Slice S0380.12: parse 'Alternative wall' window-location in pre-data slice

Cert 2636-0525-2600-0401-2296's Summary §11 Windows block lodges one
alt-wall window (1.19 m², north-facing). The PDF layout for alt-wall
rows puts the "Alternative wall" string in the slot BEFORE the W×H×A
data line — not after frame_factor where regular "External wall"
rows put it. Without this fix the extractor's
`_parse_window_from_anchors` only scanned the post-frame_factor
`middle` slice for wall tokens, defaulted to "External wall" for the
alt-wall row, and the cascade allocated the 1.19 m² opening to the
main wall instead of the alt-wall — under-deducting from main and
leaving the alt-wall gross instead of net.

Fix at `elmhurst_extractor.py:865`: also scan
`lines[before_start:data_idx]` (the pre-data slice) for "wall"
tokens. Search order:
  1. `middle` — first preference (normal layout for regular rows)
  2. `pre_data` — alt-wall rows (cert 2636)
  3. "External wall" default — no wall lodging found

Forcing function: cert 2636 walls_w_per_k moves from 20.5595 to
**20.0240 — EXACT match against worksheet (29a) Main 11.9250 + alt.1
8.0990 = 20.0240**. (Header (29a) sum is now fabric-exact; the
remaining +0.52 SAP residual on cert 2636 is in the ventilation
cascade — HTC 153.97 vs API 159.02 vs worksheet (39) avg 158.85 —
to be investigated in a follow-up slice.)

Added focused unit test
`test_summary_2636_alt_wall_window_parses_alternative_wall_location`
that pins the by-area lookup: 1.19 m² → "Alternative wall"; the
six 2.25 m² windows stay on "External wall". Guards against future
window-location parser regressions.

Pyright: 0 errors on the edited extractor + test files.

Regression suite: 685 pass + 10 fail (handover baseline 669 + 10 +
16 new GREEN tests across S0380.2..S0380.12). Cohort status:
  cert  Δ vs worksheet  spec floor?
  0380  +0.0594         ✓
  0350  +0.0458         ✓
  2225  +0.0441         ✓
  2636  +0.5167         ✗  (fabric exact; ventilation residual)
  3800  +0.0442         ✓
  9285  +0.0502         ✓
  9418  +2.5973         ✗  (Daikin)

Spec refs:
- Slice 102f-prep.10 (commit 24a7351f) — API-path equivalent
  "Alt-wall opening allocation per window_wall_type".
- SAP 10.2 §3.7 — opening (window + door) deduction from gross
  wall area, per-window allocated to the lodged wall type.

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

backend/documents_parser/elmhurst_extractor.py

@@ -862,7 +862,17 @@ class ElmhurstSiteNotesExtractor:
         # Variable-order tokens between frame_factor and Manufacturer.
         middle = [lines[j].strip() for j in range(middle_start, manuf_idx)]
         glazing_gap = next((t for t in middle if "mm" in t.lower()), None)
-        location = next((t for t in middle if "wall" in t.lower()), "External wall")
+        # Wall-location lodging. Most rows put "External wall" in
+        # `middle`; alt-wall rows (cert 2636 window-4 / cert 9418 alt-
+        # wall window) put "Alternative wall" in the PRE-data slice
+        # (between the previous window's end and W×H×A). Search both
+        # slices so either layout resolves to the correct location.
+        pre_data = [lines[j].strip() for j in range(before_start, data_idx)]
+        location = (
+            next((t for t in middle if "wall" in t.lower()), None)
+            or next((t for t in pre_data if "wall" in t.lower()), None)
+            or "External wall"
+        )
         bp_inline = next((t for t in middle if t in self._BP_INLINE_TOKENS), None)
         orient_inline = next(
             (t for t in middle if t in self._ORIENTATION_TOKENS), None

backend/documents_parser/tests/test_summary_pdf_mapper_chain.py

@@ -63,6 +63,7 @@ _SUMMARY_000903_PDF = _FIXTURES / "Summary_000903.pdf"
 _SUMMARY_000901_PDF = _FIXTURES / "Summary_000901.pdf"  # cert 3800
 _SUMMARY_000904_PDF = _FIXTURES / "Summary_000904.pdf"  # cert 9285
 _SUMMARY_000900_PDF = _FIXTURES / "Summary_000900.pdf"  # cert 2225
+_SUMMARY_000898_PDF = _FIXTURES / "Summary_000898.pdf"  # cert 2636
 
 # GOV.UK EPB API JSON for cert 001479 — the API-path counterpart of the
 # Summary_001479.pdf fixture. Together they drive the API ≡ Summary
@@ -714,6 +715,31 @@ def test_summary_0350_full_chain_sap_within_spec_floor_of_worksheet() -> None:
     assert abs(result.sap_score_continuous - worksheet_unrounded_sap) < _ASHP_COHORT_CHAIN_TOLERANCE
 
 
+def test_summary_2636_alt_wall_window_parses_alternative_wall_location() -> None:
+    # Arrange — cert 2636-0525-2600-0401-2296's §11 Windows block lodges
+    # one alt-wall window (the 1.19 m² north-facing one): the row's
+    # "Alternative wall" string appears BEFORE the W×H×A line, not
+    # after the frame_factor (the normal position for "External wall").
+    # The extractor's `_parse_window_from_anchors` was only scanning
+    # the post-frame_factor `middle` slice for wall-location tokens →
+    # defaulted to "External wall" for the alt-wall row → cascade
+    # allocated the window to the main wall instead of the alt-wall,
+    # leaving Main external walls W/K under-deducted by ~0.54 vs
+    # worksheet (29a). Fix: also scan the PRE-data slice
+    # `lines[before_start:data_idx]` for wall tokens.
+    pages = _summary_pdf_to_textract_style_pages(_SUMMARY_000898_PDF)
+    site_notes = ElmhurstSiteNotesExtractor(pages).extract()
+
+    # Act
+    epc = EpcPropertyDataMapper.from_elmhurst_site_notes(site_notes)
+
+    # Assert — the 1.19 m² window is recorded with wall_type =
+    # "Alternative wall"; all other windows stay on "External wall".
+    by_area = {round(w.window_width, 2): w.window_wall_type for w in epc.sap_windows}
+    assert by_area[1.19] == "Alternative wall"
+    assert by_area[2.25] == "External wall"  # main-wall windows unchanged
+
+
 def test_summary_2225_no_showers_lodged_resolves_to_zero_counts() -> None:
     # Arrange — cert 2225-3062-8205-2856-7204's Summary §1x Baths and
     # Showers block lodges 0 baths and ZERO showers (no shower rows at