Perf-батч #2 заменил _L1_OPEN_SQL view→inline DISTINCT ON. Тест из #1964
(test_sql_reads_physflat_view) ассертил наличие view + отсутствие FROM
objective_lots — теперь инвертирован под inline. Новый инвариант сохраняет
анти-регресс #1964 (требует physflat-дедуп через DISTINCT ON, иначе сырой
objective_lots раздул бы units_available ~2.91×) и запрещает view (full-scan).
#1964 repointed landing/supply_layers/special_indices to v_objective_lots_latest,
whose DISTINCT ON is an optimization barrier — outer WHERE can't push below it, so
the planner full Parallel-Seq-Scans 1.76M rows + external merge sorts (~6-8s) per call.
- landing._query_stats: add 1h in-process TTL cache (homepage /landing/stats was
uncached; KPI3 full-view-aggregate measured 6.7s/load). Threadpool-aware lock,
DB query outside lock, only successful result cached. Mirrors weather_cache.
- supply_layers._L1_OPEN_SQL: view->inline DISTINCT ON, stable filters
(premise_kind/district) pushed below dedup -> Bitmap on district_class_idx.
district-scoped 5.9s->0.45s (~13x).
- special_indices._ARTIFICIAL_DEMAND_SQL: view->inline 2-layer CTE; stable filters
before dedup, changing filters (is_sold/contract_date/class) AFTER dedup.
class has 248 transient physflats -> before-dedup drifted +193 on standard.
district-scoped 6.7s->0.57s (~12x), EKB-wide 8.1s->5.5s.
Result bit-identical to view, verified on prod across all classes/districts.
Uses existing covering / district_class indexes (no migration).
