orchestrate: gh pr merge --delete-branch leaks the remote slice branch on every slice (recurrence of closed #284, prose still false) #384

Found during orchestrate run prd352-20260803-015333 (PRD #352, 12 slices, all passed).

Symptom

On slice #353, gh pr merge 366 --squash --delete-branch exited 1 with:

failed to delete local branch orchestrate/slice-353: failed to run git: error: cannot delete
branch 'orchestrate/slice-353' used by worktree at '.../slice-353'

The squash-merge itself landed correctly (PR MERGED, umbrella tip moved to the squash commit). But git ls-remote --heads origin orchestrate/slice-353 still returned the branch after the merge.

The real defect

references/slice-pipeline.md step 8 says the local-branch failure is "tolerated, not a slice failure", and step 9 says "the remote half was done by --delete-branch at step 8, the local half here." That second claim is false in the worktree case: gh aborts on the local-delete error before deleting the remote ref.

Since the plugin creates a worktree for every slice and holds it until step 9, this leaks one remote branch per slice, on every run — the exact condition the "incremental reclamation" design says it prevents. It also silently breaks any non-zero-exit-code check on the merge step: a successful merge returns exit 1.

Current-code verification (this run, not inherited)

• plugins/orchestrate/skills/orchestrate/references/slice-pipeline.md:330-335 — still asserts "The --delete-branch flag reclaims the remote slice branch as part of the merge; it may additionally warn or no-op on the local branch … that warning is tolerated, not a slice failure."
• plugins/orchestrate/skills/orchestrate/references/slice-pipeline.md:420 — still asserts "This completes incremental reclamation: the remote half was done by --delete-branch at step 8, the local half here."
• plugins/orchestrate/orchestrate-mcp/src/tools/finalize-slice.ts — finalizePostMerge performs no remote operation: it checkpoints subState: merged, removes the worktree, then calls deleteLocalBranch(branch, repoPath) and returns. There is no git push --delete anywhere in the phase. So gh pr merge --delete-branch remains the only remote-reclaim attempt.

Workaround used

Explicit git push origin --delete orchestrate/slice-<N> after finalize_slice post-merge returned ok.

Suggested fix

Either:

1. Drop --delete-branch from step 8 entirely and have finalize_slice's post-merge phase reclaim both halves — it already removes the worktree first, which is precisely the ordering that makes the local delete succeed. Dropping the flag also restores the meaning of the merge step's exit code.
2. Or keep the flag but make step 8 verify the remote ref is gone with git ls-remote and delete it explicitly when it is not.

Either way, correct the step-9 prose, which currently asserts a remote reclamation that did not happen.

Relationship to existing issues (read before deduping)

• This is a recurrence of closed orchestrate: remote slice branch survives merge when local delete fails (worktree checked out) — remote cruft accumulates every run #284, which diagnosed exactly this leak (across two repos and two PRDs) and proposed the same fixes. orchestrate: remote slice branch survives merge when local delete fails (worktree checked out) — remote cruft accumulates every run #284 was closed on 2026-06-05 with no linked fix commit, and the two prose claims it flagged are byte-unchanged today (see the line citations above), as is finalizePostMerge's local-only behavior. Filed as a fresh issue rather than a comment because a closed issue cannot track the work — the same precedent orchestrate: remote slice branch survives merge when local delete fails (worktree checked out) — remote cruft accumulates every run #284 itself used when it was filed as the field regression of closed orchestrate: cleanup enforcement — incremental slice-branch reclamation + failed-worktree retention/reclaim #232. A pointer comment has been left on orchestrate: remote slice branch survives merge when local delete fails (worktree checked out) — remote cruft accumulates every run #284.
• Not orchestrate: slice→umbrella gh pr merge blocked by Claude Code auto-mode self-approval classifier, halting an autonomous run mid-wave #350 (auto-mode self-approval classifier blocking gh pr merge): a permission-layer block, unrelated to branch reclamation.
• Not finalize_slice reports false "committed-pushed" when slice worktree is checked out on the base/umbrella branch #349 (finalize_slice false committed-pushed on a wrong-branch checkout): a different phase (commit-push) and a different failure.

Acceptance criteria — documentation in-sync (P0)

The false invariant is at least half the payload here. Any code fix must land with the corrected prose in references/slice-pipeline.md (steps 8 and 9) in the same unit of work; a code-only fix leaves the contract still wrong.

Environment

orchestrate 1.6.0 (plugins/orchestrate/.claude-plugin/plugin.json:3 on development) plus the PRD #352 umbrella branch (slice-executor delegation layer). Run prd352-20260803-015333.