fix(release): run tests before mutating, restore files on abort

[adnaan]

Ports the fix that landed in the core library's release.sh (livetemplate#499) to the client, where it had never been applied.

Two problems, both hit while releasing v0.19.1

1. Files were mutated before tests ran. update_versions and generate_changelog ran ahead of build_and_test, so a failing test left VERSION, package.json, package-lock.json and CHANGELOG.md already rewritten. main() refuses to start on a dirty tree — so the next attempt fails for a reason unrelated to why the first one did.

Tests and the build now run first, on the pre-bump tree.

2. No recovery for post-write failures. Reordering can't help with anything that fails after the write — verify_build, npm pack, gh release create. An EXIT trap now restores the four release-managed files when the run aborts before committing, and stands down once commit_and_tag has run.

Ordering constraint the client has and core doesn't

verify_build asserts package.json's version equals the new one, so it genuinely must run post-bump. Moving the build ahead of the bump was only safe after confirming the dist bundle embeds no version string — otherwise dist would ship stamped with the old version. Verified: no __VERSION__ substitution, no package.json import in the TS sources. Final order:

build_and_test          # tests + build, pre-bump — fails before touching anything
update_versions         # sets release_files_written
generate_changelog
verify_build            # asserts package.json == new version, so must be post-bump
verify_package_contents
commit_and_tag          # sets release_committed — trap stands down
publish_github

Two subtleties carried over from the core fix

• Restore from HEAD, not the index. commit_and_tag stages the files before committing, so if the commit itself fails a plain git checkout -- would restore them from the index — copying the modified versions back over themselves.
• One path per invocation. git checkout HEAD -- a b c resolves the whole pathspec first and bails before touching the worktree if any entry is missing from HEAD, so a single call would restore none of them when one is untracked.

Verification

Both directions exercised against a real worktree:

# pre-commit abort
--- dirty before abort: 3 files
WARN: Release aborted before committing — restoring version and changelog files
--- after trap: 1 dirty files      (only scripts/release.sh, the actual edit)

# post-commit abort (release_committed=true)
--- VERSION after: 9.9.9           (untouched, correctly — it's in history now)

The second case matters as much as the first: a trap that fired after the commit would discard the release commit's content.

bash -n clean; --dry-run reaches the same decision points as before.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ui2cwpeGkrUfRt8rh2FgGG