fix: bound retry metadata growth in queue job serializer (#3947)

[neuralmint]

Summary

Fixes #3947 — Bound retry metadata growth when repeated failures are triggered while an agent run, task, or handler is changing lifecycle state.

Problem

The TaskScheduler.fail() method incremented task["retries"] without any upper bound. When the repeated-failures path was exercised, retry metadata could grow unbounded, causing the component to accept stale, duplicate, or policy-violating lifecycle transitions. Impact: jobs could be duplicated, starved, lost, or delivered under wrong conditions.

Fix

1. Hard cap on retry metadata (MAX_RETRY_METADATA = 100): prevents unbounded growth. fail() caps the increment; if already at the cap, it remains there (idempotent).
2. Dead letter store: tasks that exceed max_retries or the hard cap are moved to _dead_letter instead of being silently dropped. Logs explain the decision.
3. enqueue invariant enforcement: enqueue() now rejects tasks whose retry count already exceeds the hard cap, returning None.
4. Idempotent retry path: added preserve_retries parameter to enqueue() so that the fail-re-enqueue cycle preserves retry metadata instead of resetting it.
5. Scheduled task promotion (dequeue) also respects the hard cap.
6. Logging: all retry/rejection decisions are logged at appropriate levels (info for retries, warning for permanent failures).

Acceptance Criteria

• Deterministic regression test covers the repeated-failures trigger
• Queue job serializer rejects invalid transitions and preserves expected lifecycle state
• Logs explain decisions without exposing private runtime data
• Existing test suite passes unchanged

Proof

$ pytest tests/test_scheduler.py -v
========================== 10 passed in 0.13s ==========================

Closes #3947