[Performance #727] Add local-only, parity-gated per-workload scorecards

[godofecht]

Scope

Partial implementation of #727, not a closeout of the epic. Replaces the empty tools/perf_scorecard/main.flow stub with a classifier for existing local bench_harness JSON results.

It enforces verified semantic parity, timing scope, minimum repetition count and dispersion bounds before scoring. Speed wins need a >=1.05x advantage; equal-performance memory wins use a >=5% lower RSS threshold. Flow/native tax is separately retained rather than averaged away. Missing CPython baselines, noisy data, unresolved parity and empty reports never become benchmark wins. A strict mode returns nonzero for any loss, tax or unscored row.

Includes synthetic local tool tests and documentation: tests/scripts/perf_scorecard.flow, docs/project/performance-scorecards.md.

Validation

Static structure inspection: delimiters and string literal pairing passed for both Flow files. Tool execution and end-to-end compiler tests not run (local checkout/toolchain unavailable in this session). This PR is intentionally draft until local gates run.

Run locally:

./flow tool tests/scripts/perf_scorecard.flow
./flow tool bench_harness --smoke --out build/perf-measurements.json
./flow tool perf_scorecard build/perf-measurements.json --strict
./flow test-scripts

As required by #727, no hosted workflows/CI were inspected or relied upon. No performance speedup is claimed. Other #727 tasks (and #736/#766/#679/#665/#666) are not implemented by this PR.