The normalization used an unreachable all-POI-at-0m maximum, compressing real
scores into ~0.975-0.98 so every address read as "location reduces value".
Normalize against a realistic reference distance instead; strong central ~ 1.0,
poor locations stay below. Informational only (estimator does not use it).
Kept the uncalibrated-heuristic caveat. LOW audit R2 (#7a).