MC-4371

ROOT CAUSE of today's (2026-05-28) ticket bounce-storm. The shadow reviewer (mc_shadow_review.py) resolves the diff from the MC worktree branch only. When it cannot see the real diff it returns verdict=fail/uncertain, and that false verdict BOUNCES orchestrator-gated tickets back open (done->todo / done->needs_input), causing pickup to re-dispatch workers onto already-complete work — a wasteful loop. False-fail classes observed today (all verified-real work): 1. SLOT-BRANCH blindness — work committed on slot2/mc-MC-4327 (MC-4350/4348/4345): diff_source=unavailable -> verdict=fail. 2. CROSS-REPO blindness — engine committed to the ~/.claude skills git, not the MC worktree (MC-4365 radio curation: afea8fc/0187d89): invisible to the MC diff check. 3. NON-CODE deliverables — research/triage tickets (MC-4358 agent-watch, MC-4362 macro research): no diff expected, but code-review criteria applied -> fail. 4. WRONG-REPO commit existence — claimed e05be7cf "not in any branch" (MC-4335) when it is on master. FIX (proposed): - Resolve the diff from the ticket's ACTUAL branch (slot{N}/mc-MC-<id>) and any referenced repo, not just the MC worktree/master. - When diff_source is genuinely unavailable, ABSTAIN (advisory comment, low confidence) — do NOT emit verdict=fail and do NOT change a status the orchestrator set. The orchestrator gate is authoritative (3-role model: controller gates). - Detect non-code/research tickets and gate on artifact existence, not code diff. - Never auto-reopen a ticket that an orchestrator/operator moved to done. Impact: eliminates the recurring re-dispatch-onto-done-work loop + the per-cycle worker-review nag for already-gated tickets.