# Frozen sampling policy for the roof-space continuation

Written before the assessment, after the retained development probes. All prior reports remain
historical evidence. Self-authored development and testing are not an independent audit.

Supported principal inputs: the existing connected 9/11/13-cell room grid with supported storeys,
cardinal orientations, roof rise courses 1 or 2, dormer spacing zero or 7..16. Habitable roofs require
an explicit policy offering `roof-interior`. The supplied loft fits bedroom, library, or storage
functions, at least two reachable furniture groups, actual lights, and a stair landing aligned with
the ridge. Verandas may span contiguous front rooms; posts remain at most 12 cells apart. A balcony
also requires upper rooms behind its complete span. Existing support, terrain edit, and search limits
are unchanged. Unknown environmental space is still an error.

Sampling root seed is explicit (release assessment: 0). No failed input may be replaced by another.

1. Re-run the existing `examples.building_quality` recipe into this report directory: 24 building
   inputs (home/hall/high, seeds 0..7), 64 independently frozen environment/structure pairings
   (16 environment expressions times four locked designs), and all 16 bounded stress inputs.
   The same ordinary components now compose inhabited roofs. Retain every result and candidate.
2. Run `examples.building_space`: eight roof-home inputs at seeds 0..7; eight new frozen general
   environments times structure seeds 2 and 7 (16 pairings); and eight held-out compositions. The
   latter join three front rooms, a rear kitchen, and an upper room at a seeded front position,
   under a three-bay veranda and inhabited roofs. These combinations were not primitive-development
   examples. They keep explicit room functions, upper placement, and entrance locked.
3. The eight new environments use bounded gradients, phase-shifted waves, a radial field, and a
   radial blending mask, clamped to elevations 2..6. Water is independently selected from None, 4,
   and 5. Generate and hash all terrain before placing a building. Structure choice cannot change
   the environment expression. The held-out structures use these frozen inputs once each.
4. Keep the previous garden-house client through the full regression suite and the existing
   `examples.building_detail` recipe (8 seeds + 12 independent pairings), in a new output directory.

For each input record success, established incompatibility, or search-budget exhaustion separately
from defects. Record attempted and accepted contact candidates, retries, runtime, fill/excavation
costs, principal signature, and contact signature. Principal measurements omit material values,
instance names, global transforms, planting, light placement, furnishings inside the roof, and
optional porch seats. Room purpose/topology, floor shapes, structural roof shapes, veranda spans,
and actual major attachments remain. Contact signatures cannot substitute for principal variety.
The original sixteen stress environments are retained even when incompatible with unchanged bounds.

Artifacts are selected before execution: roof-home seeds 0, 2, 4, 7; new pair inputs (0,2), (3,7),
(7,2); held-out inputs 0 and 7; and the existing quality/detail recipe's fixed selections. The render
set includes long and split roofs, a tall house, a held-out veranda, and wet support adaptation.
Each complete pair is validated after reload and its structural records and current-cell queries
are compared. Preview slices are geometry-only evidence. Render at least three exterior directions,
ground interiors, roof interiors, and circulation sections, then inspect them.

A finite corpus is not a proof for every permitted composition, nor a proof of aesthetic quality.
Full existing tests remain mandatory. Native semantic replays use distinct PYTHONHASHSEED values;
this environment has no Pyodide runner, so preserve portability and report that limit explicitly.

Before release assessment, also retain the existing `examples.building_holdout` recipe as a
regression check: eight frozen transformed/blended fields times the same two older locked designs
(16 additional pairings). They are historical holdouts, not fresh independent evidence. They do not
replace any result from the new held-out veranda compositions.