Extend the read-only harness with an in-memory per-rooms asking→sold ratio
(median sold / median pred-ask, global fallback for buckets < MIN_BUCKET=20)
and a CORRECTED metrics block re-scored via the existing _compute_metrics.
--holdout-split fits on even-id / evaluates on odd-id deals (deterministic)
for an honest out-of-sample number.
Prod A/B (sample 200, holdout): overall bias +20.4%→-4.0%, MAPE 25.4%→20.2%
— a per-rooms ratio removes the systematic asking→sold bias. Still SELECT-only,
estimator untouched. +14 tests (33 total).
