Previous confidence used raw n_analogs which inflated to 'high' when
MIN_ANALOGS_PER_SOURCE quota + same-building lots concentrated count on
few unique addresses.
New logic:
- high: unique_addr >= 7 AND iqr_pct < 15%
- medium: unique_addr >= 4 OR (unique_addr >= 2 AND iqr_pct < 25%)
- downgrade one step if avg_lots_per_addr > 2.5 (anti-concentration bias)
probe-confidence-distribution.py added for tuning p75 threshold on real data.
