Prefer a MAIN-bearing prediction template so EPC-less dwellings predict 🟩

A predicted EPC is seeded by deep-copying one representative neighbour's
structure. _template chose the member whose floor area was closest to the
cohort median, ignoring building-part labels. When that member's only part
was lodged with a null identifier (mapped to OTHER), the prediction had no
MAIN part and the modelling_e2e handler rejected it as "not predictable" —
discarding an otherwise-rich same-type cohort.

Restrict the template to MAIN-bearing members (median still over the whole
cohort); fall back to closest-on-size only when none are MAIN-bearing, so an
all-unlabelled cohort is left for the handler's MAIN-part guard to reject
rather than silently relabelling real data.

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

domain/epc_prediction/epc_prediction.py

@@ -18,6 +18,7 @@ from datetime import date
 from typing import Callable, Iterable, Optional, Union
 
 from datatypes.epc.domain.epc_property_data import (
+    BuildingPartIdentifier,
     EpcPropertyData,
     MainHeatingDetail,
     SapBuildingPart,
@@ -157,13 +158,27 @@ class EpcPrediction:
         the member whose floor area is closest to the cohort median. A single
         neighbour's geometry is copied wholesale, so a size-representative
         template keeps the prediction off the cohort's size outliers (ADR-0029
-        decision 4: closest on size)."""
+        decision 4: closest on size).
+
+        The template must also present a MAIN building part: its structure is
+        copied wholesale, so seeding from a member whose only part is OTHER (the
+        gov API lodged its identifier as null) gives a prediction with no main
+        dwelling — which the modelling handler then rejects as not-predictable,
+        discarding an otherwise-rich cohort. Candidates are therefore restricted
+        to the MAIN-bearing members; the median is still taken over the whole
+        cohort (the size centre is a property of the cohort, not the template
+        pool). When no member is MAIN-bearing the whole same-type cohort is
+        unlabelled, so the closest-on-size fallback is left for that guard to
+        reject rather than silently relabelling real data."""
         members: tuple[ComparableProperty, ...] = comparables.members
         median_area: float = statistics.median(
             c.epc.total_floor_area_m2 for c in members
         )
+        candidates: list[ComparableProperty] = [
+            c for c in members if _has_main_part(c)
+        ] or list(members)
         return min(
-            members,
+            candidates,
             key=lambda c: abs(c.epc.total_floor_area_m2 - median_area),
         )
 
@@ -303,6 +318,15 @@ def _main_floor_attr(comparable: ComparableProperty, attr: str) -> Optional[int]
     return value
 
 
+def _has_main_part(comparable: ComparableProperty) -> bool:
+    """Whether a comparable carries a MAIN building part — i.e. it can seed a
+    prediction that presents a main dwelling (see `EpcPrediction._template`)."""
+    return any(
+        part.identifier is BuildingPartIdentifier.MAIN
+        for part in comparable.epc.sap_building_parts
+    )
+
+
 def _geo_weighted_floor_area(
     members: tuple[ComparableProperty, ...],
     target_coordinates: Optional[Coordinates],

tests/domain/epc_prediction/test_epc_prediction.py

@@ -176,6 +176,56 @@ def test_template_skips_a_main_less_member_so_the_prediction_has_a_main_part() -
     )
 
 
+def test_an_all_other_cohort_predicts_without_a_main_part() -> None:
+    # Arrange — every member is OTHER-only (the whole same-type cohort was lodged
+    # with null identifiers). There is no MAIN-bearing template to seed from, so
+    # the prediction must NOT silently relabel real data; it falls back to the
+    # size-closest member and yields a MAIN-less picture, which the modelling
+    # handler then rejects as not-predictable (the honest "fail" decision).
+    cohort = _cohort(
+        _epc(floor_area=80.0, identifier=BuildingPartIdentifier.OTHER),
+        _epc(floor_area=82.0, identifier=BuildingPartIdentifier.OTHER),
+    )
+
+    # Act
+    predicted: EpcPropertyData = EpcPrediction().predict(
+        PredictionTarget(postcode="LS6 1AA", property_type="2"), cohort
+    )
+
+    # Assert — no MAIN part is conjured; the picture stays as lodged.
+    assert not any(
+        part.identifier is BuildingPartIdentifier.MAIN
+        for part in predicted.sap_building_parts
+    )
+
+
+def test_template_is_the_size_closest_member_among_the_main_bearing_ones() -> None:
+    # Arrange — the size-median member is OTHER-only (80 m²). Of the MAIN-bearing
+    # members, the 78 m² one (distinctively a 2-part structure) is nearest the
+    # cohort median; the 30 m² one is far. The template must be the size-closest
+    # *MAIN-bearing* member, while the cohort median stays a property of the whole
+    # cohort (78 m², not the 54 m² median of just the MAIN-bearing subset).
+    cohort = _cohort(
+        _epc(floor_area=80.0, identifier=BuildingPartIdentifier.OTHER),
+        _epc(
+            floor_area=78.0,
+            identifier=BuildingPartIdentifier.MAIN,
+            building_parts=2,
+        ),
+        _epc(floor_area=30.0, identifier=BuildingPartIdentifier.MAIN),
+    )
+
+    # Act
+    predicted: EpcPropertyData = EpcPrediction().predict(
+        PredictionTarget(postcode="LS6 1AA", property_type="2"), cohort
+    )
+
+    # Assert — structure copied from the 78 m² MAIN-bearing member (its 2 parts),
+    # and the size estimate is the full-cohort median (78 m²), not the subset's.
+    assert len(predicted.sap_building_parts) == 2
+    assert predicted.total_floor_area_m2 == 78.0
+
+
 def test_sets_main_wall_construction_to_the_cohort_mode() -> None:
     # Arrange — the template (members[0]) is solid brick (2), but the cohort
     # majority is cavity (1). The homogeneous categorical should follow the mode,

tests/domain/epc_prediction/test_validation.py

@@ -8,6 +8,7 @@ from datetime import date
 from typing import Optional, Union
 
 from datatypes.epc.domain.epc_property_data import (
+    BuildingPartIdentifier,
     EpcPropertyData,
     MainHeatingDetail,
     SapBuildingPart,
@@ -39,6 +40,7 @@ def _comparable(
     epc.solar_water_heating = False
     epc.has_hot_water_cylinder = True
     part: SapBuildingPart = object.__new__(SapBuildingPart)
+    part.identifier = BuildingPartIdentifier.MAIN
     part.wall_construction = wall_construction
     part.wall_insulation_type = 1
     part.construction_age_band = "K"