Harden adapter conformance and prepare v0.4.0

[thomaswillner]

Outcome

Adds an independently authored adapter-conformance corpus and fail-closed validator for the experimental v0.4.0 release. The work incorporates Rastislav Drahos/DanceNitra's adapter-level findings without copying or vendoring the external runner or fixture.

Validator hardening

• five complete adapter behavior cases with exact counter-results
• stable provider-local proposition identities and active counts; no text-normalization identity
• executable anti-vacuity attacks against shared receipt acceptance, feed verification, and semantic aggregation seams
• clean runtime commit/tree plus tracked binding-source digest
• fail-closed Git and binding construction/execution/metadata timeouts
• binding or execution faults classified as invalid evidence (CLI exit 2)
• exact contributor, source, licence, conflict, AI-assistance, and independent-authorship provenance
• malformed and vacuous receipt evidence refused without traceback

Release and ownership boundary

• version metadata prepared for 0.4.0
• SECURITY support transition remains conditional until immutable v0.4.0 publication
• LLM Errata and repository-authored work remain Thomas Willner's property
• Rastislav Drahos/DanceNitra retains copyright in the external MIT fixture
• independent specification implementations retain the irrevocable royalty-free grant and required accessible attribution
• copying or commercial use of repository Reference Code still requires separate permission

Evidence

Exact implementation head: d7f5a451613c7f8bb40eba9f3e63e16255d8b494

• full local suite: 358 tests passed
• make check: PASS
• make publication: PASS
• canonical adapter report: five cases and three validator controls PASS; runtime commit/tree, canonical corpus digest, clean binding-repository identity, tracked dependency-manifest digest, and full-lifetime admitted import isolation emitted
• independent exact-HEAD spec review: PASS at d7f5a45, no Critical/Important findings
• independent exact-HEAD standards review: PASS at d7f5a45, no High/Medium findings
• exact-head CI: Python 3.11 and 3.13 PASS — https://github.com/thomaswillner/llm-errata/actions/runs/31752704702
• diff check and secret-safe pattern scan: clean

Readiness boundary

Overall verdict remains NOT_PROD_READY. G1 is PASS; G2-G6 remain BLOCKED. Passing this corpus is internal harness evidence, not independent G2 review, G4 interoperability, audited production cryptography, a Phase 3 experiment, or operational certification.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3d5dc890e4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".