Retain independent task plan and canonical digest technical reviews

Assistant: codex
Assistant-Model: gpt-6.1-sol
Assistant-Session: 01a126e1-76f8-7643-8cf3-e93336539ade

WORK-RECORDS.md

@@ -51,6 +51,7 @@
 | task | FLEX-WP-ADHOC-2026-10-11-T01 | done | — | workplans/ADHOC-2026-10-11.md |
 | task | FLEX-WP-ADHOC-2026-10-11-T02 | done | — | workplans/ADHOC-2026-10-11.md |
 | task | FLEX-WP-ADHOC-2026-10-11-T03 | done | — | workplans/ADHOC-2026-10-11.md |
+| task | FLEX-WP-ADHOC-2026-10-11-T04 | done | — | workplans/ADHOC-2026-10-11.md |
 | task | FLEX-WP-0001-T001 | done | — | workplans/FLEX-WP-0001-repo-intent-and-architecture-baseline.md |
 | task | FLEX-WP-0001-T002 | done | — | workplans/FLEX-WP-0001-repo-intent-and-architecture-baseline.md |
 | task | FLEX-WP-0001-T003 | done | — | workplans/FLEX-WP-0001-repo-intent-and-architecture-baseline.md |

docs/evidence/2026-10-11-task-execution/canonical-independent-verification.json

@@ -0,0 +1,18 @@
+{
+  "native_execution": false,
+  "synthetic_only": true,
+  "source_manifest_sha256": "c15b3704d59a3441f2ce203f1b866b62d83a49f553b50afbaa9c6fcc06105876",
+  "original_counterexample_digest": "sha256:09bef6f94ad1caf77b9cb39139c59a337b954608da4e205831962cce23050568",
+  "repaired_request_digest": "sha256:dcaff2c7d48e04a78ca49f2b0e7c19da3d01868bf432a9ac56170511ae125465",
+  "actual_go_request_digest": "sha256:dcaff2c7d48e04a78ca49f2b0e7c19da3d01868bf432a9ac56170511ae125465",
+  "repaired_approval_binding_digest": "sha256:dcaff2c7d48e04a78ca49f2b0e7c19da3d01868bf432a9ac56170511ae125465",
+  "actual_go_approval_binding_digest": "sha256:dcaff2c7d48e04a78ca49f2b0e7c19da3d01868bf432a9ac56170511ae125465",
+  "digests_match": true,
+  "independent_focused_passed": 149,
+  "independent_additional_passed": 21,
+  "additional_binary64_samples": 384,
+  "full_suite_green_claimed": false,
+  "numeric_profile": "finite binary64 floats; Python integer magnitude <= 2^53-1; all larger integers/nonfinite refused",
+  "go_algorithm_changed": false,
+  "policy_constant_flipped": false
+}

docs/evidence/2026-10-11-task-execution/compiler-canonical-counterexample.json

@@ -0,0 +1,16 @@
+{
+  "schema": "source-integration-counterexample/v1",
+  "native_execution": false,
+  "target_manifest_sha256": "41cda74ab32384c789ea8a760475671aadad4e5121335c8f60567ff53d66665b",
+  "case": "actual consumer permits purpose café<&>; full canonical request hash differs from Go HTML escaping",
+  "synthetic_only": true,
+  "actual_compiler_and_consumer_parser_accepted": true,
+  "python_request_digest": "sha256:09bef6f94ad1caf77b9cb39139c59a337b954608da4e205831962cce23050568",
+  "go_request_digest": "sha256:dcaff2c7d48e04a78ca49f2b0e7c19da3d01868bf432a9ac56170511ae125465",
+  "policy_effect": "deny",
+  "policy_reason": "initiating_provider_unadmitted",
+  "request_contains_tokens": false,
+  "fixture_sha256": "3ba1337dc913076286b6c018b2f60026d906699aa3fe6d4d50d712a64ac27c2e",
+  "go_result_sha256": "dc652292bfbb56528aa5c9b47057d64473b62951680f1a532f7ffdd662d5002d",
+  "scope_note": "Compiler catalogue fingerprint matches actual Secrets local target helper. Full request digest mismatch is separate Python/Go canonical join integration blocker, not compiler authority bypass."
+}

