perf: restore realistic benchmark suite

[mldangelo-oai]

Summary

• restore the PR benchmark workflow and report generator
• replace helper-level scan probes with persona-oriented end-to-end workloads
• condense docs/agents/performance-audit.md into a current maintainer guide

Validation

• PROMPTFOO_DISABLE_TELEMETRY=1 uv run --with pytest-benchmark pytest tests/benchmarks/test_scan_benchmarks.py tests/benchmarks/test_picklescan_benchmarks.py --benchmark-json=/tmp/modelaudit-benchmark-head.json -q
• uv run python scripts/benchmark_report.py --current /tmp/modelaudit-benchmark-head.json --summary-file /tmp/modelaudit-benchmark-summary.md
• uv run ruff check --fix modelaudit/ packages/modelaudit-picklescan/src packages/modelaudit-picklescan/tests tests/
• uv run mypy modelaudit/ packages/modelaudit-picklescan/src packages/modelaudit-picklescan/tests tests/
• PROMPTFOO_DISABLE_TELEMETRY=1 uv run pytest -n auto -m "not slow and not integration" --maxfail=1
• npx prettier --check .github/workflows/perf.yml .github/workflows/README.md scripts/README.md docs/agents/performance-audit.md CHANGELOG.md

[github-actions]

Workflow run and artifacts

Base branch does not include the benchmark suite yet; showing current results only.

Performance Benchmarks

Captured 12 benchmark results across 10 workloads.
Aggregate median across all benchmarks: 637.55ms.

Slowest benchmarks:

• tests/benchmarks/test_scan_benchmarks.py::test_scan_release_candidate_repository at 254.56ms (mixed-model-repository, release-candidate, size=547.3 KiB, files=32)
• tests/benchmarks/test_scan_benchmarks.py::test_scan_duplicate_registry_snapshot at 196.85ms (duplicate-heavy-registry, registry-snapshot, size=915.2 KiB, files=13)
• tests/benchmarks/test_scan_benchmarks.py::test_scan_suspicious_pickle_intake at 79.77ms (suspicious-pickle-intake, suspicious-intake, size=183.8 KiB, files=4)

Workload	Benchmark	Target	Size	Files	Median	Mean	Rounds
mixed-model-repository	tests/benchmarks/test_scan_benchmarks.py::test_scan_release_candidate_repository	release-candidate	547.3 KiB	32	254.56ms	268.32ms	3
duplicate-heavy-registry	tests/benchmarks/test_scan_benchmarks.py::test_scan_duplicate_registry_snapshot	registry-snapshot	915.2 KiB	13	196.85ms	196.66ms	3
suspicious-pickle-intake	tests/benchmarks/test_scan_benchmarks.py::test_scan_suspicious_pickle_intake	suspicious-intake	183.8 KiB	4	79.77ms	93.68ms	3
single-checkpoint-preflight	tests/benchmarks/test_scan_benchmarks.py::test_scan_single_checkpoint_before_load	single_checkpoint.pkl	183.0 KiB	1	37.72ms	37.82ms	3
warm-cache-rescan	tests/benchmarks/test_scan_benchmarks.py::test_scan_warm_cached_repository_rescan	release-candidate	547.3 KiB	32	34.30ms	34.27ms	3
chunked-upload-stream	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_chunked_upload_stream	chunked_stream	278.2 KiB	1	18.13ms	18.66ms	6
clean-training-checkpoint	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_clean_training_checkpoint	safe_large	278.2 KiB	1	15.12ms	15.16ms	6
padded-multi-stream-upload	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_padded_multi_stream_upload	multi_stream_padded	4.1 KiB	1	413.1us	411.9us	6
direct-malicious-upload	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_direct_malicious_upload	malicious_reduce	52 B	1	363.1us	358.6us	6
nested-payload-review	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_nested_payload_review[nested_hex]	nested_hex	130 B	1	109.6us	109.8us	6
nested-payload-review	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_nested_payload_review[nested_raw]	nested_raw	78 B	1	106.1us	107.5us	6
nested-payload-review	tests/benchmarks/test_picklescan_benchmarks.py::test_picklescan_nested_payload_review[nested_base64]	nested_base64	98 B	1	106.0us	109.4us	6