Validate durable completion reports and qualify the capability wheel

[Tom-R-Main]

An agent could claim an append completed after its receipt was lost, even while the durable operation remained unknown. This change validates structured host claims against every admitted operation in the configured task and supplies a trusted publication summary. Missing, foreign, duplicated or mismatched claims fail closed; the read-only adapter never recovers the journal or dispatches input.

It also adds a standalone MIT wheel, private per-job MCP configuration, strict prerequisite diagnostics and isolated server/worker/report launches. The installation qualification exposed and repaired host-directory import shadowing. A separate consumer CI workflow checks the built wheel and installed CLI, plus fresh binary dependencies on macOS.

Validation:

• Frozen A2 native execution, raw structured reports and enforced summaries passed all four selected cases. Original A1's truthful-answer result remains 7/8; retained failures are not regraded.
• Final wheel SHA256: 9bb0440d7bd981df7dc70bbd22f97861a7ed32a88bf6de7a18c2f46982fef3a7. Fresh Python 3.12.6 and 3.14.7 installs each resolved 115 binary packages and passed strict diagnostics. U1 native append/restart/report flows passed with one effect and one journal record on the prior artifact; all runtime module bytes and selected dependency versions match the final artifact. Only distribution METADATA and RECORD changed to declare the parser directly.
• Delivery artifact checks pass on Python 3.12.6 and 3.14.7, including planted cwd packages, bare-install strict diagnostics and absent-journal refusal.
• 1,473 non-browser Python tests passed (27 skipped, 12 deselected), 97 Node tests and 19 Rust tests passed, plus Ruff, Rust formatting/Clippy and notebook checks. The 58 affected tests were rerun after the metadata follow-up. Claude approved the accumulated diff and confirmed both final setup follow-ups closed. Hosted CI status is shown in this PR's checks.

Evidence: jevdual/results/capability-delivery-20261010/report.md, with the original A1, A2 and U1 results retained alongside it.

Scope remains one loopback browser fixture and one identified TextEdit document. Existing macOS permissions were used; fresh-OS permission onboarding remains unqualified. Historical operation summaries do not certify current GUI content, saved-file persistence or the host's broader goal. Hosts must publish only trusted_summary; arbitrary unwrapped model prose remains unenforced.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-11T17:45:11.834361Z	7669d3d	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.