fix(loop): wire the resume verdict dispatch — no more silent stall-resume no-ops (issue #154)

[robercano-ghbot]

What

Wires the action=resume issue=N verdict through the loop driver so bounded stall-resume actually spawns a driver instead of silently no-oping.

Issue #98 (PR #146) added stall detection and a bounded action=resume issue=N branch=<name> verdict to loop-tick.sh, but loop-daemon.sh's run_once dispatch only handled advance/feedback/comment-fix/ci-fix/rebase — a resume verdict fell through to the unrecognized-action branch (no driver, no ledger line), and loop-event.sh had no RESUME prompt variant (it hit the fallthrough and exit 1). Result: every overnight resume-attempt (issues #96, #106, #124, 2026-07-16 → 17) produced no progress; the resume half of #98 had effectively never run.

Fix

• loop-daemon.sh
  • driver_unit_name(): new resume issue=N arm → distinct transient unit pr-loop-driver-resume-issue<N> (NOT the advance pr-loop-driver-issue<N> shape — a distinct name keeps a resume driver from colliding with / being reattached as an advance driver).
  • verdict_from_unit_name(): exact inverse mapping (round-trips; provably no prefix collision with the advance shape).
  • run_once dispatch: resume issue=* now routes to a real driver spawn; the optional branch=... suffix is stripped before the bare resume issue=N reaches run_driver / unit-naming / the ledger.
  • Debris-check exclusion parity (the load-bearing safety property): verify_and_classify_post_exit() and reattach_orphaned_drivers() remain advance issue=*-only, so a resume verdict is never run through the destructive branch-delete debris classifier — a resume driver continues an existing feat/issue-N-* branch with WIP, which must never be deleted/misclassified. Documented with new comments so it isn't "fixed" away later.
  • Updated the ledger-header / unit-naming / reattach doc + log comment blocks to include the resume shape.
• loop-event.sh
  • Parses action=resume issue=N with an optional branch=<name> suffix (both forms), threading the branch name into the prompt.
  • New RESUME prompt variant: CONTINUE the existing feat/issue-N-* worktree/branch (never recreate / re-scope), re-run gates, route through reviewer lenses, and publish (push + open/refresh the bot PR).

Tests

• loop-daemon.test.sh: resume driver spawns (not the no-op fallthrough); branch-suffix stripping; resume branch never reaches the debris classifier (fixture deliberately omits the bot-gh.sh stub, so the query would fail loudly if it ran); pr-loop-driver-resume-issue<N> unit naming; driver_unit_name↔verdict_from_unit_name round-trip; verify_and_classify_post_exit pass-through for resume.
• loop-event.test.sh: RESUME prompt contract with and without a branch name (both parse paths), asserting resume-specific semantics (continue existing branch / never recreate / re-run gates / publish).
• loop-tick.sh unchanged. Both reviewer lenses (correctness, tests) confirmed the new tests genuinely FAIL against pre-fix code and PASS after.

Gates green: build, lint, test (loop-daemon.test.sh 113 checks, loop-event.test.sh 82 checks, plus checks.sh test + smoke-fanout.sh).

Follow-ups (non-blocking, from review)

• A dedicated reattach_orphaned_drivers startup scenario for an already-active pr-loop-driver-resume-issue<N> unit (composing pieces are unit-tested; low risk).
• An explicit "this branch pre-exists — never delete it" carve-out in the shared prompt cleanup clause for the resume path (currently self-protecting via the "you created" qualifier).

Closes #154.

🤖 Generated with Claude Code

[robercano-ghbot]

fix(loop): wire the resume verdict dispatch — no more silent stall-resume no-ops (issue #154) (not yet reviewed)