Structured repair coverage was too low for data-derived repair coefficients
(#7 v2): avito ~1% / cian ~5% / yandex 0%. When the structured field is absent
we now extract repair state from the listing description via Russian-phrase
regexes (евроремонт / без отделки / черновая / дизайнерский / косметический,
...), centralized in repair_state_normalizer.infer_repair_state_from_text().
Wired as a fallback into avito_detail, cian SERP + cian_detail, and both yandex
SERP and detail parsers — the structured field always wins; only NULLs fall back
to inference. Patterns are checked strongest-first (excellent > good >
needs_repair > standard) so mixed descriptions resolve to the latest state.
Migration 075 backfills existing rows with the same patterns; unknown stays NULL
(no fabrication).
