Resolve dominant double/triple glazing when the era is stated

Addresses PR #1402 review: the guard only claimed dominant-single, leaving
the symmetric bug open — the LLM can flatten a dominant-double/triple split
onto its minority single (e.g. "96% Double glazing 2002 or later, 4% Single").

A dominant double/triple whose era is explicit ("pre-2002" / "2002 or later")
is just as fully determined as era-free single, so the guard now claims it via
_DOMINANT_MEMBER. Only a genuinely ambiguous era ("unknown age", unstated) still
defers to the LLM — the "96% double -> None" contract now holds solely for the
era-unknown case, not the era-stated one.

Backfill script reuses the same guard, so it now corrects any dominant split;
renamed reclassify_dominant_single_glazing.py -> reclassify_dominant_glazing.py
to match. Tests cover double/triple x pre-2002/2002-or-later and the still-
deferred unknown-age case; 14 green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

domain/epc/property_overrides/glazing_mix_guard.py

@@ -5,12 +5,35 @@ from typing import Optional
 
 from domain.epc.property_overrides.glazing_type import GlazingType
 
-# "N% <base glazing type>" — the era (2002 / pre-2002) is irrelevant to whether it
-# is a mix, so only the base type is captured.
-_PERCENT_TYPE = re.compile(r"(\d+)%\s*(single|double|triple|secondary)")
+# "N% <base glazing type> [era]" — the base type decides whether it is a mix; the
+# era ("2002 or later" / "pre-2002") is captured too, so a dominant double/triple
+# can be resolved deterministically when its era is stated (see the guard below).
+# An absent or non-canonical era ("unknown age", "(SAP 9.94)") leaves the era group
+# empty, so that dominant split stays the LLM's job.
+_PERCENT_TYPE = re.compile(
+    r"(\d+)%\s*(single|double|triple|secondary)(?:\s*glazing)?"
+    r"(?:[\s,]*(pre-?2002|2002 or later))?"
+)
 # One base type at or above this share is a uniform assertion, not a mix — applied
 # as that type (a near-uniform reglaze), not deferred.
 _UNIFORM_THRESHOLD_PERCENT = 90
+# A dominant base type + stated era → the fully-determined canonical member. Single
+# glazing is era-free, so it needs no era. Double/triple with no (or non-canonical)
+# era is absent here and falls through to the LLM.
+_DOMINANT_MEMBER: dict[tuple[str, Optional[str]], GlazingType] = {
+    ("single", None): GlazingType.SINGLE,
+    ("double", "2002 or later"): GlazingType.DOUBLE_POST_2002,
+    ("double", "pre-2002"): GlazingType.DOUBLE_PRE_2002,
+    ("triple", "2002 or later"): GlazingType.TRIPLE_POST_2002,
+    ("triple", "pre-2002"): GlazingType.TRIPLE_PRE_2002,
+}
+
+
+def _canonical_era(era: str) -> Optional[str]:
+    """Normalise a matched era phrase to the key used in ``_DOMINANT_MEMBER``."""
+    if not era:
+        return None
+    return "2002 or later" if "2002 or later" in era else "pre-2002"
 
 
 def glazing_mix_guard(description: str) -> Optional[GlazingType]:
@@ -23,28 +46,36 @@ def glazing_mix_guard(description: str) -> Optional[GlazingType]:
     must resolve to ``GlazingType.MIXED`` (no overlay → the cert is kept), never to
     a single dominant type that would flatten every window (ADR-0042).
 
-    Recognises the structured ``"N% <type>, M% <type>"`` split and returns
-    ``MIXED`` when it is a genuine mix — two or more glazing types present and none
-    dominant (≥ the uniform threshold). Returns ``None`` for a uniform / near-
-    uniform assertion or an unparseable description, so the LLM classifier still
-    resolves it (a uniform reglaze is applied; varied phrasings are the LLM's job).
+    Recognises the structured ``"N% <type>, M% <type>"`` split and returns:
+
+    * ``MIXED`` for a genuine mix — two or more glazing types present and none
+      dominant (≥ the uniform threshold);
+    * the fully-determined member for a dominant near-uniform split whose type
+      carries no era ambiguity — single glazing (era-free), or a dominant
+      double/triple whose era is stated ("pre-2002" / "2002 or later"); the LLM
+      would otherwise flatten such a split onto its *minority* type;
+    * ``None`` otherwise — an unparseable description, or a dominant double/triple
+      with an unstated/non-canonical era ("unknown age") that still carries genuine
+      era ambiguity — so the LLM classifier resolves it.
     """
     matches = _PERCENT_TYPE.findall(description.lower())
-    base_types = {base_type for _, base_type in matches}
+    base_types = {base_type for _, base_type, _ in matches}
     if len(base_types) < 2:
         # One (or no) glazing type named — a uniform assertion, not a mix.
         return None
-    dominant_percent, dominant_type = max(
-        ((int(percent), base_type) for percent, base_type in matches),
-        key=lambda pair: pair[0],
+    dominant_percent, dominant_type, dominant_era = max(
+        ((int(percent), base_type, era) for percent, base_type, era in matches),
+        key=lambda triple: triple[0],
     )
     if dominant_percent >= _UNIFORM_THRESHOLD_PERCENT:
-        # One type dominates — a near-uniform assertion, not a mix. Single glazing
-        # is era-free, so a dominant-single split is fully determined: apply SINGLE
-        # deterministically (the LLM otherwise flattens it onto the minority double
-        # type). A dominant double/triple still carries era ambiguity (pre-2002 /
-        # 2002 / unknown age), so it stays the LLM's job.
-        if dominant_type == "single":
-            return GlazingType.SINGLE
-        return None
+        # One type dominates — a near-uniform assertion, not a mix. A split is fully
+        # determined (and safe to apply deterministically) when the dominant type
+        # carries no era ambiguity: single glazing is era-free, and a dominant
+        # double/triple whose era is stated ("pre-2002" / "2002 or later") is
+        # equally pinned. In those cases claim the member ourselves — the LLM
+        # otherwise flattens the split onto the *minority* type. A dominant
+        # double/triple with an unstated or non-canonical era ("unknown age") still
+        # carries genuine era ambiguity, so it stays the LLM's job (falls through to
+        # ``None``).
+        return _DOMINANT_MEMBER.get((dominant_type, _canonical_era(dominant_era)))
     return GlazingType.MIXED

