Chapter 25: A lived-in town Frozen 5 October 2026 at 17:45 BST (game main at 648feab). Implemented, reviewed and merged parts of the P7 stage (a lived-in town), which is NOT closed: P7a1 pane_glass split and seeds with its review fix; P7a2 night room stories, day glint and golden blend with its review fixes and recheck (right-block line waived at 3 by the orchestrator); P7b CPU emitter mirror, unlit panes and window families with its review fix; P7c roofscape I (sites, crowns, palette pieces, roof_dressing) with its review fix; P7d roofscape II (drawing the pieces and release cost) with the second review's corrections; P7h ivy grid fix. Owner answers: the P7b and P7d sheets approved on 5 October; orchestrator rulings on P7a2 and before P7d. Also quoted: the questions on the P7b and P7d owner sheets. Not done at the freeze: P7e ivy spill (running), P7f window boxes and Juliet rails, P7g downpipes and vent stacks, and the P7 stage close. Original source terminology is preserved. Line numbers refer to the frozen copy. Part codes: P7a1 glass split, P7a2 room stories, P7b emitter mirror and unlit panes, P7c roofscape sites and pieces, P7d drawing the roofscape, P7h ivy grid. SOURCE: docs/visual-handoff.md (frozen copy, 5 October 2026, 17:45 BST, game main at 648feab) 5: **Beautify P7h Ivy grid fix (5 October 2026):** Opus 5.5 on this Mac (report "Ivy: row stagger, JITTER 0.10, larger leaves, flat lobes, growth tint"; trees lens finding 4 revised; moved from P10c). Derived ivy no longer reads as a grid of separately outlined rosettes. Odd rows stand `STAGGER` (CELL/2) along the face from their column; on a tower they stand that far round the shell. `JITTER` goes from 0.04 to 0.10, and leaves above the foot get a seeded size from 0.95 to 1.2. The rosette's lobe normals bend 0.7 toward a flat ellipsoid behind the face, with no up lean (`crown_normals(..., flat)`), so the foliage rim falls on the patch silhouette. In each seed's growth order, the newest 25% of leaves are young (their own hue lightened 0.08) and the oldest 25% old (`IVY_OLD` 27634b). The tint is an instance colour on the existing ivy MultiMeshes, so no MultiMesh or draw call is added. `SHAPES` stays 1. Nothing is saved; scene version 2 is unchanged. 6: - **Deviation from the plan line "clearance LEAF_HALF x 1.2":** with every cell held to the largest leaf's clearance, the garden fell from 313 leaves to 109 and the lane to 0 (first attempt, 3 October). So which cells hold a leaf still follows the pre-P7h rule (`_old_fit`, read by `valid`), and seeding, growth and counts are unchanged. A grown leaf is drawn the P7h way (staggered, sized, `JITTER`) where that leaf is clear of every opening, sign and cover (`fit`, which also tests the tower shell's holes and windows where the leaf is drawn). Otherwise it is drawn as the pre-P7h leaf (`SMALL_*`). Small leaves per owner: 0 to 4 on Harbour, 0 or 1 on the quarter (probe `h/probe/owners-after.log.out` (3 October 03:44, one minute before the last ivy.gd edit; the reviewer re-ran `probe/owner_counts.gd` on the live tree into `review/owners-live.log.out` and the output is identical), 3 October); 18 of 100 odd-row leaves on the test tower. `half_of` is the P7h bound as the plan states it (nothing calls it any more: placement uses `_small_half` for the cell choice and each leaf own LEAF_HALF x size for the fit, so the check on it pins the plan value only) (LEAF_HALF x 1.2 + JITTER) and `REACH` widens the rectangles a column collects. 7: - **Side fix:** `Ivy._blocked` copied the column dictionary shallowly, so cover and box rectangles were appended to the memoised column's own `rects`. They kept blocking after the thing that made them had moved (Harbour: a house moved 0.5 m and back left the round tower's rows behind it shut). The column now gets its own copy. `test_ivy` memo_checks caught it through the staggered tower rows (in the 3 October intermediate code; the review found that with the final one-line fix removed the memo checks still pass, so no test guards it today). 8: - **world_view.gd** is outside the plan's file list. It is needed: `_sync_ivy` sets `use_colors` and `use_custom_data` on the ivy MultiMeshes and gives each instance `WorldView.ivy_colour(piece)` (the growth tint; moss stays white). Custom data matches the shader warm-up's draw. The P6a COLOR.a stamp passes through (instance alpha 1), and every ivy leaf vertex keeps a valid stamp (tested). The vine, `ivy_drape` and moss meshes are not touched. 10: Files: `scripts/ivy.gd`, `scripts/architecture.gd` (`crown_normals` flat flag, `IVY_*` constants, `ivy_tint`, rosette at the `"ivy"` decoration), `scripts/world_view.gd` (`_sync_ivy`, `ivy_colour`), `tests/test_ivy.gd` (`git diff --numstat`: 439 added, 43 removed). Snapshot: `build/backup-feel/25-P7-h/` (main 29f806f). Evidence: `artifacts/beautify/25-P7/h/`; the pre-part tree is `h/pre-tree`; backups in `h/backup/`. Resumed from an attempt cut off on 3 October: its code and tests were kept. Its captures, metrics and bench predate its last `ivy.gd` edit, so they were moved to `h/superseded-0303/` and everything was re-run in one sitting on 5 October. 11: - **Tests first:** `h/testfirst/test_ivy.log(.out)`: 174 checks, 11 failures (JITTER, stagger, sizes, clearance, flat normals, tint). After: `test_ivy` 187 checks, 0 failures (`h/suites2/test_ivy.log.out`). Full suite (`tools/suite_all.sh`, VERIFY_RES=2880x1800): 42 suites, 22092 checks, 0 failures. `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun passed (`h/full/suite_all.out`). The 1440x900 `--verify` stall is inherited and was not run. 12: - **Changed expectations:** none. The new checks are `grid_checks`, `tower_checks` and `tint_checks`, and no existing assertion changed. Expected journey count deltas: none. Integer counts, draw calls included, are identical on all 8 part views in all three rounds. Ivy counts are unchanged: Harbour 395, quarter 369 with moss 14, lane 28, garden 313, chemistry 608. 13: - **Captures** (`h/cap/pre-1..3`, `after-1..3`, alternated in one sitting, AC power, Low Power Mode 0): harbour-night, harbour-day, harbour-lane, quarter-night, quarter-day, styles-day, lane-day, lane-night. Before/after crops: `h/look/lane-day-tower.png` (a 3x4 grid of rosettes becomes an overlapping, staggered patch with lighter young tips), `lane-night-tower.png`, `quarter-day-west.png` (the dark diamond lattice loses its rings; a weaker diagonal row pattern still shows at this zoom), `quarter-day-quay.png` and `harbour-day-tower.png`. The systematic-change maps `h/look/sys23-*.png` (rounds 2 and 3, change over 8 luma in both cross pairs and within 2 in both noise pairs) show change only on ivy patches. Systematic pixels: quarter-night 0.61%, quarter-day 0.49%, harbour-lane 0.54%, lane-day 0.45%, styles-day 0 (it has no ivy). Pre-1 was an outlier against pre-2 (pre-1 to pre-2 changed 35% of the harbour-day frame, while pre-2 to pre-3 changed 17% and after-2 to after-3 7%), so round 3 was added and rounds 2 and 3 are the measure. `h/metrics/`. 14: - **Acceptance:** there is no regular grid. The leaf-mask autocorrelation rebound goes from 0.47 to 0.10 (x) and from 0.26 to 0.06 (y) on the lane-day tower, and from 0.17 to 0.05 and from 0.21 to 0.12 on the quarter-day west wall (`h/metrics/ivy-metrics.txt`). There is no ring round each rosette: leaf pixels under luma 40 fall from 62% to 39% on the quarter-day west wall and from 41% to 32% on the lane-night tower. No leaf is over an opening: test_ivy's opening checks pass, and on the test tower 0 leaves are over a hole or window. Ivy counts are identical. 15: - **Bench** (same sitting, AC power, charging, Low Power Mode 0, `h/bench/power.txt`). The release probe ran 3 alternating runs (`h/bench/release-summary.txt`). Harbour release frame, anim off: 26.54 to 26.91 ms; anim on: 32.54 to 32.35 ms. Both are within the run spread. Harbour spread total went from 58.5 to 60.1 ms, its steps ivy_columns, ivy and ivy_draw from 7.76, 5.31 and 1.48 to 8.06, 5.74 and 1.98 ms, and its anim-off release from 15 to 16 slices. Dense release frame: 118.76 to 119.08 ms, with the ivy step from 18.18 to 19.57 ms. Smoke runs used `--disable-vsync` and rounds 1 to 3, with round 0 discarded (`h/bench/smoke-summary.txt`). Dense raw_p95: 15.754 to 15.770 ms. Harbour raw_p95: 9.276 to 9.269 ms. Draw calls are 2362 and 903 in both trees. Cold `Ivy.grow` in `test_ivy` is 48.6 ms on Harbour, against 39.8 ms in the test-first run (a different sitting). The same-sitting probe from 3 October measured 43.2 to 47.4 ms cold and 32.8 ms warm on both trees. 16: - Review: filled by the orchestrator. Not done: SHAPES 2 was not added (the plan allows it only if draw calls allow, and the report did not ask for it); there is no owner sheet. Platforms: Mac run; Windows not run. 18: **Beautify P7d Roofscape II: drawing the pieces and release cost (3 October 2026):** Opus 5.5 on this Mac (report Signature D "everything drawn as clutter synths into the existing MultiMeshes" and the release-time risk; measuring part, second review required). The view now draws P7c's derived roofscape: each cluster's crown and every roof's palette pieces are decorations, recomputed on full syncs and releases only, memoised per cluster. D25 default applies (+2 ms over the Harbour animated release, memoised per owner; orchestrator ruling before P7d). **This is the owner stop after P7d: the roofscape sheet is `artifacts/beautify/25-P7/d/sheet/` (`sheet.md` and 7 pages).** 19: - **Drawing** (`scripts/world_view.gd`): 20: - `_refresh_roofs(hold)`, beside `_refresh_clutter`. A full sync takes `Composition.roofs`. A drag step (`sync_live`, a dab, the spread release frame) keeps the last pieces less those of an owner whose params or transform changed (`roof_marks`), so it never runs the pass. 21: - The spread release runs it as two queued steps after the clutter: `roofs_pass` (sliced, `roofs_begin/roofs_step`) and `roofs`. `roof_hash` is in the decoration signature. 22: - In `_sync_decorations` a piece of a kind the view already draws joins one of that kind's plain MultiMeshes (picked by its seed; never a group that draws a float or plinth stand), drawn as that group's variant and seed class. A kind not drawn yet gets exactly one new MultiMesh. 23: - The cluster pass's `state.clusters()` result is handed to the roofs pass while the objects are unchanged (`_known_clusters`). This saves 1.1 ms (Harbour) to 4.3 ms (dense) of the pass. 24: - **Memo** (`scripts/composition.gd`): `_roofs_memo` per cluster of owners. The signature hashes the ground and every object in the neighbourhood: bodies and parts within `ROOF_MEMO_REACH` 2.0 m of the cluster's boxes, landings, its wall strips and its bodies' children. `roofs()` and `roofs_begin()` take optional clusters. The stats are `roofs_cluster_hit` and `roofs_cluster_computed`. 25: - **Fixture:** `artifacts/beautify/25-P7/d/fixture/dense-bare.json` is dense with its 500 hand-placed roof parts removed (`make-dense-bare.js`). It is added to `tools/repo_evidence.sh`. 27: Files: `scripts/world_view.gd`, `scripts/composition.gd`, `tests/test_roofs.gd`, `tests/test_release_spread.gd`, `tests/test_clutter.gd`, `tools/repo_evidence.sh`, `README.md` (`git diff --numstat`: 437 added, 14 removed). Snapshot: `build/backup-feel/25-P7-d/` (main aadac85). At part start the live tree matched it (`d/diff-live-vs-25-P7-d.txt`, empty). Evidence: `artifacts/beautify/25-P7/d/`; backups in `d/backup/`; the pre-part tree is `d/pre-tree`. 28: - **Tests first:** `d/testfirst/test_roofs.log(.out)`: 184 checks, 2 failures (the missing `_refresh_roofs` and memo stats). `test_release_spread` passed 130 of 130 before the code (its new checks guard against regressions). 29: - **Tests after:** 30: - `test_roofs` 182 → **259** checks (`d/suites/`). The new checks cover: 31: - the release step order; 32: - each piece drawn at its deck point and turn in a decoration MultiMesh; 33: - no new MultiMesh for a kind already drawn (Harbour 93 → 93, quarter 135 → 135) and one per new kind (styles 28 → 31: tank, planter, utility box); 34: - the view's pass equals a cold pass; 35: - a drag step runs no roofs pass and hides only the moved roof's pieces; the release redraws them; 36: - an unchanged sync is a cache hit; `roof_dressing` false clears the roof; 37: - foliage within FOLIAGE_BUDGET on dense (874 of 3000) and dense-bare (774); 38: - the memo equals a cold pass after five edits on Harbour, the quarter and dense-bare. 39: - `test_release_spread` 108 → **130** (dense-bare added to the spread-against-whole checks). 40: - **Full suite** (`d/suite/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0): **42 suites, 22,053 checks, 0 failures.** `--verify` 2880x1800 exited 0 on try 1 with `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun was clean (the known flaky check did not fire). 1440x900 was not run (inherited stall). 41: - **Changed expectations (old → new):** 42: - `test_release_spread`: "the release frame does not run the clutter, links, foliage, planting or ivy passes" now also counts `roofs_computed`/`roofs_hit`; the digest gains `roofs`; dense-bare is added to the scenes. 43: - `test_clutter` `drawn_clutter()`: negative-id decorations, less planting, **and less roof pieces** (a site and a role). Its first run failed "a2: every other opening keeps its last drawn clutter during the drag", because the dragged house's roof pieces counted as clutter. That log was overwritten by the green rerun. 44: - **Captures** (`d/caps.sh`: the six journey views, pre-1, after-1, pre-2, after-2 in one sitting, before from `d/pre-tree`; all ok on try 1; `cap/power-*.txt` AC, LPM 0): 45: - Systematic pixels (`metrics/systematic-8-2.txt`; maps `look/sys-*.png`, looked at harbour-night, quarter-night, quarter-day): harbour-day 0.24%, harbour-lane 0.79%, harbour-night 0.11%, quarter-day 0.19%, quarter-night 0.16%, styles-day 0.33%. 46: - Most sit on the pieces; the second review found that some do not: in styles-day 2,640 of 5,308 systematic pixels are rain rings and streaks that shifted phase between the trees, and harbour-day has the same pattern (none touch a building) (harbour-night: the utility box on the main terrace and the tank by the Fish Market; quarter: three spots; noise pairs `metrics/noise-*.txt`). 47: - **Roofscape sheet** (`d/roofcaps.sh`, closer cameras, `roofcap/pre-1` and `after-1`, 02:35, AC; `d/sheet.js`; every page looked at): 48: - Harbour: 2 pieces (a tank crown on roof 10, a utility box beside the placed tank). The Fish Market roof has no crown (P7c, kept). 49: - Styles: a tank crown, a planter and two AC boxes, almost touching. 50: - The quarter: 15 sets (17 pieces), all 15 on the 3.0 m podium plazas (roofs 2, 4, 5 and 12, `probe/sites.log.out`; corrected from "13 sets" by the second review): pergola, umbrellas over tables, vents, AC boxes, planters, roof gardens. Shown from above, from the far side and at the usual angle, where most of them hide. 51: - **Journey count deltas expected** (`metrics/counts-pre1-after1.txt`): 52: - harbour-night, -day, -lane: none (both pieces join existing MultiMeshes and carry no light); 53: - quarter-day and quarter-night: emitter_sources 203 → 216, paving_streaks 181 → 194, reflection_streaks 203 → 216; 54: - styles-day: draw_calls 470 → 491, rendered_objects 546 → 567. 55: - **Acceptance** (all in one sitting against `d/pre-tree`, AC 100%, LPM 0, `bench/power.txt`): 56: - **Release rise within the owner's limit (+2 ms): met on Harbour.** Release frame, median of 3 alternating runs of 9 moves (`d/bench.sh release`, `bench/release-summary.txt`): 57: - Harbour with animations 31.63 → 31.71 ms (+0.08); without, 25.95 → 26.22 (+0.27). This sitting measures the unchanged tree at 31.6, not the plan's 26.7 (a 24 September sitting; P6b measured 30.6). 58: - Dense with animations +0.76. **Dense-bare +2.16 with and without animations**: the release frame's decoration rebuild draws 284 more instances. 59: - The spread work after it: the roofs steps take Harbour 3.1 ms, dense 14.7 and dense-bare 6.7. The Harbour spread total is 54.8 → 58.0 ms (14 → 15 slices) and its worst slice 7.60 → 7.61. 60: - **Dense p95 within +10%: met.** Raw p95 (`d/bench.sh smoke 1`, 1920x1080, `--disable-vsync`, rain 0.45, rounds 1-3; `bench/smoke-summary-t1.txt`): 61: - dense 15.077 → 15.140 ms (+0.4%), draw calls 2362 = 2362; 62: - dense-bare 13.672 → 13.934 (+1.9%), draw calls 2237 → 2270; 63: - Harbour 9.179 → 9.333 (+1.7%), draw calls 903 = 903. 64: - **Drag within 2 ms: met.** Harbour harness `step_total` median (`bench/drag-summary.txt`, rounds 1-3, no errors): handle 18.47 → 19.09 (+0.62), wall stroke 8.90 → 9.68 (+0.77), terrain raise 128.9 → 130.7 (+1.8). Harness draw calls 883 → 885 (not explained in the part). 