Chapter 26: A town that moves Frozen 6 October 2026 at 16:17 BST (game main at 58c7827). Implemented, reviewed and merged parts of the P8 stage (motion), which is NOT closed: sub-stage P8-I (P8a ambient-motion preference, motion clock and recording recipe; P8b faulty tubes, breathing pools, stopped-clock fix and light_time wrap; P8c OmniLights breathing and stuttering with their tube, with the float32 seed-rounding review fix) and its close; sub-stage P8-II (P8d rain drops one art pixel wide, with three review fixes) and its close. Owner answers: P8-I approved and P8-II approved (twice) on 6 October; orchestrator rulings after P8b. Not covered: P8g1 (wind) and later parts, which were not merged at the freeze; the P8g1 rows of owner-decisions.md are left out. Original source terminology is preserved. Line numbers refer to the frozen copy. Part codes: P8a motion setting, P8b night life in the shaders, P8c night life on the lights, P8d rain in art pixels; D17 motion setting scope, D27 faulty tubes on by default, D28 the rain veil. 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) 5: **Beautify P8-II (parts d of P8) stage close (6 October 2026):** this is the stage-end overhead of sub-stage P8-II "Motion" (step `26-P8-II` in the plan's overhead table). Opus 5.5 ran it on this Mac on the final P8-II code: `main` 2de5e65, clean (`close/head.txt`). The only part is P8d, with its three review fixes. No game code changed. The evidence is in `artifacts/beautify/26-P8-II/close/`. 6: - **Before side:** the pre-stage scratch tree `close/pre-tree`, built from `build/backup-feel/26-P8-I/`. It differs from the live tree only in P8d's files, and its `world_view.gd`, `test_wet.gd` and `test_camera.gd` equal `7f11b43` (the P8-I close). It has no `rain.gdshader` (`close/diff-pretree.txt`). 7: - **Power:** every run was on AC, 100% charged, with Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`, `bench-r2/power.txt`). Load average was 2.1-4.2. 8: - **Full suite** (`close/suite.sh` a, b, v; 11:09-11:17): **42 suites, 22,686 checks, 0 failures** (`close/suite/suite_all.out`; `test_wet` 328, `test_camera` 247). `--verify` at 2880x1800 exited 0 on try 1 in 32 s, with `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). 9: - **Journey** (`artifacts/feel-journey/26-P8-II/`; six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): 10: - Counts against 26-P8-I (`26-P8-II/counts-vs-26-P8-I.txt`) are identical in all six views, draw calls included. That matches P8d's declared delta of none, so no single-view re-run was needed. 11: - Frame >8 luma against 26-P8-I is 0.44-1.88% (`26-P8-II/frame-vs-26-P8-I.txt`). I looked at `close/look/j-1.png` and `j-2.png`: the night rain is finer and fainter, and nothing else differs visibly. 12: - `notes.json` has a `26-P8-II` entry (`close/journey-note.js`, backup `notes.json.bak-before-26-P8-II`). `index.html` and `journey-standalone.html` (108.91 MB) were rebuilt. 13: - **Art capture and protected looks** (`close/caps.sh`; 11:19-12:15, every view on try 1). The P8-I close's groups g1 (58 views) and g2 (30) are reused. Group g3 is P8-I's 8 views plus the P8d rain field (`26-P8/d/fixture/rainfield.json`) at the Harbour, quarter and lane cameras (`field-h`, `field-q`, `field-l`). That makes 99 views per set, captured as pre-1, after, pre-2, after-2. 14: - The metrics are in `close/metrics/` (`close/metrics.sh`, `frame-table.js`, `prot.js`, `consistent.js`, `elapsed.js`; regions `close/regions.json` = the P8-I close's): `ab-*`, `frame-table.txt`, `counts.txt`, `protected.md`, `consistent.txt` (maps in `close/look/consistent/`) and `elapsed-ms.txt`. Against the previous stage close, `26-P8-I/close/after`, they are `ab-p8iclose-*` and `frame-p8iclose-*`. 15: - The protected sheets are `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P8-I-close.png`, `noise-pre.png` and `noise-after.png`. I looked at page 1. 