fix(compliance): backport evaluator rejection coverage to 3.2.x

.agents/sdk-shim-ledger.json

@@ -1 +1,21 @@
-[]
+[
+  {
+    "id": "creative-build-probe-8240",
+    "title": "Portable evaluator build probe preparation",
+    "kind": "integrity-pinned-package-patch",
+    "status": "temporary",
+    "owner": "adcp compliance maintainers / issue 8240",
+    "upstream": "https://github.com/adcontextprotocol/adcp-client/pull/3332",
+    "problem": "The pinned SDK lacks portable valid build-probe preparation and the served-catalog authority prerequisite required by the merged upstream fix.",
+    "localBehavior": "Patch only reviewed SDK14.0.0-rc.53 CJS/ESM loader, runner, schema-authority and declaration preimages. A local creative_build_probe ALS scope is established before preflight and pairs served catalog selection with actual JSON capability validation (builtin or external); stock14 behavior remains outside that opt-in and for other tasks. Unlike the initial candidate, authority modules include an explicit maintenance delta and do not claim whole-module identity to15.9.5. Preserve request pins, accounting, general TaskExecutor and all four main helper payloads. Authority regressions use ledgered private scope/validator interfaces and owned temporary stock copies reconstructed by exact reverse hashes; no direct installed-package edits.",
+    "removalCondition": "Remove after the upstream creative-build-probe runner fix is merged, available in a qualified SDK release, and both full and standalone regression suites pass after upgrading the pinned dependency. Do not release the SDK from this task.",
+    "paths": [
+      "scripts/install-sdk-creative-build-probe.cjs",
+      "tests/evaluator-build-probe-authority.test.cjs"
+    ],
+    "terms": [
+      "dist/lib/testing/",
+      "dist/lib/validation/"
+    ]
+  }
+]

.changeset/evaluator-off-list-rejection-coverage.md

@@ -0,0 +1,24 @@
+---
+"adcontextprotocol": patch
+---
+
+Make both off-list evaluator rejection phases mandatory for agents that advertise evaluator support. Preserve unsupported-feature applicability, prevent later discovery from skipping the security boundary, and check all outbound methods and paths on the off-list evaluator origin.
+
+Bind each required rejection probe to an advertised build operation and a valid
+source input. Handle non-build-first and legacy catalogs without relying on
+catalog ordering or ranking features. Add an owned, integrity-checked temporary
+SDK bridge for shared full/standalone runner preparation; unavailable fixtures
+remain explicit coverage gaps with no passing security credit.
+
+Forward authored probe recipes through the manual runner, including recipe-only
+kits. Malformed understood catalog declarations retain their catalog cause as
+prerequisite failures (actionable dependency failures in hosted classification),
+instead of being mislabeled controller setup failures.
+
+Resolve capability catalogs against the validated served release while preserving
+the pinned build request and source-input contract, including provisioned legacy
+formats when the catalog uses newer canonical vocabulary.
+
+Backport the merged main fix with a separately qualified, all-target guarded
+SDK14.0.0-rc.53 adapter and the existing upstream served-catalog authority
+prerequisite. Preserve this branch's exact dependency and lockfile.

.github/workflows/sdk-response-conformance.yml

@@ -32,6 +32,7 @@ jobs:
           go-version-file: scripts/sdk-response-conformance/go/go.mod
           cache-dependency-path: scripts/sdk-response-conformance/go/go.sum
       - run: npm ci --ignore-scripts
+      - run: node scripts/install-sdk-creative-build-probe.cjs
       - run: python3 -m pip install -r scripts/sdk-response-conformance/requirements.txt
       - run: npm run test:sdk-response-conformance
       - name: Capture and validate real SDK responses

Dockerfile

@@ -5,9 +5,12 @@ WORKDIR /app
 
 # Copy package files
 COPY package*.json ./
+COPY scripts/install-sdk-creative-build-probe.cjs ./scripts/install-sdk-creative-build-probe.cjs
+COPY patches/sdk-creative-build-probe ./patches/sdk-creative-build-probe
 
 # Install ALL dependencies (including dev dependencies for TypeScript build)
-RUN npm ci --ignore-scripts
+RUN npm ci --ignore-scripts \
+ && node scripts/install-sdk-creative-build-probe.cjs
 
 # Copy source code
 COPY . .
@@ -153,10 +156,13 @@ RUN apt-get update \
  && rm -rf /var/lib/apt/lists/*
 
 COPY package*.json ./
+COPY scripts/install-sdk-creative-build-probe.cjs ./scripts/install-sdk-creative-build-probe.cjs
+COPY patches/sdk-creative-build-probe ./patches/sdk-creative-build-probe
 
