fix(extractor): drop windows-table header remnant from first window glazing type

Summary PDFs preprocessed from `pdftotext -layout` wrap the windows-table
header across several lines. The third header line's tail ("U value / g
value / Draught Proofed / Permanent Shutters") tokenises to "value value
Proofed Shutters" and lands directly above the FIRST window's data row.

Because the first window in a building part has `before_start = 0`, its
prefix block reaches back into that header remnant. The remnant is
neither an orientation nor a building-part fragment, so it survived the
pops in `_compose_window_descriptors` and leaked into glazing_type as
"value value Proofed Shutters Double between 2002 and 2021" (windows 2-3,
whose prefix starts after the previous window's manufacturer line, were
clean).

Fix: the glazing-type phrase always starts with a glazing-start word
(Single/Double/Triple/Secondary), so trim any prefix fragments preceding
that word before joining the glazing type. Orientation/bp pops still run
on the full prefix, so they are unaffected.

Reproduced from `sap worksheets/Recommendations Elmhurst Files/
cavity_wall_insulation - main wall/before/Summary_001431.pdf`. Added a
regression test driving the real `_extract_windows_from_layout` path with
the verbatim tokenised header+rows. 2306 passed (+4), pyright net-zero.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/documents_parser/elmhurst_extractor.py

@@ -1090,7 +1090,28 @@ class ElmhurstSiteNotesExtractor:
         if inline_glazing_type is not None:
             glazing_type = inline_glazing_type
         else:
-            glazing_type = " ".join([*prefix, *suffix]).strip()
+            # The glazing-type phrase always starts with a glazing-start
+            # word (Single/Double/Triple/Secondary). The FIRST window in
+            # a building part has `before_start = 0`, so its prefix block
+            # reaches back into the wrapped windows-table header; the
+            # third header line's tail tokenises to "value value Proofed
+            # Shutters" (the "U value / g value / Draught Proofed /
+            # Permanent Shutters" column titles) and is neither an
+            # orientation nor a bp fragment, so it survives the pops.
+            # Drop any prefix fragments preceding the glazing-start word
+            # so they don't leak into the glazing type.
+            glazing_start = next(
+                (
+                    idx
+                    for idx, frag in enumerate(prefix)
+                    if frag.split(" ", 1)[0] in self._GLAZING_TYPE_PREFIX_WORDS
+                ),
+                None,
+            )
+            glazing_prefix = (
+                prefix[glazing_start:] if glazing_start is not None else prefix
+            )
+            glazing_type = " ".join([*glazing_prefix, *suffix]).strip()
 
         # Building part: inline token wins; otherwise join prefix + suffix.
         if bp_inline is not None:

backend/documents_parser/tests/test_elmhurst_extractor.py

@@ -513,3 +513,112 @@ class TestLightingLedCflUnknown:
 
     def test_cfl_count_zero_when_unknown(self, result2: ElmhurstSiteNotes) -> None:
         assert result2.lighting.cfl_count == 0
+
+
+class TestWindowsLayoutHeaderRemnant:
+    """Regression for the first-window glazing-type header leak.
+
+    Summary PDFs preprocessed from `pdftotext -layout` wrap the windows
+    table header across several lines. The third header line's tail
+    ("U value / g value / Draught Proofed / Permanent Shutters") tokenises
+    to "value value Proofed Shutters" and sits directly above the FIRST
+    window's data row. Because the first window in a building part has
+    `before_start = 0`, its prefix block reaches back into that header
+    remnant, which is neither an orientation nor a building-part fragment
+    and so survived into `glazing_type` as
+    "value value Proofed Shutters Double between 2002 and 2021".
+
+    Reproduced from `sap worksheets/Recommendations Elmhurst Files/
+    cavity_wall_insulation - main wall/before/Summary_001431.pdf` (3
+    Manufacturer-data-source windows; only window 0 was corrupted).
+    """
+
+    # Faithful reproduction of the tokenised windows section (one page),
+    # captured verbatim from the Summary PDF above. The header remnant
+    # "value value Proofed Shutters" precedes window 0's wrapped glazing
+    # cell ("Double between 2002" / "and 2021").
+    _WINDOWS_PAGE = "\n".join([
+        "11.0 Windows:",
+        "Frame Frame Glazing",
+        "Building",
+        "U",
+        "g Draught Permanent",
+        "W",
+        "H",
+        "Area Glazing Type",
+        "Location",
+        "Orient. Data-Source",
+        "Type Factor Gap",
+        "Part",
+        "value value Proofed Shutters",
+        "Double between 2002",
+        "North",
+        "0.97 1.00 0.97",
+        "PVC",
+        "0.70",
+        "Main",
+        "External wall",
+        "Manufacturer 2.00",
+        "0.72",
+        "Yes",
+        "None",
+        "and 2021",
+        "West",
+        "Double between 2002",
+        "South",
+        "2.66 1.00 2.66",
+        "PVC",
+        "0.70",
+        "Main",
+        "External wall",
+        "Manufacturer 2.00",
+        "0.72",
+        "Yes",
+        "None",
+        "and 2021",
+        "East",
+        "Double between 2002",
+        "South",
+        "2.66 1.00 2.66",
+        "PVC",
+        "0.70",
+        "Main",
+        "External wall",
+        "Manufacturer 2.00",
+        "0.72",
+        "Yes",
+        "None",
+        "and 2021",
+        "East",
+        "12.0 Ventilation",
+    ])
+
+    @pytest.fixture(scope="class")
+    def windows(self) -> list[Window]:
+        return ElmhurstSiteNotesExtractor([self._WINDOWS_PAGE])._extract_windows()
+
+    def test_window_count(self, windows: list[Window]) -> None:
+        # Arrange / Act / Assert
+        assert len(windows) == 3
+
+    def test_first_window_glazing_type_excludes_header_remnant(
+        self, windows: list[Window]
+    ) -> None:
+        # Arrange / Act / Assert — no "value value Proofed Shutters" leak.
+        assert windows[0].glazing_type == "Double between 2002 and 2021"
+
+    def test_all_windows_share_clean_glazing_type(
+        self, windows: list[Window]
+    ) -> None:
+        # Arrange / Act / Assert — windows 1 and 2 were already clean;
+        # all three must agree after the fix.
+        assert [w.glazing_type for w in windows] == [
+            "Double between 2002 and 2021"
+        ] * 3
+
+    def test_first_window_orientation_unaffected(
+        self, windows: list[Window]
+    ) -> None:
+        # Arrange / Act / Assert — trimming the glazing prefix must not
+        # disturb orientation extraction (North + West fragments).
+        assert windows[0].orientation == "North-West"