Resolve solid-fuel room-heater overrides off Gas CPSU to their SAP codes 🟥

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

docs/adr/0045-gas-cpsu-dumping-ground-splits-into-solid-fuel-and-electric-boiler-archetypes.md

@@ -0,0 +1,89 @@
+# The "Gas CPSU" dumping ground splits into solid-fuel room-heater and electric-boiler archetypes
+
+## Status
+
+accepted
+
+## Context
+
+ADR-0041 named **"Gas CPSU" (SAP Table 4a 120, mains gas)** as the original
+garbage-drawer archetype the LLM classifier over-used when the heating taxonomy
+lacked a correct target. An audit of every `main_heating_system` override still
+resolving to `"Gas CPSU"` found it standing in for three *different*, unrelated
+systems — a mains-gas CPSU is neither solid fuel nor electric, so every one of
+these is mis-scored (wrong system **and**, for the electric ones, wrong fuel):
+
+| description | rows | true system |
+| --- | --- | --- |
+| Solid fuel room heaters: Open fire in grate | 16 | open fire (solid) |
+| Solid fuel room heaters: Open fire with back boiler (no radiators) | 19 | open fire + back boiler |
+| Solid fuel room heaters: Closed room heater with boiler (no radiators) | 2 | closed room heater + boiler |
+| Boiler: A rated NA | 58 | **electric** boiler |
+| Boiler: A rated CPSU | 1 | electric CPSU |
+| Boiler: C rated CPSU | 1 | *genuine* gas CPSU — correct |
+
+**The electric boilers are the sharpest mis-score.** All 58 `"Boiler: A rated NA"`
+rows carry a `main_fuel = electricity` override — "NA" is the Hyde boiler-type
+value meaning "not a gas combi / regular / CPSU", i.e. an **electric boiler**.
+Classified as Gas CPSU they are scored as a mains-gas stored-water unit. A
+lodged-cert check showed **defer-to-lodged is not viable**: 34 of the 58 lodge
+**no** main-heating code at all (fuel 0), 11 lodge electric *room heaters* (691),
+a few lodge community heating — the very gap the landlord override exists to fill.
+So the faithful move is to model the electric boiler the landlord names, informed
+by the electricity fuel we hold, not to defer to an empty cert.
+
+The correct SAP Table 4a codes are exact (confirmed against
+`domain/sap10_ml/sap_efficiencies.py`): **631** open fire (32%), **632** open fire
++ back boiler (50%), **634** closed room heater + boiler (65%); **191**
+direct-acting electric boiler (100%). All are already scored by the calculator,
+and 631/632/634 already sit in the overlay's `_SOLID_FUEL_ROOM_HEATER_CODES` set,
+so their coherent companions (house-coal fuel 33, room-heater control 2601,
+category 10) drag automatically (ADR-0035).
+
+## Decision
+
+Add the missing archetypes and reclassify the bucket. No calculator change.
+
+1. **Three solid-fuel room-heater archetypes** in `_MAIN_HEATING_CODES`:
+   `"Solid fuel room heater, open fire"` → 631,
+   `"Solid fuel room heater, open fire with back boiler"` → 632,
+   `"Solid fuel room heater, closed with boiler"` → 634. Companions drag from the
+   code (already-covered 631–636 set) — no overlay-logic change.
+2. **`"Electric boiler"` → SAP 191**, dragging **electricity (29)** via
+   `_natural_fuel_for` — an electric boiler is unambiguously electric (ADR-0041
+   "direct-acting/storage → electricity"), and this also closes the pre-existing
+   gap where 191 dragged no fuel. Meter stays Single, control None (191 is neither
+   a storage nor a wet-boiler-control system).
+3. **Deterministic `main_heating_guard` + reclassify**, TEXT-first /
+   pgEnum-deferred. The guard resolves the structured phrasings by description
+   alone: `"…open fire … back boiler"` → 632 (tested before the bare `"open fire"`
+   → 631), `"…closed … with boiler"` → 634, and `"Boiler: <band> NA"` →
+   `"Electric boiler"` (the NA type is the electric signal). `scripts/
+   reclassify_gas_cpsu.py` reuses the guard as its decision function and defers the
+   four new `main_heating_system` pgEnum-cache values.
+
+## Consequences
+
+- ~95 rows leave the Gas CPSU garbage drawer for faithful archetypes: 37 solid
+  fuel scored as solid fuel (not mains gas), 58 electric boilers scored as electric
+  (not a mains-gas CPSU). Correct Rebaselining (ADR-0039).
+- **FE-owned pgEnum additions (Dan):** four new `main_heating_system` values —
+  `"Solid fuel room heater, open fire"`, `"Solid fuel room heater, open fire with
+  back boiler"`, `"Solid fuel room heater, closed with boiler"`, `"Electric
+  boiler"`. TEXT read path fixed on merge; cache writes defer to the migration.
+- Extends [ADR-0041](0041-landlord-heating-classification-targets-a-complete-modellable-taxonomy.md)
+  (complete taxonomy — retiring the Gas CPSU garbage drawer it named) and
+  [ADR-0035](0035-coherent-heating-system-synthesis.md) (companions drag from the
+  code).
+
+### Out of scope
+
+- **Electric CPSU (SAP 192), the 1 `"Boiler: A rated CPSU"` row.** Its description
+  is identical to the genuine gas `"Boiler: C rated CPSU"` bar the efficiency
+  band — the electric/gas split is **fuel-determined**, which a pure-string guard
+  cannot see. Rather than break the guard = reclassify-decision-function symmetry
+  for one row, this is deferred to a small fuel-aware follow-up; the gas CPSU row
+  stays correctly on 120.
+- **Whether "NA" ever denotes a non-electric boiler.** It is 100% electric in the
+  corpus and the guard treats it as electric; a future non-electric "NA" would be
+  a re-open signal, not a silent mis-map (the fuel override would disagree).

tests/domain/epc/test_main_heating_system_overlay.py

@@ -78,6 +78,29 @@ def test_heating_archetypes_decode_to_their_sap_codes(
     assert simulation.heating.sap_main_heating_code == code
 
 
+@pytest.mark.parametrize(
+    ("main_heating_value", "code"),
+    [
+        # Solid-fuel room heaters the LLM funnelled into "Gas CPSU" (120, mains
+        # gas). SAP Table 4a: 631 open fire, 632 open fire + back boiler, 634
+        # closed room heater + boiler (ADR-0045).
+        ("Solid fuel room heater, open fire", 631),
+        ("Solid fuel room heater, open fire with back boiler", 632),
+        ("Solid fuel room heater, closed with boiler", 634),
+    ],
+)
+def test_solid_fuel_room_heaters_decode_off_gas_cpsu(
+    main_heating_value: str, code: int
+) -> None:
+    # Act
+    simulation = main_heating_overlay_for(main_heating_value, 0)
+
+    # Assert — its own solid-fuel code, not mains-gas CPSU (120).
+    assert simulation is not None
+    assert simulation.heating is not None
+    assert simulation.heating.sap_main_heating_code == code
+
+
 @pytest.mark.parametrize(
     ("main_heating_value", "code"),
     [