Closes a systematic +0.02..+0.07 SAP over-prediction on every triple-
glazed cert in cohort 2 (13 of 38) and removes a silent-default
failure mode flagged via cert 3336-2825-9400-0512-8292 (+0.0674 Δ).
Root cause: `_map_elmhurst_window` (datatypes/epc/domain/mapper.py)
was passing the Elmhurst-lodged glazing-type string verbatim into
`SapWindow.glazing_type` (declared `Union[int, str]`). The §5 (66)..
(67) daylight-factor cascade at
`domain/sap10_calculator/worksheet/internal_gains.py:512` requires
`isinstance(w.glazing_type, int)` to look up Table 6b col light g_L —
string lodgings silently fell through to the `_G_LIGHT_DEFAULT = 0.80`
(double-glazed) branch. Cert 3336 (Triple glazed, worksheet "Window,
Triple glazed") got g_L = 0.80 instead of the correct 0.70, inflating
C_daylight from 1.072 to 1.041 → lighting kWh under-predicted by
−4.53 kWh/yr → total fuel cost under by −1.17 GBP → ECF Δ −0.0049 →
SAP continuous over by +0.0674.
Fix: `_ELMHURST_GLAZING_LABEL_TO_SAP10` dict + `_elmhurst_glazing_
type_code` helper translate the Elmhurst Summary §11 lodged strings
to the SAP 10.2 Table U2 integer codes the cascade keys on:
"Single" → 1
"Double pre 2002" → 2
"Double between 2002 and 2021" → 3
"Double with unknown install date" → 3
"Double with unknown 16 mm or install date more" → 3
"Double post or during 2022" → 5
"Triple post or during 2022" → 6
"Triple post or during" → 6 (year-trunc.)
"Secondary" → 7
Two regex passes strip the layout noise the extractor sometimes folds
into the glazing-type token: a `(?:Part )?value value Proofed Shutters`
prefix (from adjacent column headers) and a ` Summary Information` /
` Alternative wall…` suffix. Verified against the union of cohort-1
(7 certs) + cohort-2 (38 certs) + test-fixture (9 PDFs) glazing
labels: 18 distinct surface forms, all closed by the dict + noise
patterns; one window in cert 2636's Summary_000898.pdf lodged the
year-truncated "Triple post or during" — added as an alias for code 6
per worksheet "Triple glazed" lodging.
Strict-enum gate: `_elmhurst_glazing_type_code` raises
`UnmappedElmhurstLabel("glazing_type", label)` (Slice S0380.15
pattern, extended to the new helper) when the label is None or not
in the dict — surfaces mapper-coverage gaps at extraction time rather
than masking them as a SAP precision floor.
Cohort-2 Summary-path delta progression (38 certs):
bucket before slice 2 after slice 2
exact (<1e-4) 11 11
<0.005 0 5 ← 9421 +0.0012, 2536 +0.0016, 9370 +0.0017, 0100 +0.0028, 2800 +0.0044
0.005-0.07 15 10 ← all triple-glazed
0.07-0.5 5 5
0.5-1 4 4
1-5 1 1
5+ 2 2
RAISES 0 0
3336 (user's flag) closes from +0.0674 → +0.0400 — the residual is
the remaining systematic offset the next slice will investigate.
Tests added (3):
- `test_summary_3336_triple_glazed_windows_route_to_code_6` — pins
the mapper output for the user's flagged cert.
- `test_summary_000474_double_glazed_windows_route_to_code_3` —
exercises the DG branch + the year-unknown alias mapping.
- `test_summary_mapper_raises_on_unmapped_glazing_type_label` —
strict-enum coverage gate via mutated site notes.
Tests updated (1):
- `test_first_window_glazing_type` (test_elmhurst_end_to_end.py):
asserts int code 5 (DG low-E argon — "Double post or during 2022")
not the string verbatim. The string-passthrough behaviour was
always a latent bug; this test was the only direct pin on it.
Pyright net-zero per file:
- datatypes/epc/domain/mapper.py: 32 (baseline 32)
- backend/documents_parser/tests/test_summary_pdf_mapper_chain.py: 0
- backend/documents_parser/tests/test_elmhurst_end_to_end.py: 0
Regression baseline: 694 pass + 10 fail (= prior 691 + 10 + 3 new).
Triple-glazed original-cohort certs are now closer to worksheet too;
the ±0.07 chain tests on the original cohort still hold, and a future
slice tightens them once the next-largest residual is closed.
Spec refs:
- SAP 10.2 Table U2 — glazing-type integer enum.
- SAP 10.2 Table 6b col light — light-transmission g_L by glazing
type (triple 0.70, double-glazed variants 0.80, single 0.90).
