Independent review found two MAJOR defects in POST /api/v1/trade-in/coverage:
MAJOR-1: the probe cohort WHERE clause was missing three predicates present
in estimator._COMMON_WHERE / Tier W (novostroyki guard, geo_precision !=
'city', price_rub > 0) — the free probe could answer "ok" at points where
the paid estimator's own 1000m radius tier sees zero real analogs. Prod
example: 56.868904/60.837955, 2 rooms, 50 m2 gave n_listings=22/status=ok
while the estimator's cohort at the same radius was 0 (all 54 rows were
novostroyki). Added the three predicates verbatim from estimator.py, plus
both a static SQL-text regression test and a real-Postgres integration test
(skip_allowlist.txt, same _live_session() pattern as test_gar_flats_loader)
that inserts novostroyka/geo_precision=city/price=0 rows and asserts they
are not counted.
MAJOR-2: median_listing_age_days was computed from days_on_market, which on
prod is populated almost exclusively by one source (yandex) — thin cohorts
produced a "median" over 1-2 listings. Added n_with_age to the response
(honest count of listings the median is based on); median is now null below
COVERAGE_MIN_AGE_SAMPLES=5, and values above COVERAGE_MAX_AGE_DAYS=365 (near
-certainly dead listings, per prod: 15% of fresh yandex rows exceed 365d,
max 4261d) are excluded as outliers before the percentile is computed.
MINOR: city_hint was trusted at face value and echoed back verbatim — a
client could pass city_hint="Екатеринбург" with coordinates in Серов and get
threshold=8/status=ok. _resolve_coverage_city now prioritizes the SQL
cohort's mode city (ground truth) over the client hint, falling back to hint
only when the cohort is empty (where status is forced not_covered anyway).
Unmatched cities no longer echo the raw client string in the city field.
3e9af2fdef