65: - **Journey counts change only by the roof synths: met** (the list above; no other draw or instance count moved; the second review also saw `light_select_calls` and `audio.listens` go 4 → 5 in both styles-day pairs, timing-side fields). 66: - **Foliage within FOLIAGE_BUDGET on dense: met** (874 of 3000; the roof kinds are not foliage kinds). 67: - **Memo** (`d/suite/suites/test_roofs.log.out`, `ROOFS_MEMO`, and the profile `d/probe/roofprof-1.log.out`): 68: - dense-bare, one house moved: 96 clusters reused, 4 recomputed, 9.2 ms against 63.5 cold; 69: - Harbour: 8 of 9 reused, 3.9 against 4.7 ms; 70: - the quarter is one cluster (22 ms, no gain); 71: - what is left on dense is the preparation that reads every object (clusters 4.3 ms, part chunks 9.1 ms). 72: - An optimisation round was taken (the clusters hand-off: Harbour roofs steps 4.2 → 3.1 ms, dense 18.5 → 14.7; first try in `bench/try1/`). 73: - Review: filled by the orchestrator. 74: - **Not done:** 75: - No 1440x900 `--verify`. 76: - No editor toggle for `roof_dressing` (D26 default). 77: - The part chunks of the preparation are not memoised (9 ms on dense, spread). 78: - The dense-bare release-frame rise (+2.16 ms, a stress fixture) is reported, not reduced. 79: - The scratch tree `d/pre-tree` is left in place. 80: - **For the owner:** the quarter's podium sets, the Fish Market roof without a crown, and the two AC boxes almost touching on the styles roof. 81: - Platforms: Mac run; Windows not run. 83: **Beautify P7c Roofscape I: sites, crowns, palette pieces, `roof_dressing` (3 October 2026):** Opus 5.5 on this Mac (report Signature D "roofscape grammar": a crown accent on each cluster's tallest roof, then clamp(A/7, 1, 6) style-palette pieces, the `roof_dressing` opt-out; architecture lens F1). The composition layer now derives a roofscape for every flat roof: its site, one crown per cluster, and its palette pieces. **Nothing draws it yet** (P7d does), so no picture changes. D26 (plan default) applies: the param has no editor toggle; it is JSON only, with a README note. The owner gate before P7d (release-time limit) is next; no P7d work was started. 84: - **Sites** (`Composition.roofs`, `roofs_begin`/`roofs_step`): 85: - Every flat building or tower's exposed deck, as `ClusterWalls.caps` cuts it for its cluster. A lone terrace is a cluster of one, so lone and merged roofs take one path. 86: - The pieces `caps()` cuts one roof into stay one site: its polygons, and its boundary edges with the seams split out. 87: - Each site carries its area, storeys over the ground, top and deck heights, and the tallest flag: one per cluster, by highest top, then larger area, then lower id. 88: - Every boundary edge carries its parapet drop and what it faces: ground, water, a lower roof (with the drop to it), or wall, where a taller body stands and there is no parapet. 89: - **Crown:** one per cluster, on the tallest roof. It is a mast or a tank by seed, at a convex corner `CROWN_IN` 0.8 m from both edges, clear of every blocker. A mast or tank already placed on that roof counts as the crown (`crowned_by`), and none is derived. An opted-out tallest roof keeps the crown off its neighbours too. The tallest site records what became of its crown (`crown`: derived, placed, opted out, or no clear corner), and `roofs()` lists the tallest roofs with no clear corner in `crownless` (review fix). 90: - **Pieces:** n = clamp(round(A/7), 1, 6) sets from the style palette (canal: roof garden, planter, umbrella over a café table, a vent twice, utility box, pergola; victorian: planter, chimney; modern: utility box, planter, pergola; medieval: flag, planter; japanese: bonsai, planter). 91: - **Deviation 1:** pieces placed on the roof by hand count toward n (`wanted` = n − placed, never below 0), so a hand-dressed roof is not doubled up. 92: - They are packed through `pack()` in the building's own frame, wholly inside the 0.55 m inset (`pack` takes a new optional `fits` callable; no existing caller passes it, and the random sequence is unchanged). 93: - They keep clear of: the stair head of a canal terrace over 40 m², each terrace flight's head, connection landings, every placed part on the deck (the chimney blockers, planters, pots and the rest), and the cat's spot. A placed part keeps clear its oriented footprint, a capsule along its longer side; a wall part, or a part off the deck, that reaches no more than `ROOF_PARAPET_BAND` 0.35 m into the deck hangs on the parapet and is no blocker (listed in the site's `hung`) (review fix). The cat's spot comes from the new `Creatures.terrace_spot`, which the cat's own nap search now calls too (a refactor). 94: - **Deviation 2 (not in the plan's list):** they also keep clear of the doors of bodies standing on the deck (a stacked room's door, 0.9 m round a point 0.7 m out). 95: - A set that finds no room is replaced once by the palette's smallest. 96: - **`roof_dressing`** on a building or tower: absent = automatic; `false` = bare; `true` = automatic; a non-negative integer reseeds. `validate_document` refuses anything else ("Invalid roof dressing."), and `restore` turns a saved integer back into an int. 97: - **Job:** the work runs in units of one cluster's caps, one site, or one site's dressing, with the objects read in chunks of 64 under the new `DioramaState.shape_pass()` (parents' boxes memoised for one unit). P7d can therefore put it in the spread release. Caching uses a hash of objects, terrain and heights (`roofs_computed`/`roofs_hit` in `cache_stats`). 99: Files: `scripts/composition.gd` (Roofs section; `pack` optional `fits`; `reset_cache`), `scripts/scene_state.gd` (validation, normalisation, `shape_pass`), `scripts/creatures.gd` (`terrace_spot`), `tests/test_roofs.gd` (new, in `tools/test.sh` and `tools/Test.ps1`), `README.md`. `git diff --numstat`: 612 added, 13 removed in tracked files, plus 572 lines of new test. Snapshot: `build/backup-feel/25-P7-c/` (main 4091ce7). At part start the live tree matched it (`diff -rq` on scripts, shaders, tests, data, scenes and tools: empty, `c/diff-live-vs-25-P7-c.txt`). Evidence: `artifacts/beautify/25-P7/c/`; backups are in `c/backup/`; the pre-part tree is `c/pre-tree`. 100: - **Tests first:** `c/testfirst/c-testfirst-test_roofs.log(.out)`: 11 checks, 11 failures; every section is gated on the missing functions. 101: - **New suite `test_roofs`: 182 checks** (159 before the review fix). It covers: 102: - constants and palettes (existing kinds only); 103: - a lone house: 30 m², one storey, four parapets dropping 2.6 m to the ground, n 4, one crown; 104: - a turned house; 105: - a quay house with parapets over both water and ground; 106: - two wall-to-wall houses one storey apart: one cluster, the low roof's wall side, a 2.6 m drop onto the low roof, one crown; 107: - overlapping equal roofs: union 48 m², one tallest; 108: - a stacked room: the terrace is one 36 m² site whose boundary is 42 m with 14 m of wall, and the crown goes on the room; 109: - a covered terrace is no site; 110: - the crown at 0.8 m on 12 seeds, both kinds; 111: - a placed tank acts as the crown; 112: - counts by area (n 1, 2, 4, 6, 6); 113: - every style draws from its own palette; 114: - the umbrella stands over its table; 115: - each blocker kind, including a stacked room's door, kept clear over 8 to 10 seeds; 116: - the param: false, true, six integers giving at least 4 layouts, validation, the JSON round trip; 117: - Harbour, styles and the quarter save byte for byte the same after `roofs()`; 118: - cache, determinism, invalidation, and the sliced job; 119: - the four stock scenes keep every invariant; 120: - (review fix, `band_checks`) a facade sign and string lights at the roof are hung, not blockers, and the house keeps the bare house's crown and pieces; a turned bench is a capsule along its long side; a fence inside the deck is still a blocker; Harbour's Fish Market roof is crownless only because parts on its deck hold all four corners, and is crowned beside its sign once they are lifted. 122: The suite runs in about 2 s. 123: - **Full suite** (`c/suite/suite_all.out`, VERIFY_RES=2880x1800; AC 100%, Low Power Mode 0): **42 suites, 21,931 checks, 0 failures** (before the review fix; after it, see Review fix). 124: - `--verify` 2880x1800 exited 0 on try 1 with `PIXELPUNK_INTEGRATION []`. 125: - The creature_reactions rerun was clean; the known flaky check did not fire. 126: - 1440x900 was not run (the inherited stall). 127: - Counts: 41 → 42 suites and 21,772 → 21,931 checks (+159 `test_roofs`); no other count moved. 128: - `git status` after the suite shows no tracked artifact rewritten. 129: - **Changed expectations:** none. One comment in `scene_state.gd` was reworded so that `test_composition`'s source check ("scene_state.gd does not call the composition layer", which greps for `Composition.`) still holds; its first run failed on that comment (`c/suites/aff-test_composition.log.out`). 130: - **Debug images** (looked at): `c/look/roofs-harbour.png`, `roofs-styles.png`, `roofs-quarter.png`. `test_roofs` rewrites them into the untracked `c/debug/` on every run, so the copies in `look/` are the evidence. They show: 131: - decks in grey, with the tallest in warm grey; 132: - parapets coloured by what they face; 133: - blockers as rings; 134: - the crown as a gold disc and the pieces as white rings. 135: - **Stock scenes** (`c/suite/suites/test_roofs.log.out`; per site in `c/probe/roof_probe-2.log.out`): 136: - **Harbour:** 5 sites in 3 clusters. One derived crown (a tank on the 3x3 m roof 10) and one placed crown (tank 47 on roof 13). **The third cluster's tallest roof, the Fish Market (roof 9, 6x4 m), has no crown** (`crownless` ["roof:9"]): its four corners 0.8 m in are held by the planter, flower patch, umbrella, chimney, flag and the cat's spot, all placed on its deck. Before the review fix its FISH MARKET sign was also a 3.68 m disc over the whole deck. One piece (a utility box on roof 13) of 3 wanted. Most of its roofs are already dressed by hand (11 placed parts counted), as the report's challenge foresaw. 137: - **Styles:** the modern roof gets a tank crown, a planter and two utility boxes. 138: - **Quarter:** one cluster of 18 sites. No derived crown (4 roofs are crowned by placed masts or tanks). 15 sets (17 pieces) of 16 wanted, almost all on the 3.0 m podium decks 2, 4, 5 and 12. Roof 2, for example, gets a pergola, an umbrella over a table, two vents, a planter and a utility box. 139: - **Dense:** 100 sites and 0 pieces, because every terrace already carries 4 placed parts and an antenna. 140: - **Acceptance:** 141: - **`test_roofs` green on Harbour, styles, the quarter and dense: met.** 142: - **Roof composition time on dense, cold:** 38,037 / 38,189 / 39,137 µs over three probe rounds (`c/probe/prep_probe-2.log.out`) and 40,314 µs in the suite. It runs in 311 units; the largest is 3,961-4,441 µs, which is the `state.clusters()` unit (3.9-4.2 ms alone). The quarter takes 20.8 ms in 43 units, the largest 4.7 ms; Harbour 4.1 ms; styles 1.2 ms. A warm call is a cache hit. 143: - **Captures** (`c/caps.sh`; the six journey views, one sitting, pre-1, after-1, pre-2 and after-2, before from `c/pre-tree`; all ok on try 1; `cap/power-*.txt` AC, LPM 0): 144: - Integer counts, draw calls included, are identical in all 6 views (`metrics/counts-pre1-after1.txt`). 145: - Systematic pixels (`metrics/systematic-8-2.txt`; changed by more than 8 in both cross pairs and by at most 2 in both noise pairs): harbour-day 3,622 (0.23%), harbour-lane 704, quarter-night 112, harbour-night 28, quarter-day 20, styles-day 8. The maps `look/sys-harbour-day.png` and `sys-harbour-lane.png` (looked at) put them on the rain rings and inside the chimney smoke, which run on TIME. Nothing on a building. 146: - Noise pairs: `metrics/noise-pre.txt`, `noise-after.txt`. 147: - **Bench** (`c/bench.sh`, c/pre-tree against live; dense and Harbour at 01:00; `--disable-vsync`, 1920x1080, rain 0.45; round 0 discarded, then rounds 1-6; raw fields; `bench/summary-6rounds.txt`; `bench/power.txt` AC 100%, LPM 0, load 2.7-4.1): 149: | Scene | raw p50 | raw p95 | paired p95 | 150: |---|---|---|---| 151: | Dense | 13.50 → 13.50 | 15.14 → 15.15 | −0.04 | 152: | Harbour | 7.61 → 7.69 | 9.10 → 9.10 | −0.04 | 154: Draw calls are identical (2362, 903). Harbour's p50 is slower in 5 of 6 pairs (paired median +0.15), but its pre rounds alone span 7.05-7.82, and no runtime path calls the new code. This is read as order or noise; it was not chased. 155: - **Journey count deltas expected:** none. Nothing drawn changes. 156: - **For the P7d gate and sheet (not decided here):** 157: - (1) Deviation 1, that hand-placed pieces count toward n. 158: - (2) The quarter's podium decks would gain 15 sets. 159: - (3) Dense gets no roof pieces, so P7d's "dense p95 within +10%" line will not exercise roof synths unless a fixture with bare roofs is added. 160: - (4) The 4-4.7 ms largest unit (`state.clusters()` on dense, a cap cut or site in the quarter) against the release budget. 161: - (5) Harbour's Fish Market roof 9 stays without a crown, because its hand-placed deck parts hold every corner. The plan's rule is "one crown per cluster on its tallest roof"; the drop is kept and reported (`crownless`). The owner may want a crown off the corners, or one of the parts moved. 162: - **Review fix (3 October 2026):** The review found one must-fix. Every placed part was a blocker disc of half its longest side, wall parts outside the parapet included. So Harbour's FISH MARKET `neon_bar` (7.36 m, hung outside roof 9) became a 3.68 m disc over the whole deck, and the cluster's tallest roof lost its crown without any report. 163: - **Fix (`scripts/composition.gd`):** 164: - A part's blocker is now its oriented footprint (local box through its transform), a capsule along its longer side, radius half its shorter side plus 0.1 m. A square part keeps the same disc as before. 165: - A wall part, or a part off the deck, that reaches no more than `ROOF_PARAPET_BAND` 0.35 m into the deck (battlements 0.32 m, canal coping 0.27 m) hangs on the parapet: it is no blocker and is listed in `hung`. 166: - `pack()` gets disc blockers as obstacles and capsules through `fits`. 167: - The tallest site carries `crown`, and `roofs()` carries `crownless`. The debug image draws capsules. 168: - The preparation now takes the part's box from its own transform instead of a second `bounds()` call. A first version cost +5 to +8 ms on dense, and this put it back. 169: - **Tests first:** `c/fix/testfirst/test_roofs.log(.out)`: 169 checks, 25 failures, then a script error on the missing capsule. The signed house had no crown and no pieces. 170: - **After:** `test_roofs` 182 checks, 0 failures (`c/fix/suites/`). Full suite (`c/fix/suite/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0): **42 suites, 21,954 checks, 0 failures**. `--verify` 2880x1800 exited 0 on try 1 with `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun was clean. 1440x900 was not run (inherited stall). 171: - **Changed expectations (old → new):** 172: - `invariants()` measures a part blocker as a capsule (was a disc round its point). 173: - Every tallest site must state its crown and match `crownless` (new). 174: - **Layouts unchanged:** the derived crowns and pieces are identical before and after the fix in Harbour, styles, the quarter and dense (`c/fix/probe/{pre,live}-{1,2}.derived`, from `c/fix/probe/fix_probe.gd` on the pre-fix tree `c/fix/prefix-tree` and the live tree, ABAB). Only blocker shapes and the new `hung` lists differ (`c/fix/probe/layout-diff.txt`). For example, Harbour roof 8 now hangs the flower patch of roof 7 on its edge, and the quarter's roof 2 hangs roof 12's shrub. 175: - **Dense roof time, cold, ABAB** (`c/fix/probe/live-*.log.out` against `pre-*.log.out`, AC, Low Power Mode 0, load 2.6): after 38.7-40.4 ms against before 38.6-41.1 ms; 40.7 ms in the suite. The largest unit is 4.1-4.3 ms against 3.9-4.3 ms. 176: - **Reviewer probe re-run** (`c/fix/probe/rv_probe2-after.log.out`): roof 9's blockers are now planter, flower patch, umbrella, chimney, flag and the cat; the sign and string lights are gone. Its corners are hit only by those. 177: - **Debug images** (looked at): `c/fix/look/roofs-harbour.png`, `roofs-styles.png`, `roofs-quarter.png`. The large ring over the Fish Market roof (`c/look/roofs-harbour.png`) is gone; long parts show as capsules. 178: - Nothing draws the roofscape and nothing at runtime calls it, so no capture or bench changes; neither was re-taken. 179: - Review: filled by the orchestrator. 180: - **Not done:** 181: - No drawing (P7d). 182: - No 1440x900 `--verify`. 