Slice S0380.42: Decimal HALF_UP per-window areas per RdSAP10 §15 — closes cert 1536

Cert 1536 lodged window dimensions including (0.65 × 0.70) × 3
windows. In float arithmetic 0.65 × 0.70 = 0.45499999999999996,
which the `_round_half_up(float, dp)` helper snaps to 0.45 vs the
spec answer 0.46 (Decimal: 0.65 × 0.70 = 0.4550 exact, HALF_UP at
2 d.p. = 0.46). The shortfall of 0.01 m² × 3 windows = 0.03 m²
under-counted as ~0.073 W/K of conduction loss vs the worksheet's
windows_w_per_k = 25.6354 — closing the cert 1536 residual at
+0.00152 to <2e-6.

Same class of bug as the S0380.34/35 living-area / gross-wall /
party-wall closures (Decimal HALF_UP at the 0.005 boundary that
float drops). RdSAP10 §15 (p.66) lists "all element areas (gross)
including window areas: 2 d.p." — Decimal is the only arithmetic
that matches that boundary deterministically.

Three cascade sites now use Decimal HALF_UP for per-window areas:

- heat_transmission.py: `_decimal_round_half_up_product(W, H, 2)`
  replaces `_round_half_up(W × H, 2)` at the windows_w_per_k cascade
  AND at the per-bp window-area accumulation (the wall-net deduction
  branch must agree with the conduction branch for cascade-internal
  consistency, per the existing comment at line 575-583).

- internal_gains.py: `_decimal_window_area_2dp(W, H)` replaces the
  inline `_round_area_2dp(W × H)` in the daylight factor `g_l`
  sum so §5 (66)..(67) sees the same per-window areas as §3 (27).

- solar_gains.py: same Decimal helper replaces `_round_area_2dp` in
  `_wall_window_solar_gain_monthly_w` so §6 (74)..(81) area = (27).

The `_round_area_2dp` helpers were inlined per-module in pre-S0380.42
work; this slice deletes them since the Decimal-aware product
replaces all call sites. `_round_half_up` stays in heat_transmission
for non-product per-element area calls (single-value rounds).

Test impact:
  - Cohort-2 cert 1536 API path: +0.00152 → -1e-6 (<1e-4 ✓).
    Moves from _COHORT_2_API_OPEN to _COHORT_2_API_CLOSED. Cohort
    distribution: 37/38 exact (was 34/38 at start of session);
    only cert 2102 (-6.30 secondary-heating routing) remains open.
  - Cohort-2 cert 0300/9380 unchanged (already <1e-4 after S0380.41).
  - Cohort-1 ASHP 9/9 unchanged: <1e-4 on both paths.
  - Elmhurst 6-cert worksheet sweep: unchanged (lodges
    `window_width=area, window_height=1.0` per the Elmhurst lodging
    convention — Decimal(area) × Decimal(1.0) = Decimal(area), no
    rounding shift).

Test suite: 750 pass + 0 fail. Pyright net-zero per touched file
(heat_transmission 13/13; internal_gains 4/4 pre-existing; solar_gains
0/0; chain test 0/0).

Spec citation: RdSAP 10 Specification §15 "Rounding of data" p.66 —
"All element areas (gross) including window areas and conservatory
wall area: 2 d.p." Decimal is the float-precision-stable arithmetic
that matches this rule at the .005 boundary.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/documents_parser/tests/test_summary_pdf_mapper_chain.py

