Record Elmhurst-validation provenance finding (#1649)

Two outputs of validating the Wythenshawe cohort against accredited Elmhurst:

1. Harness note: the DB-oracle residual tail is not engine defects. The two
   worst under-raters, 6 Barry Road (ours 60.12, oracle 62) and 87 Alderue
   (ours 61.80, oracle 63), were re-keyed into accredited Elmhurst on our exact
   inputs — Elmhurst returned 60 and 62, matching our engine and both below the
   lodged oracle. So our engine computes correct RdSAP; the lodged certs carried
   a lighter input set (dropped the surveyed 17.4 m2 solid-brick alternative
   wall, +1.7 SAP). It's input-provenance drift, not a calculation gap — don't
   chase these as engine bugs or tune the engine to close them. Corrects the
   earlier "spec-literal alt-wall" note.

2. epc-to-elmhurst skill gotcha: give Elmhurst the RAW surveyed room height, not
   the engine's storey height. The mapper already adds +0.25 m to upper floors
   (_UPPER_FLOOR_HEIGHT_ADD_M), so feeding sap_floor_dimensions.room_height_m
   makes Elmhurst double-add the joist void (~1 SAP too low — it under-scored
   6 Barry 60->59 until the first-floor height was corrected 2.76 -> raw 2.51).

Comment/doc only; no behaviour change (harness still 89.2%/0.297, pyright clean).

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

.claude/skills/epc-to-elmhurst-rdsap-inputs/SKILL.md

@@ -115,6 +115,16 @@ one-line reason. Absent ≠ zero everywhere — note which.
 - **Water heating** 903 = Electric immersion off-peak → Elmhurst "Water Heater"
   category, not "Boiler Circulator" (901 = from main system).
 - **Windows** are synthesised from `glazed_area` band × TFA — not real geometry.
+- **Room height — give the RAW surveyed room height, NOT the engine's storey
+  height.** Elmhurst adds the RdSAP +0.25 m upper-storey (joist-void) allowance
+  itself. Our mapper already applies that +0.25 to upper floors
+  (`_UPPER_FLOOR_HEIGHT_ADD_M`), so `sap_floor_dimensions[*].room_height_m` for
+  a first floor is *already* `raw + 0.25`. Reading that value onto the sheet
+  makes Elmhurst add a SECOND 0.25 → storey height too tall → ~4-5 m² extra wall
+  + volume → **~1 SAP too low** (it under-scored 6 Barry Road 60→59 until the
+  first-floor height was corrected from 2.76 back to the raw 2.51). Pull heights
+  from the survey/raw `building_measurements`, not the transformed
+  `sap_floor_dimensions`; the ground/lowest floor takes the raw value either way.
 - **Age band — pick by construction YEAR, not the engine's band letter.** The
   live RdSAP-10 tool bands differently from `mapping.md`'s older table: it
   offers `…K 2007-2011`, **`L 2012-2022`**, **`M 2023 onwards`**. A 2020 build

backend/documents_parser/tests/test_pashub_sap_accuracy_wythenshawe.py

@@ -55,13 +55,26 @@ _MANIFEST_PATH = _FIXTURES_DIR / "manifest.json"
 # 2026-07-19 (#1649 A-F): with the six from_site_notes fixes in place, the
 # **148 DB-oracle-matched fixtures** (39 UPRN + 109 postcode+house-number; 20
 # unmatched, excluded) score **within-0.5 89.2% (132/148), MAE 0.297**. The
-# residual tail is not fixable engine defects: the DB oracle is a *rounded
-# integer* SAP graded against a continuous engine (a +0.5-to-band cluster of
-# spec-correct filled-cavity dwellings sits at that rounding floor), plus a
-# couple of spec-literal-vs-accredited-engine gaps (e.g. 6 Barry Road's
-# uninsulated solid-brick alternative wall at the RdSAP Table 6 U=1.7 default).
-# Treat floor/ceiling as a regression tripwire, not a target; re-baseline
-# (coverage growth is not loosening) as PasHub sources the 20 unmatched SAPs.
+# residual tail is NOT fixable engine defects — two independent causes, both
+# established, not inferred:
+#   1. Integer-oracle rounding: the DB oracle is a *rounded integer* SAP graded
+#      against our continuous engine, so a band of spec-correct dwellings sits
+#      just over a rounding boundary (within-1.0 is 98%).
+#   2. Input-provenance drift between the site-notes survey and the lodged cert.
+#      The two worst under-raters, 6 Barry Road (ours 60.12, oracle 62) and 87
+#      Alderue (ours 61.80, oracle 63), were re-keyed into accredited Elmhurst
+#      on our exact inputs: Elmhurst returned **60 and 62** — matching our
+#      engine, and both *below* the lodged oracle. So our engine computes
+#      correct RdSAP; the lodged cert simply carried a lighter input set (it
+#      dropped the 17.4 m² solid-brick alternative wall the surveyor recorded
+#      via brick-bond evidence, +1.7 SAP). Accredited engines agree on identical
+#      inputs — a 2-point gap is an input difference, not a calculation one.
+# So do NOT chase these as engine bugs, and do NOT tune the engine to close them
+# (that would move the gov-API corpus off 78.9% and diverge from accredited
+# Elmhurst). Treat floor/ceiling as a regression tripwire, not a target;
+# re-baseline (coverage growth is not loosening) as PasHub sources the 20
+# unmatched SAPs. The clean accuracy proof is engine-vs-accredited on identical
+# inputs (done for 6 Barry / 87 Alderue), not engine-vs-lodged-integer.
 _MIN_WITHIN_HALF: float = 0.88
 _MAX_SAP_MAE: float = 0.31