fix(finish): report MERGED as success instead of failing on branch cleanup

[NagyVikt]

Summary

• fix(finish): report MERGED as success instead of failing on branch cleanup

Test plan

• verified locally

[NagyVikt]

Decision: APPROVE — 0 CRITICAL, 0 HIGH, 0 MEDIUM, 2 LOW.

Caveat stated up front: this is a self-review by the change's author, not an independent one. The repo's automated review gate (gx review / the removed cr.yml) no longer dispatches, so it is the strongest review available here.

Makes a landed merge report as success. The root cause is that git branch -d requires an ancestor link to HEAD or full upstream coverage, and a squash merge — this flow's default via gh pr merge --squash — leaves neither once the remote branch is dropped. That refusal was fatal while every other post-merge cleanup step only warns.

Verified the premise rather than assuming it: in an isolated repo, a squash merge leaves f.txt byte-identical in main, and git branch -d still errors "not fully merged". So this fired on ordinary squash merges, not just the rebase case seen on #695.

The force-delete is gated on GitHub's merged-PR record for that exact head SHA, not on a patch-id guess. That distinction matters: git cherry reports + (no upstream equivalent) for every commit of a squash-merged multi-commit branch, so patch-id matching would have been wrong here.

Behavior tested against the real extracted function, both directions:

Case	Result
Squash-merged, GitHub confirms head landed	force-deleted, returns 0
Head in no merged PR (rebase / post-merge commits)	branch kept, returns 1, caller warns and exits 0

Caught and fixed one self-inflicted regression mid-review: collapsing the two closing summaries lost the distinct branch/remote vs branch/worktree wordings that tests match on — 8 failures, now restored verbatim and re-verified.

LOW-1: read_merged_pr_for_head mutates PR_STATE / PR_MERGED_AT / pr_url as a side effect. Harmless today — the only consumer is the banner, which already printed by then (confirmed: no pr_url reads after line 1533) — but a latent trap if someone later reads those globals during cleanup.

LOW-2: when a branch is kept, the prune drops --delete-branches for all dirty lanes that run, not just the protected one. Conservative in the safe direction (branches survive rather than getting -D'd), and gx cleanup sweeps the rest.

Tests: 845 total, 41 failing — failing set md5-identical to main (f003cdc8…), zero new. New regression test covers merge-lands-but-delete-refused. openspec validate --specs 133/133; change validates --strict. Symlink parity OK.