# Continued individual-building assessment, Hearth 1.4.0

The buildings now have fitted rear glazing around service masonry, framed flower boxes, distinct
domestic stations, visible porch guards, ceiling timbers and planted borders. A new lightweight
garden-house program composes inhabited wings around an open rear court. These changes use shared
components and public mounting interfaces. The kernel, transport, preceding extension package and
all 200 earlier tests are unchanged. No serious flaw in the original requested architecture was
found; the defects encountered were component planning and binding errors.

This is self-authored development evidence. It is not an independent audit, a new Pyodide run or a
claim that the buildings match the references' craftsmanship. All eight original lodge images were
opened and compared with the final renders. Reference generators and snapshots were not used or edited.

## Executed results

The full [regression suite](full-regression.txt) passes **230 tests in 194.81 seconds**: the original
200 plus 30 new cases. The new tests cover transformed public mounts, actual furniture functions
and NBT, substrate and overhead support, whole plants, window-box replacement and stale ownership,
permission overflow, headroom preservation, complete rollback, facade limits and stable independent
branches. Existing door, route, lighting, roof, environment, dependency, shared-ownership and
extension tests remain intact.

The [sampling policy](sampling-policy.json) was recorded before the assessments. Every specified
input was attempted once, without resampling failed inputs. Environment expressions and semantic
digests were frozen independently before placement. The older 64-pair and 16-pair holdout recipes
are regression inputs in this continuation. Six new environment expressions crossed with two
structure seeds add 12 independent pairings; those became regression evidence after development.
There is no claim of an untouched external holdout.

| Final batch | Valid inputs | Accepted / attempted candidates | Principal signatures | Contact signatures | Fill range | Seconds |
|---|---:|---:|---:|---:|---:|---:|
| Three retained lodge clients, seeds 0-7 | 24/24 | 48/48 | 18 | 23 | 0-121 | 42.972 |
| Sixteen frozen environments x four locked designs | 64/64 | 84/128 | 4 | 45 | 0-256 | 77.606 |
| Earlier holdout recipe rerun | 16/16 | 16/32 | 2 | 12 | 49-257 | 21.309 |
| New garden-house client, seeds 0-7 | 8/8 | 16/16 | 8 | 4 | 130-202 | 37.346 |
| Six new frozen environments x two locked designs | 12/12 | 19/24 | 2 | 12 | 0-127 | 46.514 |
| Bounded stress, no resampling | 5/16 | 6/32 | 4 | 5 | 74-143 | 7.296 |

Supported inputs total **124/124**. Candidate acceptance is separately **183/248**. There are no
final supported-domain defects or budget-exhausted cases. The stress input success rate is **5/16**;
its other **11 inputs** exceed the declared site-relief bound of five cells. Their measured relief
is 6-19 cells. Both declared contact alternatives reject those inputs. This establishes incompatibility
with the selected contract and strategies, not impossibility for other components. No limit was
widened, locked building moved, entire site flattened or failure silently replaced. External retry
counts are zero; the two contact alternatives and their individual outcomes are retained in each row.

Fill counts include bounded bearing work. The rows also retain excavation, replaced cells, newly
occupied sitework, strategy and runtime. Timings include validation and, for selected inputs,
paired export/audit; they are observations from this host, not performance guarantees. These finite
samples do not prove general validity or broad environmental coverage. Contact adaptation still
uses the existing stepped and pier strategies.

Authoritative final records are [release/results.json](release/results.json),
[release/summary.json](release/summary.json),
[release-prior-holdout/results.json](release-prior-holdout/results.json), and
[release-new-composition/results.json](release-new-composition/results.json).

## Structural variety and identity

The 32 client inputs have **26 observed principal signatures**: retained home 4/8, hall 7/8,
high house 7/8, and new garden house 8/8. The principal measurement uses actual normalized room,
roof and attachment geometry and room connections. Surface materials, planting, names and global
placement do not supply additional principal signatures. Contact is counted separately. The new
20-input batch has ten interior-station signatures and six opening-layout signatures, reported as
detail observations rather than principal architectural variety. No estimate of total combinations
is made.

The [selected-pair audit summary](paired-audit-summary.json) records **22 full artifact pairs**,
**5,635 structural nodes**, **5,213 nonempty-component queries** and **895 inventory-bearing cells**.
Each export is checked against its generating scene for geometry digest, complete nodes, typed
relations, dependencies, operations, port uses, displaced records, reverse indexes, exact current
cell sets, representative queries and all inventory NBT. Parsed block names and explicit property
mappings must agree; only property order is ignored.

The unchanged default command generated **five buildings, 128,055 blocks and 1,005 components** in
`output.litematic` with its matching companion. The seed-1 command with a spaced output path generated
three buildings, 91,211 blocks and 658 components. Both were regenerated and their complete records,
including all **1,663 descendant-cell queries**, matched reload in
[default-pair-audit.json](default-pair-audit.json). Two seed-0 executions with Python hash seeds 3
and 8675309 have the same semantic digest in [replay.json](replay.json):

`71808152734967b7b001211fb66a486f72ae6092a16453ed99b04f7e611ef6a0`

The runnable lodge inspection example printed a real roof-window glass chain and its identical
reloaded query ([inspection.json](inspection.json)). New installation tests additionally recover
window-box -> window -> wall -> room -> building and ceiling timbers -> room -> building. These
relationships arise from actual bindings, including installation outside the host's child program.

