feat(ba-propose): tell the origin ticket what shipped

[azevedo]

Risk: high — a skill-invocation guard was removed, and this command gains an irreversible write outside the repo

/ba-propose reads the origin ticket, legitimately drifts from it during a run, and lands that drift in docs/ and a PR body — while the ticket itself, the thing the team actually reads, stays exactly as it was written. It now posts one append-only comment after the PR opens, carrying the shipped PR URL and the deviation texts gathered from the commit range. Getting there also forced two changes worth knowing about: the composition rows that read issue context had to be re-gated, because a context that exists but is empty silently changes what they render, and ba-propose itself became model-invokable, so it can run from a follow-up turn instead of printing an invocation for you to retype.

Motivation

The flow had exactly one outward effect: the PR. Everything it learned along the way — that a retry had to be jittered, that a clock-skew allowance had to be 30s — reached reviewers and docs/, never the ticket. At typo tier it reached nobody: the reviewer body suppresses deviations at that tier, and the "Linear rollup" the spec named as the fallback guarantee was never implemented by any step.

Two roadmap items land together here — the feature itself, and the bug that six documented lines across three files already described it as working. Refs #92, #91.

Where to look

• skills/ba-propose/SKILL.md — the new ## Ship-Time Ticket Write-Back section owns the whole operation: ref-shape routing, a closed six-literal disposition enum evaluated first-match-wins, a pre-check that observes local presence and makes no network call, a bounded wait, a no-retry rule, and the sanitization contract. The call site in 5e is two lines and zero branches.
• The removed disable-model-invocation line — one frontmatter key, plus three prose mirrors that asserted this skill could not be invoked. Reviewing the frontmatter alone will miss those.

Breaking changes

BREAKING: skills/ba-propose/SKILL.md no longer declares disable-model-invocation: true. The skill is now model-invokable, so a session can run it — including on its own initiative — where the Skill tool previously refused. Its outward effects are now gated by the standing rule that hard-to-reverse and outward-facing actions get confirmed at the point of action, rather than structurally by the loader. That is a deliberate trade, not an oversight: the residual risk is a session auto-triggering it off its own description on a bare "commit this". CLAUDE.md carries the reasoning and explicitly warns against extending it to ba-execute or ba-review, whose blast radius is an open-ended sequence of edits rather than one confirmable action.

Testing instructions

There is no automated test for prompt behaviour — the shipped artifact is prose — so this was decided by fixture A/B rather than by argument: seven planted fixtures across two arms (main vs. proposed), one subagent per cell, no repository access, every verdict attributed to a quoted source sentence or explicitly marked inconclusive.

Round 1 failed its own pre-committed decision rule, and the design changed twice as a result:

• Two cells independently found the GitHub write route was dead by construction — refs were normalized to a repo-qualified shape that the router rejects as cross-repo, so the GitHub half of the feature was unreachable and four disposition rows were dead text.
• Two more found the Step 4 preview announced posts that the provenance gate would then refuse, which reads as a promise.

Round 2 passed all three conditions. Score table and per-cell attribution: docs/research/2026-08-25-ship-time-write-back-ab-research.md.

To verify by hand: run node scripts/selfcheck-invariants.mjs && node scripts/check-invariants.mjs (both green), then the live run below.

Testing limitations

The A/B simulates a tracker from planted fixture facts; no subagent made a network call. A confirmed write and a rejected write have never been observed against a real tracker, so those two dispositions are unverified — which is why the receipt's posted literal is defined by an observable (the call exits zero and the response carries a comment id or URL) rather than left to the word.

Closing that gap needs a fresh session — a session cannot dry-run the skill body it loaded at start — shipping a trivial diff twice against two distinct freshly-created scratch issues. Distinct because the behaviour under test writes, which makes repeated runs non-independent. The research doc reserves a section for the observed receipt lines and carries a standing note explaining why a verbatim receipt line must not appear anywhere else in that file.

CI covers the invariants only: 8/8 checks and 57/57 selfchecks pass, including a new needle that pins the retired Linear rollup claim so it cannot be reintroduced by a future edit.

Proof: pending — add tests / QA notes / screenshots before merge

Alternatives considered

• An injectable write port would have let the fixture A/B reach the confirmed-write and rejected-write paths directly, keeping the call site just as small. Rejected at brainstorm lock as an element of a already-rejected design; the accepted cost is that those two outcomes are only ever observed by hand.
• Fixes <ref> for GitHub cross-refs was the obvious reading once a GitHub issue ref could reach that row for the first time — but a closing keyword auto-closes the issue on merge, a new outward effect that no requirement asked for, on a row whose unconditional keyword is already a known separate defect. It renders Refs instead; Linear keeps Fixes, where the keyword is inert.