---
Status: accepted
---

# Non-separated conservatory: mapper-owned §6.1 split, persisted 1:1, with round-trip fidelity as the contract

Decided in a `/grill-with-docs` session (2026-06-24) while fixing the gov-API
conservatory path (repro cert 718138, `RdSAP-Schema-19.0`). Extends
[ADR-0015](0015-mappers-own-cert-normalization.md) (mappers own cert
normalization) and echoes [ADR-0035](0035-coherent-heating-system-synthesis.md)
(coherence is a synthesis-time obligation, never a calculator normalisation).

## Context

The gov register lodges a **Non-Separated Conservatory** (`conservatory_type = 4`)
as a glazed "building part" carrying only `{floor_area, room_height,
double_glazed, glazed_perimeter}` — no construction age band, no wall/roof, no
floor dimensions. Treated as a generic fabric building part it is **mis-scored**
(its floor heat-loss is counted as a normal part) and **cannot persist** (the
`epc_building_part` columns `construction_age_band` / `wall_construction` are
NOT-NULL). The correct RdSAP treatment folds it into the dwelling under §6.1; a
**Separated Conservatory** (`type 2/3`) is disregarded under §6.2 and lodges no
glazed part. See CONTEXT.md for both terms.

Two placements were available for the §6.1 "split" (detect the glazed part, lift
it out of the fabric parts, carry it as the dwelling's conservatory):

- in the **mapper**, producing a normalized `EpcPropertyData` with
  `sap_conservatory` set and the glazed part excluded from `sap_building_parts`
  (the existing 21.0.1 choice, commit `d501535c`);
- at the **calculator / persistence-reload boundary**, re-deriving the split
  each time the picture is scored.

The production flow is *map → save → reload-from-DB → score*. The mapper split
alone is invisible to the scored picture, because the reload (`_compose`) did not
reconstruct `sap_conservatory` and the glazed part was excluded from the
persisted `sap_building_parts` — so the conservatory was silently dropped on the
path that actually scores. This is the same latent gap the 21.0.1 fix carried.

## Decision

1. **The §6.1 split lives in the mapper.** It is cert-shape knowledge, not
   physics (ADR-0015). `EpcPropertyData` — `sap_conservatory` populated, glazed
   part excluded from `sap_building_parts` — is the canonical normalized shape.
   We do **not** re-derive the split in the calculator or at reload; a global
   coherence pass over the assembled picture is rejected for the same reason
   ADR-0035 rejected one for heating.
2. **Persist the conservatory 1:1 on `epc_property`.** `SapConservatory` is a
   single optional value per dwelling, so it is five nullable columns on
   `epc_property` (alongside the existing scalar conservatory flags), not a child
   table. Save writes them; `_compose` rebuilds `SapConservatory` from them. The
   columns are FE/Drizzle-owned — documented in
   `docs/migrations/epc-property-round-trip-fidelity.md`, applied there, never by
   this repo.
3. **Round-trip fidelity is the contract that makes the split correct.** Because
   the split lives only in the mapper, the *only* thing guaranteeing the scored
   (reloaded) picture matches the mapped one is that `EpcPropertyData` survives
   `save → get` unchanged. We treat that as an enforced invariant: the
   deep-equality round-trip test (Model#1129) gains a non-separated-conservatory
   case, and a structural field-coverage guard fails if any `EpcPropertyData`
   field is neither reconstructed by `_compose` nor on a documented allow-list —
   turning "latent until a fixture exercises it" into an explicit decision.

## Consequences

- A future reader sees a conservatory — a "building part" in the source —
  persisted as flat columns on `epc_property` and absent from the dwelling's
  building parts. That is deliberate: it is a §6.1 dwelling attribute, not a
  fabric part.
- The persistence + guard are schema-version-agnostic, so they close the same
  latent gap for the already-merged 21.0.1 path, not only 19.0.
- The structural guard documents `sap_roof_windows` and the `solar_hw_collector_*`
  fields as known round-trip gaps (their FE columns are tracked follow-ups), so
  they are visible rather than silently dropped.
- The per-schema mapper piece (declare the four glazed fields + filter + fold)
  still has to be ported to the remaining gov mappers (17.0/17.1/18.0/20.0.0/
  21.0.0) one validated cert at a time; this ADR covers the shared shape, not
  that rollout.