feat(replay): add capture plan orchestration

[PhilipJohnBasile]

Split from #336 as one independently reviewable system.

This PR adds engine-neutral replay orchestration only:

• bounded capture discovery and deterministic seeded selection
• metadata filters, public-fingerprint deduplication, and replayability classification
• stale-plan validation against the complete current selection
• atomic, private-by-default replay receipt persistence
• no candidate execution, evaluators, request capture, promotion, or server wiring

The deterministic replay engine remains in #351. Privacy-safe request capture remains in #356. This PR imports neither implementation. A caller hands the orchestrator report and decision mappings after its replay engine finishes.

Measurement on this commit:

python scripts/bench_replay_orchestrator.py
1,000 captures, 128 selected cases
plan selection, 30 iterations: p50 50.151 ms, p95 51.384 ms, p99 51.676 ms
freshness validation, 100 iterations: p50 51.003 ms, p95 55.209 ms, p99 58.789 ms
atomic receipt write plus freshness check, 50 iterations: p50 54.083 ms, p95 58.031 ms, p99 59.897 ms
source digests across plans: 1
stale mutation rejected: yes, 57.013 ms
receipt files: 50, temporary files remaining: 0

Validation:

python -m pytest tests/test_replay_orchestrator.py -q
9 passed

ruff check mtplx/replay_orchestrator.py tests/test_replay_orchestrator.py scripts/bench_replay_orchestrator.py
All checks passed!

python -m compileall -q mtplx/replay_orchestrator.py tests/test_replay_orchestrator.py scripts/bench_replay_orchestrator.py
git diff --check

The existing macOS CI job now includes tests/test_replay_orchestrator.py.