Document category evidence-sharing experiment - #52

What changes?

Adds a reproducible research-only experiment testing whether category evidence sharing suppresses useful men's/women's differences. It sweeps 81 level/slope sharing combinations and a common-curve control, selects candidates on validation swimmers, and checks the frozen shortlist against the main, later-season and regional holdouts.

The report and aggregate receipt support retaining the current settings: removing sharing raises average error on all three holdouts, while the validation winner offers no consistent improvement. Production calculations and the shipped model remain unchanged; swimmer-level pairs and the fitted-object cache remain local and ignored.

Validation

• 28 relevant research tests passed (23 evaluator tests and 5 experiment tests).
• Reconstructed baseline cells exactly match the shipped empirical snapshot; vectorized baseline predictions match scalar predictions on every validation and main-test row.
• Input hash matches the pinned evaluation receipt; report table matches the aggregate results; git diff --check passed.
• Application/native suites were not run because this adds standalone offline research without changing consumer or engine behavior.

Review checklist

• The change is focused and relevant documentation is updated.
• Validation appropriate to the change is described above.
• Applicable accessibility, offline, and privacy behavior has been checked.
• Applicable data attribution, metadata, coverage, and rounding behavior are documented.