Record Flow 9.1.0 qualification and canary evidence

[vriesd]

The Flow 9.1.0 release now has a passed exact-artifact canary and a sealed OpenAI-only qualification bundle. The bundle independently regrades 38 of 38 fixed-target attempts as passing and binds the packed artifact SHA-256 0b7f7a66bf1910b5d0da3a235f891d532b4aa352a2dbd25e398325c18e051b24.

Four original campaign transcripts contained the recording host's home path in diagnostic tool output. The source report is preserved byte-for-byte as an extensionless snapshot. A redaction receipt lists each original and derived transcript hash and the 12 output fields changed. The sealed bundle contains the privacy-safe derived report and transcripts. Recorded decisions, attempt outcomes, and the 38/38 score did not change. The original transcripts remain local and are not published.

Validation: bun run check passed with 1,591 tests, 20 skips, and no failures. bun run scripts/release-metadata.ts --tag v9.1.0 --artifact <exact tarball> --canary evals/canary/9.1.0.json independently verified the committed bundle and exact canary. This PR does not tag or publish the release.

[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 2026-09-27T06:51:28.821766Z	b6a3a03	PR opened

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.

[vriesd]

Independent read-only evidence review at b6a3a033dd65b81b21d1d4bee90ee47590ee8e53: PASS.

The packed artifact and bundled artifact both match SHA-256 0b7f7a66bf1910b5d0da3a235f891d532b4aa352a2dbd25e398325c18e051b24. All 272 sealed objects match their recorded sizes and hashes. The original report matches the committed snapshot byte for byte. The redacted bundle changes exactly the receipt's 12 diagnostic output fields across four transcripts; all 38 recorded outcomes remain unchanged. Strict bundle regrading returns VERIFIED at 38/38 with the exact canary.

This verifies retained evidence. It does not rerun paid model calls.