$shadow_password" \
+        -e POSTGRES_DB=crank \
+        postgres:16-alpine >/dev/null
+
+    shadow_ready=false
+    attempt=1
+    while [ "$attempt" -le 30 ]; do
+        if docker logs "$shadow_postgres" 2>&1 | grep -q 'PostgreSQL init process complete' \
+            && docker exec "$shadow_postgres" pg_isready --username "$shadow_admin" --dbname crank >/dev/null 2>&1; then
+            shadow_ready=true
+            break
+        fi
+        sleep 1
+        attempt=$((attempt + 1))
+    done
+    if [ "$shadow_ready" != true ]; then
+        echo "Shadow PostgreSQL did not become ready" >&2
+        return 1
+    fi
+
+    docker exec "$shadow_postgres" psql --username "$shadow_admin" --dbname crank \
+        --set ON_ERROR_STOP=1 \
+        --command "create role ${shadow_user} login password '${shadow_password}' nosuperuser nocreatedb nocreaterole noreplication" \
+        --command "alter database crank owner to ${shadow_user}"
+
+    timeout 10m docker run --rm --network "$shadow_network" \
+        -e PGPASSWORD="$shadow_password" \
+        -v "$backup_dir:/backup:ro" \
+        postgres:16-alpine \
+        pg_restore --host "$shadow_postgres" --username "$shadow_user" --dbname crank \
+            --no-owner --no-privileges --single-transaction --exit-on-error \
+            /backup/postgres.dump
+
+    shadow_database_url="postgres://${shadow_user}:${shadow_password}@${shadow_postgres}:5432/crank"
+    timeout 10m docker run --rm --network "$shadow_network" \
+        -e CRANK_DATABASE_URL="$shadow_database_url" \
+        "$admin_image" crank-migrate apply
+    shadow_preflight="$(timeout 2m docker run --rm --network "$shadow_network" \
+        -e CRANK_DATABASE_URL="$shadow_database_url" \
+        "$admin_image" crank-migrate preflight)"
+    printf '%s\n' "$shadow_preflight" | grep -Fxq '{"status":"current","version":13}'
+
+    shadow_cleanup
+    trap - EXIT HUP INT TERM
+    echo "Shadow migration validation passed"
+}
+
 rollback() {
     echo "New release failed readiness; restoring previous deployment" >&2
     if [ ! -f .env.previous ] || [ ! -f docker-compose.previous.yml ]; then
@@ -119,16 +231,17 @@ rollback() {
     if [ "$cache_backend" = "valkey" ] || [ "$cache_backend" = "redis" ]; then
         compose_profiles="--profile cache"
     fi
-    compose up -d --remove-orphans
+    compose_up
     wait_for_stack health
 }
 
 compose config -q
 create_backup
 compose pull
+validate_backup_migration
 
-if ! compose up -d --remove-orphans || ! wait_for_stack ready; then
+if ! compose_up || ! wait_for_stack ready; then
-    compose ps >&2 || true
+    show_failure_diagnostics
     rollback
     exit 1
 fi

scripts/docker-http-healthcheck.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+port="${1:?port is required}"
+path="${2:?path is required}"
+
+if [[ ! "$port" =~ ^[0-9]{1,5}$ ]] || (( port < 1 || port > 65535 )); then
+    exit 64
+fi
+
+if [[ "$path" != /* || "$path" == *$'\r'* || "$path" == *$'\n'* ]]; then
+    exit 64
+fi
+
+# Bash provides /dev/tcp without adding a network client package to the runtime
+# image. Compose still applies its own five-second timeout to the whole probe.
+exec 3<>"/dev/tcp/127.0.0.1/${port}"
+printf 'GET %s HTTP/1.1\r\nHost: 127.0.0.1\r\nConnection: close\r\n\r\n' "$path" >&3
+
+IFS=$'\r' read -r -t 2 status <&3
+case "$status" in
+    'HTTP/1.0 200 '*|'HTTP/1.1 200 '*) ;;
+    *) exit 1 ;;
+esac