@@ -1912,6 +1912,7 @@ _COHORT_2_API_CLOSED: list[tuple[str, float]] = [
     ("0390-2066-4250-2026-4555", 65.3253),
     ("0464-3032-0205-4276-3204", 80.4533),
     ("0652-3022-1205-2826-1200", 70.9577),
+    ("1536-9325-5100-0433-1226", 65.8928),    # S0380.42 closure
     ("2007-3011-9205-8136-3204", 68.3914),
     ("2031-3007-0205-1296-3204", 64.1734),
     ("2130-3018-4205-4686-5204", 71.3158),
@@ -1958,13 +1959,6 @@ _COHORT_2_API_CLOSED: list[tuple[str, float]] = [
 #     API mapper likely lodges the secondary fuel differently. Probe
 #     the API JSON's `secondary_heating` block first.
 _COHORT_2_API_OPEN: list[tuple[str, float, float]] = [
-    # S0380.41 partially closed this cert: residual moved from +0.4445
-    # to +0.0015 via the same RdSAP 21 → SAP 10.2 glazing-type alias
-    # that closed 0300/9380 cleanly. A sub-2e-3 secondary tail remains
-    # — likely a windows-area Decimal-rounding boundary case in the
-    # cert's specific dimensions; investigate per the cohort closure
-    # convention in Slice S0380.42.
-    ("1536-9325-5100-0433-1226", 65.8928, 65.894324),
     ("2102-3018-0205-7886-5204", 63.8732, 57.570156),
 ]
 

domain/sap10_calculator/worksheet/heat_transmission.py

@@ -101,6 +101,19 @@ def _round_half_up(value: float, dp: int) -> float:
     return -floor(-value * factor + 0.5) / factor
 
 
+def _decimal_round_half_up_product(a: float, b: float, dp: int) -> float:
+    """a × b in Decimal arithmetic, HALF_UP-quantised at `dp` decimal
+    places. Mirrors `_round_half_up(a * b, dp)` but lands on the exact
+    .005 spec boundary that float multiplication drops (e.g. window
+    dimensions 0.65 × 0.70 = 0.4550 exact / 0.45499... in float —
+    float `_round_half_up` snaps to 0.45 vs spec answer 0.46). Used
+    for per-window area rounding under the RdSAP10 §15 "round per
+    element area" convention; the §15 Σ-then-round counterpart is
+    `_decimal_round_half_up_sum`."""
+    d = Decimal(str(a)) * Decimal(str(b))
+    return float(d.quantize(Decimal(10) ** -dp, rounding=ROUND_HALF_UP))
+
+
 _WALL_INSULATION_NONE: Final[int] = 4
 _DEFAULT_DOOR_AREA_M2: Final[float] = 1.85
 _DEFAULT_STOREY_HEIGHT_M: Final[float] = 2.5
@@ -476,8 +489,8 @@ def heat_transmission_from_cert(
     if windows_have_per_window_u:
         windows_w_per_k_total = 0.0
         for w in epc.sap_windows or []:
-            a_w = _round_half_up(
+            a_w = _decimal_round_half_up_product(
-                float(w.window_width) * float(w.window_height), _AREA_ROUND_DP
+                float(w.window_width), float(w.window_height), _AREA_ROUND_DP
             )
             u_raw_w = float(w.window_transmission_details.u_value)  # type: ignore[union-attr]
             u_eff_w = (
@@ -583,9 +596,8 @@ def heat_transmission_from_cert(
         # cascade-internal consistency.
         for w in epc.sap_windows:
             idx = _window_bp_index(w.window_location, len(parts))
-            area = _round_half_up(
+            area = _decimal_round_half_up_product(
-                float(w.window_width) * float(w.window_height),
+                float(w.window_width), float(w.window_height), _AREA_ROUND_DP,
-                _AREA_ROUND_DP,
             )
             window_area_by_bp[idx] += area
             if _window_on_alt_wall(w):

domain/sap10_calculator/worksheet/internal_gains.py

@@ -24,19 +24,25 @@ factor for L2a daylighting calc).
 from __future__ import annotations
 
 from dataclasses import dataclass
+from decimal import Decimal, ROUND_HALF_UP
 from enum import Enum
-from math import cos, exp, floor, pi
+from math import cos, exp, pi
 from typing import Final, Optional
 
