Assignment ADR-0053/delivery/008, coordinator 01a0e095-f279-7fd0-a8ef-92ce8f9547fc/local. Root requested local diagnosis before considering another production window. Delivery007 c9 UNKNOWN, observed404, original audit/result, expiry and separate cleanup are preserved. Delivery008 made no production proof HTTP request, signature/window, role activation, env change, deployment or retained deployment deletion. No Phase C or business/recipient call occurred. Root remains sole canonical publisher/indexer.
Accepted local reproduction
The retained sourceac04747 artifact passes an enabled, correctly signed request through its actual materialized Linux Node24.13.0 ARM64 Vercel launcher. It materializes343 exact file mappings, changes to the function-root directory as the real launcher requires, resolves the SDK through native createRequire/resolve, verifies SDK3.8.1 and all five original executable hashes, and returns200/no-store/passedtrue. Three real rows (approved/reserved/completed) exist in a fresh isolated PostgreSQL17.9 fixture. Native role/privilege/owner/clock/metadata reservation SQL runs against the fixture's restricted role; the SDK itself is real, with explicit local RSA/JWKS/request-context and token-fetch doubles. One SDK exchange double, one JWKS double, two real runtime audit SQL calls and zero unexpected network calls were recorded. Bad signature, runtime/config drift, default-off and SDK tampering return404 without an audit query.
The isolated cluster is .tmp/local008/pgcluster, loopback127.0.0.1:55482, databaseproof008 and roleproof_runtime008. Only this new cluster was stopped after final evidence; its data is retained. No older database, fixture or listener was reset/stopped. The native PG transport double checks exact public CA and rejectUnauthorizedtrue options, then bridges only to real local PostgreSQL. This is not a deployed TLS handshake, Vercel workload/epoch/resource or production role attestation.
Prototype qualification: the first CJS-only pg hook did not intercept native import, so its404 is excluded from every causal claim. Network isolation was not demonstrated for that prototype; it may have attempted an unauthenticated pooler connection using a nonexistent synthetic fixture username/password. No real credential was loaded and no production SQL success was observed. A subsequent JWKS-double and local transport-deadline issue were also excluded. The accepted harness patches the actual CJS export before ESM observes it and forbids all nonlocal sockets. Thus the production404's exact cause did not reproduce in accepted local fixtures and remains unknown.
Narrow source changes and distinct artifact
Local source commit 5f9ea287c19e5d9466f6b85eea362302f9f33eb9 adds only temporary proof-stage observability and stricter approval span, plus local diagnosis tooling/tests. diagnostic-log.ts emits three fixed fields: event, stage and outcome. It logs only when proof-enabled and production, never request data, tokens, URLs, database strings, keys, headers, stack traces or arbitrary error messages. Denials still return404/no-store; default-off is silent. Adapter resolution/config/target stages, signed approval/runtime binding, native SDK integrity, reservation, claims/SDK entry and completion have finite stage codes. Existing approval/source/target/runtime binding, explicit TLS, role privileges, durable audit cap, cancellation and SDK checks remain in place.
The private app and local owner writer now reject exp-nbf greater than120. The writer captures numeric HTTP status and bounded character-allowlisted x-vercel-id/x-vercel-request-id values before asserting status/body/audit results. Unknown remains conservative; no retry or replacement approval is introduced. This does not retroactively change c9's original121-second encoded span.
Exact diff against pre008 source8975d56 is docs/adr0053-workload-proof/review/local008-source.patch, SHA256 710599acdf4bb5446b2a30cf9cd64ae69c97f763a410af536ee4180dcee0c41a. Distinct .tmp/linux/diagnostic-observability/ Build Output was produced from clean exact5f9 source, with fake unavailable database inputs, no env pull, proof/fanout/external controls off and no migrations. Verification records691 entries (572 regular,119 contained links), six Node24/configuredARM64 functions, private proof maxDuration30, zero native addons, unchanged five cron definitions, all five SDK hashes, embedded source and zero actual env files. Tracked public .env.example LF hash11bdf511 remains exact to its Git blob and packaged bytes, with only this public-placeholder allowance.
| Evidence | SHA256 |
|---|---|
| New artifact-manifest.json | e2fdd3e9d91f499d53e1c10765e4795029562e56ef3e99c21680db9732c604f6 |
| New artifact-verification.json | 45c4a04dd065760b155634038a6b904ce0058963c1d29b5136263f9c3c75fbb5 |
| New native packaged matrix | 64e1bd2fea2db7d225b1c20c57627c744c0810b7240e2dc1fa213afea4530b36 |
| Full review packet | acd90cea673781d6d503e60fd208013a1d52fe4318da38ee796a8f499a750f2a |
Artifact evidence is under .tmp/linux/diagnostic-observability/.tmp/workload-proof-build/. Matrix is docs/adr0053-workload-proof/results/2026-10-03-local008-new-packaged-matrix.json; packet is docs/adr0053-workload-proof/review/local008-review-packet.json. The packet pins23 source/tool/diff/plan/evidence/artifact/native-runtime files. Harness expansion after5f9 is local tool source, not a claim that the built artifact contains that later tooling revision.
The distinct new packaged launcher also returns200/no-store with three real audit rows. Eleven native negatives pass: bad signature,121-second encoded span, source drift, runtime/config drift, default-off, replay and each of the five individually tampered SDK files. All but replay perform zero audit queries; replay makes one denied reservation query and does not invoke SDK again. Fixed stage logs identify approval, reservation and SDK-integrity denials. Default-off produces no stage log. Total runtime audit calls are3 (reserve, complete, denied replay), with one SDK/JWKS double and zero unexpected network activity. Approval span is exactly120. This positive fixture passage remains separate from genuine Vercel attestation.
Validation:996 tests pass,15 skipped; eight audit-lock unit cases pass; typecheck and runtime smoke pass. No booking, financial reconciliation or operator page flow changed, so production lock audit and e2e were not run. Native packaged cases supplement source tests. Both old artifacts were independently rescanned: retained ac04747 manifestd20025 and all572 regular bytes match; cleanup5c36 manifest1540b1 and all508 regular bytes match. Links remain in their preserved outputs. Only local build dependencies moved to the new build checkout; old Build Output bytes were not amended.
Next proof and cleanup, review only
Concrete plan is docs/adr0053-workload-proof/review/local008-next-proof-plan.md, pinned by the packet. It requires independent root review of this exact source/artifact/fixture/tool package and a distinct custody entry/profile before any new authority. Immutable old f48 custody must not be reused because it pins old source/artifacts and encodes121seconds. Delivery008 creates no new production claim or executable authorization.
Any separately authorized future attempt must encode total span at most120, use one credential/server-clock fifteen-minute maximum with no extension/rekey, pin current actual default/canonical/all observed alias targets before staging, enforce exactOID/revoked/zero-session/TLS/effective-role/five-denial and lifetime gates, create/remove only seven authoritative env IDs, require exact READY/uploaded new artifact pins and scoped promotion, and make at most one guarded request. Preserve original UNKNOWN with no retry; use only safe status/correlation and fixed provider stage evidence for diagnosis. Restore exact READY cleanup6kMS, original alias/env/settings baseline and revoked role, drain only reviewed exactPID/full-precision start, hold unknown sessions for root, separately reconcile audit/route absence/protection and existing approval expiry, then close custody. Retained diagnostic deletion remains held for exact root retirement review.
Production workload/exchange, migration, epoch/role, consumer/resource/provider, financial and activation gates remain open. Frozen G2, old runtimes and both deadlines are preserved. Cleanup6kMS remains the approved production cleanup baseline; no new deployment or production mutation was made by this local review package.