fix: validate metadata-only release readiness evidence

[leehack]

Summary

Closes #485.

• Add an explicit, evaluator-verified core-patch release metadata route without fabricating changed production tests or classifying artifact-consumer releases as standard risk.
• Require literal exact-base/head Git blobs and a complete fixed-path inventory: only the next core patch scalar, matching generated local chat lock version, current release notes and four maintained install docs. Companion versions/pins are already prepared and must not change here.
• Preserve independent exact-head audit, zero P1/zero unresolved-thread gates, strict schema/row commands, and external authenticated auditor/App/environment/ruleset boundaries.

Scope and safety

The metadata route requires release-metadata-verification, its exact strict-release-verifier/test command, PASS evidence and exactly the unchanged test/unit/tooling/verify_release_docs_companion_pins_test.dart. The verifier and test must be identical regular Git blobs at base/head. The evaluator does not execute candidate scripts or authenticate a claimed test run; independent review must inspect actual command evidence.

Any hook/runtime/pin/binding/Swift/workflow/policy change, broader companion release, renamed/added/deleted/nonregular path, dependency/SDK/source/hash/path/inventory lock change, stale snippet, historical changelog rewrite or executable MDX activation cannot use this route. Documentation prose remains independently reviewed. Normal high-risk changes retain the changed-production-test requirement.

This fixes a pre-existing policy mismatch exposed by release PR #482, not a runtime regression caused by that release or Apple guard #484. Release #482 stays draft and is not modified by this corrective PR.

Validation

• PASS: canonical Flutter3.47.1/Dart3.13.1 preparation, whole format (590 files unchanged) and analysis.
• PASS: final serial full VM 2,046 tests /77 fixture-dependent skips.
• PASS: focused evaluator 102 tests; combined readiness/classifier/matrix/release-verifier contracts 151/151, including real temporary Git repository exact-blob and stale-head tests.
• PASS: Node20 production docs build and strict links; git diff check.
• PASS: live read-only replay of release chore: prepare 0.8.23 release #482 exact head 16eaa80fcdb7578599d8757056cfb18536a642cb / base ccc2d35633e3ca6347a697f19fde4944e0a8dbd0 reaches unverifiedPrerequisites instead of missingTestPath without changing its diff or inventing a changed test.
• Adversarial coverage includes code/pin/workflow/policy mixtures, pubspec/lock mutation, file modes, stale/self audit, P1/threads, verifier/test tampering, wrong/malformed evidence, historical docs, fenced/inline code-to-MDX activation, Docusaurus executable fences, and unavailable external prerequisites.
• Real model/device smoke: N/A to developer-only evidence tooling. No public inference, pins, artifacts or runtime behavior changes.

High-risk regression review

• Classification: high-risk / regressionPolicy.
• Implementation task: current_release_prep, isolated /private/tmp/llamadart-readiness-485.
• Exact head: a8bb6ff318d2ef028b81f549cd1aefe7929abf07; base ccc2d35633e3ca6347a697f19fde4944e0a8dbd0.
• Independent Astra final exact-head review: PASS; independently reran 102 evaluator tests. Fence activation, JSX template-literal and script-body review findings were fixed with durable adversarial regressions before readiness.
• Current-head hosted checks: 12/12 SUCCESS, including replacement CI run 34159032954 and the read-only advisory. Linux coverage and aggregate passed. Review threads: 0 total /0 unresolved. No automated approving review is claimed.
• Known PR-caused P1 regressions: zero after independent review.
• Separate pre-existing tooling issue Exclude generated website output from maintained writer source scan #487 tracks the maintained-source scanner reading generated website build files. One local full-VM run overlapped docs generation and hit an ENOENT in that unchanged scanner; the authoritative serialized rerun passed 2,046 tests. This PR does not weaken or modify that scanner.

Explicit boundary

Successful local high-risk evaluation still returns only unverifiedPrerequisites (exit2), isReady=false. No payload, flag, credential or environment variable can produce operational-ready output. Protected external enforcement remains separately governed and unconfigured by design; this PR neither enables nor weakens it.

No mark-ready, merge, publication, tag, dispatch, settings or runtime-pin action is included.