fix(tradein/estimator): scope year_built NULL cohort-bypass away from avito (#2013)

Cohort/era WHERE clause carried `OR year_built IS NULL`, meant to let avito's
frequently-missing year_built field through the filter. In practice this let
ANY-era avito listing bypass the cohort filter entirely once cohort_year_min/
max were set — pollution bug: khrushchevka target could pull in a modern avito
comp with no year parsed, and vice versa.

Narrow the escape to `OR (year_built IS NULL AND source <> 'avito')` in both
places that must stay in sync:
- _COMMON_WHERE (Tier S/H, ~estimator.py:4210)
- Tier W inline SQL copy (~estimator.py:4640)

Non-avito NULL-year rows (cian/domklik/etc without a parsed year) still pass
through untouched; avito NULL-year rows now must satisfy the same
year_built BETWEEN cohort_year_min AND cohort_year_max test as everything
else.

Added tests/test_estimator_avito_cohort_null_2013.py: SQL-source assertions
that both cohort blocks carry the new guard, stay in sync (exactly 2
occurrences), and that the old unscoped escape is gone. A full DB-backed
integration test was impractical — this suite's estimator tests run against
a dummy DATABASE_URL with no live DB.

tradein-mvp/backend/app/services/estimator.py

@@ -4217,9 +4217,13 @@ _COMMON_WHERE = """
     -- Tier W имеет свою копию этого блока в inline SQL. Если cohort_year_min IS NULL —
     -- фильтр прозрачен. CAST обязателен — psycopg3 prepared statement не выводит
     -- тип $N при IS NULL в predicate (см. PR #518 fix).
+    -- #2013: year_built IS NULL escape изначально был для avito (missing-year
+    -- listings), но пропускал ЛЮБОЙ NULL-year comp мимо cohort/era фильтра —
+    -- пуллюция any-era avito в comp-пул. Сужаем escape: только не-avito NULL
+    -- проходит прозрачно; avito NULL-year обязан пройти year_built BETWEEN.
     AND (
         CAST(:cohort_year_min AS integer) IS NULL
-        OR year_built IS NULL
+        OR (year_built IS NULL AND source <> 'avito')
         OR year_built BETWEEN CAST(:cohort_year_min AS integer)
                           AND CAST(:cohort_year_max AS integer)
     )
@@ -4637,9 +4641,11 @@ def _fetch_analogs(
                   -- (хрущёвка vs новостройка). Если cohort_year_min IS NULL —
                   -- фильтр прозрачен. CAST обязателен — psycopg3 prepared statement
                   -- не выводит тип $N при IS NULL в predicate (см. PR #518 fix).
+                  -- #2013: сужаем year_built IS NULL escape — не должен пропускать
+                  -- avito any-era comp мимо cohort фильтра (sync с _COMMON_WHERE).
                   AND (
                       CAST(:cohort_year_min AS integer) IS NULL
-                      OR year_built IS NULL
+                      OR (year_built IS NULL AND source <> 'avito')
                       OR year_built BETWEEN CAST(:cohort_year_min AS integer)
                                         AND CAST(:cohort_year_max AS integer)
                   )

tradein-mvp/backend/tests/test_estimator_avito_cohort_null_2013.py

@@ -0,0 +1,89 @@
+"""Regression test for #2013 — avito NULL-year listings bypassing cohort filter.
+
+Bug: the cohort/era WHERE clause carried an `OR year_built IS NULL` escape
+meant for avito's frequently-missing year_built field. In practice this let
+ANY-era avito listing (not just genuinely unknown-year ones) into the comp
+pool once cohort_year_min/max were set — pollution bug (#2013).
+
+Fix: narrow the escape to `OR (year_built IS NULL AND source <> 'avito')` so
+avito NULL-year rows are held to the same year_built BETWEEN test as every
+other source, while non-avito NULL-year rows (e.g. cian/domklik legacy data
+without a parsed year) still pass through untouched.
+
+The clause exists in TWO places that must stay byte-identical in intent:
+  - `_COMMON_WHERE` (Tier S / Tier H, shared string constant)
+  - the Tier W inline SQL copy (own CTE, per code comments "must stay in sync")
+
+Wiring a real Postgres-backed integration test is impractical here (this
+suite's estimator tests run against a dummy DATABASE_URL with no live DB —
+see tests/test_estimator_pure_units.py / test_estimator_cohort.py). Instead
+we assert directly on the SQL source: both cohort blocks must contain the new
+guard, and the old unscoped `OR year_built IS NULL` escape must be gone from
+both (guarding against a partial fix that only patches one copy).
+"""
+
+import os
+
+# Settings requires DATABASE_URL at init time. Set dummy DSN before any app import.
+os.environ.setdefault("DATABASE_URL", "postgresql+psycopg://test:test@localhost:5432/test")
+
+from pathlib import Path
+
+from app.services import estimator
+
+ESTIMATOR_SRC = Path(estimator.__file__).read_text(encoding="utf-8")
+
+GUARD = "OR (year_built IS NULL AND source <> 'avito')"
+
+
+def _cohort_block(anchor: str) -> str:
+    """Extract the `AND ( ... )` cohort-filter block that follows `anchor`.
+
+    Balances parens from the first `AND (` after `anchor` so we grab exactly
+    the cohort predicate group, not neighbouring clauses.
+    """
+    idx = ESTIMATOR_SRC.index(anchor)
+    tail = ESTIMATOR_SRC[idx:]
+    start = tail.index("AND (")
+    open_idx = tail.index("(", start)
+    depth = 0
+    for i, ch in enumerate(tail[open_idx:], start=open_idx):
+        if ch == "(":
+            depth += 1
+        elif ch == ")":
+            depth -= 1
+            if depth == 0:
+                return tail[start : i + 1]
+    raise AssertionError(f"unbalanced parens extracting cohort block after {anchor!r}")
+
+
+def test_common_where_cohort_block_has_avito_guard() -> None:
+    """Tier S/H (_COMMON_WHERE): NULL-year avito rows must not bypass cohort."""
+    block = _cohort_block("_COMMON_WHERE = ")
+    assert "cohort_year_min" in block
+    assert "cohort_year_max" in block
+    assert GUARD in block
+
+
+def test_tier_w_inline_cohort_block_has_avito_guard() -> None:
+    """Tier W inline copy: same guard, kept in sync with _COMMON_WHERE."""
+    block = _cohort_block("-- Когортный фильтр (PR 9)")
+    assert "cohort_year_min" in block
+    assert "cohort_year_max" in block
+    assert GUARD in block
+
+
+def test_avito_guard_present_in_exactly_both_copies() -> None:
+    """Exactly two occurrences: _COMMON_WHERE + Tier W inline, kept in sync."""
+    assert ESTIMATOR_SRC.count(GUARD) == 2
+
+
+def test_old_unscoped_null_escape_is_fully_removed() -> None:
+    """No leftover `OR year_built IS NULL` (unscoped) predicate anywhere.
+
+    Catches a partial fix that patches only one of the two WHERE copies.
+    The one remaining textual mention of the bare phrase is a prose comment
+    (ends in a period, not immediately followed by a newline like the SQL
+    predicate form did), so it does not match this pattern.
+    """
+    assert "OR year_built IS NULL\n" not in ESTIMATOR_SRC