Recorder: checkpointed finalization and recovery for interrupted long-running post-processing

[mark-e-deyoung]

Summary

Implement checkpointed/fault-tolerant recorder finalization so long-running post-processing can resume after interruption.

Problem

Recorder finalization currently runs as a one-shot chain (events load/adjust, subtitles generation, concat, mux). If interrupted, reruns may repeat expensive work and can leave partially generated artifacts.

Scope

• Add explicit phase checkpoints for finalization stages.
• Persist checkpoint state in session metadata.
• Add recovery command/API path that resumes from last completed phase.
• Ensure idempotent behavior for repeated recovery attempts.
• Emit operation progress/heartbeat per phase.

Acceptance Criteria

• Interruption during finalization can be resumed without restarting from scratch.
• Recovery operation is idempotent.
• Progress is visible through operation status endpoints.
• Tests cover interruption/recovery path and phase idempotency.

Notes

Deferred from temporal-correctness Wave 1 implementation.

[mark-e-deyoung]

Release triage classification (2026-02-25): Release Blocker for downstream consumers.

Reason:

• Interrupted recorder finalization can break artifact availability/consistency expected by dependent projects.

Recommendation:

• Prioritize before release cut.

[mark-e-deyoung]

Release-gate disposition update (2026-03-01): this item remains valuable but is deferred as non-blocking for the immediate cut because current recorder finalization behavior is stable under existing CI/e2e coverage. We should schedule it as the next hardening increment before broadening long-run/post-processing workloads.