chore: add derivative measurement analysis tools

[binaryaaron]

Summary

• Restores the derivative measurement tools on top of feat: anonymizer measurement instrumentation and benchmark tooling #177 so the base PR stays focused on capture/export/basic analysis.
• Adds DataDesigner trace analysis (analyze_dd_traces.py) and staged detection output analysis (analyze_staged_detection_output.py).
• Adds benchmark strategy comparison and screening tools (compare_strategy_pairs.py, screen_strategy_comparisons.py) plus signature-delta extraction.
• Adds benchmark-only direct and staged detection probes for local performance experiments.
• Adds benchmark-only detection strategy plumbing in run_benchmarks.py, including parser compatibility and native runtime configuration for direct OpenAI-compatible endpoint experiments.
• Expands the measurement tool README with usage patterns for the derivative tools.

These tools are experiment infrastructure. They are not public Anonymizer defaults, and candidate strategies still need workload-specific safety, leakage, provenance, parser, and reliability checks before promotion.

Stack

• Base: feat: anonymizer measurement instrumentation and benchmark tooling #177 (binaryaaron/perf-epic)
• This PR: derivative measurement probes, trace analysis, strategy comparisons, and signature-delta tooling

Validation

• uv run --frozen pytest tests/test_measurement.py tests/engine/test_ndd_adapter.py tests/tools/test_measurement_tools.py tests/tools/test_benchmark_output_analysis.py tests/tools/test_detection_artifact_analysis.py tests/tools/test_compare_strategy_pairs.py tests/tools/test_dd_parser_compat.py tests/tools/test_dd_trace_analysis.py tests/tools/test_detection_strategies.py tests/tools/test_direct_detection_probe.py tests/tools/test_extract_signature_deltas.py tests/tools/test_screen_strategy_comparisons.py tests/tools/test_staged_detection_output_analysis.py tests/tools/test_staged_detection_probe.py -q
  • Result: 188 passed, 9 existing DataDesigner model-config deprecation warnings
• uv run --frozen ruff check tools/measurement/run_benchmarks.py tools/measurement/analyze_dd_traces.py tools/measurement/analyze_staged_detection_output.py tools/measurement/compare_strategy_pairs.py tools/measurement/dd_parser_compat.py tools/measurement/detection_strategies.py tools/measurement/direct_detection_probe.py tools/measurement/extract_signature_deltas.py tools/measurement/screen_strategy_comparisons.py tools/measurement/staged_detection_probe.py tests/tools/test_compare_strategy_pairs.py tests/tools/test_dd_parser_compat.py tests/tools/test_dd_trace_analysis.py tests/tools/test_detection_strategies.py tests/tools/test_direct_detection_probe.py tests/tools/test_extract_signature_deltas.py tests/tools/test_measurement_tools.py tests/tools/test_screen_strategy_comparisons.py tests/tools/test_staged_detection_output_analysis.py tests/tools/test_staged_detection_probe.py
• uv run tools/codestyle/format.sh --check
• git diff --cached --check
• CLI smoke:
  • uv run python tools/measurement/analyze_dd_traces.py --help
  • uv run python tools/measurement/analyze_staged_detection_output.py --help
  • uv run python tools/measurement/compare_strategy_pairs.py --help
  • uv run python tools/measurement/screen_strategy_comparisons.py --help
  • uv run python tools/measurement/direct_detection_probe.py --help
  • uv run python tools/measurement/staged_detection_probe.py --help
  • uv run python tools/measurement/extract_signature_deltas.py --help
  • uv run python tools/measurement/run_benchmarks.py --help

Notes

• Probe outputs and raw trace sidecars can contain prompts, model outputs, secrets, or PII. Treat them as sensitive local debugging artifacts.
• Runtime endpoint/model names are intentionally captured through config/env metadata for experiment reproducibility, but raw endpoint URLs should not become portable benchmark fixtures.