scripts/restore-community.sh

@@ -32,6 +32,15 @@ compose() {
     # shellcheck disable=SC2086
     docker compose $compose_profiles "$@"
 }
+compose_up() {
+    postgres_scale=""
+    if [ "$(env_value POSTGRES_HOST postgres)" != "postgres" ] \
+        && compose config --services | grep -Fxq postgres; then
+        postgres_scale="--scale postgres=0"
+    fi
+    # shellcheck disable=SC2086
+    compose up -d --remove-orphans $postgres_scale
+}
 
 compose stop admin-api mcp-server ui
 
@@ -40,13 +49,18 @@ postgres_port="$(env_value POSTGRES_PORT 5432)"
 postgres_db="$(env_value POSTGRES_DB crank)"
 postgres_user="$(env_value POSTGRES_USER crank)"
 postgres_password="$(env_value POSTGRES_PASSWORD)"
+docker run --rm \
+    -v "$backup_dir:/backup:ro" \
+    postgres:16-alpine \
+    pg_restore --list /backup/postgres.dump >/dev/null
 docker run --rm --network host \
     -e PGPASSWORD="$postgres_password" \
     -v "$backup_dir:/backup:ro" \
     postgres:16-alpine \
     pg_restore --host "$postgres_host" --port "$postgres_port" \
         --username "$postgres_user" --dbname "$postgres_db" \
-        --clean --if-exists --no-owner --no-privileges /backup/postgres.dump
+        --clean --if-exists --no-owner --no-privileges \
+        --single-transaction --exit-on-error /backup/postgres.dump
 
 admin_container="$(compose ps -aq admin-api)"
 storage_root="$(env_value CRANK_STORAGE_ROOT /var/lib/crank/storage)"
@@ -54,7 +68,7 @@ docker run --rm --volumes-from "$admin_container" \
     -v "$backup_dir:/backup:ro" alpine:3.21 sh -eu -c \
     "find '$storage_root' -mindepth 1 -delete; tar -C '$storage_root' -xzf /backup/artifacts.tar.gz"
 
-compose up -d --remove-orphans
+compose_up
 attempt=1
 while [ "$attempt" -le 45 ]; do
     if curl --fail --silent http://127.0.0.1:3000/ >/dev/null \

tests/unit/test_authenticated_product_smoke.py

@@ -17,6 +17,48 @@ def load_smoke_module():
 
 
 class AuthenticatedProductSmokeTests(unittest.TestCase):
+    def test_client_attaches_csrf_only_to_browser_api_mutations(self) -> None:
+        smoke = load_smoke_module()
+
+        class Response:
+            status = 200
+            headers = {}
+
+            def read(self, _limit):
+                return b"{}"
+
+        class Opener:
+            def __init__(self):
+                self.requests = []
+
+            def open(self, request, timeout):
+                self.requests.append((request, timeout))
+                return Response()
+
+        client = smoke.Client("http://crank.test", 5)
+        opener = Opener()
+        client.opener = opener
+        client.csrf_token = "a" * 32
+
+        client.request_json("POST", "/api/admin/workspaces/ws/operations", {})
+        client.request_json("POST", "http://mcp.test/v1/ws/agent", {})
+
+        self.assertEqual(opener.requests[0][0].get_header("X-csrf-token"), "a" * 32)
+        self.assertIsNone(opener.requests[1][0].get_header("X-csrf-token"))
+
+    def test_login_keeps_server_issued_csrf_token(self) -> None:
+        smoke = load_smoke_module()
+
+        class FakeClient:
+            csrf_token = None
+
+            def request_json(self, *args, **kwargs):
+                return smoke.JsonResponse(200, {}, {"csrf_token": "b" * 32})
+
+        client = FakeClient()
+        smoke.login(client, "owner@crank.test", "safe-password")
+        self.assertEqual(client.csrf_token, "b" * 32)
+
     def test_operation_payload_uses_internal_upstream(self) -> None:
         smoke = load_smoke_module()
 
