heal-sandbox-test — T2 auto-repair regression fixture

NOT a Store actor. A deliberately-trivial, self-contained actor whose only job is to prove the autonomous T2 code-fix loop (departments/05_heal/skills/agy_fix.py) works end-to-end on real Apify infra. Bond made it a PERMANENT fixture (2026-06-19) so the loop can be re-proven anytime the heal system changes.

It parses a tiny embedded HTML catalog (3 <div class="product"> cards) with parsel and pushes one row per product — no network, no _shared, no anti-bot → fully deterministic, so a deliberate break is unambiguous and the fix is verifiable on a cloud smoke run.

Re-prove the T2 loop (≈3 min, our lane)

1. Break it — change the selector in src/main.py: for node in sel.css("div.product") → sel.css("div.item") (now finds 0 products).
2. Make the repair task (data/repair_queue/heal-sandbox-test.task.json) with actor, bare_name, repo_dir, optional last_run_id/error_hint — or let autorepair.detect() enqueue it.
3. Run T2 live:

   python3 departments/05_heal/skills/agy_fix.py --task data/repair_queue/heal-sandbox-test.task.json --live
   

4. Expect: agy GENERATES the corrected file → _apply_blocks writes it → forbidden-diff guard PASS → pytest → adversarial review APPROVE → deploy → cloud smoke PASS (3 records) → RESOLVED_T2, src/main.py healed back to div.product. The record lands in data/insights/incidents.jsonl.

A green run proves: agy-as-generator returns (no headless hang), the money-line guard only allows src/, the deploy+rollback chain works, and the cloud smoke gate is honest. First proven 2026-06-19 (flushed 4 production-fatal bugs in the T2 fixer — see memory agy-fix-generation-mode).