Multi-worker supervisor inside a single invocation

[splaice]

Context

PR #58 / #60's supervised path is single-worker only — when `workers > 1` the run() function falls back to the unsupervised `ProcessPoolExecutor` path which has no hard timeout. For non-MPS workloads (embedding-only re-runs, future Docling parallelism, CPU-bound extractors) proper multi-worker supervision would unlock real parallelism.

Not urgent because MPS contention makes parallel marker workers strictly bad anyway (#64).

Proposal

Extend `_run_supervised_single_worker` into `_run_supervised_workers(n)`:

• Spawn N `_subprocess_worker` children, each with its own UUID-based `claimed_by`.
• Single supervisor loop polls `_find_stuck_extracting` per-worker and SIGKILLs only the offending child.
• Respawn killed children individually.
• Same process-group kill semantics for grandchild reaping.

`run()` dispatches to this when `workers > 1` AND `extract_timeout_s > 0`.

Acceptance criteria

• Multi-worker supervised path exists.
• N parallel children share work via SKIP LOCKED + claimed_by isolation.
• Killing one worker leaves the others running.
• Unit + integration tests for multi-worker kill scenarios.

Out of scope

This issue does NOT advocate using multiple MPS workers — see #64. The motivating use case is CPU-bound or non-MPS extractors.

🤖 Generated with Claude Code