- RdSAP 10 §11 Windows — Summary lodging of glazing type as a
type+install-date phrase.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Three orthogonal issues surfaced by the full project test sweep:
1. Dockerfile.test: install poppler-utils alongside postgresql.
The 20× `pdfinfo: No such file or directory` failures in
test_summary_pdf_mapper_chain.py traced to the CI test image
missing the poppler-utils system package (pdfinfo + pdftotext).
`_summary_pdf_to_textract_style_pages` shells out to these for
layout-preserving PDF text extraction. Pure-Python alternatives
(pymupdf, pypdf) don't reproduce pdftotext -layout's row-major
table cell ordering, which the Elmhurst Summary extractor depends
on. So system poppler is the right fix; added to apt-get install
with an explanatory comment.
2. test_from_rdsap_schema.py::test_total_floor_area: expected 55.0,
got 45.82. Slice 95 (commit f502db8c) changed the API mapper to
compute total_floor_area_m2 from the precise sum of per-bp
sap_floor_dimensions[*].total_floor_area rather than the lodged
scalar. The synthetic 21_0_1.json fixture has lodged total_floor_
area=55 + a single fd of 45.82 (per-bp sum doesn't match lodged).
Updated the expected to 45.82 with a comment explaining the
Slice 95 per-bp-sum precedence.
3. test_elmhurst_end_to_end.py::test_emitter_temperature: expected
"Unknown", got int 1. Pre-existing failure (confirmed by checking
out commit 985a59e1 and reproducing). `_elmhurst_emitter_
temperature_int` in datatypes/epc/domain/mapper.py converts the
Elmhurst Summary §14 "Design flow temperature: Unknown" to SAP10.2
Table 4d code 1 (high-temp / ≥45 °C, worst-case for unmeasured
boilers). The int encoding mirrors the API mapper's MainHeating
Detail.emitter_temperature for cross-mapper field parity. Test
updated to expect 1 (with comment) since the conversion is the
correct production behaviour.
Verified:
- Layer 4 1e-4 gate (test_api_001479_full_chain_sap_matches_worksheet_
pdf_exactly) still GREEN.
- Wider domain sweep (domain/sap10_calculator + domain/sap10_ml):
1654 passed / 20 failed, exact pre-fix baseline.
- All three originally-failing tests now PASS.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Sibling migration to the sap10_calculator move — `domain.ml` now lives
at the root-level layout (`domain/sap10_ml/`) matching the pattern
already used by `domain.addresses`, `domain.tasks`, `domain.postcode`,
and `domain.sap10_calculator`.
Changes:
- `git mv packages/domain/src/domain/ml → domain/sap10_ml` (19 files;
history preserved).
- Subpackage rename: `domain.ml` → `domain.sap10_ml`. 32 references
rewritten across .py and .md files: 11 internal + 21 external
(datatypes/epc/domain/mapper.py, 14 files in domain/sap10_calculator,
2 backend tests, 2 ADRs, 1 README, 1 design doc).
- Path-string updates: `pytest.ini` testpath
`packages/domain/src/domain/ml/tests` → `domain/sap10_ml/tests` so
ML tests stay in the default auto-discovered sweep. `CONTEXT.md`
also updated.
`packages/domain/src/domain/` is now empty — the workspace `domain/`
tree has been fully migrated. Together with the `domain/__init__.py`
deletions from the sap10_calculator commit (29ac35cc), `domain` is
now a single root-level namespace package with subpackages
{addresses, sap10_calculator, sap10_ml, tasks} + the standalone
`postcode.py` module.
Verified:
- Focused sweep (backend mapper-chain + sap10_calculator worksheet
e2e + golden fixtures): 99 passed / 19 failed — identical baseline.
- Wider sweep (all sap10_calculator + sap10_ml): 1654 passed / 20
failed (same pre-existing failures).
- domain/sap10_ml/tests: 210/210 PASSED at new path.
- Pyright net-zero: heat_transmission.py 13, cert_to_inputs.py 35,
mapper.py 33, rdsap_uvalues.py 1 (all unchanged from baseline).
Note: `packages/domain/pyproject.toml` still declares
`packages = ["src/domain"]` for the hatchling wheel — that target
directory is now empty and the wheel build is effectively a no-op.
Retiring the workspace package or repointing the wheel is a follow-up.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Three mapper extensions, validated by 000516 closing to 1e-4:
1. Roof-window separation by U-value threshold. Elmhurst Summary PDFs
pool roof windows into the §11 vertical-window table with no type
marker. The U-value is the only reliable signal — vertical glazing
in the cohort tops out at 2.80 W/m²K, while Table 24 roof windows
start at 3.0+. `_is_elmhurst_roof_window` filters U > 3.0 into
`sap_roof_windows`; the rest flow through the `sap_windows` path.
