docs(skill): bank corpus-1000 findings (cost triangulation, software divergences, assumed-roof rule)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

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

@@ -169,6 +169,34 @@ Pattern: `with E.session() as (ctx,page): E.goto(...); E.set_text/set_select(...
   lines 17/18). This drove the first campaign mapper fix — see Banked findings.
 
 ## Banked findings (fold new ones in here as the corpus grows)
+- **Register cost-triangulation (no Elmhurst needed):** the register lodges
+  `heating_cost_current` / `hot_water_cost_current` / `lighting_cost_current`
+  at RETAIL prices — absolute £ never match SAP Table-12 spec prices, but for
+  same-fuel end-uses the RATIOS localize a divergence without a worksheet.
+  Normalize by lighting: `(lodged_hw/lodged_light) / (our_hw/our_light)` ≈ 1
+  means our water cascade agrees with the lodging software; »1 flags the
+  divergent end-use. CAVEAT: the register's "hot water cost" INCLUDES the
+  SAP (64a) electric-shower electricity — add `instant_shower` cost to our
+  hw bucket before comparing on any electric-shower cert, else every
+  shower-outlet-type-2 cert false-positives at ratio ~2 (corpus-1000 band
+  triage, 36 certs, mean ratio 2.0 → 1.04 after adding (64a)).
+- **Lodging-software conventions diverge from Elmhurst on two known axes**
+  (mark such corpus certs ⚠, don't tune): (a) out-of-range-PSR database
+  heat pumps — SAP 10.2 N2 footnotes 44/45 mandate the reciprocal extension
+  toward 100% at 2× the record's largest PSR (Elmhurst does this, golden
+  case 56); at least one software bills 100% direct electric + standard
+  schedule instead (corpus 4510053280, lodged 51 vs engine/Elmhurst-method
+  75). (b) a lodged 0 extract fans — Elmhurst renders 0 as "unknown" →
+  Table 5 age-band default (worksheet case 46); LIG-21.0 lodgements
+  (calculation_software_version 10.2.2.0) honour the 0 literally (corpus
+  quadruplet at Δ −2.56, fans account for 1.42).
+- **"insulated (assumed)" roof ≠ observed retrofit:** RdSAP renders
+  "(assumed)" when insulation PRESENCE is an age-band assumption — Table 18
+  age default governs, for BOTH the "ND" and "NI" thickness sentinels
+  (which are lodging-software noise). The §5.11.4 50 mm row (U 0.68) is
+  only for the bare "insulated" description (observed retrofit, depth
+  unknown). Fixed in u_roof (corpus 74.2% → 75.5%); unit-pinned in
+  test_rdsap_uvalues.
 - **MAPPER GAP — cylinder insulation thickness dropped (RdSAP-17.0+):** the mapper
   carries `cylinder_size` + `cylinder_insulation_type` but NOT
   `cylinder_insulation_thickness` → `EpcPropertyData.sap_heating.cylinder_