Propagate Closes #N from constituent PRs into the release PR

[sebastientaggart]

What changed

Reworks how issue close-intent flows from constituent feature PRs into the /deploy release PR, so issues reliably auto-close when a release ships.

submit-for-review.md — dev-target PRs now use Closes #N when a PR fully resolves an issue (and Related to #N for context-only references) instead of the old Issue #N link. Closes #N is inert on a non-default-branch merge, so it does not prematurely close — it records recoverable close-intent for /deploy to harvest. Trunk mode gains the Related to #N option too.

deploy.md — Step 2 now parses a close set (every Closes #N line across constituent PRs, recorded per-PR for verbatim reproduction) separately from a reference set (Related to #N plus legacy Issue #N). The Step 5 HUMAN GATE shows two distinct lists — "Issues that will close on merge" vs "Referenced but NOT closing" — so a dropped close is visible at gate time. The release PR body reproduces the close set verbatim and adds Related to #N lines for references.

status.md — added related to #<number> to the in-progress detection patterns.

Docs — branching.md, docs/skills/submit-for-review.md, and docs/skills/deploy.md updated to describe the new convention.

Generated adapters (.claude / .cursor / .agents / .gemini) regenerated via sync.py --force.

Why

Previously, dev-target PRs recorded issue links as Issue #N, which carries no machine-recoverable close-intent. Weeks later /deploy could not distinguish "this should close" from "this is just referenced," and merged both into one set — so the release PR's auto-close set came out incomplete and issues silently stayed open after release.

Notes

• Legacy Issue #N PRs already merged to dev degrade to a visible entry in the gate's reference list — no migration machinery; manual close of any straggler is acceptable for these small private projects.
• The ticket attributed part of the bug to /start, but start.md has no closing-keyword handling — all PR-body enforcement lives in /submit-for-review, so start.md is unchanged.

Test plan

These are prompt-skill markdown files with no executable suite. make check (placeholder + permission validation via sync.py --validate) passes, and the source/adapter copies are in sync.

Closes #186

[sebastientaggart]

Review Summary

Verdict: APPROVE

This PR introduces a Closes #N (inert until reaching the default branch) / Related to #N convention for dev-targeting PRs, and teaches /deploy to parse a close-set and reference-set separately. The core premise is correct against GitHub's real behavior: closing keywords only auto-close when a PR merges into the repository's default branch, so Closes #N on a dev-targeted PR is genuinely inert and safe to record as close-intent.

Findings

• [NOTE] The propagation chain is internally consistent across the three skills: /submit-for-review records Closes #N / Related to #N, /deploy partitions constituent-PR references into a close-set (reproduced verbatim into the main-targeting release PR) and a reference-set (emitted as Related to #N, no auto-close), and /status recognizes both keywords. The "reproduce every Closes #N line verbatim, omitting none" instruction is repeated identically in both the two-branch and three-branch /deploy Step 6 variants, so neither promotion path drops a close.

• [NOTE] The close-set/reference-set computation lives in /deploy Step 2 under a single {{#if BRANCH_DEV}} block, which covers both two-branch and three-branch modes (BRANCH_DEV is set in both). Step 4 and both Step 6 variants reference that same computed data, so the three-branch (BRANCH_TEST) path is covered, not just two-branch.

• [NOTE] The /status change adds related to #<number> (and was already matching bare #<number> and issue #<number>) to the "In progress" bucket detection. Because the bare-#<number> match already caught any PR mentioning an issue, this addition is explicitness rather than a behavior change — a context-only Related to reference would already have classified the issue as in-progress. No regression.

• [NOTE] Legacy Issue #N is deliberately routed to the reference-set (not guessed into the close-set) and surfaced at the human gate as "referenced but NOT closing — confirm none of these should actually close." This is a safe fail-closed choice: it never silently auto-closes an issue whose close-intent can't be recovered, and the gate gives the operator a chance to catch a straggler.

• [NOTE] Generated adapters under .claude/, .cursor/, .agents/, .gemini/ mirror their skills/github-agile/*.md sources. Verified by running sync.py (reports all skills up to date, clean working tree afterward) and by spot-checking .claude/commands/deploy.md against the source's rendered BRANCH_DEV path. This repo is two-branch mode (BRANCH_DEV=dev, BRANCH_PROD=main, BRANCH_TEST=""), so the adapters correctly render the BRANCH_DEV branch; the trunk-mode template edits in the sources do not appear in this repo's adapters, as expected.

The change is documentation/prompt-only, touches no authentication, payments, secrets, production config, or destructive operations, so the sensitive-area gate does not apply. No correctness, consistency, or sync issues found.