fix regression tests

.claude/skills/expand-sap-accuracy-corpus/worklist.md

@@ -67,10 +67,19 @@ prediction, and the skipped certs surface for follow-up:
   cert numbers, so the gaps can be closed deliberately. +3 tests; 0 new pyright.
   Live-verified on BN11 4EP: cohort now builds 35 + records the 1 skip.
 
-**✅ DONE — 2 safe generic mapper fixes (+regression tests, 0 new pyright):**
+**✅ DONE — generic mapper fixes (+regression tests, 0 new pyright):**
 - `_normalize_sap_schema_16_x`: `setdefault("insulated_door_count", 0)` (the
-  original prod crash) AND `setdefault("multiple_glazed_proportion", 100)` (16.3
-  cert 0418-3986-7250-2884-7970; ML-only field the SAP calc ignores; modal 100).
+  original prod crash).
+- ⚠️ REVERTED `setdefault("multiple_glazed_proportion", 100)` — it made an
+  otherwise-unmappable 16.3 cert (0418-3986-7250-2884-7970) join the EPC-prediction
+  **frozen-fixture donor pool** and tipped near-tie similarity matches, regressing
+  the component-accuracy gate (`has_hot_water_cylinder` 30→29/36; `door_count`
+  residual 23→25/36 — both ≫ the gate's 1e-3 float tolerance). The field is ML-only
+  (SAP calc never reads it) and the cohort skip-and-report path handles the cert,
+  so we leave it fail-loud instead of synthesising a value. Re-baselining the gate
+  was rejected (loosens a tighten-only gate for a real regression). For Khalim: if
+  16.x certs missing `multiple_glazed_proportion` should map, derive it (single→0 /
+  double→100) rather than a flat default, and re-measure the gate.
 - `_with_recorded_performance` co2/consumption/rating: `float(co2)` → crashed on
   certs lodging `co2_emissions_current` as a Measurement dict `{'value':3.5,
   'quantity':'tonnes per year'}` (16.x cert 2308-4997-7262-0137-9930, surfaced as