2. Table-24 roof-window U-value lookup. The cohort lodges Manufacturer
U=3.10 for the 000516 roof window, but the worksheet's (27a) line
(U_eff=2.99) reverse-engineers to a raw U=3.40 — the RdSAP10
Table 24 "Double pre 2002" roof-window default. `_elmhurst_roof_
window_u_value` keyed on glazing-type captures the +0.3 W/m²K step;
falls back to the lodged U for glazing types not yet in the table.
3. `SapWindow.window_width × window_height = lodged Area` convention.
The Elmhurst Summary PDF carries lodged W (2 d.p.) × lodged H
(2 d.p.) AND a precomputed Area (2 d.p., not always equal to
product after rounding). The cascade reads only the W×H product
across §3 / §5 / §6, so flattening to `(area, 1.0)` keeps the
downstream area aligned with the worksheet's rounded value rather
than reconstructing W×H with its own rounding drift (e.g. 1.22 ×
1.76 = 2.1472 m² vs lodged 2.15 m²). The existing
`test_first_window_*` tests pinning literal W/H were updated to
pin the area product (the cascade-relevant invariant).
Cohort state after this slice:
000474 0.0000 ✓ Slice 47
000477 +1.1161 Elmhurst floor_ach quirk
000480 0.0000 ✓ Slice 50
000487 +1.1844 extractor still drops most §11 windows
000490 0.0000 ✓ Slice 49
000516 0.0000 ✓ THIS SLICE
4/6 closed at 1e-4. 756 tests pass; pyright net-zero (35 baseline).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The full Summary→ElmhurstSiteNotes→EpcPropertyData→cascade→SAP chain now produces unrounded SAP 62.52 for cert U985-0001-000474 vs the worksheet PDF's 62.2584 — inside the 0.5 tolerance the user accepts on the API-cert residual cohort. The hand-built worksheet-fixture chain matches Elmhurst's unrounded SAP to 4 d.p. (62.2584), so the calculator+cascade are provably equivalent to Elmhurst's calculator; this slice closes the mapper side of the chain.
Mapper changes drop the string-versus-int impedance mismatch that prevented the cascade from consuming Elmhurst-coded values:
- construction_age_band: `_strip_code('B 1900-1929')` → 'B' (was '1900-1929')
- wall_construction: `_elmhurst_wall_construction_int('CA Cavity')` → 4 (was string 'Cavity')
- wall_insulation_type: `'A As Built'` → 4 (was string 'As Built')
- party_wall_construction: same int-mapping treatment
- main_fuel_type: `_elmhurst_main_fuel_int('Mains gas')` → 26 (the Table 12 fuel code; was string)
- heat_emitter_type: `'Radiators'` → 1 (was string)
- main_heating_control: `_elmhurst_sap_control_code('SAP code 2106, ...')` → 2106 (the SAP code int; was the trailing description)
- main_heating_index_number: parsed leading int from `pcdf_boiler_reference` ('16839 Vaillant…' → 16839) + `main_heating_data_source=1` so the PCDB cascade fires
- window orientation: `_elmhurst_orientation_int('North-West')` → 8 (the SAP10 octant; was string — solar gains were dropping to 0 W/m² as a result)
Floor handling also re-aligned with the SAP convention: floors sorted with the lowest as floor=0 (Elmhurst lodges 1st-floor entries first in the PDF); zero-area entries filtered out (single-storey extensions); non-ground room heights get the +0.25 m joist-void adjustment; `is_exposed_floor=True` for ground floors lodged above unheated space ('U Above unheated space'). `total_floor_area_m2` now sums across main + extensions.
Three regression pins on the new path:
- sap_building_parts == 3 (multi-bp)
- sap_windows == 7 (layout-style window parser)
- unrounded SAP within 0.5 of 62.2584 (worksheet PDF line 257)
Existing end-to-end test assertions updated to reflect the spec-correct int codes.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Threads the strict BuildingPartIdentifier type (introduced in a8b443f6)
through the two remaining backend touchpoints:
- EpcBuildingPartModel.from_*: SQLModel column expects a string, so
unbox the enum with .identifier.value before binding to the DB.
- documents_parser end-to-end tests: swap bare-string equality
("main" / "extension_1") for identity checks against the enum
members (BuildingPartIdentifier.MAIN / EXTENSION_1).
Documents_parser test pack passes (105/105). No dedicated SQLModel test
covers EpcBuildingPartModel.from_*; the .value line is exercised
transitively via db_writer.py / local_runner.py in production.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
