docs: record PV-array + floor-flag round-trip closures 🟪

Document Slice 2 in the persistence-fidelity migration log, plus the 7
calculator-read nested gaps the recursive guard newly surfaced (tracked
follow-up).

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

docs/migrations/epc-property-round-trip-fidelity.md

@@ -58,14 +58,57 @@ now fails if any `EpcPropertyData` field is neither reconstructed by `_compose`
 allow-list — so the gaps below are no longer "latent until a fixture exercises them", they are
 explicit and tracked.
 
-**Still open (follow-up issues):** the remaining §2 structural tables (room-in-roof detail, PV
-arrays, **roof windows** — `sap_roof_windows`) + §3 nested-wall fields
-(`SapAlternativeWall.u_value`/`wall_thickness_mm`) + `SapFloorDimension` exposed-floor flags + the
+**Still open (follow-up issues):** the remaining §2 structural tables (room-in-roof detail,
+**roof windows** — `sap_roof_windows`) + §3 nested-wall fields
+(`SapAlternativeWall.u_value`/`wall_thickness_mm`) + the
 solar-HW collector columns (`solar_hw_collector_orientation` / `_pitch_deg` / `_overshading`, §3.1).
 Each is on the structural-guard allow-list with a reason until its columns land.
 
 ---
 
+## Slice 2 — PV arrays + floor heat-loss flags (#TBD)
+
+Closed two **scoring-relevant** round-trip gaps that were corrupting the modelling baseline by up
+to ~12 SAP points on a real Hyde property (725634, an ASHP + solar dwelling whose DB-read picture
+scored ~12 points below the in-memory cert it was modelled from):
+
+- **§2 `epc_photovoltaic_array` child table** — `sap_energy_source.photovoltaic_arrays` (a list of
+  `peak_power` kWp + `pitch`/`overshading`/`orientation` SAP codes) had no table, so every array was
+  dropped on save and the dwelling lost its PV generation credit. Now persisted (one row per array,
+  ordered by `array_index`) and reconstructed in `_to_energy_source`. **Round-trip test asserts
+  deep-equality** (`test_photovoltaic_arrays_round_trip`).
+- **§3 `epc_floor_dimension` flags** — `is_exposed_floor` + `is_above_partially_heated_space` had no
+  columns, so a `True` flag round-tripped back to the `False` default and silently flipped the
+  floor's heat-loss path. Now persisted and reconstructed in `_to_floor_dimension`. **Round-trip
+  test** `test_floor_dimension_heat_loss_flags_round_trip`.
+
+DB migration: FE branch `feature/epc-pv-and-floor-heatloss-schema` (additive — new table + two
+`NOT NULL DEFAULT false` columns; no backfill). Run `drizzle-kit generate`/migrate **before**
+deploying this backend code.
+
+### Recursive-guard gaps (newly surfaced)
+
+Generalising the ADR-0036 structural guard to recurse into **nested** dataclasses (it previously
+saw only `EpcPropertyData`'s top-level fields — which is exactly how the PV-array and floor-flag
+drops stayed latent) surfaced **7 pre-existing, calculator-read** nested fields with no FE column.
+These are silent-drop gaps of the **same class as the PV bug** and want their own follow-up
+(verify scoring impact per field, then add FE columns):
+
+| Field | Lives on |
+|---|---|
+| `community_heating_boiler_fuel_type`, `community_heating_chp_fraction` | `MainHeatingDetail` |
+| `is_sheltered` | `SapAlternativeWall` |
+| `wall_insulation_thermal_conductivity` | `SapBuildingPart` |
+| `pv_diverter_present` | `SapEnergySource` |
+| `cylinder_volume_measured_l` | `SapHeating` |
+| `air_permeability_ap50_m3_h_m2` | `SapVentilation` |
+
+(Plus dormant, not-calculator-read nested fields: `SapBuildingPart.{floor_u_value, roof_u_value,
+wall_u_value, wall_is_basement, rafter_insulation_thickness}`, `SapAlternativeWall.is_basement`,
+`SapHeating.cylinder_heat_loss` — allow-listed as dormant.)
+
+---
+
 ## 1. Type fidelity — convert `Union[int, str]` code columns to JSONB
 
 These columns hold SAP/RdSAP categorical codes that are **`int` from the gov API** and **`str`