@@ -136,6 +178,8 @@ class AuthenticatedProductSmokeTests(unittest.TestCase):
                     return smoke.JsonResponse(200, {}, {"operation_id": "op_safe", "version": 7})
                 if path.endswith("/operations/op_safe"):
                     return smoke.JsonResponse(200, {"ETag": '"safe-etag"'}, {"id": "op_safe"})
+                if path.endswith("/agents/agent_safe"):
+                    return smoke.JsonResponse(200, {"ETag": '"safe-agent-etag"'}, {"id": "agent_safe"})
                 if path.endswith("/publish") and "/operations/" in path:
                     return smoke.JsonResponse(200, {}, {"published_version": 7})
                 if path.endswith("/agents"):
@@ -158,11 +202,14 @@ class AuthenticatedProductSmokeTests(unittest.TestCase):
         self.assertEqual((agent_id, agent_version, published_agent_version), ("agent_safe", 3, 3))
         binding = next(payload for _, path, payload, _ in client.requests if path.endswith("/bindings"))[0]
         self.assertEqual(binding["operation_version"], 7)
+        binding_request = next(request for request in client.requests if request[1].endswith("/bindings"))
+        self.assertEqual(binding_request[3]["headers"], {"If-Match": '"safe-agent-etag"'})
         publish = next(request for request in client.requests if request[1].endswith("/operations/op_safe/publish"))
         self.assertEqual(publish[3]["headers"], {"If-Match": '"safe-etag"'})
 
-    def test_admin_test_run_uses_created_version_and_rejects_failed_outcome(self) -> None:
+    def test_admin_test_run_uses_created_version_and_reports_safe_typed_failure(self) -> None:
         smoke = load_smoke_module()
+        trace_id = "0123456789abcdef0123456789abcdef"
 
         class FakeClient:
             def __init__(self, ok):
@@ -171,7 +218,22 @@ class AuthenticatedProductSmokeTests(unittest.TestCase):
 
             def request_json(self, method, path, payload=None, **kwargs):
                 self.payload = payload
-                return smoke.JsonResponse(200, {}, {"ok": self.ok, "errors": [{"message": "secret-canary"}]})
+                return smoke.JsonResponse(
+                    200,
+                    {},
+                    {
+                        "ok": self.ok,
+                        "trace_id": trace_id,
+                        "errors": [
+                            {
+                                "code": "outbound_target_rejected",
+                                "stage": "adapter",
+                                "message": "secret-canary",
+                                "context": {"url": "https://private.invalid/token-canary"},
+                            }
+                        ],
+                    },
+                )
 
         passing = FakeClient(True)
         smoke.run_operation_test(passing, "ws", "op", 9)
@@ -179,7 +241,188 @@ class AuthenticatedProductSmokeTests(unittest.TestCase):
 