docs/reviews/2026-10-11-platform-typed-plan-technical-review.md

@@ -0,0 +1,70 @@
+# Independent technical review: Platform typed plan compiler
+
+Reviewer: Access source integration agent access_task_execution_policy.
+Scope: technical component interoperability only; no Platform owner approval.
+Candidate manifest /tmp/platform-typed-plan-stage-v2/manifest.json SHA256
+41cda74ab32384c789ea8a760475671aadad4e5121335c8f60567ff53d66665b.
+Verified all four candidate file hashes. Code SHA256
+0a96c39a71c6fcca5ea695d1a46ce20d7345b27a507851d4d46e1543fd97e656;
+maker tests SHA256 432f2a316a94b586286426906504cae9368e74cc6ba7a32d788c37924cad8224.
+
+Result: source component correspondence established within the stated boundary;
+whole execution readiness refused. Independently reproduced 40 maker checks and
+10 additional checks. Tests use synthetic approved copies of public owner data,
+not an approved actual CCR or matching native catalogue. No policy source constant
+was flipped, no credential/native call or owner checkout write occurred.
+
+The actual thirteen-field Secrets build_action_request catalog_target equals
+Compiler.proposal's owner-derived projection and local catalogue fingerprint.
+Actual ExecutionContract.parse accepts the six generated GET/PUT/GET ACL then
+GET/PUT/GET auth-role method/path/payload hashes. Compiler.before_call returns
+bytes whose hashes equal those descriptor entries; GET bodies hash empty bytes.
+Normal fixture's actual Python full/approval-free request digest matches Go.
+Actual Go/OPA evaluates the unchanged closed source package as deny,
+initiating_provider_unadmitted. This is not a runtime allow test.
+
+Independent forged Plan schema/operation/source_pins/catalog_target/budget/cleanup
+metadata all refused on before_call. After an observed pinned owner auth TTL
+change and re-pinning, an old Plan refused before role call. Supplied root/body,
+wrong operation, replay-readiness, alternate fields/path and local observed source
+indirection/drift counterexamples likewise refuse in the independently rerun maker
+suite. Source trust/runtime custody and mutation of Python memory remain external
+requirements explicitly stated by the maker; the compiler is not a sandbox.
+
+## Concrete additional integration counterexample
+
+A valid catalogue consumer purpose café<&> is accepted by the actual catalogue
+validator, compiler, typed descriptor and request builder. Its local target hash
+matches Secrets. But the actual Python canonical request digest differs from Go
+RequestDigest: Go JSON Marshal escapes HTML characters, Python compact UTF8 JSON
+does not. Exact evidence is recorded in
+../evidence/2026-10-11-task-execution/compiler-canonical-counterexample.json.
+This is a Python/Go canonical join blocker, not permission escalation. The current
+hard provider refusal still denies both. Baseline ASCII fixture matches; a broad
+claim of arbitrary metadata canonical interoperability is not established.
+Keep the counterexample and fix/test canonical consumer serialization before any
+real request/approval/PDP consume adoption. Do not conflate descriptor-local
+catalogue hashing (intentionally separate) with the PDP canonical request hash.
+
+## Boundaries still required
+
+Compiler request_digest is an opaque prefixed hash, not authenticated authority.
+Full descriptor construction must additionally bind owner grant/baseline/readback,
+estate/origin/deadline/source bundle and verified initiating representation. It
+must not fabricate task_execution_projection or member/MFA/verified flags. Current
+Access source admission remains literally false; authenticated KeyCape origin,
+User/Tenant facts and server projection are absent. Source compiler cannot supply
+them. Native CCR42 is still in-flight/unadmitted and no live matching catalogue
+was demonstrated.
+
+The future owner mediator must enforce the real full canonical request/claim/PDP
+and trusted consume joins, owner-private journal phase/order/budget and durable
+pending intent, unchanged exact source/body bytes immediately at dispatch, actual
+native token TTL/use/endpoint scope and custody, baseline conflict/readback and
+revoke/cleanup/unknown recovery. before_call rederives bytes but has no dispatcher,
+clock deadline, phase authority, transaction or native CAS. Component source tests
+cannot close Secrets T06 or RPF35-T02/T12, or certify authenticated native access.
+
+This review uses the existing completed Access T04 source-contract scope and does
+not sign as Platform's owner or accept another owner/provider boundary. No new
+operational workplan is created.