183: - No daytime bench: the part changes no runtime path and the plan's P7c table has no bench. 184: - The scratch tree `c/pre-tree` was left in place. 185: - The report's "any roof standing ≥2 storeys over a neighbour also gets a crown" was not taken; the plan says one crown per cluster. 186: - Platforms: Mac run; Windows not run. 188: **Beautify P7b CPU emitter mirror, unlit panes, window families (3 October 2026):** Opus 5.5 on this Mac (report Signature D: "mirror the story on the CPU emitter without consuming rng", the shop story, the silhouettes under curtains; plan header: unlit panes into the pane material with an off flag). A lit window's light and spill pools now follow its room's story: cool 8fb4ff at 0.45, dim 0.4x. Curtains hide the silhouette. An unlit pane is glass with an off flag instead of a sunlit `plain` box. Lit shopfronts tell a shop story, and lit ribbon facets tell their story at night. **Not asked for by the plan: the day glass of `window_body` is now twice as light (`STORY_DAY_GLASS` 2.0)**, because the moved unlit panes would otherwise have joined the black slots (see Captures). **This is the stage's owner stop: the room-story sheet is `artifacts/beautify/25-P7/b/sheet/` (`sheet.md` and 8 pages).** 189: - **CPU half:** `scripts/room_story.gd` (`RoomStory`, new class). It holds the story from the pane's own seed (float32-rounded, never the rng), `light()`, `pool()` and `veiled()`. `Architecture.window`, `styled_bay_legacy`, `ClusterArchitecture.styled_window`, the japanese tower bays and the modern ribbon's lights all key on it. A ribbon light keys on its facet's pane seed: tower seed + `ribbon_facet_key`, using the new `Architecture.ribbon_pane_centre`, which `modern_ribbon` also uses. 190: - **Flags in UV2.y** (the fixture order, which the dusk threshold clamps, so the turn is unchanged): `ORDER_OFF` -1 for an unlit pane, `ORDER_SHOP` -2 for a lit shopfront pane. They are set only around the pane and its crown; `Openings.arch_head` takes an optional 4th `pane` entry so the room behind keeps order 0. 191: - **Moved:** unlit panes from `plain` to `window` (flush canal panes, their arched crowns, styled windows, legacy bays, oriel side lights, japanese tower bays), with colours 2c2f4a/456474 kept. Dark recessed and unlit ribbon panes (`pane_glass`) take the flag. Ribbon panes get UVs (`RIBBON_UVS`). 192: - **Shaders:** `window_body` draws an off pane as `vcol x STORY_OFF_NIGHT` (0.6) at night, with no core lift. `glass_body` tells a ribbon pane apart by its alpha (≤ `STORY_RIBBON_ALPHA` 0.24): no day story, so towers-day ribbon columns have 0 systematic pixels; at night a lit facet's story starts from `STORY_RIBBON_ROOM` (WARM). A shop pane gets `story_shop` (x1.15, shelves at 0.35 and 0.6, goods) and only the glint by day. The door bay of a wide front keeps the ordinary story. 194: Files: `scripts/room_story.gd` (+ `.uid`), `architecture.gd`, `cluster_architecture.gd`, `openings.gd`; `shaders/room_story.gdshaderinc`, `window_body.gdshaderinc`, `glass_body.gdshaderinc`; `tests/test_dusk.gd`, `test_building_tools.gd`, `test_cluster_walls.gd`; `tools/repo_evidence.sh` (review fix). Snapshot: `build/backup-feel/25-P7-b/` (main 7753bb8). At part start the live tree matched it (`diff -rq` on scripts, shaders, tests, data, scenes and tools was empty). Evidence: `artifacts/beautify/25-P7/b/`; backups are in `b/backup/`. 195: - **Tests first:** `b/testfirst/b-testfirst-test_dusk.log(.out)` (332 checks, 13 failures; the CPU checks are gated behind the missing script) and `b-testfirst-test_building_tools.log(.out)` (691, 3). 196: - **New checks:** 197: - `test_dusk` `story_mirror_checks` (+31 checks with the changed ones), covering: 198: - the constants mirrored, with shares within 5 points over 1,000 seeds; 199: - RoomStory agreeing with the twin on ≥999 of 1,000 seeds; 200: - every lit flush, recessed and styled window's light and pools following its pane's story, with all six stories seen; 201: - silhouettes and figures left out only under curtains, the rng state unchanged; 202: - unlit panes flagged and never lit, the room behind not flagged; 203: - ribbon UVs, the off flag on unlit facets, and each facet's light keyed on its own seed; 204: - shop panes only on lit ground bays; 205: - **all 297 stock clusters and connections drawing the same geometry as the pre-P7b tree** against `b/probe/geo-pre.json` (written by `b/probe/geo_hashes.gd` on `b/pre-tree`; the live pre-code run was identical): everything outside the panes, pools and light colours byte for byte, and the panes as the same triangles and colours apart from silhouettes; 206: - no unlit pane left in `plain`. 207: - `test_building_tools` `unlit_pane_checks` (+3): every style and family, wide fronts, legacy, oriel, towers. 208: - Two of my own test-first lines were changed after the first captures. Off at night was `pow(vcol,2.2) x 0.5`, which put a 456474 pane at p50 21 against 63 before; it is now `vcol x 0.6`, with a check that it is never brighter than its drawn colour. A day-glass check was added. 209: - **Full suite** (`b/suite/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0): **41 suites, 21,772 checks, 0 failures**. 210: - Counts changed: dusk 321 → 352 and building_tools 688 → 691. 211: - `--verify` 2880x1800 exited 0 on try 1, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun was clean; the known flaky check did not fire. 212: - 1440x900 was not run (the inherited stall). 213: - The first run (`b/suite-first/`) failed `test_cluster_walls` 5 checks: the changed expectation below. 214: - **Changed expectations (old → new):** 215: - `test_dusk` P5c1 (a window lights in its room colour; a styled window in its glass colour; a recessed window, and its pool, in its room colour) → the same, unless the pane's story is cool (8fb4ff). 216: - P5c1 "two identical canal houses draw the same geometry" → the same, counted without silhouettes. 217: - P7a1 "every stock cluster and connection byte for byte with pane_glass folded into glass" (against pre-P7a1) → only the row count is still checked; the P7b geometry check replaces it against pre-P7b. 218: - P7a2 gate `told = (UV == 0 && alpha < LIT_MIN)` and "no ribbon pane tells a story or wears a day blind" → `told_day = 1 - ribbon` (by alpha) and "by day no ribbon pane…". The `glass_story_told` twin now reads `STORY_RIBBON_ALPHA`. 219: - `test_cluster_walls` P0d pinned plank pieces 301/126/351/664/307 → 294/120/344/640/300. Unlit panes are no longer `plain`, so they no longer count as timber plank pieces; plank events are unchanged. 220: - **Captures** (`b/caps.sh`; one sitting, pre-1, after-1, pre-2, after-2, before from `b/pre-tree`; 15 views, all 60 on try 1; `cap/power-*.txt` AC, LPM 0). The views are the six journey views, quarter-golden, glade-day, lane-night, styles-night, towers- night/day/golden, and shops- night/day. `b/probe/shops.json` is a new look fixture of three wide canal fronts, because no stock scene has a shopfront. 221: - Integer counts, draw calls included, are **identical in all 15 views in both pairs** (`metrics/counts-*`). 222: - Noise: `metrics/noise-pre.txt` and `noise-after.txt`. Systematic change: `metrics/systematic-8-2.txt`, maps in `look/sys-*.png`. It is largest at shops-night 7.4%, lane-night 2.4%, harbour-lane 2.2%, towers-day 1.8% (the neighbours' lifted styled glass) and styles-day 1.2%. 223: - Railings: systematic pixels are 0 in most rows. harbour-night has 12 and 8, quarter-day 12: window panes or pools beside the rails, not looked at pixel by pixel. 224: - Tuning crops (looked at): `look/t1-*.png` and `t2-*.png`. Sheet pages (looked at: 02, 04, 06, 07). 225: - **Acceptance:** 226: - **Each cool or dim pane's light matches its pane: met.** Unit checks pass for every family, with all six stories seen. In the stock scenes, 7 cool and 23 dim of the quarter's 106 lit panes light to match, and Harbour has 5 and 11 of 113 (`metrics/stories-and-lights.txt`). 227: - **Geometry identical apart from the moved material: met.** 297 of 297 rows match (`b/probe/geo-compare.txt`). Silhouette triangles went from 26,160 to 19,632, all under curtains. 228: - **Styles-day modern panes no longer black slots: met, only with the day-glass lift** (`metrics/panes.txt`): the modern house's top-left panes p50 42.4 → 89.5, p10 29.4 → 62.8. Moving the unlit panes alone darkened them. The victorian unlit pane, sunlit light blue before (p50 142.5), is now glass at 84.7. At night that pane went from 116.3 (lamp-lit `plain`) to 73.8. 229: - **p95 within noise: met** (below). 230: - P5c1's room_tone lerp on styled glass was not redone. 231: - **Bench** (`b/bench.sh`, both story parts against **pre-P7a1** (`a1/pre-tree`); round 0 discarded, rounds 1-3; `--disable-vsync`, 1920x1080, rain 0.45; `bench/summary.txt`; `bench/power.txt` AC 100%, LPM 0; raw medians, pre → live): 233: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 234: |---|---|---|---|---| 235: | Dense | 13.42 → 13.46 | 15.17 → 15.21 | 13.19 → 13.20 | 14.91 → 14.82 | 236: | Harbour | 8.05 → 8.03 | 9.24 → 9.19 | 8.11 → 8.01 | 9.13 → 9.04 | 238: Paired p95 medians are +0.03 (dense night), +0.02 (dense day), −0.06 and −0.20 (Harbour). Draw calls are identical (2362, 903). 239: - **Journey count deltas expected:** none; integer counts are identical in all views. The pictures change: lit panes and pools at night, all window glass by day, ribbon facets at night. 240: - **Owner sheet:** `b/sheet/sheet.md` and `p7b-01` to `p7b-08`. It asks for six decisions: D24 (room stories on, the plan default); the right block at 3 story kinds (keep, or re-key per column); the day-glass amount; unlit panes at night; the P7a2 night-alpha deviation; and the shop story. 241: - **Review fix (3 October 2026):** the review found one must-fix. The new fixture `artifacts/beautify/25-P7/b/probe/geo-pre.json`, which `test_dusk` loads (`GEO_PRE`), was missing from the fixture list in `tools/repo_evidence.sh`, and `artifacts/` is gitignored. On a fresh clone `test_dusk` would therefore fail "the pre-part geometry fingerprints are on disk" and skip the 297-row geometry check. Fix: the file is now in the fixture `add` list (`tools/repo_evidence.sh` line 21), and `--dry-run` stages it. 242: - Clone simulation (`b/review/clone-sim.sh`): the tracked and untracked files outside `artifacts/`, the tracked `artifacts/` files and what `repo_evidence.sh` stages, copied, imported and run through `test_dusk`. Without the file: 348 checks, 1 failure, the one the review named (`b/review/clone-sim/test_dusk-without.log.out`). With the fixed list: 352 checks, 0 failures (`test_dusk-with.log.out`). The copies were deleted afterwards. Both runs also log a pre-existing `ERROR` for the missing `assets/ui/hologram/terrain_flowers.svg`; the live suite logs it too, and it is not a failure. 243: - Full suite rerun (`b/review/suite/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0): 41 suites, 21,772 checks, 0 failures. `--verify` 2880x1800 exited 0 on try 1, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun was clean. No code, captures or bench changed, so nothing was re-captured or re-benched. 244: - Review: filled by the orchestrator. 245: - **Not done:** 246: - No shop emitter change (the plan names only cool and dim). 247: - The `core_twin` in `test_dusk` still uses the old day glass; its checks are ratios. 248: - No 1440x900 `--verify`. 249: - `artifacts/building-tools/tests.json` was rewritten by the suite. I also once restored it with `git checkout` after a test run, which I should not have done. 250: - The scratch tree `b/pre-tree` was left in place. 251: - Platforms: Mac run; Windows not run. 253: **Beautify P7a2 Night room stories, day glint, golden blend (2 October 2026):** Opus 5.5 on this Mac (report Signature D "room stories" and "a diagonal glint by day"; the day glint and golden blend taken from P7b). A lit pane now shows one of six rooms, picked from its own fixture seed: a lamp hot-spot, side curtains, a pulled-down blind, plants on the sill, a dim room or a cool (TV) room. By day the panes show cloth, blinds and a diagonal glint; the golden hour blends both. **One acceptance line is missed: 3 story types on the quarter-night right block, target 5. The part is not done until the orchestrator or owner waives that line or the seed is re-keyed** (see Measurements and Review fixes). 254: - **Shader:** `shaders/room_story.gdshaderinc` (new). s = fract(UV2.x × 7.13 + 0.37); shares lamp 30%, curtains 25%, blind 15%, plant 10%, dim 12%, cool 8%. The cool room is static (no TIME). Curtains and blinds raise the glass alpha to 0.85 (at night more stories raise it: see Glass specifics). The blind is solid at pixel size 2 and has 5 slats only at pixel size 1 (`story_fine`). The glint is p.x + p.y in [0.55, 0.62], +0.08. 255: - **Who draws it:** `window_body` paints it on every pane it shades (window, window_cyan, window_pink). `glass_body` paints it only under `#define PANE_STORY`, which the `pane_glass` wrappers set and `glass`/`glass_build` do not. Railing, bridge and terrace glass therefore never get a story. 256: - **Pane coordinates:** read from BoxMesh's 3 × 2 atlas, with v running down. The atlas probe is `a2/probe/atlas_probe.gd` and its log. A surface without UVs reads the pane's top middle. On `pane_glass` that is now only the arched crowns (alpha 0.30 or 0.45); the UV-less ribbon panes are gated out (review fix below). 257: - **Glass specifics:** 258: - Only the front face shows a story. The far face of the two-sided glass mirrored the glint into an X, seen in `a2/look/t2-qd-arch.png` and fixed in `t3`. 259: - The night story starts from 0.65 of the pane colour (`STORY_GLASS_ROOM`), so a dim room reads dim. 260: - At night a lit pane's alpha rises to 0.51 over a lamp room and to 0.85 under cloth, under leaves and over a dim or cool room (`story_night` hands back that cover). **This goes beyond the plan's tests line, which names only curtains and blinds at 0.85.** The dim and cool rooms and most of a lamp room no longer show the recessed room detail (back wall, lamp box, picture) through the 0.30 glass (`a2/review2/qn-right-zoom.png`). Recorded as a deviation for the owner gate after P7b; the alpha is the owner's taste call and was not changed. 261: - Lit is read from vertex alpha: 0.30 lit recessed, 0.45 dark, 0.22 ribbon. Ribbon panes are also drawn without UVs (`append_quad` with no uvs, architecture.gd 4790), so glass_body tells them no story at all, day or night: `told = 0` for UV-less glass under alpha 0.25. Their story waits for P7b (UVs or the facet key, and the off flag). 262: - **View:** `WorldView.sync_environment` sets `story_fine` (pixels ≤ 1) on the window materials, `pane_glass` and the build variants. 264: Files: `shaders/room_story.gdshaderinc` (+ `.uid`), `window_body.gdshaderinc`, `glass_body.gdshaderinc`, `pane_glass.gdshader`, `pane_glass_build.gdshader`, `scripts/world_view.gd`, `tests/test_dusk.gd`. Snapshot: `build/backup-feel/25-P7-a2/` (main c43ae4b). At part start the live tree matched it (`diff -rq` on scripts, shaders, tests, data, scenes and tools was empty). Evidence: `artifacts/beautify/25-P7/a2/`. 265: - **Tests first:** `a2/testfirst/a2-testfirst-test_dusk.log(.out)` shows 11 failures, 306 checks. 266: - **New checks** (`test_dusk.gd` `room_story_checks`, 22 at implementation, 28 after the review fixes): 267: - the constants and s formula; 268: - static, with no TIME; 269: - both bodies include the story; glass only under PANE_STORY; railing wrappers never define it; 270: - alpha 0.85 under cloth by day; at night the alpha story_night sets (review fix); 271: - no story on a modern tower's UV-less ribbon panes; the arched crowns keep theirs (review fix); 272: - shares within 5 points over 1,000 hash01 seeds; 273: - curtains at the sides, narrowing at the tie-back; 274: - the blind solid at size 2, 5 slats at size 1, covering 30-55%; 275: - the view sets `story_fine` on all four materials and the build variants; 276: - all six stories among the quarter's panes. 277: - **Full suite after the review fixes** (`a2/fix/suite/suite_all.out`; VERIFY_RES=2880x1800; AC power, Low Power Mode 0, `a2/fix/suite-power.txt`): **41 suites, 21,738 checks, 0 failures** (dusk 315 → 321). `--verify` 2880x1800 exited 0 on try 1 in 33 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 278: - **Full suite at implementation** (`a2/suite/suite_all.out`; VERIFY_RES=2880x1800; AC power, Low Power Mode 0, `a2/suite-power.txt`): **41 suites, 21,732 checks, 0 failures** (P7a1 21,710; dusk 293 → 315). `--verify` exited 0 on try 1, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun was clean; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 279: - **Changed expectations (old → new):** `test_dusk` P7a1 wrapper checks: "pane_glass(_build).gdshader is glass(_build)'s wrapper, code lines identical" → "identical apart from `#define PANE_STORY`". Review fix: "P7a2: only curtains and blinds cover the glass" → "P7a2: by day only curtains and blinds cover the glass (story_day's cover is story_cover)". The check is the same; the old name claimed it for the night glass as well. No other assertion changed. 