         with self.assertRaises(smoke.SmokeError) as raised:
             smoke.run_operation_test(FakeClient(False), "ws", "op", 9)
-        self.assertNotIn("secret-canary", str(raised.exception))
+        self.assertEqual(
+            str(raised.exception),
+            "stage=adapter code=outbound_target_rejected trace_id=0123456789abcdef0123456789abcdef",
+        )
+        for forbidden in ("secret-canary", "private.invalid", "token-canary", "context"):
+            self.assertNotIn(forbidden, str(raised.exception))
+
+    def test_admin_test_run_fails_closed_for_malformed_diagnostics(self) -> None:
+        smoke = load_smoke_module()
+        valid_trace_id = "0123456789abcdef0123456789abcdef"
+        valid_failure = {"code": "upstream_timeout", "stage": "upstream"}
+        malformed_results = (
+            None,
+            [],
+            "secret-canary",
+            {"trace_id": valid_trace_id, "errors": [valid_failure]},
+            {"ok": None, "trace_id": valid_trace_id, "errors": [valid_failure]},
+            {"ok": 0, "trace_id": valid_trace_id, "errors": [valid_failure]},
+            {"ok": "false", "trace_id": valid_trace_id, "errors": [valid_failure]},
+            {"ok": False, "errors": []},
+            {"ok": False, "errors": {}},
+            {"ok": False, "trace_id": valid_trace_id, "errors": [None]},
+            {"ok": False, "errors": ["secret-canary"]},
+            {"ok": False, "errors": [valid_failure]},
+            {
+                "ok": False,
+                "trace_id": valid_trace_id[:-1],
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id + "0",
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": "0" * 32,
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id.upper(),
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": "a" * 31 + "١",
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": "a" * 31 + "\n",
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": 42,
+                "errors": [valid_failure],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "", "stage": "upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "a" * 65, "stage": "upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "Upstream_timeout", "stage": "upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_таймаут", "stage": "upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": 42, "stage": "upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_timeout", "stage": ""}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_timeout", "stage": "a" * 65}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_timeout", "stage": "Upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_timeout", "stage": "вверх"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_timeout", "stage": 42}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "unsafe-value://secret-canary", "stage": "upstream"}],
+            },
+            {
+                "ok": False,
+                "trace_id": valid_trace_id,
+                "errors": [{"code": "upstream_timeout", "stage": "unsafe stage secret-canary"}],
+            },
+            {
+                "ok": False,
+                "trace_id": "not-a-trace-id-secret-canary",
+                "errors": [{"code": "upstream_timeout", "stage": "upstream"}],
+            },
+        )
+
+        class FakeClient:
+            def __init__(self, result):
+                self.result = result
+
+            def request_json(self, *args, **kwargs):
+                return smoke.JsonResponse(200, {}, self.result)
+
+        for result in malformed_results:
+            with self.subTest(result=result), self.assertRaises(smoke.SmokeError) as raised:
+                smoke.run_operation_test(FakeClient(result), "ws", "op", 9)
+            self.assertEqual(str(raised.exception), "stage=operation_test code=outcome_not_ok")
+            self.assertNotIn("secret-canary", str(raised.exception))
+
+    def test_admin_test_run_accepts_maximum_length_safe_diagnostics(self) -> None:
+        smoke = load_smoke_module()
+        identifier = "a" * 64
+        trace_id = "0123456789abcdef0123456789abcdef"
+
+        class FakeClient:
+            def request_json(self, *args, **kwargs):
+                return smoke.JsonResponse(
+                    200,
+                    {},
+                    {
+                        "ok": False,
+                        "trace_id": trace_id,
+                        "request_preview": {"credential": "secret-canary"},
+                        "response_preview": {"token": "secret-canary"},
+                        "errors": [{"code": identifier, "stage": identifier}],
+                    },
+                )
+
+        with self.assertRaises(smoke.SmokeError) as raised:
+            smoke.run_operation_test(FakeClient(), "ws", "op", 9)
+        rendered = str(raised.exception)
+        self.assertEqual(rendered, f"stage={identifier} code={identifier} trace_id={trace_id}")
+        self.assertLessEqual(len(rendered.encode("ascii")), 256)
+        self.assertNotIn("secret-canary", rendered)
+
+    def test_admin_test_run_success_ignores_malformed_diagnostics(self) -> None:
+        smoke = load_smoke_module()
+
+        class FakeClient:
+            def request_json(self, *args, **kwargs):
+                return smoke.JsonResponse(
+                    200,
+                    {},
+                    {
+                        "ok": True,
+                        "trace_id": "secret-canary",
+                        "errors": "secret-canary",
+                        "request_preview": {"token": "secret-canary"},
+                    },
+                )
+
+        smoke.run_operation_test(FakeClient(), "ws", "op", 9)
 
     def test_edit_and_archive_use_fresh_operation_preconditions(self) -> None:
         smoke = load_smoke_module()

tests/unit/test_collect_capability_baseline.py

@@ -57,6 +57,7 @@ class CapabilityBaselineCollectorTests(unittest.TestCase):
     def test_required_openapi_tests_fail_closed_when_missing_skipped_or_flaky(self) -> None:
         required = ["OpenAPI required scenario"]
         reports = [
+            {"suites": []},
             {"suites": [{"specs": [{"title": "another scenario", "tests": [{"status": "expected", "results": [{"status": "passed", "retry": 0}]}]}]}]},
             {"suites": [{"specs": [{"title": required[0], "tests": [{"status": "skipped", "results": []}]}]}]},
             {"suites": [{"specs": [{"title": required[0], "tests": [{"status": "flaky", "results": [{"status": "failed", "retry": 0}, {"status": "passed", "retry": 1}]}]}]}]},
@@ -70,6 +71,73 @@ class CapabilityBaselineCollectorTests(unittest.TestCase):
                 self.assertEqual(candidate["execution_verdict"], "fail")
                 self.assertFalse(candidate["accepted"])
 