The [native import guard](portability.json) generated and validated garden-house seed 2/environment
4100 while forbidding MCIO, NumPy, PIL, SciPy and litemapy imports. This supports the preserved
standard-library generation boundary. It is not a substitute for a fresh browser execution; the
user's earlier external Pyodide results remain historical evidence.

## Visual review and fixes

Five selected complete sources each have three exterior views and two cutaways: home 0, hall 6,
high 1, garden house 2 and wet adaptation input 00. All **25 release images were opened and inspected**.
The [view index](../../previews/building-detail/README.md) and
[per-view findings](../../previews/building-detail/visual-review.json) retain source digests, camera
angles, slice coordinates, strengths and visible limitations. They are raw proc renders, without
paintovers or substituted image generation. Cutaways are geometry-only aids, not completed semantic
artifacts. Cut trunks, severed foliage and open roof sections in them are expected slicing effects.

The prior spruce visibility concern was resolved by mesh and tint diagnostics: selected fences,
tables and chairs had geometry but blended into similarly coloured floors. The
[diagnosis](render-diagnosis.md) preserves those experiments. Production dark-oak guards and domestic
stations are now clearer in the ordinary renders; no renderer patch was needed.

Rear windows now flank chimneys rather than leaving an entire blank facade. Two-cell framed boxes
fit their explicitly enlarged envelopes. Shared exclusions prevent porch and roof intersections.
Ceiling timbers fit beneath actual floor plates while avoiding protected stairs. Kitchen counters,
reading shelves, storage and seating make wider rooms more legible. Planted borders and tiered
conifers connect the building base to its grounds. The new U-shaped composition materially changes
occupied massing and internal room adjacency across seeds, rather than merely changing decoration.

The wet sample preserves the home plan over bed-supported piers and validates the local approach
endpoint. It intentionally omits `GardenCourt` and does **not** include a shore connection or claim
to be a completed settlement. An enclosing network must discharge that external connection obligation.

## Retained development failures

Earlier evidence directories remain untouched. This continuation also retains its own intermediate
logs and artifacts. Directories named `first-*`, `final`, `new-composition` and `prior-holdout` are
historical development runs; only `release*` names identify the final assessed implementation.

The first full run had seven failures and 223 passes ([log](full-regression-first.txt)). Four failures
exposed ceiling timbers bound as a connection, which did not give the intended room ancestry. The
component now uses installation through a zero-replacement room grant. Two failures were new-test
assumptions: a connected garden path legitimately changes its approach endpoint's ownership, and
an invalid declared input is rejected before negotiation. Those assertions were corrected without
changing an earlier test or lowering its requirements.

The remaining failure was the retained settlement seed-4 regression: a steep lower roof could reach
a facade two levels above it. The shared facade planner had considered only the immediately lower
level. It now uses all lower roof profiles, conservatively records optional gable-side omissions,
and preserves required room and roof intent. A 52-test focused rerun and the entire 230-test suite
then passed. The complete corpus was rerun after that fix.

Earlier focused runs also caught a porch/tall-window projection conflict, initial window-box framing
issues and mistakes in new test queries. Logs remain here. These were ordinary geometry-planning,
binding and test implementation defects, not reasons to bypass transactions or change the original
architecture. No reproducible final supported-domain failure is knowingly hidden by selecting a
different seed.

## Remaining quality and capability limits

The release is more complete and varied, but remains **below the lodge references' visual richness**.
Room bays and window rhythms are regular; roof ridges/dormers repeat; chimney caps are heavy; the
landscape lacks the references' dense terraces, rock edges and small-scale storytelling. Warm light
sources exist, but the proc renderer does not reproduce the references' lighting treatment. This
is not solely a rendering gap: finer proportions and more varied architectural forms remain work.

The building planner still uses 9/11/13-cell bays, six-cell floor spacing, bounded gable profiles,
conservative roof abutments and two contact strategies. Optional windows can be omitted behind a
roof envelope. Attics are sealed nonhabitable voids, not accessible furnished rooms. The open rear
court has outdoor access, but no direct interior garden door in this client. Borders have an explicit
optional minimum of zero. Discrete movement, lighting, support and cover checks are not a complete
Minecraft simulation or general structural mechanics. Stress relief failures remain unresolved.

Ordinary clients need no private geometry or manual provenance. New component authors still need
the documented public port and physical-model conventions. No new city planner, browser runtime,
viewer, reference-family dispatcher or second-round connection demonstration was added.

## Rerun

Use fresh directories to retain earlier matching evidence:

```sh
python -m pytest -q test
python generate.py
python generate.py --seed 1 --output "samples/my seed 1.litematic"
python -m examples.verify_replay --seed 0 --output reports/my-replay.json
python -m examples.building_quality --seed 0 --output reports/my-quality --artifacts samples/my-quality
python -m examples.building_holdout --seed 0 --output reports/my-prior-holdout --artifacts samples/my-prior-holdout
python -m examples.building_detail --seed 0 --output reports/my-detail --artifacts samples/my-detail
python -m examples.lodge_inspection --seed 0 --environment-seed 0 --output samples/my-inspection.litematic
python -m examples.building_views --seed 2 --input samples/building-detail/release-new-composition/court-2.litematic --output previews/my-court --size 1200
```

The public additions and compatibility details are in [api-delta.md](api-delta.md), and every original
requirement is mapped in the supplemental ledger in [GENERATOR.md](../../GENERATOR.md).