 # --ignore-scripts blocks arbitrary postinstall lifecycle scripts from the full
 # dep tree; native deps are rebuilt explicitly per-package.
 RUN npm ci --omit=dev --ignore-scripts \
+ && node scripts/install-sdk-creative-build-probe.cjs \
  && npm rebuild sharp \
  && npm cache clean --force
 

package.json

@@ -117,10 +117,10 @@
     "test:pagination-invariant": "node --test --test-force-exit --test-timeout=30000 tests/lint-pagination-invariant.test.cjs",
     "test:version-envelope": "node --test --test-force-exit --test-timeout=30000 tests/lint-version-envelope.test.cjs tests/compact-version-envelope.test.cjs",
     "test:test-dynamic-imports": "node --test --test-force-exit --test-timeout=30000 tests/lint-test-dynamic-imports.test.cjs",
-    "test:sdk-shims": "node --test --test-force-exit --test-timeout=30000 tests/lint-sdk-shims.test.cjs tests/sdk-runtime-compat.test.cjs",
+    "test:sdk-shims": "node --test --test-force-exit --test-timeout=30000 tests/lint-sdk-shims.test.cjs tests/sdk-runtime-compat.test.cjs tests/sdk-creative-build-probe-bridge.test.cjs",
     "test:tenant-routing": "vitest run --config server/vitest.config.ts server/src/training-agent/tenants/tenant-smoke.test.ts",
     "test:format-identity-boundaries": "node --test --test-force-exit --test-timeout=30000 tests/lint-format-identity-boundaries.test.cjs",
-    "test:sdk-runner-capability-gates": "node --test --test-force-exit --test-timeout=30000 tests/sdk-runner-capability-gates.test.cjs",
+    "test:sdk-runner-capability-gates": "node --test --test-force-exit --test-timeout=30000 tests/sdk-runner-capability-gates.test.cjs && node --test --test-timeout=30000 tests/evaluator-allowlist-runner.test.cjs tests/evaluator-build-probe-authority.test.cjs tests/evaluator-build-probe-cleanup.test.cjs",
     "test:callapi-state-change": "node --test --test-force-exit --test-timeout=30000 tests/lint-callapi-state-change.test.cjs",
     "test:cdn-artifacts-worker": "node --test --test-force-exit --test-timeout=30000 tests/artifact-cdn-worker.test.cjs",
     "test:sign-protocol-tarball": "node --test --test-force-exit --test-timeout=30000 tests/sign-protocol-tarball.test.cjs",
@@ -197,7 +197,8 @@
     "eval:addie-fixed-traces-direct-full-suite": "tsx --import dotenv/config server/tests/manual/fixed-trace-direct-full-suite-eval.ts",
     "eval:addie-gemini-direct-ablation": "tsx --import dotenv/config server/tests/manual/gemini-direct-ablation-eval.ts",
     "eval:addie-matched-v4-plan": "tsx server/tests/manual/matched-v4-evaluation-plan.ts",
-    "eval:addie-matched-v4-authorized": "tsx --import dotenv/config server/tests/manual/matched-v4-authorized-execution.ts"
+    "eval:addie-matched-v4-authorized": "tsx --import dotenv/config server/tests/manual/matched-v4-authorized-execution.ts",
+    "postinstall": "node scripts/install-sdk-creative-build-probe.cjs"
   },
   "dependencies": {
     "@adcp/sdk": "14.0.0-rc.53",

patches/sdk-creative-build-probe/README.md

@@ -0,0 +1,57 @@
+Temporary SDK bridge for adcp#8240. Upstream source and executable tests:
+https://github.com/adcontextprotocol/adcp-client/pull/3332
+
+Qualified upstream source: `9f932f174e8125e00ec55d64db2a292cba2be7f9`.
+
+The manifest contains bounded insertions into SDK **14.0.0-rc.53** loader, runner and
+matching declarations, plus the shared resolver compiled from upstream source.
+It does not replace the whole newer SDK runner. Both public CJS and ESM entry
+points receive the same preparation, including standalone steps. Existing global
+accounting and required-phase/vacuous-pass guards remain unchanged.
+
+`node scripts/install-sdk-creative-build-probe.cjs` validates package identity,
+all target preimages/postimages, unique insertion anchors and deterministic helper
+bytes before writing. A repeated complete installation is a no-op; drift or a
+mixed state fails loudly. Normal installs run this as postinstall. Docker's
+`--ignore-scripts` stages explicitly invoke only this trusted installer, as does
+the SDK conformance workflow. Tests reconstruct stock bytes and exercise both
+public module entry points before and after explicit installation.
+
+The automatic source recipe relies on existing format completeness semantics;
+operator recipes assert account-specific valid input for every route matched by
+the existing fixture DSL. Capability catalogs follow the SDK-authorized served release; build selectors,
+source manifests and dispatch retain the run pin. Missing source/preconditions yield visible coverage
+gaps, not passing security credit. See the authored runner test-kit contract.
+
+Qualification/removal: upgrade to a published SDK containing the upstream fix,
+remove this directory, installer, lifecycle/Docker/workflow calls and owned shim
+ledger entry, then run the complete authored fixture and bridge/standalone
+regressions with the public replacement. This task publishes no release.
+
+To refresh after a separately reviewed maintenance qualification, rebuild that SDK commit, copy
+only creative-build-probe.{js,mjs,d.ts,d.mts}, and regenerate the bounded insertion
+blocks and SHA-256 pre/postimage hashes against a pristine npm SDK 14.0.0-rc.53 install.
+Review the resulting manifest and rerun integrity/public-entry regressions.
+
+Maintenance prerequisite: SDK 14 does not export served-release schema selection.
+The manifest ports the SDK 15.9.5 selection additions with a reviewed maintenance
+delta: an AsyncLocalStorage scope enabled only by the locally authored
+`creative_build_probe` directive, before capability preflight. Catalog selection
+and its capability response-schema backstop use that same scope. Within it, the
+actual authorized JSON schema governs even for builtin roots; stale generated
+Zod disagreement remains visible as a warning. Invalid JSON fails and unavailable
+schema produces the existing failing `no_schema_registered` harness diagnostic.
+These modules are not byte-identical to SDK 15.9.5. Unrelated OAuth or error changes
+are excluded.
+The four evaluator helper payloads remain byte-identical to main. All 25 target
+preimages/postimages are qualified against the exact locked npm SDK 14 tarball.
+The branch lockfile and dependency pin are unchanged. No newer runner is copied.
+
+Scope: only opted creative-probe capability preflight and capability response
+checks use the added policy. Other storyboards and non-capability checks retain
+stock SDK 14 semantics, including external-root authority. Explicit false scopes
+shadow parent opt-in. `servedReleaseSchemaRoots` is deferred/inert for other
+storyboards and outer bundle discovery; mixed bundles forward it to each opted
+storyboard. A direct call of the new probe helper is itself an explicit opt-in.
+SDK 14's general TaskExecutor is untouched and remains pin-only. This is not a
+claim of whole-SDK parity with 15.9.5.