+    def test_required_scope_ignores_unrelated_flaky_tests(self) -> None:
+        required = "OpenAPI required scenario"
+        report = {
+            "suites": [
+                {
+                    "specs": [
+                        {
+                            "title": required,
+                            "tests": [{"status": "expected", "results": [{"status": "passed", "retry": 0}]}],
+                        },
+                        {
+                            "title": "Unrelated wizard scenario",
+                            "tests": [
+                                {
+                                    "status": "flaky",
+                                    "results": [
+                                        {"status": "failed", "retry": 0},
+                                        {"status": "passed", "retry": 1},
+                                    ],
+                                }
+                            ],
+                        },
+                        {"title": "Unrelated failure", "tests": [{"status": "unexpected", "results": [{"status": "failed", "retry": 0}]}]},
+                        {"title": "Unrelated skip", "tests": [{"status": "skipped", "results": []}]},
+                        {"title": "Unrelated malformed", "tests": [{"status": "expected", "results": None}]},
+                    ]
+                }
+            ]
+        }
+        result, output, temporary = self.run_playwright(report, [required])
+        self.addCleanup(temporary.cleanup)
+        self.assertEqual(result.returncode, 0, result.stderr)
+        candidate = json.loads(output.read_text(encoding="utf-8"))
+        self.assertEqual(candidate["execution_verdict"], "pass")
+        self.assertTrue(candidate["accepted"])
+        self.assertEqual(candidate["summary"], {"passed": 1, "failed": 0, "flaky": 0, "skipped": 0, "not_run": 0})
+
+    def test_final_failure_after_retry_is_failed_not_flaky(self) -> None:
+        for status in ("unexpected", "flaky"):
+            with self.subTest(status=status):
+                report = {
+                    "suites": [
+                        {
+                            "specs": [
+                                {
+                                    "tests": [
+                                        {
+                                            "status": status,
+                                            "results": [
+                                                {"status": "failed", "retry": 0},
+                                                {"status": "failed", "retry": 1},
+                                            ],
+                                        }
+                                    ]
+                                }
+                            ]
+                        }
+                    ]
+                }
+                result, output, temporary = self.run_playwright(report)
+                self.addCleanup(temporary.cleanup)
+                self.assertEqual(result.returncode, 0, result.stderr)
+                candidate = json.loads(output.read_text(encoding="utf-8"))
+                self.assertEqual(candidate["execution_verdict"], "fail")
+                self.assertEqual(candidate["summary"]["failed"], 1)
+                self.assertEqual(candidate["summary"]["flaky"], 0)
+
     def test_raw_report_content_never_reaches_candidate_or_error(self) -> None:
         canary = "Bearer secret-canary /home/private/workspace https://private.invalid?q=secret"
         report = {"suites": [], "errors": [{"message": canary}], "stdout": [canary]}
@@ -84,6 +152,8 @@ class CapabilityBaselineCollectorTests(unittest.TestCase):
         reports = [
             {"errors": [{"message": "fatal"}], "suites": [{"specs": [{"tests": [{"status": "expected", "results": [{"status": "passed"}]}]}]}]},
             {"suites": [{"specs": [{"tests": [{"status": "expected", "results": ["not-an-object"]}]}]}]},
+            {"suites": [{"specs": [{"tests": [{"status": "expected", "results": None}]}]}]},
+            {"suites": [{"specs": [{"tests": [{"status": "expected", "results": [{"status": "passed", "retry": True}]}]}]}]},
         ]
         for report in reports:
             with self.subTest(report=report):