docs/reviews/2026-10-11-secrets-canonical-go-technical-review.md

@@ -0,0 +1,65 @@
+# Independent technical review: Secrets Go canonical digest repair
+
+Reviewer: Access source integration agent access_task_execution_policy.
+Disposition: exact bounded consumer source repair technically accepted;
+no Secrets owner assent, package activation or native execution admission.
+Manifest /tmp/secrets-canonical-go-stage/manifest.json SHA256
+c15b3704d59a3441f2ce203f1b866b62d83a49f553b50afbaa9c6fcc06105876.
+Verified all six candidate hashes. authorization.py staged SHA256
+d58e5fd01ca2470a5d11a6b4f7e9d262a0c1565637e07b35466faa30e4991d14.
+
+## Independent source evidence
+
+In an isolated full public source copy overlaid with exact candidate files,
+149 focused tests passed: real Go oracle, immutable execution descriptor and four
+builder preservation, action authorization, claim/consume joins and normal replay.
+21 additional independent tests passed, including 384 separately seeded binary64
+bit samples, nextafter neighbors at fixed/scientific formatting boundaries,
+HTML/Unicode/control/separator characters in nested keys/arrays, negative zero,
+unsupported numeric refusals and semantic mutation/approval-free joins.
+
+The original actual typed compiler counterexample (catalogue consumer purpose
+café<&>) is repaired. Its original Python digest
+sha256:09bef6f94ad1caf77b9cb39139c59a337b954608da4e205831962cce23050568
+now becomes exactly the actual unchanged Go result
+sha256:dcaff2c7d48e04a78ca49f2b0e7c19da3d01868bf432a9ac56170511ae125465.
+Both full and approval-free digests match. Safe summarized evidence is in
+../evidence/2026-10-11-task-execution/canonical-independent-verification.json.
+No human-policy source constant was flipped and no native provider was contacted.
+
+## Numeric contract and algorithm boundaries
+
+The admitted numeric profile is explicitly finite binary64 floats and Python
+integers in [-(2^53-1),2^53-1]. Larger integers refuse rather than hash an exact
+Python integer while Go's HTTP decoder rounds it. NaN and infinities refuse.
+Shortest binary64 formatting, Go exponent/fixed thresholds and signed zero are
+exercised against the actual exported Go API after actual JSON decoding. This
+is not a claim that arbitrary Go-native json.Number/big integers, Decimal objects,
+invalid scalar Unicode or other non-JSON values are supported.
+
+Only the consumer full Access request and Access approval-free request digest
+serialization changes. Descriptor-local UTF8 catalogue target fingerprints,
+Approval's distinct claim digest and owner body hashes remain separate; no Go
+canonical algorithm, served claim, immutable original decision or policy changes.
+Existing approvals must not be rewritten: an old mismatching request needs a
+fresh properly bound request/approval. Advisory purpose hashing parity does not
+make an enforcement API accept an otherwise rejected top-level purpose.
+
+Very large integer timestamps, including proposed epoch-nanosecond projection
+fields, are outside this profile. Future provider projection must select a faithful
+cross-runtime representation instead of silently rounding those fields. The prior
+human projection is unadmitted and its source package remains closed; this review
+provides no admission for that shape.
+
+## Validation limits
+
+This review does not claim the full repository suite is green. The frozen maker
+receipt retains full-suite custody failures under uid65534 in /tmp and one unchanged
+baseline reproduction. Only one of the 16 full-suite failures was reproduced against unchanged
+baseline. The other 15 were not independently classified as baseline failures.
+The coordinator corrected an earlier broader statement; no broader baseline
+reproduction is claimed. The focused
+candidate suite and independent counterexamples passed; private file custody
+behavior is not weakened to satisfy the test environment. No deployment, credential
+request, native proof, default migration, ownership acceptance or operational task
+closure follows from this serializer repair.