280: - **Captures** (`a2/caps.sh`; one sitting, order pre-1, after-1, pre-2, after-2, diag; 80 views, all on try 1; power in `a2/cap/power-*.txt`, AC, LPM 0). There are 16 views: a1's 15 plus glade-day. Before is `a2/pre-tree`. 281: - Integer counts, draw calls included, are **identical in all 16 views in both pairs** (`metrics/counts-*`). Draw calls are unchanged; P7a1's extra ribbon-roof transparent surface was already in both trees. 282: - Whole-frame change: `metrics/ab-1.txt` and `noise-*.txt`. 283: - Systematic change (`a2/systematic.js` 8/2, `metrics/systematic-8-2.txt`, maps `look/sys-.png`) is largest at lane-night 3.99% and harbour-night 2.49% (panes and their canal mirror). It is quarter-night 1.53% and styles-day 0.16%. 284: - Crop sheet: `look/sheet-p7a2.png` (looked at). Tuning crops: `look/t1-`, `t2-`, `t3-` and `t4-*.png`, plus `look/glade-day-mid.png`, `harbour-night-mid.png` and `t4-dimpane.png` (all looked at). 285: - **Measurements:** 286: - **At least 5 story types on the quarter-night right block: MISSED, 3** (lamp 1,216 px, dim 3,204, curtains 1,887; cool 4 px and blind 8 px are slivers). 287: - Method: the diagnostic tree `a2/diag-tree` is the live code with `story_night` returning one flat key colour per story. It is not shipped. `a2/storycount.js` compares it with after-1 (`metrics/storycount.txt`, `look/diag-qn-right.png`). 288: - The block's 7 visible lit panes happen to be 3 dim, 3 curtains and 1 lamp under the plan's fixed formula. 289: - The wider right side (1000-1520 × 280-620) shows 4 types. The tower region shows 5. The whole frame shows all 6 in quarter-night, harbour-night, lane-night, quarter-2045 and towers-night. 290: - Not pushed: changing 7.13/0.37 to pass would be tuning the measurement. The choice for the owner is left open: accept, or re-key the seed (say, per window column). 291: - **Nothing on railing, bridge or terrace glass: met.** Systematic pixels inside the railing regions are 0 in 9 of 12 view×region rows. 292: - harbour-golden/-night `railings_terrace` have 4 and 8 of 1,136 px. That is the lit window beside the rail (`look/rail-harbour-night.png`). 293: - quarter-night `railings` has 8 of 16,840 px: one wet-paving art pixel by a lamp (`look/rail-quarter-night.png`). 294: - **Probe pane checked:** `look/t3-qd-arch.png` shows the blind at the top, the curtains at the sides and one glint per pane. 295: - **Bench** (`a2/bench.sh`; pre-tree against live, round 0 discarded, rounds 1-3; `--disable-vsync`, 1920x1080, rain 0.45; `bench/summary.txt`; `bench/power.txt` AC 100%, LPM 0; raw medians, pre → live): 297: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 298: |---|---|---|---|---| 299: | Dense | 13.39 → 13.48 | 15.13 → 15.20 | 13.14 → 13.18 | 14.84 → 14.87 | 300: | Harbour | 7.95 → 7.66 | 9.08 → 9.07 | 8.06 → 7.85 | 8.96 → 9.07 | 302: Paired medians are +0.02/+0.08 (dense night) and +0.03/−0.01 (dense day). Draw calls are identical (dense 2362, Harbour 903). Within noise. 303: - **Start-up (`--startup-report`)**, owed by P7a1 (`bench/startup-summary.txt`, medians of rounds 1-3, AC, LPM 0): 305: | Tree | ready_ms | warm_ms | all_ms | entries | 306: |---|---|---|---|---| 307: | pre-P7a1 | 6281 | 4260 | 10486 | 131 | 308: | P7a1 | 6367 | 4334 | 10899 | 135 | 309: | P7a2 | 6418 | 4399 | 10958 | 135 | 311: P7a1 cost about +86 ms ready, +412 ms all (four more warm-up entries). P7a2 cost about +51 ms ready, +65 ms warm. These runs were sequential triplets, not interleaved pairs. 312: - **Journey count deltas expected:** none. Integer counts are identical in all six journey views. The journey pictures do change, in the window panes. 313: - **Review fixes (2 October 2026, evening).** At part start the live tree differed from `build/backup-feel/25-P7-a2` only in this part's files. Backups of the four edited files are in `a2/fix/backup/`. 314: - **Ribbon panes (must-fix):** modern tower ribbon panes are drawn without UVs, so `story_pane` returned (0.5, 0.97) everywhere. By day every blind-kind facet (about 15%) wore a full-height beige blind at alpha 0.85, and about 34% of that showed at the golden hour. No facet could glint. The earlier line here saying ribbons get "the day cloth and glint" was false. 315: - Fix in `glass_body.gdshaderinc`: `told = 0` for UV-less glass under `STORY_LIT_ALPHA_MIN` (0.25), so the TOWER_PANE ribbon, at 0.22, tells no story. The arched crowns (0.30 or 0.45) keep the fallback. The comment in `room_story.gdshaderinc` was updated. 316: - Tests first: `a2/fix/test_dusk-testfirst.log` has 321 checks and 1 failure (the gate). `a2/fix/test_dusk-fixed.log` has 321 and 0. 317: - Re-captured in one sitting from `a2/pre-tree` and the live tree (`a2/fix/caps.sh`; pre-1, after-1, pre-2, after-2; 7 views; all on try 1; AC, LPM 0, `a2/fix/cap/power-*.txt`). Integer counts, draw calls included, are identical in all 7 views in both pairs (`a2/fix/metrics/counts-*`). 318: - The noise pairs are 6.1-10.3% this sitting (`a2/fix/metrics/noise-*.txt`). 319: - Systematic change (8/2) in the modern tower box [770,250,1000,470] (`a2/fix/sys.js`, `metrics/systematic-8-2.txt` against `systematic-unfixed-8-2.txt`): towers-day 2,168 → 144 px, towers-golden 2,520 → 416 px, towers-night 632 → 640 px. The remaining pixels are on the neighbouring towers at x 760-800 and 960-1000. In the ribbon columns (x 800-960) they are 40 px by day and 0 at the golden hour (`metrics/towerbox-columns.txt`). 320: - Looked at `a2/fix/look/towers-day-modern.png`, `towers-golden-modern.png` and `towers-night-modern.png`, each showing pre, fixed and unfixed. The fixed ribbon matches the pre-tree, and the beige panels of the unfixed build are gone. 321: - **Night alpha (must-fix):** the check "only curtains and blinds cover the glass" tested the GDScript twin of `story_cover` and not what the glass does at night. 322: - The test now mirrors `story_night`'s cover with the twins `story_night_cover`, `story_plant` and `story_pane`. It asserts the shipped alphas: lamp 0.51; curtain, blind, leaf, dim and cool 0.85; gaps 0.30. It also checks the shader sets those covers. 323: - The deviation from the plan's tests line is recorded above under Glass specifics for the owner gate. The alpha values were not changed. 324: - **Right block, 3 types against 5 (must-fix, twice):** not fixed here, because it is a decision for the orchestrator or owner. The reviewer counted it independently and got 3 (`a2/review2/stories-qn-right.txt`). The reviewer's Monte Carlo puts the chance that 7 independent panes show 5 or more types at about 28% under the plan's fixed formula. 325: - The constants were not tuned. **The part cannot be called done until the line is waived or the seed is re-keyed** (for example, per window column). 326: - Not re-benched: the fix adds one compare per glass fragment, and the draw calls are unchanged. 327: - **Recheck 2 (2 October 2026, late).** The recheck closed the ribbon-pane and night-alpha fixes. It left one finding open: the right block still shows 3 story types against 5 (lamp, dim, curtains). Two reviewers counted 3 independently, and the recheck agreed that the fix pass was right not to tune the seed. 328: - The finding asks for a decision, not more code, so this pass changed no code and no constants. Its only edit is this entry. 329: - The code files' hashes are in `a2/recheck2/code-sha1.txt`. The files were last written before the review-fix full suite (`a2/fix/suite/suite_all.out`, 41 suites, 21,738 checks, 0 failures, `--verify` 2880x1800 exit 0), so that run still covers this tree. No suite, capture or bench was re-run. 330: - **Open for the orchestrator or owner:** waive the line (the tower region shows 5 types and the whole quarter-night frame shows all 6, `metrics/storycount.txt`), or re-key the seed, for example per window column. Re-keying is a design change to the plan's fixed formula `fract(UV2.x × 7.13 + 0.37)`. Under amendment 7 the part is not done until one of the two happens. 331: - Review: filled by the orchestrator. 332: - **Not done:** 333: - The right-block line (above): needs a waiver or a re-key. 334: - Ribbon panes have no story at all, day or night, since the review fix. They need pane UVs (architecture.gd, outside this part's files) or P7b's facet key and off flag. The styled glass's unlit `plain` panes also wait for P7b. 335: - The CPU emitter does not mirror the story yet (P7b). 336: - No 1440x900 `--verify`. 337: - The scratch trees `a2/pre-tree` and `a2/diag-tree` were left in place. `a2/pre-tree` is the before side for P7b's bench, together with `a1/pre-tree`. 338: - `artifacts/building-tools/tests.json` was rewritten by the suite. 339: - Platforms: Mac run; Windows not run. 341: **Beautify P7a1 `pane_glass` split and seeds (2 October 2026):** Opus 5.5 on this Mac (report Signature D "gate it away from the glass balustrades"; architecture lens F0 revised). A window's glass now has its own material key, `pane_glass`, so P7a2's room stories can go on panes and never on railing, bridge or terrace glass. Mechanical part: nothing is meant to look different. 342: - **What moved to `pane_glass`:** the recessed window's pane and its arched crown (`Architecture.window`), and the modern tower's ribbon panes (`modern_ribbon`). Japanese tower bays were already `window`/`plain`; nothing moved there. 343: - **What stays `glass`:** the railings, which keep their colours 0.4/0.7/0.76/0.4 and 0.44/0.72/0.77/0.45. These are the helix guard (`_helix_quad`), the modern building parapet, the modern tower's flat-roof edge, the modern bridge rail and the cluster terrace guard (`cluster_architecture.gd`). 344: - **Registries:** the new key is in `Architecture.materials()` (the shader loop, render_priority -1 as the glass, the P1c2 order), in `BuildAnim.MATERIALS` (appended, which gives it a build variant), in `Dusk.SEEDED` and `SEEDED_SET`, and in the `sync_profile` loops (`albedo`, `windows`, `day_albedo`; `dusk` through SEEDED). `ShaderWarmup` picks it up from `world.materials`. 345: - **Shaders:** `shaders/pane_glass.gdshader` and `pane_glass_build.gdshader` (new; `.uid` from a headless import) are code-identical wrappers over the unchanged `glass_body.gdshaderinc`. 346: - **Per-facet seeds:** each ribbon pane's UV2.x is now the tower seed plus `Architecture.ribbon_facet_key(world centre)`, a whole number 1..64 (`RIBBON_FACET_KEYS`). The new `MeshBatch.seed_offset` carries it; emitters never take it. Dusk reads only the fractional part, so the ribbon still lights at the tower's turn (float32 error under 4e-6). P7b's emitter mirror should key a facet's light on the same helper. 348: Files: `scripts/architecture.gd`, `build_anim.gd`, `dusk.gd`, `mesh_batch.gd`, `world_view.gd`; the two new shaders; `tests/test_dusk.gd`, `test_build_anim.gd`, `test_building_tools.gd`, `test_meet.gd`, `test_release_spread.gd`, `test_warmup.gd`; `tools/repo_evidence.sh` (review fix). Snapshot: `build/backup-feel/25-P7/`. Before the first edit the live tree matched it (`diff -rq` on scripts, shaders, tests, data, scenes and tools was empty). Evidence: `artifacts/beautify/25-P7/a1/`. 349: - **Tests first:** the failing logs are in `a1/testfirst/` (`a1-testfirst-.log.out`): 350: - dusk: 4 failures; 351: - build_anim, release_spread, building_tools: 1 each; 352: - warmup and meet: 2 each. 353: - **New checks** (`test_dusk.gd` `pane_glass_checks`): 354: - the registries, priority -1 and the build variant; 355: - the wrappers are code-identical to glass's; 356: - every glass uniform is equal on `pane_glass` and its build variant at 01:00, 05:45, 06:45, 13:00, 19:30, 20:45 and 22:00; 357: - recessed square and arched panes, lit and dark, are `pane_glass`; 358: - ribbon panes: one seed per pane, a whole number plus the tower seed (to 1e-5), at least half distinct per band; 359: - railings stay `glass`; 360: - in 12 stock scenes and fixtures, `glass` holds only railing colours and `pane_glass` none, and no connection draws `pane_glass`; 361: - **every stock building cluster and connection (297) draws byte for byte as before with `pane_glass` folded into `glass`**, against `a1/probe/pane-pre.json`. That file was written on the pre-part tree by `a1/probe/pane_hashes.gd`; the probe on live gives an identical `pane-live.json`, and the pre-code live run was identical too (determinism). 362: - **Full suite** (`a1/suite/suite_all.out`, VERIFY_RES=2880x1800, AC power, Low Power Mode 0): **41 suites, 21,710 checks, 0 failures** (P6 close 21,672). Changes: dusk 273→293, build_anim 422→431, meet 1392→1397, warmup 34→36, building_tools 687→688, release_spread 107→108. `--verify` exited 0 on try 1, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun was clean; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 363: - **Changed expectations (old → new):** 364: - `test_dusk` P5c1 recess pane: `rb.groups.glass` → `pane_glass`. 365: - `test_building_tools` modern_tower_glazing: ribbon panes counted in `glass` → `pane_glass`; the dark-tone check now covers both keys. 366: - `test_meet` modern_glass_checks: ribbon ray hit `glass` → `pane_glass`; helix guard: plus "draws no pane_glass". 367: - `test_release_spread`: the "no other priority moves" exemption list gains `pane_glass`, plus a check that it draws where the glass does. 368: - `test_build_anim`: the body of `pane_glass` is `glass_body.gdshaderinc` (the generic `_body` rule would expect `pane_glass_body`); the shader-moved list gains `pane_glass`. 369: - **Captures** (`a1/caps.sh`; one sitting, 21:42-21:48, order pre-1, after-1, pre-2, after-2, all 60 on try 1; AC, Low Power Mode 0 in `cap/*/capture-meta.txt`). There are 15 views: the six journey views plus quarter-golden, harbour-golden, quarter-2045, lane-night and styles-, towers- golden/night/day. 370: - Integer counts, draw calls included, are **identical in all 15 views in both pairs** (`metrics/counts-*`). 371: - Whole-frame changes: `metrics/frame-table.tsv`. Regions (windows_right, windows_tower, right_block, railings, railings_terrace, modern_block, …): `metrics/regions-table.tsv`. Every pre→after change sits inside the spread of the two noise pairs. Example: quarter-night windows_right 1.17%/0.65% against noise 1.82%/0%; windows_tower 0%. 372: - `systematic.js` counts pixels that changed more than 8 in both cross pairs and at most 2 in both noise pairs (`metrics/systematic-8-2.txt`, maps `look/sys-.png`). Building views are at most 0.06%. styles-golden (0.58%), styles-night (0.86%) and towers-day (0.62%) are higher, but the maps (looked at: styles-night, styles-golden, towers-day) show them only on TIME-driven cloud edges and rain rings on the water, never on a building. Styles has no `pane_glass` at all, so this is the trees' TIME phase at capture, not the split. 373: - The crop sheet `look/sheet-p7a1.png` (looked at): quarter right block at night, golden and 20:45, the lit tower window, the styles modern block and the Harbour terrace rail are the same by eye. 374: - **Bench** (`a1/bench.sh`; pre-tree against live, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, rain 0.45; `bench/summary.txt`, `bench/power.txt` AC 100% Low Power Mode 0; raw p50/p95 medians, pre → live): 376: | Scene | 01:00 p50 | 01:00 p95 | 19:30 p50 | 19:30 p95 | 377: |---|---|---|---|---| 378: | Dense | 13.41 → 13.44 | 15.09 → 15.18 | 15.13 → 15.14 | 16.94 → 16.93 | 379: | Harbour | 7.78 → 7.56 | 9.27 → 9.06 | 7.26 → 7.80 | 9.00 → 9.05 | 381: Draw calls are identical (dense 2362 / 3126, Harbour 903 / 818). Within noise. 382: - **Journey count deltas expected:** none. 383: - Review: filled by the orchestrator. 384: - **Review fix (2 October 2026):** the review found one must-fix. `pane_glass_checks` read the fixture `artifacts/beautify/25-P7/a1/probe/pane-pre.json`, but `artifacts/` is gitignored and `tools/repo_evidence.sh` did not list the file. On a fresh checkout or in a cloud session, test_dusk would then fail its "pre-part fingerprints are on disk" check and skip the 297-row identity check. 385: - Fix: the file is now on the fixtures line of `tools/repo_evidence.sh`, next to B2q's `stairs-pre.json`. `--dry-run` lists it (`a1/fix/repo_evidence-dry.txt`). It reaches Git when the orchestrator runs the script and commits. No game or test code changed. 