feat: Phase 4 parallel file-disjoint workers (#37)

[caoergou]

The final phase of the #37 roadmap: run multiple specs concurrently, safely, by scheduling on file-disjointness rather than lock coordination — and only now that Phase 1 stall detection exists to bound the cost of missing progress. Builds on merged Phases 1–3.

File-disjoint scheduling (spec_queue.py)

Specs declare a ## Files scope. get_parallel_batch(specs_dir, N) greedily selects a set of ready specs whose scopes don't overlap (glob- and directory-prefix aware), up to N workers. A spec with no declared scope runs alone — correctness over parallelism. Ready specs are already dependency-independent, so scheduling only has to reason about file scope.

Shared verification gate (verification.py)

The deterministic gate — acceptance criteria + backpressure via subprocess, plus the tamper-guard hash — is extracted into one module that both the sequential engine and the parallel orchestrator call. There is exactly one definition of "verified"; the engine now delegates to it (no behavior change, existing tests green).

Parallel orchestrator (parallel.py) + owloop run --workers N

Each worker runs one target spec in its own git worktree on its own branch: implement → shared gate → commit. The orchestrator then merges each passed branch into the base branch — because the batch is file-disjoint, those merges never conflict. A round with zero progress counts toward a stall (StopReason.STALLED) so a parallel run can't spin forever. A fresh adapter is created per worker (adapters hold per-run streaming state). --workers 1 (default) leaves the sequential engine untouched.

Docs & specs

Spec template + generator now teach the ## Files scope and disjoint decomposition; CLAUDE.md documents the file-disjoint principle and the shared gate.

Tests & checks

• tests/test_parallel.py — real worktrees + merges, overlap→sequential, stall, preflight.
• file-disjoint scheduling tests in tests/test_spec_queue.py.
• Full suite 307 passed; ruff + mypy clean.

This closes out the last phase of #37. Parallelism deliberately relies on decomposition (the ## Files contract), never lock coordination.