 from datatypes.epc.domain.epc_property_data import EpcPropertyData, SapWindow
 
 
-def _round_area_2dp(value: float) -> float:
+def _decimal_window_area_2dp(width: float, height: float) -> float:
-    """Half-away-from-zero rounding to 2 d.p. matching heat_transmission.
+    """W × H in Decimal arithmetic, HALF_UP-quantised at 2 d.p. Per
-    RdSAP 10 §15 "Rounding of data" (p.66): "All element areas (gross)
+    RdSAP10 §15 "Rounding of data" (p.66) "All element areas (gross)
-    including window areas: 2 d.p." Inlined rather than imported so this
+    including window areas: 2 d.p." — uses Decimal so the lookup lands
-    module doesn't reach into heat_transmission's private helpers."""
+    on the exact .005 spec boundary that float multiplication drops
-    return floor(value * 100.0 + 0.5) / 100.0
+    (e.g. 0.65 × 0.70 = 0.4550 exact / 0.45499... in float — float
+    rounding snaps to 0.45 vs spec 0.46). Matches `heat_transmission.
+    _decimal_round_half_up_product` so the daylight factor's per-window
+    areas agree with the fabric cascade."""
+    d = Decimal(str(width)) * Decimal(str(height))
+    return float(d.quantize(Decimal("0.01"), rounding=ROUND_HALF_UP))
 
 _DAYS_PER_YEAR: Final[float] = 365.0
 _APPLIANCES_E_A_COEFF: Final[float] = 207.8
@@ -600,7 +606,7 @@ def _daylight_factor_from_cert(
     # including window areas: 2 d.p." — mirrors solar_gains and heat_
     # transmission so G_L sees the same area as the fabric cascade.
     wall_g_l_numerator = sum(
-        _round_area_2dp(float(w.window_width) * float(w.window_height))
+        _decimal_window_area_2dp(float(w.window_width), float(w.window_height))
         * _g_light(w) * _frame_factor(w) * z_l
         for w in epc.sap_windows
     )

domain/sap10_calculator/worksheet/solar_gains.py

@@ -30,8 +30,9 @@ coefficient sets (N, NE/NW, E/W, SE/SW, S).
 from __future__ import annotations
 
 from dataclasses import dataclass
+from decimal import Decimal, ROUND_HALF_UP
 from enum import Enum
-from math import cos, floor, radians, sin
+from math import cos, radians, sin
 from typing import Final
 
 from datatypes.epc.domain.epc_property_data import EpcPropertyData, SapWindow
@@ -43,12 +44,15 @@ from domain.sap10_calculator.climate.appendix_u import (
 from domain.sap10_calculator.worksheet.internal_gains import OvershadingCategory
 
 
-def _round_area_2dp(value: float) -> float:
+def _decimal_window_area_2dp(width: float, height: float) -> float:
-    """Half-away-from-zero rounding to 2 d.p. matching heat_transmission.
+    """W × H in Decimal arithmetic, HALF_UP-quantised at 2 d.p. Mirrors
-    RdSAP 10 §15 "Rounding of data" (p.66): "All element areas (gross)
+    `_round_area_2dp(W × H)` but lands on the exact .005 spec boundary
-    including window areas: 2 d.p." Inlined rather than imported so this
+    that float multiplication drops (e.g. 0.65 × 0.70 = 0.4550 exact /
-    module doesn't reach into heat_transmission's private helpers."""
+    0.45499... in float — float rounding snaps to 0.45 vs spec 0.46).
-    return floor(value * 100.0 + 0.5) / 100.0
+    Matches `heat_transmission._decimal_round_half_up_product` so solar
+    gains' per-window areas agree with the fabric cascade."""
+    d = Decimal(str(width)) * Decimal(str(height))
+    return float(d.quantize(Decimal("0.01"), rounding=ROUND_HALF_UP))
 
 
 # Table 6d first column — winter solar access factor Z for heating gains.
@@ -331,7 +335,7 @@ def _vertical_window_gain_monthly_w(
     # RdSAP 10 §15 "Rounding of data" (p.66): "All element areas (gross)
     # including window areas: 2 d.p." — matches heat_transmission's per-
     # window area rounding so solar gains and conduction agree on area.
-    area = _round_area_2dp(float(w.window_width) * float(w.window_height))
+    area = _decimal_window_area_2dp(float(w.window_width), float(w.window_height))
     g_perp = _g_perpendicular(w)
     ff = _frame_factor(w)
     return tuple(