386: - Re-run (`a1/fix/run.sh`; AC power, Low Power Mode 0, `a1/fix/power.txt`): test_dusk passed 293 checks with 0 failures (`a1/fix/test_dusk.log.out`). The full suite (`a1/fix/suite/suite_all.out`, VERIFY_RES=2880x1800) ran 41 suites and 21,710 checks with 0 failures. `--verify` exited 0 on try 1 with `PIXELPUNK_INTEGRATION []`, and the creature_reactions rerun was clean. The counts match the first run. 387: - Captures and the bench were not repeated: the fix touches no game code. 388: - **Not done:** 389: - The start-up warm-up time was not measured (`--startup-report`); the plan's risk table lists P7a1 for it. `test_warmup` covers the new material. 390: - No 1440x900 `--verify`. 391: - The scratch tree `a1/pre-tree` was left in place for P7a2/P7b (whose bench is against pre-P7a1). 392: - `artifacts/building-tools/tests.json` was rewritten by the suite (687 → 688). 393: - Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 5 October 2026) 196: ## Orchestrator ruling on P7a2 (2 October 2026, owner not asked) 198: | Part | Ruling | 199: |---|---| 200: | P7a2 | The plan line "at least 5 story types on the quarter-night right block" is **waived at 3** (lamp, dim, curtains; two reviewers counted the same). The plan's fixed seed formula gives a block of about 5-6 independent seeds (two-light windows share one fixture seed), so 5 types is not structurally guaranteed, and the constants (STORY_SEED_SCALE 7.13, STORY_SEED_OFFSET 0.37) were not tuned to hit a count. The tower region shows 5 types, the wider right side 4 and the whole frame all 6. The owner decides on the P7b room-story A/B sheet (quarter-night right block and tower): keep, or re-key the seed per window column (a small part with new tests, captures and counts). | 202: ## Orchestrator rulings before P7d (3 October 2026, owner not asked) 204: | Item | Ruling | 205: |---|---| 206: | D25 release-time limit | The plan default applies (the owner has not answered, and the P7 gates are not among the owner's blocking stops): +2 ms over today's 26.7 ms Harbour animated release, memoised per owner. P7d reports the number plainly. | 207: | P7c deviation: hand-placed roof pieces count toward n (wanted = n - placed) | Kept. Dense therefore derives 0 pieces (400 placed by hand) and Harbour 1 set; the quarter derives 15 sets, 13 of them on its 3.0 m podium decks. P7d's "dense p95 within +10%" line needs a bare-roof fixture to mean anything: P7d builds one (the dense scene with the hand-placed roof parts removed, under artifacts/, added to tools/repo_evidence.sh) and benches it as well as dense.json. | 208: | P7c: Harbour Fish Market roof gets no crown (hand-placed parts on all four corners) | Kept and recorded in the P7c entry and the P7d gate notes. | 209: | P7d on the quarter's podium decks (a protected, hand-composed scene) | P7d draws what the composition derives, and the roofscape sheet at the stage stop shows the quarter, Harbour and styles-day before and after; the owner may veto the podium sets there. | 211: ## P7 sheets approved (owner, 5 October 2026, in chat) 213: "All P7 changes approved", given while P7h was running, after the P7b room-story sheet (`artifacts/beautify/25-P7/b/sheet/`) and the P7d roofscape sheet (`artifacts/beautify/25-P7/d/sheet/`) had been sent. Read by the orchestrator as approval of everything on those two sheets and of P7a1 to P7c, as built: 214: - **P7b (six decisions):** D24 room stories on by default; the quarter-night right block stays at 3 story kinds (the seed is not re-keyed); the daytime window glass lift (STORY_DAY_GLASS 2.0); unlit panes at night at their own colour x 0.6, including that unlit panes away from lamps go flat grey-teal; the P7a2 night-glass alpha deviation (0.85 over dim and cool rooms, 0.51 over a lamp room); the shop story. 215: - **P7d (four questions):** the quarter's 15 derived sets stay on its 3.0 m podium decks, including the night lights they add (13 emitters and reflection streaks); the Fish Market roof stays without a crown; the two nearly touching AC boxes on the styles roof stay; the D25 reading (a rise of +2 ms over the sitting's own pre-part figure) stands. 216: Not covered, because it does not exist yet: P7h, P7e, P7f and P7g. Those get the P7 stage sheet at the close. SOURCE: artifacts/beautify/25-P7/b/sheet/sheet.md (room-story owner sheet, 3 October 2026) 1: # P7 room stories: owner sheet after P7b (3 October 2026) 3: This is the P7 stage's owner stop. Before (left on every page) is P7a2 as committed (main 7753bb8, so room stories are already in). After (right) is P7b. Both were captured in one sitting on this Mac (AC power, 100%, Low Power Mode 0). Numbers are luma 0-255. P7c (roof dressing) may run while you look. 5: Pages: 6: - `p7b-01-overview-night.png`: the quarter and Harbour at night 7: - `p7b-02-quarter-right-block-tower.png`: quarter right block and tower at night, right block at golden hour 8: - `p7b-03-harbour-night.png`: Harbour courtyard and upper windows at night 9: - `p7b-04-styles.png`: styles by day and at night 10: - `p7b-05-quarter-day.png`: the quarter and Harbour by day 11: - `p7b-06-modern-ribbon.png`: the modern tower's glass ribbon at night, golden hour and by day 12: - `p7b-07-shopfronts.png`: shopfronts (a test scene; no stock scene has one) 13: - `p7b-08-glade-lane.png`: Glade by day and the lane at night 15: ## What P7b changed 17: 1. **Light that matches the room.** A cool (TV) room now lights the street pale blue (8fb4ff, energy 0.45). A dim room lights it at 0.4 of the usual strength. The window's spill pool on the wall and the paving takes the same light. The quarter has 106 lit panes: 7 are cool and 23 dim. Harbour has 113: 5 cool and 11 dim. 18: 2. **Silhouettes.** A window with curtains no longer shows the small figure standing in it. About a quarter of the figures are gone (26,160 to 19,632 triangles across the stock scenes). Nothing else about any building moved: every stock building and bridge (297) draws the same geometry as before. 19: 3. **Unlit windows.** Until now an unlit window was a flat dark box that the sun lit. It is now glass: by day it looks like the lit windows, with curtains, blinds and the glint, and at night it stays dark. 20: 4. **Day glass is lighter (please look at pages 4 and 5).** Lit windows by day were the dark slots in the audit. Moving the unlit ones into the same glass would have made them dark too. Before, they were the light-blue sunlit panes. So all window glass by day is now twice as light. On the modern house in styles by day, the panes' middle value went from 42 to 90 and the darkest tenth from 29 to 63. One unlit victorian pane, which the sun lit light blue before, went from 142 to 85: it is now glass like its neighbours. The recessed canal windows are unchanged by day. The amount is one number (`STORY_DAY_GLASS`, 2.0). 21: 5. **Modern tower ribbon (page 6).** At night, a lit facet of the glass ribbon now tells its room's story: amber lamp, curtains, or a dim or cool room. By day the ribbon is exactly as before (0 changed pixels in its columns). At golden hour a little of the night story shows. 22: 6. **Shopfronts (page 7).** A lit shop bay on a wide canal front shows a brighter room, two dark shelves and coloured goods at night. By day it shows only the glint. The bay with the shop door keeps the ordinary story so the shelves do not cross the door. None of the stock scenes has a shopfront, so this page uses a small test scene. 24: ## The room stories on the quarter's right block (unchanged by P7b) 26: P7b did not change which story a window tells. P7a2's count still stands: 27: - the right block at night: 3 kinds (lamp, dim and curtains; target 5, waived at 3 by the orchestrator); 28: - the wider right side: 4 kinds; 29: - the tower: 5 kinds; 30: - the whole quarter-night frame: all 6 kinds, and Harbour at night has all 6. 32: ## What you decide 34: 1. **D24: room stories on by default** (the plan's default is yes). Keep them on, or off. 35: 2. **Right block at 3 story kinds:** keep it as it is, or give each window column its own seed. Re-keying is a small separate part with new tests and captures. 36: 3. **The lighter day glass** (point 4): keep 2.0, or choose a lighter or darker value. 37: 4. **Unlit windows at night** are their own dark colour times 0.6. A nearby lamp no longer lights them as it did the old box. One victorian pane next to a lamp went from 116 to 74. Unlit panes away from lamps also change: they lose the moonlit blue and become one flat grey-teal whichever way they face (the modern house unlit side pane goes from 192f5c to 314c5d, and the upper tower panes on Harbour go from vivid blue to sea-green grey; crops review/zoom-hn-upper.png and review/zoom-sn-modern.png under artifacts/beautify/25-P7/b/). Added by the orchestrator after the review. Keep this? 38: 5. **P7a2 deviation, still open:** at night, a lit recessed pane's glass covers the room behind it more than the plan's test line said (0.85 over dim and cool rooms, 0.51 over a lamp room). Keep it? 39: 6. **Shop story:** keep the shelves and goods as shown. 41: Cost: none measurable. Dense and Harbour against the tree before P7a1, at night and by day, are within noise: dense p95 15.17 to 15.21 ms at night and 14.91 to 14.82 by day. Draw calls are unchanged. SOURCE: artifacts/beautify/25-P7/d/sheet/sheet.md (roofscape owner sheet, 3 October 2026) 1: # Beautify P7d roofscape sheet (owner stop after P7d, 3 October 2026) 3: Before = the P7c tree (`d/pre-tree`, main aadac85), after = P7d (live), one sitting at 02:35 BST, AC power 100%, Low Power Mode 0 (`roofcap/power-*.txt`). Cameras in `d/roofcaps.sh`. Whole frames at 0.75x, zooms nearest-neighbour. 5: 1. `p7d-01-numbers.png`: release, static-frame and drag numbers (sources: `bench/release-summary.txt`, `bench/smoke-summary-t1.txt`, `bench/drag-summary.txt`). 6: 2. `p7d-02-harbour-day.png`, 3. `p7d-03-harbour-night.png`: Harbour's two derived pieces (tank crown on roof 10 beside the Fish Market; utility box beside the placed tank on the main terrace). The Fish Market roof itself has no crown (its corners are held by hand-placed parts; P7c, kept). 7: 4. `p7d-04-styles-day.png`: the modern roof: tank crown, planter, two AC boxes (almost touching). 8: 5. `p7d-05-quarter-top.png`, 6. `p7d-06-quarter-back.png`, 7. `p7d-07-quarter-usual.png`: the quarter's 15 sets (17 pieces), all 15 on its 3.0 m podium decks (the plazas): pergola, umbrellas over tables, vents, AC boxes, planters, roof gardens. 10: Questions for the owner: 11: - Night lights (added by the orchestrator after the review): the quarter derived pieces carry lights (two umbrellas with 4 bulbs each = 8, a pergola with 3 lanterns, two cafe tables with 1 lantern each), so quarter-night gains 13 light emitters and 13 reflection streaks on the canal (emitter_sources 203 to 216, paving_streaks 181 to 194, reflection_streaks 203 to 216). The systematic map shows no change on the water itself, only on the pieces, but the canal mirror is a protected look. 12: - Keep the derived sets on the quarter's podium plazas, or opt those decks out (`roof_dressing: false` on the podium bodies, a JSON-only change to `artifacts/artwork/artwork.json`; or a rule that skips decks lower than one storey)? 13: - Keep the Fish Market roof without a crown? 14: - The release limit (D25 default +2 ms): Harbour +0.08 ms with animations, +0.27 without; the bare-roof dense stress fixture +2.16 ms. ADDED 6 October 2026: the rest of the P7 stage. Frozen 6 October 2026 at 16:17 BST (game main at 58c7827). P7e ivy spill with its review fix; P7f window boxes and Juliet rails with the cost trim of 6 October (planting leaf vertices halved, greener colour); P7g derived downpipes and vent stacks with its review fix round; the P7 stage close (committed at 98a2e79) with the whole-stage bench after the trim (ee8379a); the orchestrator ruling on P7f and the owner's approval of the whole of P7 on 6 October. Part codes: P7e ivy spill (curtains), P7f window boxes and rails, P7g pipes and vents. Local paths and run IDs are replaced by placeholders; moves "to ~/.Trash/" read "to the macOS Trash". SOURCE: docs/visual-handoff.md (frozen copy, 6 October 2026, 16:17 BST, game main at 58c7827; the working file also held uncommitted P8g1 additions, which were not frozen or used) 346: **Beautify P7 stage close (5 October 2026):** the stage-end overhead of P7 "A lived-in town" (step `25-P7`), run by Opus 5.5 on this Mac on the final P7 code: `main` c6227d4, clean (`close/head.txt`). The parts are P7a1, P7a2, P7b, P7c, P7d, P7h, P7e, P7f and P7g. No game code changed. Evidence is in `artifacts/beautify/25-P7/close/`. 347: - **Before side:** the pre-stage scratch tree `close/pre-tree`, from `build/backup-feel/25-P7/`. It equals `main` d27c025 (the P6 sign-off) on scripts, shaders, tests, data, scenes, tools and `project.godot`; only the three `.ps1` tools differ, in line endings (`close/diff-pretree-vs-d27c025.txt`). Its class cache was refreshed with a headless import (`close/import-pre.sh`; RoomStory gone from the cache). 348: - **Power:** AC, 100% charged, Low Power Mode 0 for every run (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). Load average 1.8-4.1. 349: - **Full suite** (`close/suite.sh` a, b, v; 22:41-22:48): **42 suites, 22,375 checks, 0 failures** (`close/suite/suite_all.out`). `--verify` at 2880x1800 exited 0 on try 1 in 33 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 350: - **Journey** (`artifacts/feel-journey/25-P7/`, six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1). Counts against 24-P6 (`25-P7/counts-vs-24-P6.txt`) change **exactly by the sum of the parts' declared deltas**: 351: - harbour-night, -day and -lane: draw calls +4 (893 → 897; lane 737 → 741), ivy 395 → 440; 352: - quarter-day and -night: draw calls 1008 → 1011, ivy 369 → 465, emitters and reflection streaks 203 → 216, paving streaks 181 → 194; 353: - styles-day: draw calls 470 → 497, ivy 0 → 40, foliage multimeshes 1 → 4. 354: - The stage capture re-ran the six views with identical counts (`close/metrics/counts.txt`), so no single-view re-run was needed. 355: - Frame >8 luma against 24-P6 is 7.3-11.1% (`close/metrics/journey-frame-24P6-25P7.txt`); by eye (`close/look/j-1.png`, `j-2.png`): rooms in the windows, a darker lawn in the Harbour courtyard at night (below), roof tanks and planters, ivy curtains, window boxes. 356: - `notes.json` has a `25-P7` entry (`close/journey-note.js`, backup `notes.json.bak-before-25-P7`); `index.html` and `journey-standalone.html` (107.93 MB) were rebuilt. 357: - **Art capture and protected looks** (`close/caps.sh`; one sitting 22:50-23:25, every view on try 1). Group g1 is the P6 close's 58 views; group g2 is 30 P7 views and close-ups taken from the part scripts (styles night and golden, towers, shops, Amsterdam, castle, courtyard, roof, ivy, box and pipe close-ups). Order: pre-1 g1, pre-1 g2, after g2, after g1, pre-2 g1, pre-2 g2, after-2 g2, after-2 g1. That is 88 views per set. Metrics are in `close/metrics/` (`close/metrics.sh`, regions `close/regions.json` = the P6 close's): `ab-*`, the noise pairs, `frame-*`, `frame-table.txt`, `counts.txt` and `protected.md` (`close/prot.js`), plus `ab-p6close-*` and `frame-p6close-*` against `24-P6/close/after`. Protected sheets: `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P6-close.png`, `noise-pre.png` and `noise-after.png`. Pages 1, 3 and 4 were looked at. 358: - **Counts:** pair 1 and pair 2 agree in every fixed view; the orbit views vary inside one tree, as at P6. The P6 close capture and today's pre-1 have identical counts in every fixed view. 359: - **Canal mirror columns:** within noise. p50 43.6 → 43.9 and 52.1 → 52.2. Changed >8 is 29% against noise of 7-32% (quarter-night) and 5-19% (quarter-21:00). 360: - **HDR neon cores:** 361: - fish sign within noise (1.4% / 0.1%); 362: - bar sign 8.0% in pair 1 but 0.1% in pair 2, so noise; 363: - lane sign p50 143.0 → 140.2 in both pairs (7.8% / 8.0% against noise of 0.3% / 1.1%), cause not established; 364: - the quarter tower window, a protected look, changed on purpose: it now shows a room story (p50 112.7 → 96.0 in both pairs, noise 0%; P7a2/P7b, approved). 365: - **05:45 and 20:45 skies:** within noise (p50 equal to ±0.3). 366: - **Glade/Canal separation:** Glade's night zenith (`#001d45`), horizon, water, day lawn and bank are identical. Harbour's night zenith p50 and mean are identical, but 23% / 25% of its pixels change by more than 8 (noise 0.2% / 3.4%). This is probably the cloud dither's TIME phase at a later capture moment (P7 adds release work before the capture); not proven. 367: - **Approved tree style:** unchanged. Grove-day crowns p50 55.1 = 55.1, conifers 55.6 = 55.6, quarter undersides 27.7 = 27.7. The grove frame changes on 0.12-0.17%, from window boxes and ivy on its buildings (grove draw calls +6). 