feat(replay): add deterministic counterfactual core

[PhilipJohnBasile]

Splits the deterministic counterfactual replay core out of #336.

This PR contains only the offline replay system:

• ordered, bounded candidate execution
• private execution deduplication and redacted public fingerprints
• mutation-isolated candidate and evaluator inputs
• candidate and evaluator timeouts
• deterministic regression policy decisions
• content-redacted report serialization
• a standalone no-model measurement

It does not change live serving, request capture, replay orchestration, model selection, promotion state, dashboard code, or workflows.

Measurement on commit f96817c, macOS 27.0 arm64, Python 3.14.6, 10 repeats with 128 cases and 4 evaluators:

Metric	Result
Mean suite time	26.916 ms
Suite p50	19.964 ms
Suite p95	59.264 ms
Mean throughput	4,755.6 cases/s
Candidate executions	64
Deduplicated cases	64
Unique report digests	1
5 ms timeout caller wall	7.112 ms

Reproduce:

python scripts/bench_deterministic_replay.py \
  --cases 128 \
  --evaluators 4 \
  --repeats 10 \
  --output /tmp/deterministic-replay.json

Validation:

python -m pytest tests/test_deterministic_replay.py tests/test_trace_parity.py
17 passed in 0.19s

python -m compileall -q mtplx/deterministic_replay.py tests/test_deterministic_replay.py scripts/bench_deterministic_replay.py
PASS

ruff check mtplx/deterministic_replay.py tests/test_deterministic_replay.py scripts/bench_deterministic_replay.py
PASS

git diff --check
PASS

This measurement covers offline control-plane cost and determinism. It makes no live throughput or quality claim.