[andersalm]

Part of #89. From the 6 Oct audit: package run 37492620275 built source 67c268c while develop CI for 67c268c had failed.

Change.

• release-publish.sh prepare refuses a source unless develop's own ci.yml push run for that exact commit succeeded, with every required job success. A skipped required job also refuses.
• policy and seed print the source, package run and CI run together.
• Develop push runs no longer cancel each other, and they always run custody-harness-smoke. Every develop commit therefore gets a complete run the gate can use. PRs keep their path filter and cancelling.
• AGENTS.md: one Acceptance record per owning issue (candidate, exercised checks, review links with SHA, installed evidence, remaining gate, next owner) is the only place for current status.

Tests: release-publish 25 OK, ci-release-policy 18 OK.

Review: Claude reviewer FAIL (filter base under cancellation; a filter failure could pass the required custody check; two status places). Fixed by removing the skip handling, then PASS on 5c4801f. Left as is (P3): main, tag and dispatch runs also no longer cancel.

Candidate e68e9ae = reviewed 5c4801f + a develop merge (no conflicts). Written by Codex (gpt-6.1-sol).

[andersalm]

CI/build coherence review for #194 found a conflict between this gate and pre-merge release packaging. On published train 8679214a, release-package.yml:61-68 permits a new open PR head H. A normal merge makes develop commit M, while the package receipt retains H. release-publish.sh:57-62,143 then requires successful develop-push CI at H. PR CI at H and develop CI at M cannot satisfy that exact-SHA gate. The current local train f7224ac7 has the same release sources.

Until qualified input-equivalence admission is reviewed, prepare release artifacts from the final merged commit. Keep the complete-green requirement and original producer provenance. Pre-merge packaging becomes useful only when the integration policy and publication evidence selector agree on which tested artifact can publish. The full findings and measured costs are in #194. This review changed no code and started no build.