368: - **Whole frames:** before against after changes 4-24% of pixels by more than 8 in building views, against noise of 0-7% (`frame-table.txt`). 369: - **Whole-stage bench after the P7f cost trim (6 October; supersedes the close's figures below, which are kept):** `f/fix2/bench/` (`summary.txt`, `release-summary.txt`, `drag-summary.txt`). One sitting with the pre-stage tree, the pre-trim close code and the trim interleaved, 3 rounds, AC, Low Power Mode 0. Details are in the P7f entry, "P7f cost trim". 370: - Dense raw p95 against the pre-stage tree: 15.09 → 16.09 (**+6.6%**) at 01:00, 14.63 → 15.79 (**+7.9%**) at 13:00, 16.66 → 17.71 (**+6.3%**) at 19:30. This is under the +10% line and under the +8% target. 371: - The pre-trim code read +8.9% / +11.4% / +9.3% in the same sitting. 372: - Dense-bare: +13.3% / +13.0% / +11.9%. 373: - Harbour animated release: 29.65 → 31.35 ms (**+1.70**, within D25). 374: - Drag: handle +0.32; terrain raise 125.98 → 128.45 (**+2.47**, still over P7d's 2 ms line). P7f's share of the raise fell from +1.85 to +1.14. 375: - Draw calls are unchanged by the trim. 376: - **Whole-stage A/B bench at the close (superseded by the trim figures above)** (`close/bench.sh`; `bench-t1.out`, `-t13.out`, `-t19.5.out`, `bench-release.out`, `bench-drag.out`; `bench/summary.txt` from `close/bench-summary.js`, `bench/release-summary.txt`, `bench/drag-summary.txt`). Setup: pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync` before `--path`, 1920x1080, rain 0.45. Dense `artifacts/weather-pass9/dense.json` and Harbour at 01:00, 13:00 and 19:30. Every run was on try 1 with 491 raw samples and vsync disabled. Raw p50 and p95 medians, pre → live (paired median; rounds where live was slower): 378: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 379: |---|---|---|---|---|---|---| 380: | Dense | 13.03 → 14.55 (+1.50, 3/3) | 15.05 → 16.61 (+1.56, 3/3) | 12.63 → 14.21 (+1.54, 3/3) | 14.58 → 16.21 (+1.60, 3/3) | 14.81 → 16.38 (+1.55, 3/3) | 16.71 → 18.22 (+1.51, 3/3) | 381: | Harbour | 6.50 → 6.54 (+0.08, 2/3) | 8.30 → 7.09 (-1.23, 1/3) | 6.15 → 6.24 (+0.08, 3/3) | 7.30 → 7.87 (+0.61, 3/3) | 6.24 → 6.34 (+0.10, 3/3) | 7.66 → 7.44 (-0.22, 1/3) | 383: - Draw calls: dense 2362 → 2583 (19:30: 3126 → 3433); Harbour 903 → 907 (19:30: 818 → 822). 384: - **Dense p95 is +10.3% at night, +11.2% at 13:00 and +9.0% at 19:30: at or over the plan's +10% line.** The parts' own benches account for it: P7f +1.29 to +1.36 ms (box planting), P7e +0.28, P7d +0.06, and the rest about 0. Under the orchestrator's P7f ruling (5 October), this makes a cheaper box planting (fewer shells or lower-poly lobes) the directed fix. Not done here: no game code was changed. 385: - **Release probe** (headless, 9 house moves, 3 alternating runs; release frame median pre → live): 386: - Harbour with animations 29.48 → 31.49 ms (**+2.01, against the D25 limit of +2 ms**); without animations 24.27 → 25.76 (+1.49); 387: - dense 98.79 → 102.82 with animations, 108.55 → 113.66 without; dense-bare +4.6 / +5.6; 388: - the Harbour spread total 51.1 → 57.4 ms (13 → 15 slices), worst slice 7.11 → 7.94. 389: - **Drag harness** (Harbour, `step_total` median): handle 17.00 → 17.37, wall stroke 8.33 → 8.55, slider 8.32 → 8.76, grass stroke 10.87 → 11.56, terrain raise 125.04 → 128.17 (**+3.1, over the 2 ms P7d line**; the per-round medians overlap: pre 124.9-125.1, live 125.9-129.6). Harness draw calls 883 → 889. No errors. 390: - **Owner sheet** `artifacts/beautify/25-P7/close/sheet/`: before (`pre-1`) on the left and after on the right, colour then greyscale. Crops come from `close/crops.js` and pages from `close/sheet.sh`. Pages: `p7-01-overview.png`, `p7-02-P7a2-P7b-rooms-night.png`, `p7-03-P7a2-P7b-glass-day.png`, `p7-04-P7c-P7d-roofs.png`, `p7-05-P7h-P7e-ivy.png`, `p7-06-P7f-boxes-rails.png`, `p7-07-P7g-pipes-vents.png`, `p7-08-existing-scenes-1.png`, `p7-09-existing-scenes-2.png`, `p7-10-harbour-night-court.png`. `sheet.md` has one plain paragraph per part with its numbers, then existing scenes, the bench and the suite. Pages 2, 5, 6, 7 and 10 were looked at; 1, 3, 4, 8 and 9 were made but not opened. 391: - Flagged for the owner (not changed): 392: - At Harbour night the lawn by the tower is darker: some windows became dim or cool rooms and their emitters now match (P7b). Mean luma -1.3 over the courtyard crop (page 10). 393: - P7f box greenery reads as dark grey-black lumps, not green, at Harbour and quarter zoom (page 6). *Fixed by the P7f cost trim (6 October):* the planting is now a lighter green leaf. Pages 1, 6, 8, 9 and 10 were re-made; the old pages are in `close/sheet/pre-trim/`. 394: - P7e's curtain on the white styles modern block reads as dark clumps (page 5). 395: - P7g's pipes read clearly on the courtyard scene but are hard to tell from the existing copper pipes on the quarter's right block (page 7). 396: - **Not done / open:** 397: - The owner has not seen this sheet. P7h, P7e, P7f and P7g have had no owner look. 398: - Dense p95 over +10% (above): the directed fix has not been made. *Superseded: the P7f cost trim (6 October) brought dense p95 to +6.6% / +7.9% / +6.3%.* 399: - The Harbour animated release sits at +2.01 ms against D25's +2 ms, and the drag terrain raise is +3.1 ms. Both are reported, not fixed. *After the trim: +1.70 ms and +2.47 ms. The terrain raise is still over P7d's 2 ms line; P7f's share is +1.14.* 400: - Journey (checked after the trim): draw calls and rendered objects are identical before and after the trim in the six journey views. The journey PNGs in `artifacts/feel-journey/25-P7/` still show the pre-trim planting. 401: - The lane sign core (-2.8 p50) and the Harbour night zenith pixel churn are not explained. 402: - Drag baseline and journey counts were not re-based: the plan's overhead line "re-based once" is left to the orchestrator. 403: - No 1440x900 `--verify`. 404: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 405: - Scratch trees: `close/pre-tree` and the part trees `a1/`, `a2/`, `b/`, `c/`, `d/`, `e/`, `f/`, `g/` and `h/pre-tree` were moved to the macOS Trash with `mv` once their numbers were recorded, as `beautify-25-P7--pre-tree-235149` (`close` for the stage tree). The diagnostic tree `a2/diag-tree` was left in place. 406: - Platforms: Mac run; Windows not run for any P7 part (P6e, the Windows kit, is still deferred by the owner). 408: **Beautify P7g Derived downpipes and vent stacks (5 October 2026):** Opus 5.5 on this Mac (plan section "P7g"; report Signature D; lens F8 revised). Flat-roofed canal and victorian buildings of 2+ storeys with glazed windows and an open top now derive their services (`Architecture.services`, drawn by `emit_services`): clamp(ceil(footprint / 30 m2), 1, 2) downpipes, hopper under the coping to a shoe 0.15 m over the foot, and on 3+ storeys one vent stack from the first floor to 1.2 m (`PIPE_ABOVE`) over the parapet with a 6-sided copper cowl. Nothing is stored. The services are emitted last in the building mesh, only into the batch's existing `painted_metal` group (P1f), so the pre-P7g geometry is a byte-identical prefix and no surface or draw call is added. The bores are painted between `SERVICE_PAINT` a3a89f and b8a68f (pale stone-grey, never darker than STONE), hashed from (seed, side, x); collars and cowls stay copper. (Before the review fix the bore tint was PIPE lerped toward COPPER by 0.2-0.45; see the fix round below.) 409: - **Placement:** 410: - Candidate piers between bay columns; side walls (1, 3) before front and back, a fresh side before a used one. 411: - Two passes: roomy piers (0.28 m clear of a light, 0.25 m of an arched light) first, then tight ones (0.13 m). 412: - Kept clear of door steps, balconies, oriels, P7f boxes and rails (0.15 m), wall mounts (0.4 m), terrace flights, Openings holes, neighbouring buildings and connections (occlusion samples every 0.5 m), and round towers (the axis at least `SERVICE_TOWER_CLEAR` 0.25 m outside the shell or eaves over the shared height; added in the review fix round). 413: - A pipe whose foot would stand above the ground must land on a neighbour's roof, or it is dropped. 414: - Merged blocks plan from their `ClusterWalls.panels`. 415: - **Deviations:** 416: - **Corner pipes.** The roomy pass and the landing rule left Harbour and dense with no pipe, so a building with no free pier takes its pipe at a back corner of a side wall, 0.14 m (`PIPE_STANDOFF`) off both faces, with band-and-stay collars to the corner edge and the hopper centred. Back corners are preferred because door clutter sits on the front. 417: - **Painted parts.** Collars, hopper, shoe and cowl are drawn in painted metal (lighter copper tints), not bare metal, so no draw call is added. 418: - **Param.** `services` is bool only (no reseed integer); solid-window buildings are excluded. 419: - **The param:** `services`, JSON only (D26): absent or true is automatic and false draws none. Any other value is refused in `scene_state.gd`. Save version stays 2. A README note sits beside the window_boxes paragraph. 420: - Files: `scripts/architecture.gd`, `scripts/cluster_architecture.gd`, `scripts/scene_state.gd`, `tests/test_building_tools.gd`, `tests/test_dusk.gd`, `README.md`. `tools/repo_evidence.sh` is unchanged: the tests read no fixture under `artifacts/` (the pins' provenance is `g/probe/sig.gd` and `g/probe/sig-pre.log.out`). 421: - Snapshot: `build/backup-feel/25-P7-g/` (main 512d2d1, P7f). At part start the live tree matched it (`g/diff-live-vs-25-P7-g.txt`, empty). 422: - Evidence: `artifacts/beautify/25-P7/g/`. The pre-part tree is `g/pre-tree`, backups are in `g/backup/` and probes in `g/probe/`. An earlier round taken before the corner change is in `g/superseded/`. 423: - **Tests first** (`g/testfirst/test_building_tools.log`): 783 checks, 1 failure (no services API). 424: - **Tests after** (`g/full2/suites/`): `test_building_tools` 782 → **897**, `test_dusk` 352 → 352, both 0 failures. After the review fix round (`g/fix/suites/`): `test_building_tools` 897 → **900**, `test_dusk` 352, both 0 failures. 425: - **Full suite** (`g/full2/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0): **42 suites, 22,372 checks, 0 failures**. `--verify` 2880x1800 exit 0 on try 1, `PIXELPUNK_INTEGRATION []`. The 1440x900 `--verify` stall is inherited and was not rerun. After the review fix round, on the final code (`g/fix/full/`, `g/fix/full-{a,b,v}.out`, AC, LPM 0): **42 suites, 22,375 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1, `PIXELPUNK_INTEGRATION []`; 1440x900 not rerun (inherited stall). 426: - **Changed expectations (old → new):** 427: - `test_dusk`: the P7b stock geometry pin now also sets `services` false on every building (the pin itself is unchanged). 428: - `test_building_tools`: the P7f "window_boxes false equals the pre-tree pin" check now also sets `services` false and erases it afterwards (the `P7F_PRE` pins are unchanged). 429: - **Starters changed** (`g/probe/scenes-8.log.out`, downpipes/vents): harbour 2/0, quarter 6/4, courtyard 4/2, castle 4/2, amsterdam 5/5, dense 100/0. styles and demo0-3 are unchanged. 430: - groups unchanged in every scene. 431: - Vertices: harbour +504, quarter +2,556, courtyard +1,764, castle +1,476, amsterdam +2,808, dense 4,073,322 → 4,098,522 (+25,200, +0.62%). 432: - Clearances: dress_in 0, links_in 0 and mounts_in 0 in every scene (the nearest mount beside a run is 0.92 m away). 433: - Glass hits 0 in every scene (`probe/glass-4.log.out`). 434: - Taste items, not fixed: courtyard has 1 clutter clash (a firewood stack by owner 4's back-corner downpipe); ivy leaves sit inside the pipe volume in castle (6 climbing, 6 spill) and amsterdam (9 spill). 435: - **Journey count deltas expected:** draw_calls 0 in every view. Pre and after are identical in all 8 capture views (amsterdam 469, castle 1070, courtyard 747, harbour 897, quarter 1011, styles 497), and in smoke (dense 2583, Harbour 907). Vertex counts rise as listed above. 436: - **Captures** (`g/caps.sh`, `g/closeups.sh`; pre-1, after-1, pre-2, after-2, close-pre and close-after, all in one sitting; AC, LPM 0; `cap/power-*.txt`). 437: - Systematic pixels (`metrics/systematic.txt`; >8 luma in both cross pairs, ≤2 in both noise pairs; maps `look/sys-*.png`): quarter-day 0.27%, quarter-night 0.15%, harbour-day 0.10%, harbour-night 0.09%, courtyard-day 0.23%, castle-day 0.14%, amsterdam-day 0.19%, styles-day 0.0008% (control, no services). 438: - Before/after pairs: `look/pair-*.png` and `look/close-*.png`. Vent cowls read on the quarter roofline (`pair-quarter-day-right.png`). The Fish Market corner downpipe reads attached (`close-harbour-pipes.png`). The courtyard tower corner vent stands proud of the corner against the sea (`close-court-corner.png`, a taste call). 439: - **Acceptance:** 440: - **Pipes above luma 40 by day in quarter-day: met after the review fix round** (quarter-day median 50.2; numbers in the fix round below). The first round missed it (`metrics/pipe-luma.txt`, `g/pipeluma.js` over the systematic pixels). quarter-day: median 8.4, p25 5.1, 22.5% of pixels above 40. Other views: harbour-day 21.7, courtyard-day 10.7, castle-day 21.4, amsterdam-day 46.9. 441: - The bores are P1f `painted_metal` (dark red) and read near-black. The copper cowls read 70-120. 442: - P1f's hand pipes measure about p50 11 in the same view, so this follows from the P1f material. The first round raised it for the owner; the review asked for a lighter paint or a waiver, and the fix round took the lighter paint (P7g bores only; P1f's hand pipes are unchanged and still dark). 443: - **Tests line: met.** 444: - Pipe and vent counts per rule. 445: - None on blocked piers, within 0.4 m of a mount or planter, or within 0.14 m of a P7f box or rail (30 reseeds, and the Harbour merged blocks). 446: - `services` false equals the pre-tree pins for 9 fixtures. 447: - No surface added. 448: - **No extra draw calls: met** (above). 449: - **Bench** (`bench/smoke-summary-t1.txt`, 1920x1080, `--disable-vsync`, rain 0.45, means of rounds 1-3, `bench/power.txt` AC LPM 0): 450: - dense raw p95 16.274 → 16.114 ms (−1.0%), raw p50 14.377 → 14.256 ms (−0.8%), +25,200 vertices (+0.62%); 451: - Harbour raw p95 7.707 → 7.645 ms, raw p50 6.580 → 6.584 ms. Both are within noise. 452: - **Release probe** (`bench/release-summary.txt`, 3 rounds; each figure is the mean of the three rounds' median release frames, in ms. This line said "median ms" until it was corrected on 5 October after the stage-close handoff check: the figures were always the summary's means, and the per-round values are in `bench/release-{pre,live}-{off,on}-{1,2,3}.out`): 453: - Harbour, animations off: 25.93 → 25.70; on: 31.94 → 31.64. 454: - dense, animations off: 112.70 → 115.25 (+2.3%); on: 103.03 → 103.53. 455: - dense-bare, animations off: 46.19 → 48.09 (+4.1%); on: 38.85 → 39.37. 456: - Planning costs about 107 ms CPU over dense's 100 buildings and 2.5 ms on Harbour (`probe/cost-3.log.out`). It runs at mesh build. 457: - **Review fix round (5 October 2026, Opus 5.5 on this Mac; evidence `g/fix/`, backups of the reviewed files in `g/backup/prefix/`).** The review raised two must-fix findings: 458: - **A corner downpipe ran inside a round tower.** `_service_neighbours` took buildings and connections only. Courtyard building 1's back-corner downpipe (side 1, x 2.14) ran its whole length inside tower #3's shell (axis 1.865 m from the tower axis, radius 2.0; `g/review/clash2.log.out`), and slivers poked out of the tower wall. 459: - Fix: towers are now neighbours (`SERVICE_NEAR_TOWER`). A run is refused when its axis comes within the tower's radius (with its roof's eaves) plus `SERVICE_TOWER_CLEAR` 0.25 m over the height the two share. The neighbour entries are now tagged building, tower or link instead of a bool. 460: - Result: courtyard building 1's downpipe moves to the other back corner (side 3, x −2.14; `g/fix/probe/where-4.log.out`). No other starter changes. Counts stay harbour 2/0, quarter 6/4, courtyard 4/2, castle 4/2, amsterdam 5/5, dense 100/0, with identical groups and vertex counts (`g/fix/probe/scenes-9.log.out`). The review probe now finds 0 tower hits (`g/fix/probe/clash3.log.out`). It still lists 2 quarter hits on an umbrella's and a blossom's bounding boxes, which were not findings and are not fixed. 461: - Picture: `g/fix/look/court-tower-old-vs-fix.png` (the review's after-1 beside the fix's after-1). The dark sliver at 786,322 is gone, with 0 pixels changed in both cross pairs at 780-795 x 300-369. 462: - **The luma acceptance was missed** (quarter-day median 8.4). The cause: a bore darker than the wall it stands on takes the presentation outline (`presentation.gdshader`) on itself. 