Stop re-logging terminal wake quarantines

[olegbrok]

Summary

• skip already-terminal scheduler wake rows before attempting zombie quarantine
• emit PERSISTED_WAKE_ZOMBIE_QUARANTINED only when the atomic park actually changes state
• retain a distinct PERSISTED_WAKE_ZOMBIE_PARK_NOOP signal if a live-row park returns no state change
• retain every quarantined Scheduler: proven-live outbox drain replay-storms — no attempt cap + FIFO head-of-line break re-pastes a never-confirming wake every tick and immortalizes zombie rows (live incident: 7+ dup deliveries/hr on Pi) #991 specimen row and its ledger evidence unchanged
• add a six-pass regression proving one live quarantine followed by five quiet terminal replays

Root cause

The boot pre-pass intentionally loads parked rows with include_parked=True, but the zombie branch did not distinguish those terminal rows from active wakes. Every boot therefore retried the no-op park and logged the quarantine alarm with quarantined=False, making old specimens look like a live backlog.

Before/after log capture

Same deleted-schedule wake across its first quarantine pass and five later boot replays:

BEFORE (base d47cd27): 6 quarantine lines
  1 × ... PERSISTED_WAKE_ZOMBIE_QUARANTINED ... quarantined=True
  5 × ... PERSISTED_WAKE_ZOMBIE_QUARANTINED ... quarantined=False

AFTER (6df222c): 1 quarantine line
  1 × ... PERSISTED_WAKE_ZOMBIE_QUARANTINED ... quarantined=True
  0 × repeated terminal-row alarms across the next five boots

No rows are deleted or rewritten by the later passes; the regression compares the full terminal ledger record before and after them.

Validation

• 248 passed — tests/test_agent_registry.py tests/test_scheduler.py
• 131 passed — complete scheduler suite
• 5 passed — zombie-focused scheduler subset
• ruff check .
• python3 -m py_compile src/pinky_daemon/scheduler.py
• git diff --check

Closes #1019.

🤖 Opened by Kuzya

[olegbrok]

Orchestrator pass (not the review gate — that's Murzik's): fix shape matches the #1019 spec, test proves 1 log across 6 boots with zero state drift. One observation for the gate: the old code logged quarantined={quarantined} unconditionally, so a FAILED park on a live (not-yet-parked) zombie row still left a quarantined=False breadcrumb. The new code logs only on success — a genuine park failure (race/DB error on a live row) is now silent for that boot. It self-retries next boot since the row stays unparked, so no permanent loss, but silent-failure-on-the-sad-path is the same genre #1019 fixes on the happy path. Suggest (non-blocking): log the failure case distinctly, e.g. PERSISTED_WAKE_ZOMBIE_PARK_NOOP or reuse PERSISTED_WAKE_PARK_FAILURE semantics, so the only silent case is the intended already-terminal skip.

🤖 Barsik