Performance

Performance work in fincore follows a reproducible methodology (see docs/quality/performance-methodology.md for the full protocol). The short version:

1. Every workload is deterministic — a fixed seed and a SHA256 input digest so a run can prove it measured the same input across commits and platforms.
2. Baselines are platform-labelled (linux-x86_64, darwin-arm64) and never compared across unmatched platforms.
3. An optimization is accepted only when it preserves numeric and shape semantics (unchanged output digest) and improves a profiled case.

Profiling a hotspot

python scripts/profile_hotspots.py --scenario medium --output build/hotspots-before.json

This writes a JSON report (top cumulative functions, wall seconds, cold-import seconds, peak RSS) and a Markdown summary beside it. Run it again after an optimization and diff the pair.

Factor benchmarks

Factor-analysis performance is release-blocking only when a matching approved platform baseline exists. Platform baselines live in benchmarks/factor-analysis-baselines/; a pending (candidate-only) baseline is never selected, so an absent approval is explicit rather than a silent pass. See docs/quality/factor-benchmark-approval.md for the promotion protocol.