463: - Fix: the bores are painted `SERVICE_PAINT` [a3a89f, b8a68f], hashed between the two, with relative luminance 0.383 and 0.394 against STONE's 0.259. The material stays `painted_metal`, and collars and cowls stay copper. 464: - A darker pair, [8f9a92, a8957f], measured quarter-day median 37.6 in its own sitting and missed the target (`g/fix/paint2/metrics/pipe-luma.txt`), so it was not taken. 465: - **Tests first** (`g/fix/testfirst/test_building_tools.log`): 898 checks, 2 failures: no `SERVICE_PAINT`, and "courtyard downpipe owner 1 side 1 x 2.14 in tower #3". 466: - **Changed expectations (old → new):** 467: - `test_building_tools` check 7: "bores are PIPE lerped toward COPPER by 0.2-0.45" → "bores lie between `SERVICE_PAINT[0]` and `[1]`", plus "the paint is never darker than STONE". 468: - New check 7b: no pipe or vent axis comes within a tower's radius plus the bore in the courtyard or the castle (12 items). 469: - Total 897 → 900. 470: - **Captures** (`g/fix/caps.sh`, `g/fix/closeups.sh`, `g/fix/capall.sh`): pre-1, after-1, pre-2, after-2, close-pre and close-after, all in one sitting from `g/pre-tree`. Power was AC with LPM 0 (`g/fix/cap/power-*.txt`). 471: - Draw calls are identical pre and after in all 8 views. 472: - Systematic pixels (`g/fix/metrics/systematic.txt`): quarter-day 0.23%, quarter-night 0.18%, harbour-day 0.17%, harbour-night 0.07%, courtyard-day 0.26%, castle-day 0.23%, amsterdam-day 0.24%. 473: - styles-day is 0 px, with a pre→after >8 share of 0.42% against noise pairs of 0.21% and 0.43% (`g/fix/metrics/noise-*.txt`). 474: - **Acceptance, re-measured** (`g/fix/metrics/pipe-luma.txt`, `g/fix/pipeluma.js` over the systematic pixels, which include the outline and shade the pipe now casts on the wall beside it): 475: - **quarter-day: median 50.2 against the target of over 40: met.** p25 is 39.4, and 69.0% of pixels are above 40. 476: - Other views: harbour-day 56.6, courtyard-day 51.7, castle-day 52.8, amsterdam-day 113.1, harbour-night 53.6 and quarter-night 37.7. 477: - The review's sample box (quarter-day 1206-1222 x 200-221) read (17,5,5) to (23,11,11). With the final paint the same pixels read (49,52,45) to (55,59,49). 478: - The weakest places left are the vent at the quarter's right corner against the sky (region 1200,172-1225,207, median 39.9) and a 40 px sliver at 1228,276 (32.6). The corner downpipe on the shaded face over the sea reads median 49.6 (region 1230,272-1259,385), against about 33 for the wall behind it. 479: - Look: `g/fix/look/quarter-day-right-old-vs-fix.png` (reviewed after-1 beside the fix's after-1) and `g/fix/look/pair-quarter-day-right.png` (pre-1 beside after-1). The bores read as pale stone-grey pipes with copper collars and cowls, with no black stripes. Whether the paint suits the roofscape is a taste call for the stage sheet. 480: - Files changed in the fix round: `scripts/architecture.gd`, `tests/test_building_tools.gd`, `README.md` (round towers and paint), and this entry. 481: - Review: (filled by the orchestrator). 482: - Not done: 483: - Bench and release probe not rerun in the fix round: the fix moves one courtyard pipe and changes the bore colour only, so groups and vertex counts are identical to the reviewed round in every starter (`g/fix/probe/scenes-9.log.out` against `g/probe/scenes-8.log.out`). 484: - The `court-tower` close-up camera in `g/fix/closeups.sh` missed the tower (it frames a roof terrace); the tower check uses the courtyard-day crops instead. 485: - The drag harness and start-up were not run. 486: - The taste items above are not fixed. 487: - Platforms: Mac run; Windows not run. 489: **Beautify P7f Window boxes and Juliet rails by bay column (5 October 2026):** Opus 5.5 on this Mac (plan section "P7f"; report Signature D; lens F6). Canal and victorian buildings with glazed windows now dress their upper windows by bay column. The choice is hashed from (seed, side, column x) through a local `RandomNumberGenerator`, never the geometry rng (`Architecture.column_dressing`). A column gets a window box under every window from the first floor up on 35% of columns. On 12% of the end or centre columns of a front of 3+ storeys it gets a Juliet rail instead; the box chance on those columns is 0.35/0.88, so boxes stay 35% of every column class. The items are collected while the walls draw and emitted last (`emit_dressing`), so everything drawn before P7f is a byte-identical prefix of each mesh group. 490: - **The box:** a trough (stone trim on canal, timber on victorian), soil, two brackets, leaf lobes and shells on the `leaf` material (the P6a stamp), 2-3 flower accents, and on 40% of boxes with room 2-3 trailing strands. 491: - **The rail:** `railing()` 0.22 m out with two stays. 492: - **Where nothing is drawn:** within 0.4 m of a wall mount (`dressing_mounts`, hidden mounts skipped), on the canal balcony window, or in the victorian oriel column. 493: - **Merged blocks:** they dress per panel with the lone front's column key (`dress_offset`), so a merged block matches the lone building. 494: - **The param:** `window_boxes`, JSON only (D26): absent or true is automatic, false draws none, and an integer 0..SEED_OVERRIDE_MAX reseeds. Validated in `scene_state.gd`; save version stays 2. A README note sits beside the roof_dressing paragraph. 495: - **Leaf build variant (the default decision):** 496: - `shaders/foliage.gdshader` is now a wrapper over the new `foliage_body.gdshaderinc` (the old body unchanged, plus a `BUILD_REVEAL` block). 497: - The new `foliage_build.gdshader` defines `BUILD_REVEAL`. `BuildAnim.MATERIALS` gains `leaf`. 498: - The window code (pane_glass, room stories, glass_body, window_body) is untouched. 499: - **Deviations:** 500: - The trough top is at sill - 0.02 (`BOX_DROP`) and 0.32 m deep (`BOX_DEPTH`; report 0.24), because the sill trim stands 0.31 proud and hid the planting (tuning captures `cap/close-after-0*`). 501: - One lobe per metre of box. 502: - Trailing strands stop at the floor line (victorian), at floor - 0.30 (canal), and never hang over a shopfront. 503: - Laundry: `composition.gd` `_laundry` skips a window that gets a box (outside the plan's file list). Without it, 128 dense laundry ends sat in boxes. 504: - Leaf surfaces on cluster meshes are not in the foliage lamp plant list. 505: - Files: 506: - `scripts/architecture.gd`, `cluster_architecture.gd`, `composition.gd`, `scene_state.gd` and `build_anim.gd`; 507: - `shaders/foliage.gdshader`, plus the new `foliage_body.gdshaderinc` and `foliage_build.gdshader` (and their `.uid` files); 508: - tests `test_building_tools`, `test_cluster_walls`, `test_build_anim`, `test_warmup`, `test_dusk`, `test_foliage`, `test_foliage_merge`, `test_look`, `test_owner_0924` and `test_tree_growth`; 509: - `README.md` and `tools/repo_evidence.sh` (adds `f/scenes/*.json`). 510: - Snapshot: `build/backup-feel/25-P7-f/` (main f198880). At part start the live tree matched it (`f/diff-live-vs-25-P7-f.txt`, empty). 511: - Evidence: `artifacts/beautify/25-P7/f/`. The pre-part tree is `f/pre-tree`, backups are in `f/backup/` and probes in `f/probe/`. 512: - **Tests first** (`f/testfirst/`): `test_building_tools` 694 checks, 3 failures; `test_cluster_walls` 255, 3; `test_build_anim` 435, 4; `test_warmup` 37, 1. The laundry checks were written after the code. 513: - **Tests after** (`f/full2/suites/`): `test_building_tools` 691 → **782**, `test_cluster_walls` 249 → **255**, `test_build_anim` 431 → **444**, `test_warmup` 36 → **37**, all 0 failures. Those four add +111; the full suite adds +112. 514: - **One more check count moved, unexplained** (added 5 October after the stage-close handoff check): `test_weathering` 507 → **508**, 0 failures (`e/fix1/full/suite_all.out` against `f/full2/suite_all.out`). The move is steady: every P7 suite log up to P7e reads 507, and every one from P7f on (`f/full`, `f/full2`, `g/full1`, `g/full2`, `g/fix/full`, `close/suite`) reads 508. `tests/test_weathering.gd` is not in 512d2d1 and was last changed in P3d (743ca37, 30 September). It builds its test buildings through `ClusterArchitecture.build`, so it runs `scripts/architecture.gd` and `scripts/cluster_architecture.gd`, which P7f changed. Its `stream_checks` makes one check per material group of a 6 x 5 x 6.5 m canal building, so a material group that P7f adds to that building would add one check, but the suite logs record only the count, so this has not been confirmed. Which check is new was not identified. 515: - **Full suite** (`f/full2/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0, `f/full2-power.txt`): **42 suites, 22,257 checks, 0 failures** (P7e 22,145, +112). `--verify` 2880x1800 exit 0 on try 1, `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun was clean. 1440x900 was not run (inherited stall). 516: - **Changed expectations (old → new):** 517: - `test_dusk`: the P7b stock geometry pin now sets `window_boxes` false on every building first (the pin itself is unchanged). 518: - The foliage shader text checks read `foliage_body.gdshaderinc`, not `foliage.gdshader`: `test_foliage` (4 sites), `test_foliage_merge` (2), `test_look`, `test_owner_0924` and `test_tree_growth`. 519: - `test_building_tools`: the oriel top-floor check uses |x| < 1.2 (was 1.3), because a neighbouring column's box edge reaches x 1.295. 520: - **Starters changed** (`f/probe/scenes.log.out`, boxes/rails): harbour 11/0, quarter 15/4, styles 3/0, courtyard 9/2, castle 13/4, amsterdam 79/14, demo0 4/0, demo2 11/0, dense 432/0. demo1 and demo3 are unchanged. 521: - clutter_in 0, oriel_in 0, laundry_in 0 in every scene. 522: - Ivy inside a box volume: harbour 1 climbing and 1 spill, amsterdam 6 spill, dense 6 spill. These are taste items, not fixed. 523: - Laundry lines (`probe/laundry_count-{pre,live}.log.out`): dense 180 → 145; harbour, quarter, styles, courtyard, castle and amsterdam unchanged. 524: - Cluster vertices: amsterdam 315,777 → 461,721; dense 3.35 M → 4.07 M (+21.7%). 525: - **Journey count deltas expected** (`f/metrics/counts-pre1-after1.txt`; both noise pairs are identical): 526: - draw_calls and rendered_objects: amsterdam +4, castle +6, courtyard +5, harbour views and lane +4, quarter +3, styles +3 (one leaf surface per cluster with boxes); 527: - dense draw calls 2365 → 2583 (smoke bench). 528: - **Captures** (`f/caps.sh`, `f/closeups.sh`; pre-1, after-1, pre-2 and after-2, plus close-pre and close-after, all in one sitting; AC, LPM 0; `cap/power-*.txt`). 529: - Systematic pixels (`metrics/systematic.txt`; maps `look/sys-*.png`): amsterdam-day 1.14%, amsterdam-night 1.41%, castle 0.04%, courtyard 0.22%, harbour-day 0.06%, harbour-lane 0.28%, harbour-night 0.06%, lane-day 0.43%, lane-night 0.23%, quarter-day 0.08%, quarter-night 2.05%, styles 0.05%. 530: - Much of quarter-night's 2.05% is a cloud band in the sky, which P7f does not draw. Its cause is not established. 531: - Before/after pairs: `look/pair-*.png` and `look/pair-close-*.png` (quarter-r16 shows a box and a Juliet rail; amsterdam-near shows the boxes with trailing strands). 532: - **Acceptance:** 533: - **Shares within ±5%: met** (`probe/shares.log.out`). Rule: boxes 34.9% of 80,000 columns, rails 12.07% of 48,000 end or centre columns. Drawn over 200 reseeds: canal 12x10 boxes 36.2%, rails 11.8%; victorian 10x8 boxes 35.7%, rails 11.3%. 534: - **Existing geometry identical: met.** With `window_boxes` false, the 9 fixtures equal the pre-tree pins (`P7F_PRE`, `probe/sig-pre.log.out`). With the dressing on, the pre-P7f vertices are an exact prefix of every group. 535: - **Boxes reveal with the wall: met** (`test_build_anim` `p7f_reveal_checks`). The leaf is keyed, wears `build_materials.leaf`, starts inside the wall's reveal and hands back to the settled leaf. 536: - **p95 within +10%: met** (`bench/smoke-summary-t1.txt`, 1920x1080, `--disable-vsync`, rain 0.45, rounds 1-3 medians, `bench/power.txt` AC LPM 0): 537: - dense raw p95 15.991 → 17.309 ms (+8.2%; paired +1.29, +1.36, +1.27), raw p50 14.31 → 15.65; 538: - Harbour raw p95 9.261 → 9.428 ms (+1.8%). 539: - **Start-up** (`bench/startup-summary.txt`, medians of rounds 1-3): ready_ms 6782 → 6950, warm_ms 4569.5 → 4726.1 (+157), entries 135 → 136 (the leaf build variant). 540: - **P7f cost trim (5 October; run 5-6 October by Opus 5.5 on this Mac).** 541: - **What the orchestrator asked for**, after the P7 close (main 98a2e79): the whole-stage dense p95 was +10.3% / +11.2% / +9.0%, with P7f's box planting the largest share. The ask: 542: - at least 40% fewer leaf vertices per box, without changing the box shares, places, mounts, clearances, the hash seeding or the pre-P7f geometry; 543: - planting that reads green, not as grey lumps, at Harbour and quarter zoom; 544: - the whole-stage bench re-run in one sitting. 545: - Evidence is in `artifacts/beautify/25-P7/f/fix2/`. At the start the live tree was clean at 98a2e79. Against `build/backup-feel/25-P7-f`, it differed only by P7f, P7g and the close (`fix2/diff-live-vs-25-P7-f.txt`), as expected. 546: - **What changed (planting only):** 547: - Every box leaf sphere (lobes, shells, strands) is a new 72-vertex low sphere instead of the 144-vertex `leaf` sphere. It is `MeshBatch.foliage(..., low)`: `_low_leaf_mesh` is SphereMesh 6 x 2 rings with its zero-area pole triangles dropped. 548: - It is passed through a new optional `canopy(..., low_poly)`. Every other caller draws exactly as before. 549: - `BOX_LEAF` goes 376a49 → 5a9654, with a lighter ramp `BOX_RAMP` 0.22 (was 0.35 on the planting and 0.15-0.35 on the strands). The tint's jitter is -0.05..0.08 (was ±0.1). 550: - The rng draws exactly as before (one draw per tint), so every box, trail and accent is unchanged. 551: - **Cause of the grey:** the planting's own vertex colours. A dark, cool-ramped green on a shaded wall is lit by ambient alone. Fixed in the planting only; no shader was touched. 552: - **Tried and dropped:** crown-normal "mound" normals (`TreeGrowth.CROWN_BLEND` 0.75). They made no visible change (`fix2/look/t5-*.png`) and their per-vertex pass cost CPU on every rebuild: dense cluster build +82 ms against +26 ms without, compared with no dressing (`probe/scenes-live.log.out`, `scenes-live2.log.out`). Their first bench is kept in `fix2/bench-mound/`; the Harbour animated release there was +2.21 ms. 553: - **Vertices** (`probe/trim-prefix.log.out` → `trim-live2.log.out`, `scenes-prefix.log.out` → `scenes-live2.log.out`): 554: - leaf per box exactly halves: fixtures canal12 1231.4 → 615.7 and vict10 1392 → 696; 60 reseeds canal 1369.3 → 684.7 and victorian 1487.2 → 743.6 (−50%; the line was −40%); 555: - dense leaf 586,944 → 293,472; dense cluster vertices 4,098,522 → 3,805,050 (P7f's addition 727,128 → 433,656); 556: - Harbour leaf 11,808 → 5,904; Amsterdam 115,488 → 57,744. 557: - Boxes, rails, trails and clash counts are identical in every starter. 558: - **Tests first** (`fix2/testfirst/test_building_tools.log.out`): 930 checks, 12 failures. These were the 11 vertex budgets plus a normals check (the mound). 559: - **After:** `test_building_tools` 900 → **930**, 0 failures. Full suite (`fix2/full2/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0, `suite2-power.txt`): **42 suites, 22,405 checks, 0 failures** (close 22,375). `--verify` at 2880x1800 exited 0 on try 1 with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun was clean. 1440x900 was not run (inherited stall). 560: - **Changed expectations (old → new):** 561: - P7g's pin: the whole-batch `P7G_PRE` → `P7G_PRE_REST`, every group but `leaf`. It was pinned on 98a2e79 with services false, which drew `P7G_PRE` exactly (`probe/trim-prefix2.log.out`). 562: - The test-first normals check (mean normal y ≥ 0.2; measured 0.163) was removed with the mound. 563: - New checks: 564: - `TRIM_REST`: everything but the leaf group is unchanged; 565: - `TRIM_REC`: the same boxes, trails, drops and accents; 566: - the per-box vertex budgets; 567: - a colour check (luma ≥ 0.42, green lead ≥ 0.15; measured 0.491). The colour check was written after the look tuning. 568: - `artifacts/building-tools/tests.json` was rewritten by the suite. 569: - **Draw calls:** identical in all 20 views, against the pre-trim tree and within each noise pair (`metrics/counts-pre1-prefix1-live1.txt`). **Journey count deltas: none.** The journey records no vertex counts. The 25-P7 journey PNGs still show the pre-trim planting. 570: - **Captures:** 571: - Sets: one sitting, AC, Low Power Mode 0 (`cap/power-*.txt`). Sets pre-1/2 (P7e tree), prefix-1/2 (98a2e79) and live-1/2, 20 views each, all on try 1. 