# Building quality development evidence

Initial architectural review found no serious blocker. The kernel has not been edited in this phase. Read `review.md` and `baseline-hashes.json` for the starting review and module identity.

All eight original lodge images were opened as aesthetic references only. Source snapshots were not used. The first fresh baseline and superseded development pairs/previews remain in separate directories.

Observed defects and general fixes:

- Dormer caps initially climbed the main roof too far. The mount now describes a five by three opening with a level local ridge; the scoped cut opens into an explicitly nonhabitable attic. Gable walls, dormer faces and glass are actual nested components.
- Porch posts conflicted with projected shutters. The rear bearing is the actual host wall; the front columns and upper deck have actual support rules. Sheltered facades suppress competing relief. Balcony guardrails are checked as actual blocks.
- Raised facade beams were mistaken for ground by approach elevation queries. The approach now measures below a bounded overhead exclusion, retains nearby upward ground steps, and the building's route check includes the public street endpoint.
- Roof abutments intersected upper shutters and vegetation. The shared roof-run planner now informs the facade planner: walls meeting a lower gable receive no concealed window, and walls above a lower eave use a flush opening. This is an architectural adjacency rule, not a seed special case.
- A sparse tree candidate grid missed legal planting locations. Landscape now exposes its sampling step and height domain; domestic courts use a two-cell grid while retaining three required trees and the existing separation rule.
- Wider raised beds exposed missing internal substrate above perimeter footings. A bounded soil lift now supplies actual continuous support under the planting cells; stone edging remains a separate supported component.
- The first recorded assessment validated all 24 clients but only 31 of 64 adaptation inputs. Every one of the 33 adaptation failures contained an actual chimney-support diagnostic in the pier candidate; these are implementation defects, even though the bounded solver surfaced them as `budget-exhausted` alongside an incompatible stepped strategy. A small general bearing-course adapter now checks actual anchors and a taxicab span bound of two, filling at most six shaft-bearing cells. No site limit, building position, principal room layout, or validator was relaxed. The original failed assessment is retained.
- Final visual inspection exposed porch fence arms pointing beyond the rail line. Connections now follow actual neighboring rail cells; all four rotations verify both connected neighbors and absence of dangling arms. The preceding successful corpus, pairs and views remain under `before-railing-*`. All assessments, CLI replay and regression checks were rerun after this correction.
- The new inspection CLI successfully generated/exported/reloaded but initially failed to print a set-valued query as JSON. It now serializes sets in sorted order, matching the existing inspection example. Its final recorded invocation exits successfully.

The roof weather check starts at each intended course height and examines actual geometry, while exact dormer boundaries are checked separately. Eaves outside the principal roof footprint can be trimmed against existing geometry. This is not a fluid simulation. Chimneys are supported exterior masonry flues, not a simulated heating system. Attic voids are not counted as usable rooms.

Some balcony fence segments are poorly represented in the full proc renders. The decoded schematic
contains the correct guards, confirmed both by MCIO and the renderer's own parser (`proc-probe.json`).
An isolated fence render is retained as `previews/building-quality/render-diagnostic.png`. The exact
cause of the full-scene visual discrepancy is unresolved; do not use those pixels as proof of guards.
Actual-state validation and deliberate missing-guard tests are the structural evidence. The renderer
and its assets were not edited, and no schematic geometry was substituted to improve a picture.