fix(#2980): hotfix cherry-pick only picks commits that change what ships (#2981)

* fix(#2980): pre-skip workflow-file cherry-picks in release-sdk hotfix loop

The default GITHUB_TOKEN issued to the release-sdk run lacks the
`workflow` scope, so the prepare job's `git push origin "$BRANCH"` is
rejected by GitHub when any cherry-picked commit modifies a file under
`.github/workflows/`:

  ! [remote rejected] hotfix/X.YY.Z -> hotfix/X.YY.Z
    (refusing to allow a GitHub App to create or update workflow ...
     without `workflows` permission)

Pre-#2980 behavior: the auto_cherry_pick loop happily picked
workflow-file commits, then the trailing push exploded with no clear
signal which commit was the culprit. v1.39.1 hit this on PR #2977
(run 25232010071) — earlier release-sdk fixes (#2965, #2967, #2970)
had been skipped on conflict so their workflow-file changes never
reached the push step, masking the bug; #2977 was the first
workflow-file commit to apply cleanly and the push immediately
exploded.

Fix: pre-pick guard in the cherry-pick loop. Inspect each candidate
commit's file list via `git diff-tree --no-commit-id --name-only -r`
BEFORE attempting the pick. If any path matches `^\.github/workflows/`,
skip the commit, emit a `::warning::` annotation naming the dropped
commit, and append to a new `WORKFLOW_SKIPPED` bucket. The run summary
surfaces this bucket in its own section, distinct from `CONFLICT_SKIPPED`
(real merge conflicts) and `POLICY_SKIPPED` (feat/refactor exclusions),
so operators reviewing the run never confuse the remediation paths.

The loud-warning piece is non-negotiable: silent drops were explicitly
rejected as a failure mode during the option-1/2/3 tradeoff discussion.
If a workflow-file fix genuinely needs to ship in a hotfix, the
operator applies it manually on the hotfix branch using a token with
`workflow` scope, or lands it on main and re-cuts the release.

Regression covered by tests/bug-2980-skip-workflow-file-cherrypicks.test.cjs
(5 assertions: pre-pick guard exists, uses `git diff-tree`, emits
`::warning::`, lands in dedicated bucket, surfaces in summary).

The bug-2964 test's 4 KB window after the cherry-pick-loop anchor was
nudged to 6 KB to accommodate the new pre-pick scaffolding — the test's
own comment had already anticipated this kind of growth (citing #2970's
merge-commit pre-skip as prior precedent).

Closes #2980

* refactor(#2980): replace workflow-file pre-skip with shipped-paths filter

The previous commit on this branch caught only the .github/workflows/*
subset of the bug, treating the symptom (push rejection on workflow-file
changes) rather than the root cause (the fix:/chore: filter is too broad
— it picks any commit with that conventional-commit type even when the
diff cannot affect the published npm package).

CI-only fixes (release-sdk.yml itself, hotfix tooling, test-only
commits) shouldn't flow through hotfix runs at all — they cannot change
what `npm install get-shit-done-cc@X.YY.Z` produces. The
.github/workflows/* push rejection is just the loudest of these
"shouldn't have been picked" cases; tests/, docs/, .planning/ commits
get picked silently with the same lack of effect on consumers.

Replace the workflow-file pre-skip with a shipped-paths filter:

- New scripts/diff-touches-shipped-paths.cjs reads package.json `files`,
  plus package.json itself (always-shipped per `npm pack` semantics),
  and exits 0 iff any input path is in the shipped set. Lockfile is
  not shipped (npm pack excludes it unless explicitly in `files`).
- Workflow loop now pipes `git diff-tree --no-commit-id --name-only -r`
  through the classifier; on exit 1 the commit is skipped and
  appended to a new NON_SHIPPED_SKIPPED bucket (replaces
  WORKFLOW_SKIPPED).
- Run summary surfaces NON_SHIPPED_SKIPPED as informational — no
  ::warning:: annotation. A non-shipping commit cannot affect the
  package, so a yellow alert would imply remediation is possible
  and would mislead operators.