572: - Systematic pixels, prefix → live: 0.002-0.64% (`metrics/systematic-prefix-live.txt`). On the maps (`look/sys-trim-*.png`), the systematic pixels sit on the boxes, plus water-ripple and rain phase dots. 573: - Planting colour over the systematic green pixels, prefix → live (`metrics/box-colour.txt`): 574: - quarter-r16: #010b04 → #032809, green lead +7.5 → +31.4; 575: - harbour-day: #081e11 → #103615, luma 24 → 44; 576: - quarter-night: #0d1b1c → #0c301a, green lead −1.1 → +22.0; 577: - harbour-r2-night: #1f2325 → #233621. 578: - Look crops at 3-6x (`look/t5-*.png`; pre-trim / mound / final) and the tuning crops (`look/t2-*` to `t4-*`). The planting reads green by day and olive-green under the lamp term at night, at Harbour and quarter zoom. On a sunlit wall (amsterdam-near) it is a leaf green, not lime. 579: - **Whole-stage bench** (`fix2/bench.sh`, `summary.js` → `bench/summary.txt`, `release-summary.txt` and `drag-summary.txt`; 01:01-01:38 on 6 October; AC, Low Power Mode 0 at every start and end, `bench/power.txt`): 580: - Sides, interleaved: pre-stage `stage-tree` (25-P7), pre-trim `prefix-tree` and live. Rounds 1-3, raw p95 medians, `--disable-vsync`, 1920x1080, rain 0.45. The release probe and the drag harness also ran `nobox-tree`: the live code with `dressing_seed` returning −1, which gives P7f's own share. 582: | Dense raw p95 | 01:00 | 13:00 | 19:30 | 583: |---|---|---|---| 584: | stage → live | 15.09 → 16.09 (**+6.6%**) | 14.63 → 15.79 (**+7.9%**) | 16.66 → 17.71 (**+6.3%**) | 585: | pre-trim, same sitting | 16.43 (+8.9%) | 16.30 (+11.4%) | 18.21 (+9.3%) | 586: | live − pre-trim, paired | −0.43 −0.34 −0.31 | −0.43 −0.51 −0.52 | −0.48 −0.50 −0.49 | 588: - Target under +8% at all three times: **met**; 13:00 is close at 7.9%. 589: - Dense-bare: +13.3% / +13.0% / +11.9% (pre-trim +15.9% / +16.2% / +14.7%). Dense-bare is dense without its hand-placed roof parts, so P7d's roofscape is drawn there. 590: - Harbour p95: within noise. The per-round spread is up to 1.5 ms at 7-8 ms. 591: - Draw calls: dense 2362 → 2583 (19:30: 3126 → 3433), Harbour 903 → 907, the same before and after the trim. 592: - **Harbour animated release** (D25): 29.65 → 31.35 ms (**+1.70**, within +2; pre-trim +1.80). P7f's share is +1.35 (pre-trim +1.45). Without animations +1.38. Dense with animations +2.67 (pre-trim +3.05). 593: - **Drag harness** `step_total`: 594: - handle 17.06 → 17.38 (+0.32; pre-trim +0.21). The nobox tree reads 17.55, so P7f's share is none. 595: - terrain raise 125.98 → 128.45 (**+2.47, still over P7d's 2 ms line**; pre-trim +3.18). P7f's share fell from +1.85 to +1.14 ms. The other +1.33 (nobox − stage) is not P7f's. 596: - Drag draw calls 883 → 889, the same before and after the trim. 597: - **Owner sheet:** pages 1, 6, 8, 9 and 10 of the P7 close sheet were re-made from a new sitting (`fix2/sheet.sh`; before = the pre-stage tree, after = the trim). They are copied into `close/sheet/`, with the old pages in `close/sheet/pre-trim/`, and `sheet.md` is updated. Pages 6 and 8 were looked at; 1, 9 and 10 were made but not opened. 598: - Scratch trees `fix2/stage-tree`, `pre-tree`, `prefix-tree` and `nobox-tree` are left in place. 599: - Review: (filled by the orchestrator). 600: - Not done: 601: - The release probe and the drag harness were not run; the plan's acceptance names only p95. 602: - The dense +8.2% is close to the limit. 603: - Platforms: Mac run; Windows not run. 605: **Beautify P7e Ivy spill (P7e and P10d merged) (5 October 2026):** Opus 5.5 on this Mac (plan section "P7e Ivy spill"; built on the P7h ivy MultiMeshes and instance tint). `Ivy.grow` now hangs derived curtains of ivy from planted roofs and from wall planters and balconies with pots, through one mechanism (`_spill`): a lip leaf and a crest leaf at the edge, then strands down the wall below, widest-reaching in the middle and frayed at the tips. Curtains are memoised per owner (`_spill_memo`, key: column count, blocker key, sources). They are counted after moss and before base ivy, capped at `SPILL_SHARE` 0.25 of the room left, and stop at the first curtain that does not fit (the rest are counted as `spill_skipped`). 606: - **Sources** (`_spill_sources`, which reads `Composition.roofs` and is a cache hit in the release order): a roof qualifies when it holds a plant (derived or placed: roof_garden, planter, pot, flower_patch, shrub, bonsai, wildflowers, tree, blossom, conifer) and has an edge over open ground or water with a drop of at least 2.4 m and a length of at least 1.2 m. It gets 1-2 curtains (seeded), less one for each placed drape. Curtains never hang over a neon, neon bar, neon strip, awning, door or arch (`_spill_clear`, `_spill_doors`). Strand length is `clamp(0.35 x drop, 1.2, 3.5)` (`spill_length`). 607: - **Deviations:** 608: - spill may hang over windows (the plan forbids only neon, awning and door); it stands `SPILL_HANG_OUT` 0.14 m off the wall there; 609: - `SPILL_WIDTH` is 4-8 strands, not the report's 3-8; 610: - the lip and crest leaves are free leaves, not cells (jettied and top-storey faces have no valid cell under the eave); 611: - the minimum drop is 2.4 m; 612: - a placed drape counts towards the 1-2 curtains; 613: - `roof_dressing` false clears the spill; 614: - the growth order makes the tips young and the crest old, not the report's "new-growth tint at the top"; 615: - the budget early stop; 616: - `world_view.gd` and `architecture.gd` are unchanged (spill leaves are ordinary ivy pieces in the existing `ivy_g` MultiMeshes). 617: - `Composition.roofs` takes an optional `known_sig`, so the spill reuses the release's roofs result. 619: Files: `scripts/ivy.gd`, `scripts/composition.gd`, `tests/test_ivy.gd`, `tests/test_roofs.gd` (`git diff --numstat`: 935 added, 15 removed; 997 and 17 after the review fix). Snapshot: `build/backup-feel/25-P7-e/` (main 648feab). At part start the live tree matched it (`e/diff-live-vs-25-P7-e.txt`, empty). Evidence: `artifacts/beautify/25-P7/e/`; the pre-part tree is `e/pre-tree`; backups are in `e/backup/`; probes are in `e/probe/` (`accept.gd`, `spills.gd`, `drape_sig.gd`, `qsrc.gd`). No new fixture: the tests read artwork.json, styles.json and dense.json, which are already in `tools/repo_evidence.sh`. 620: - **Tests first:** `e/testfirst/test_ivy.log(.out)`: 188 checks, 1 failure (`spill_length`; `spill_checks` returns after that guard, so its other checks did not run before the code). The 3 `test_roofs` checks were written after the code. 621: - **Tests after:** 622: - `test_ivy` 187 → **236** (`e/suites/test_ivy.log.out`; 237 after the review fix, below). The new `spill_checks` cover: 623: - one mechanism for roofs, planters, wall planters, balconies with pots and roof gardens; 624: - 1-2 curtains per qualifying roof, none over a neon, awning or door; 625: - the length clamp; 626: - the budget order (after moss, before base ivy); 627: - placed drapes draw exactly as before (`DRAPE_SIG_PRE` 4002483335, from `probe/drape-pre.log.out`); 628: - the per-owner memo (moving one planted house recomputes only its spill, 1 computed and 2 replayed, and the result equals a cold pass). 629: - `test_roofs` 259 → **262** (the `known_sig` hit; another hash recomputes; every roof that spills holds a plant). 630: - Full suite (`e/full/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0): **42 suites, 22,144 checks, 0 failures.** `--verify` 2880x1800 exit 0 on try 1, `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun was clean. 1440x900 was not run (inherited stall). The `terrain_flowers.svg` not-found backtrace in `test_release_spread` and verify is pre-existing (the same line is in `h/full`). 631: - **Changed expectations (old → new):** 632: - `test_ivy` seed_checks: "No leaf above the eave" → "No climbing leaf above the eave" (spill leaves start at the lip). 633: - "A house on stone paving ... takes no ivy" → "... takes no climbing ivy" (P7e: its roof may spill). 634: - scene_checks: "no leaf overlaps a window, door or arch" → "no climbing leaf overlaps a window, door or arch" and "no spill leaf overlaps a door or arch". 635: - scene_checks fixture: it runs `C.roofs(s)` before it times the cold pass, as the release order does. 636: - **Journey count deltas expected** (`e/metrics/counts-pre1-after1.txt`; pre-1 → pre-2 and after-1 → after-2 are identical): 637: - harbour-night, -day, -lane and lane-day: foliage.ivy 395 → 440, foliage.total 616 → 661; 638: - quarter-day and -night: foliage.ivy 369 → 465, foliage.total 454 → 550; 639: - styles-day: draw_calls 491 → 494, rendered_objects 567 → 570, foliage.multimeshes 1 → 4, foliage.ivy 0 → 40, foliage.total 185 → 225; 640: - dense: 3 more draw calls (2362 → 2365, smoke bench). 641: - **Captures** (`e/caps.sh`; pre-1, after-1, pre-2, after-2 in one sitting; AC, LPM 0; `cap/power-*.txt`). The seven views are harbour-night, -day, -lane, quarter-night, -day, styles-day and lane-day. Close-ups are in `cap/close-pre` and `cap/close-after`. 642: - Systematic pixels (`metrics/systematic-8-2.txt`; maps `look/sys-*.png`): harbour-day 0.03%, harbour-lane 0.13%, harbour-night 0.02%, lane-day 0.12%, quarter-day 0.48%, quarter-night 0.29%, styles-day 0.49%. Some of the styles pixels are rain-ring phase over the water. 643: - Looks: `look/styles-modern.png` (a tapering curtain from the planted parapet), `look/lane-day-r9.png` (a curtain under a wall planter), `look/quarter-night-r6.png` (a battlement fringe and a curtain). 644: - `look/quarter-night-right.png`: the right block is unchanged. Its roofs face lower roofs or already carry placed drapes (`probe/qsrc.log.out`). 645: - **Acceptance** (`probe/accept.log.out`): 646: - **Curtains a third of the way down: met on the median.** The median of each roof curtain's median strand reach is 0.31 of the drop, over 26 curtains (min 0.04, max 0.46). The fringe-only curtains (Harbour roof 7 at 0.04, quarter roofs 4 and 6 at 0.06) have their strands cut by covers or signs. Part curtains (wall planters): median 0.35 over 2. 647: - **Tapered** (middle strands reach further than the outer ones): Harbour 3/3, styles 1/1, quarter 4/6, chemistry 4/4, lane 5/5, dense 9/9. 648: - **Budget respected:** every scene is within BUDGET 2000. Dense spill is 272 leaves (the 25% cap), with 56 trimmed and 94 sources skipped. 649: - Spill pass cost (`probe/spills-6.log`): cold Harbour 2.1 ms, quarter 4.5, dense 7.0. Warm per house move: Harbour 0.27 ms, quarter 2.6-3.2, dense 2.6. 650: - **Bench** (same sitting, AC 100%, LPM 0, `e/bench/power.txt`; `e/bench.sh`): 651: - **Release (D25, +2 ms on Harbour): met.** Release frame, median of 3 alternating runs (`bench/release-summary.txt`): Harbour anim off 24.65 → 24.54 ms (-0.11), anim on 29.80 → 29.88 (+0.08). The Harbour spread total is 55.1 → 56.5 ms and its worst slice 7.36 → 7.82. The ivy step is 5.21 → 5.99 ms and ivy_draw 1.86 → 2.11. 652: - Dense release frame: +1.65 ms (anim off) and +1.98 (anim on). Dense-bare: +1.58 and +1.75. The dense ivy step is 17.8 → 23.4 ms (+5.6); it is now dense's worst slice, 22.94 → 23.44. The dense-bare worst slice is 17.36 → 21.32. 653: - **Static frame** (`bench/smoke-summary-t1.txt`, 1920x1080, `--disable-vsync`, rain 0.45, rounds 1-3): dense raw p95 15.047 → 15.325 ms (+1.8%; paired +0.29, +0.00, +0.23), draw calls 2362 → 2365. Harbour raw p95 7.258 → 7.778 (+7.2%), but the paired deltas are +0.93, -0.21 and -0.01: one noisy pair. Its raw p50 is 6.532 → 6.549 and its draw calls are 903 → 903. 654: - **P7h notes for the stage sheet:** the quarter-day west wall still shows a weaker diagonal lattice. The P7h fit keeps the old rule (0 to 4 small leaves per owner on Harbour). 655: - **Review fix (5 October 2026, must-fix 1 of 1):** the review found that spill leaves skipped every rectangle before `open_n`, and that included the victorian oriel bays `_building_columns` adds to the base rects. On a flat-roofed victorian house with a wide front, a roof curtain hung at `SPILL_HANG_OUT` 0.14 m inside an oriel that stands `ORIEL_DEPTH` 0.40 m proud (review probe `e/review/oriel.log.out`: 3 leaves inside with 3 storeys, 8-9 with 4). No shipped starter was affected (`e/review/oriel_scenes.log.out`). 656: - **Fix:** `_building_columns` collects the oriel rectangles apart (`proud`), adds them last and records `base_open`, the rect count before them. `_blocked` copies the rects up to `base_open`, adds the placed windows, sets `open_n`, then appends the oriels before the covers and the other parts. Climbing ivy sees the same rectangles as before, in a different order (order does not change a fit). Wall and tower columns have no `base_open` and keep the old order. 657: - **Test first:** new `oriel_spill_checks` in `test_ivy` (the review probe's four houses: 3 and 4 storeys, 8 and 10 m fronts). Before the fix: `e/fix1/test_ivy-fail.log.out`, 237 checks, 1 failure (23 of 186 front spill leaves inside an oriel). After: `e/fix1/test_ivy-fix.log.out`, **237 checks, 0 failures** (0 of 159). 658: - **Re-measured:** review probes re-run on the live tree, `e/fix1/oriel.log.out` (inside_oriel 0 in all four cases; front spill leaves 38, 37, 42 and 42, down from 40, 40, 53 and 53) and `e/fix1/oriel_scenes.log.out` (byte-identical `OS` lines to the review's: amsterdam 243, courtyard 51, castle 58, harbour 45, quarter 102, styles 40, demos 0). The `P7E_SPILL` and `B4C_COST` leaf counts in `test_ivy` are unchanged for Harbour, quarter, styles, dense, garden, chemistry and lane, so the journey count deltas above stand. 659: - **Full suite** (`e/fix1/full/suite_all.out`, VERIFY_RES=2880x1800, AC 100%, Low Power Mode 0, `e/fix1/power.txt`): **42 suites, 22,145 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1, `PIXELPUNK_INTEGRATION []`; the `test_creature_reactions` rerun was clean. 660: - **Captures** (`e/fix1/caps.sh`; scene `e/fix1/scene/oriel.json` from `e/fix1/make_oriel.gd`, the 4-storey 10 m case; before from `e/fix1/bug-tree`, the live tree with the pre-fix `ivy.gd` from `e/fix1/backup/`; same sitting, AC, LPM 0). `e/fix1/oriel-side-pair.png` and `oriel-front-pair.png` (2x crops, before left): before, leaves sit on and through the oriel's dark cap; after, the cap is clear, the strands stop above it and the side strands hang beside it. Pixels (`e/fix1/metrics.txt`): before against after 2.50% / 2.01% changed (front / side), >24 levels 1.06% / 1.03%; noise pairs 1.58-2.21% / 2.47-2.54%, >24 levels 0.42-0.45% / 0.48-0.51%. Draw calls 145 and 127, unchanged. foliage.ivy 110 → 99 on that scene. 661: - **Changed expectations:** none changed; one check added. Bench: not re-run (the fix reorders one array copy per column; `test_ivy` B4C_COST cold ms Harbour 47.2 → 48.2, dense 172.8 → 178.1, within run-to-run spread of the earlier logs, for example 177.4 in `e/review/test_ivy.log.out`). 662: - Review: filled by the orchestrator. Not done: no owner sheet; the quarter's right block gains no curtain (its roofs face lower roofs or carry drapes, as designed); the Harbour p95 pair noise was not re-run. Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 6 October 2026, 16:17 BST, game main at 58c7827) 218: ## Orchestrator ruling on P7f (5 October 2026, owner not asked) 220: | Item | Ruling | 221: |---|---| 222: | Dense p95 +8.2% (15.991 to 17.309 ms; box planting adds about 1,360 leaf vertices per box, 432 boxes on dense, +21.7% cluster vertices) against the plan's +10% | Accepted as acceptance-compliant, reported. No speculative trim now: the P7 stage close benches the whole stage against the pre-stage tree, and if dense p95 there is over +10% a cheaper planting (fewer shells or lower-poly lobes) is a directed fix then. Release-time for P7f was not measured by the part (reviewer's single round: Harbour +0.51 ms anim off, +1.77 ms anim on, within D25); the stage close runs the 3-round release bench. | 224: ## P7 approved (owner, 6 October 2026, in chat) 226: "Move on to p8 when done. P7 approved." Given while the directed fix on P7f's box planting was still running and before the P7 stage sheet (`artifacts/beautify/25-P7/close/sheet/`) had been sent. Read by the orchestrator as approval of the whole of P7 as it will stand after that fix: P7e (ivy spill), P7f (window boxes and Juliet rails), P7g (downpipes and vent stacks), P7h (ivy grid fix) and the stage close, including the looks listed for the sheet (ivy curtains reading as dark clumps on the white styles block, pipes close to the existing copper pipes on the quarter right block, the darker Harbour courtyard lawn at night, the lane sign core 143 to 140, the courtyard corner vent) and the cost lines (Harbour animated release at +2.01 ms against D25, terrain-raise drag +3.1 ms). Dense p95 stays the orchestrator's: the trim must bring it under +10% or it is reported again. P8 follows when the fix finishes.