[bradfordlynch]

The Official Rules incorporate src/examples/scoring.py as the controlling scoring methodology, so it must carry the full contest calculation, not just the percentile transform.

• Verbatim float-for-float ports of the production OLS: OLSFit, ols_fit (univariate surprise-only benchmark), ols_fit2 (two-regressor model via the centered 2×2 normal equations). Same guards as production: None on too-few points, zero variance, or a singular system — and deliberately no univariate fallback.
• Rules §5 imputation: impute_predictions computes the submission-specific mean over its timely, valid predictions on the common Scored Events set before filling, then imputes it into every missed event. No coverage ⇒ None (mean undefined, not prize-eligible).
• score_submission: both metric families for one submission — the no-fill fit on its own predicted events and the _imputed family over the full common sample, with the imputed surprise-only benchmark fit once on the full common sample (submission-invariant). Emits n_obs (pre-imputation coverage, the tie-break), all seven metrics per family, imputed_mean, imputed_event_count.
• rank_submissions: contest ordering — metric (descending; ascending for the MSE family), then higher pre-imputation coverage, then submission id (display-only; residual exact ties share the combined prize per the Rules — no random tie-break). Undefined metric ⇒ unranked, dropped.
• 20 new tests: exact-value OLS recovery, degenerate-fit guards, mean-before-fill arithmetic, full-coverage family equality, benchmark invariance across submissions, tie-break ordering.

Note for merging: §5 pins a frozen version of this file — the post-merge commit hash on main is what should be stamped into the Rules doc before the Aug 9 registration deadline.