capture: make request records privacy-safe by default

[PhilipJohnBasile]

This extracts only the request-capture privacy boundary from #336.

Request captures are still off unless MTPLX_REQUEST_CAPTURE_DIR is set. Once enabled, the default record now replaces prompt and completion token IDs with counts and stable SHA-256 digests. Prompt text, response text, messages, and exception text are also omitted by default, while nested credentials are redacted.

The public completion_token_ids(tokens) helper and its existing one-argument call remain supported. It now returns the completion count and digest by default. Raw completion IDs require their own MTPLX_REQUEST_CAPTURE_INCLUDE_COMPLETION_TOKENS=1 opt-in. That flag is separate from MTPLX_REQUEST_CAPTURE_INCLUDE_PROMPT_TOKENS=1, and both paths enforce independent limits.

Measurement

Command: python scripts/measure_request_capture.py --records 1000

Measured source: fe5b5e59a14e0066f3191951483e3a65b9c55592 on Apple arm64 with Python 3.14.6.

• Capture disabled: 0.658 ms total, 0.658 microseconds per record
• Default privacy capture: 595.733 ms total, 595.733 microseconds per record
• Throughput: 1,678.604 records/s
• Output: 1,000 files, 1,349,890 bytes total, 1,349.89 bytes per record
• Payload per record: 256 prompt tokens, 64 completion tokens, 736 prompt characters, 400 response characters

The measurement verifies all 1,000 records completed, raw prompt and completion IDs were absent, prompt/response/message/secret content was absent, and token counts plus digests were present. Validation runs outside the timed section. The committed receipt is request-capture-1000.json.

Validation

• python -m pytest tests/test_request_capture.py -q: 18 passed
• python -m compileall -q mtplx/request_capture.py scripts/measure_request_capture.py tests/test_request_capture.py: passed
• python -m ruff check mtplx/request_capture.py scripts/measure_request_capture.py tests/test_request_capture.py: passed
• git diff --check origin/main...HEAD: passed
• U+2014 and U+2013 scan across all four changed files: no matches

Exclusions

This PR does not change server call sites, runtime registries, replay systems, dashboards, API routes, workflow files, model behavior, or scheduling. The measurement covers request-capture persistence overhead only. It does not claim inference throughput or model-quality improvement.

[youssofal]

Thanks. This draws the right line: capture stays off unless MTPLX_REQUEST_CAPTURE_DIR is set, and the default record then holds counts and digests instead of the prompt and completion token IDs. It is not in 2.11.2. It is on the list for the next release under your name. The trace tooling reads the request log, not the capture files, so nothing changes there.

[PhilipJohnBasile]

@youssofal Agreed on the next-release placement. The boundary stays unchanged: capture is off unless configured, counts and digests are the enabled default, and raw prompt/completion token IDs require separate explicit opt-ins. No request-log/trace-tooling changes are added here. The PR is available for source review without expanding the scope or treating the release deferral as a defect.