fix(session): retry recoverable memory extraction failures

[CryoThrust]

Summary

• allow queued session commits to retry recoverable Phase 2 memory extraction failures
• preserve terminal handling for corrupt archives, missing messages, and cancelled commits
• add regression coverage for replaying an archive after a transient extraction failure

Fixes #4419

Validation

• python -m py_compile openviking/session/session.py tests/unit/session/test_session_commit_resume.py
• git diff --check
• targeted pytest is blocked locally by a missing apscheduler dependency

[linhongyu510]

The new retry branch is not reachable through the real queue lifecycle after the original extraction failure. _run_memory_extraction() records .failed.json and returns normally; resume_queued_commit() therefore returns True, SessionCommitProcessor.on_dequeue() completes normally, and QueueManager.process_one() unconditionally calls ack(), deleting the persistent message. Nothing invokes resume_queued_commit() a second time for that archive, and a later empty user commit creates no replacement work. The new test fabricates the failed marker and directly calls resume_queued_commit(), so it proves that a hypothetical replay would work but not that any replay is scheduled. Please add a processor/queue-level regression that fails extraction once, observes the first message lifecycle, then proves the same archive is actually re-enqueued/retried and reaches .done without a process restart. The implementation likely needs an explicit recoverable retry signal/re-enqueue before ack, with bounded retry/backoff semantics, rather than only handling a marker on an otherwise unreachable second invocation.