scripts/lisasrequest/reclassify_dominant_glazing.py

@@ -1,13 +1,15 @@
-"""Backfill glazing overrides where a >= 90%-single dwelling was flattened onto its
-*minority* double type.
+"""Backfill glazing overrides where a dominant (>= 90%) split was flattened onto
+its *minority* type.
 
 Sibling to ``scripts/reclassify_mixed_glazing.py``. That script rescued genuine
 mixes (no type >= 90%) to ``MIXED``; it deliberately left near-uniform rows to the
-LLM classifier. But for a dominant-*single* split ("4% Double glazing 2002 or
-later, 96% Single Glazing") the LLM had latched onto the 4% minority and written a
-"Double glazing…" value. ``glazing_mix_guard`` now resolves a dominant-single
-split deterministically to ``SINGLE`` (single glazing is era-free), so this
-backfill fixes the rows written before that guard change.
+LLM classifier. But for a dominant split the LLM sometimes latched onto the
+minority type — a dominant-*single* dwelling ("4% Double glazing 2002 or later,
+96% Single Glazing") written as "Double glazing…", or the symmetric dominant-double
+case written as "Single glazing". ``glazing_mix_guard`` now resolves any
+era-unambiguous dominant split deterministically — SINGLE (era-free), or a
+double/triple whose era is stated — so this backfill fixes the rows written before
+that guard change.
 
 Uses the SAME guard as the live path, so the backfill and the classifier cannot
 drift. SCOPED TO ONE PORTFOLIO (``--portfolio``, default 796 = Hyde) and only
@@ -20,8 +22,8 @@ row changed (property_id, uprn, old value, new value) so the change is reversibl
 Idempotent — only rows whose stored value differs from the guard's target member
 are touched.
 
-    python -m scripts.lisasrequest.reclassify_dominant_single_glazing            # dry run
-    python -m scripts.lisasrequest.reclassify_dominant_single_glazing --apply    # write
+    python -m scripts.lisasrequest.reclassify_dominant_glazing            # dry run
+    python -m scripts.lisasrequest.reclassify_dominant_glazing --apply    # write
 """
 
 from __future__ import annotations
@@ -107,7 +109,7 @@ def main() -> int:
 
     if args.apply and audit:
         out = _REPO_ROOT / "scripts" / "lisasrequest" / (
-            f"reclassify_dominant_single_glazing_{args.portfolio}_audit.csv"
+            f"reclassify_dominant_glazing_{args.portfolio}_audit.csv"
         )
         with out.open("w", newline="") as fh:
             w = csv.writer(fh)

tests/domain/epc/test_glazing_mix_guard.py

@@ -38,19 +38,45 @@ def test_dominant_single_glazing_resolves_to_single(description: str) -> None:
     assert result is GlazingType.SINGLE
 
 
+@pytest.mark.parametrize(
+    ("description", "expected"),
+    [
+        # >= 90% double with a *stated* era is era-pinned — the LLM otherwise
+        # flattens it onto the 4% minority single (the symmetric bug to the
+        # dominant-single case above).
+        ("96% double glazing 2002 or later, 4% single glazing", GlazingType.DOUBLE_POST_2002),
+        ("95% double glazing pre-2002, 5% single glazing", GlazingType.DOUBLE_PRE_2002),
+        ("90% double glazing 2002 or later, 10% single glazing", GlazingType.DOUBLE_POST_2002),
+        # Triple is era-bearing too — resolve when the era is stated.
+        ("92% triple glazing pre-2002, 8% single glazing", GlazingType.TRIPLE_PRE_2002),
+        ("91% triple glazing 2002 or later, 9% single glazing", GlazingType.TRIPLE_POST_2002),
+    ],
+)
+def test_dominant_era_stated_double_or_triple_resolves_deterministically(
+    description: str, expected: GlazingType
+) -> None:
+    # Act
+    result = glazing_mix_guard(description)
+
+    # Assert — a dominant double/triple whose era is stated is fully determined, so
+    # the guard claims it rather than trusting the LLM.
+    assert result is expected
+
+
 @pytest.mark.parametrize(
     "description",
     [
         "100% double glazing 2002 or later",  # uniform — one type
-        "96% double glazing 2002 or later, 4% single glazing",  # >= 90% double -> era ambiguity
+        "90% double glazing unknown age, 10% single glazing",  # dominant but era unstated
         "80% double glazing 2002 or later, 20% double glazing pre-2002",  # same base type
         "some double glazing and a bit of single",  # unparseable — no percentages
         "",
     ],
 )
-def test_uniform_or_unparseable_glazing_defers_to_the_llm(description: str) -> None:
-    # The guard only claims a genuine percentage mix; a uniform assertion (applied
-    # as that type) or a varied phrasing is left for the LLM classifier.
+def test_uniform_or_ambiguous_glazing_defers_to_the_llm(description: str) -> None:
+    # The guard only claims a genuine mix or a fully-determined dominant split. A
+    # uniform assertion, a dominant split whose era is unstated ("unknown age" — the
+    # era is genuinely ambiguous), or a varied phrasing is left for the LLM.
 
     # Act
     result = glazing_mix_guard(description)