patches/sdk-creative-build-probe/creative-build-probe.d.mts

@@ -0,0 +1,34 @@
+import type { FixtureMatchClause, StoryboardContext } from './types.mjs';
+export interface CreativeBuildProbeSpec {
+    context_prefix: string;
+    message: string;
+}
+export interface CreativeBuildProbeInputFixture {
+    match: FixtureMatchClause[];
+    /** Operator-provisioned valid input for every route matched by this recipe. */
+    input: Record<string, unknown>;
+}
+export type CreativeBuildProbeResolution = {
+    status: 'resolved';
+    capabilityId?: string;
+    formatId?: Record<string, unknown>;
+    context: StoryboardContext;
+    source: 'generation' | 'test_kit';
+    requirements: FixtureMatchClause[];
+} | {
+    status: 'unsatisfied';
+    detail: string;
+} | {
+    status: 'failed';
+    detail: string;
+};
+export declare function validateCreativeBuildProbeSpec(value: unknown): asserts value is CreativeBuildProbeSpec;
+/**
+ * Uses only actual catalog entries and explicitly provisioned test-kit inputs.
+ * Global generation support does not identify a generation route in a mixed
+ * catalog: the automatic brief recipe is therefore used only when the catalog
+ * has exactly one build route (including legacy unaddressable build entries),
+ * and a complete, understood declaration proves there are no technical seeds.
+ * Other modes require a matching operator recipe, not inferred source assets.
+ */
+export declare function resolveCreativeBuildProbe(creative: unknown, spec: CreativeBuildProbeSpec, testKit?: unknown, adcpVersion?: string, governanceEnforcement?: unknown): CreativeBuildProbeResolution;

patches/sdk-creative-build-probe/creative-build-probe.d.ts

@@ -0,0 +1,34 @@
+import type { FixtureMatchClause, StoryboardContext } from './types';
+export interface CreativeBuildProbeSpec {
+    context_prefix: string;
+    message: string;
+}
+export interface CreativeBuildProbeInputFixture {
+    match: FixtureMatchClause[];
+    /** Operator-provisioned valid input for every route matched by this recipe. */
+    input: Record<string, unknown>;
+}
+export type CreativeBuildProbeResolution = {
+    status: 'resolved';
+    capabilityId?: string;
+    formatId?: Record<string, unknown>;
+    context: StoryboardContext;
+    source: 'generation' | 'test_kit';
+    requirements: FixtureMatchClause[];
+} | {
+    status: 'unsatisfied';
+    detail: string;
+} | {
+    status: 'failed';
+    detail: string;
+};
+export declare function validateCreativeBuildProbeSpec(value: unknown): asserts value is CreativeBuildProbeSpec;
+/**
+ * Uses only actual catalog entries and explicitly provisioned test-kit inputs.
+ * Global generation support does not identify a generation route in a mixed
+ * catalog: the automatic brief recipe is therefore used only when the catalog
+ * has exactly one build route (including legacy unaddressable build entries),
+ * and a complete, understood declaration proves there are no technical seeds.
+ * Other modes require a matching operator recipe, not inferred source assets.
+ */
+export declare function resolveCreativeBuildProbe(creative: unknown, spec: CreativeBuildProbeSpec, testKit?: unknown, adcpVersion?: string, governanceEnforcement?: unknown): CreativeBuildProbeResolution;