The classifier in a separate .cjs file (rather than inline bash
heredoc) is so its rules — directory-prefix vs exact-match,
package.json-always-shipped, lockfile-not-shipped — are unit-testable
in tests/bug-2980-hotfix-only-picks-shipping-changes.test.cjs (11 new
assertions: 4 static workflow + 6 classifier behavioral + 1 mixed-
diff edge case).

Why this dissolves the original push-rejection bug: workflow files
aren't in `files`, so workflow-only commits are skipped pre-pick.
The push step never sees them.

If a workflow-file fix genuinely needs to ship in a hotfix release
(extremely rare — the hotfix workflow is read from main's ref, not
the hotfix branch's), the operator applies it manually using a token
with `workflow` scope. The pre-skip puts that requirement in the run
summary explicitly.

Closes #2980

CHANGELOG.md

@@ -8,6 +8,7 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ### Fixed
 
+- **`release-sdk` hotfix only cherry-picks commits that change what actually ships** — the `fix:`/`chore:` filter in `Prepare hotfix branch` was too broad: it picked any commit with that conventional-commit type regardless of whether the diff could affect the published npm package. CI-only fixes (release-sdk.yml itself, hotfix tooling, test-only commits) were getting cherry-picked into hotfix branches even though they cannot change the tarball — and the subset touching `.github/workflows/*` then caused the prepare job's `git push` to be rejected by GitHub because the default `GITHUB_TOKEN` lacks the `workflow` scope, aborting the run. v1.39.1 hit this on PR #2977 (run [25232010071](https://github.com/gsd-build/get-shit-done/actions/runs/25232010071)). The loop now pre-skips any candidate commit whose `git diff-tree` output doesn't intersect the npm tarball's shipped paths (entries in `package.json` `files`, plus `package.json` itself, which `npm pack` always includes). Skipped commits land in a new `NON_SHIPPED_SKIPPED` summary bucket framed as informational — non-shipping commits cannot affect the package, so the skip needs no operator action. The shipped-paths classifier lives in `scripts/diff-touches-shipped-paths.cjs` so its rules (file-OR-directory prefix matching `npm pack` semantics, the always-shipped rule for `package.json`, the lockfile-not-shipped rule) are unit-testable. Regression covered by `tests/bug-2980-hotfix-only-picks-shipping-changes.test.cjs`. (#2980)
 - **`release-sdk` hotfix workflow fails on real run with `npm error Version not changed`** — the `release` job's `Bump in-tree version (not committed)` step ran `npm version "$VERSION"` without `--allow-same-version`, so it errored on real (non-dry-run) hotfix runs because `prepare` had already committed the bump on the hotfix branch. The release job's checkout `ref` is asymmetric — `BRANCH` (already bumped) on real runs vs `BASE_TAG` (older version) on dry-runs — which is why dry-run never caught the bug. Both `npm version` calls in that step now pass `--allow-same-version`, matching the existing pattern in `release.yml:326`. (#2976)
 - **`gsd-sdk query agent-skills` emits raw `<agent_skills>` block instead of JSON-wrapped string** — workflows that embed via `$(gsd-sdk query agent-skills <agent>)` were receiving a JSON-quoted string literal mid-prompt (e.g. `"<agent_skills>\n…"`), silently breaking all `<agent_skills>` injection into spawned subagents. The CLI dispatcher now honors an opt-in `format: 'text'` field on `QueryResult` and writes such results raw via `process.stdout.write`; `--pick` always returns JSON regardless. (#2917)
 - **`sketch --wrap-up` now dispatches correctly** — `/gsd-sketch --wrap-up` was silently no-oping because the flag dispatch wiring was omitted when the micro-skill entry point was absorbed in #2790. (#2949)