[zomux]

The PSR reconstruction used for the additional-model experiments existed only in the local checkout, leaving readers without its formula or an executable implementation. Add scripts/score_psr.py and docs/benchmark/metrics.md, link the guide from the README and scoring documentation, and explicitly retain the historical-reconstruction caveat.

The script preserves fractional checkpoint credit and equal task weighting. It supports arbitrary model names and relocated result folders, checks exact task coverage, rejects duplicate/invalid rows, and invalidates cached inputs when traces change. The guide also explains SR, CCE, Rounds/Chats, verifier provenance, and split/judge comparability.

Validation:

• Eight offline PSR tests cover the worked example, clipping, booleans, round limits, weighting, complete/partial coverage, invalid inputs, relocated traces, and cache invalidation.
• Recollected all 900 locally archived experiment trajectories: checkpoint credits and outcomes match exactly; task PSR differences are at most 5.55e-17, and all 12 split summaries match within 1e-12.
• All 29 benchmark tests pass with runner/judge dependencies installed; all four scoring/runner CLI help checks pass.
• Documentation file links and git diff --check pass.

No game/model calls were made, and historical result files were not modified. This does not certify equivalence to the unavailable historical PSR program or change the newer adjacent-task SR verifier protocol.