16: - **A first after-2 sitting was discarded** and moved to `close/after-2-odd/`. In it, many views ran capped at about 120 fps: 52 of 99 have a mean_fps more than 1.4x below the same view in after, against 200-540 fps in the other three sets. Two of them came out wrong: grove-day was framed differently (29% of the frame changed), and disc-gg-sun145 had 764 rendered objects against 847. Single-view re-runs (`close/rerun.sh`, `close/rerun/`: live twice, pre once) gave 305/381 and 771/847 every time, so the capture was at fault, not the code. The cause is not known. It looks like a display or window state change; no other Godot was running (every run waits for none). All three groups of after-2 were then captured again (12:00-12:15) with no outliers. The tables above use that second capture. 17: - **Counts:** identical in every fixed view in both pairs. The orbit views vary within one tree, as at P6, P7 and P8-I (`qn-orbit` and `quarter-orbit`). The P8-I close's after against today's pre-1 differs only in `quarter-orbit`. 18: - **Frame changes:** in 6 of 99 views, both before→after legs are above both noise legs: harbour-0545, harbour-2100, harbour-night, quarter-2200, styles-golden and styles-night. The pixels that change in all four cross pairs and in neither noise pair cover 0.66-3.9% of those frames (`consistent.txt`). They are rain rings on open water, water reflections and cloud dither, all on `TIME` (`close/look/consistent-4.png`, looked at). The drops cannot cause a consistent change, because every still has its own drop positions. In this sitting the live captures were not systematically later (`elapsed_ms` median difference 0 ms; later in both pairs in 15 of 99 views). So the P8-I close's `TIME`-phase explanation fits less well here, and the cause is still not explained. 19: - **Canal mirror columns:** p50 44.0 → 43.8 (44.1 → 43.8) in quarter-night and 52.5 → 52.2 (52.6 → 51.9) at 21:00. The changed-pixel shares are mixed: 9.7% / 29.5% against noise of 21.2% / 9.0%, and 10.2% / 29.0% against 19.7% / 16.5%. These are reflections plus finer drops over the water. 20: - **HDR neon cores:** within noise. Fish 66.3 → 66.4, bar 70.2 → 70.6 (pre-2 70.6), lane sign 140.1 → 140.2, tower window 96.0 = 96.0. 21: - **05:45 and 20:45 skies:** within noise (p50 within ±0.3). 22: - **Glade/Canal separation:** Glade night zenith `#001d45`, horizon and water are identical, and Glade lawn and bank 0%. The Harbour night zenith changes 22.7% / 23.9% in both pairs against noise of about 1%. But the P8-I close's after, which has the same code as the pre-tree, run from the live folder, matches today's after to 0.18%. So the difference is between the scratch tree and the live folder (cloud dither phase), not P8d. The hex is `#2b337f` on both sides. 23: - **Approved tree style:** unchanged. Grove crowns 55.0 → 55.1, conifers 55.6 = 55.6, quarter undersides 27.7 = 27.7, golden grove 48.5 = 48.5. 24: - **Whole-stage A/B bench** (`close/bench.sh`; `bench-t1.out`, `-t13.out`, `-t19.5.out`; `bench/summary.txt` from `close/bench-summary.js`; 12:24-12:37). The setup was the P8-I close recipe: pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync` before `--path`, 1920x1080, rain 0.45, `--motion-live`. The scenes were dense `artifacts/weather-pass9/dense.json` and Harbour, at 01:00, 13:00 and 19:30. Every run has 491 raw samples with vsync disabled. One run needed a second try (harbour-pre-1 at 19:30 left no `performance.json`). Raw p50 and p95 medians, pre → live (paired median; rounds where live was slower): 26: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 27: |---|---|---|---|---|---|---| 28: | Dense | 13.933 → 13.927 (−0.006, 0/3) | 15.695 → 15.735 (−0.046, 1/3) | 13.681 → 13.616 (+0.000, 1/3) | 15.410 → 15.355 (−0.035, 1/3) | 16.786 → 16.934 (+0.134, 3/3) | 18.682 → 19.122 (+0.440, 2/3) | 29: | Harbour | 6.623 → 6.625 (+0.005, 2/3) | 8.091 → 8.497 (+0.406, 3/3) | 6.251 → 6.285 (+0.039, 3/3) | 7.264 → 7.500 (+0.423, 2/3) | 7.577 → 7.628 (+0.047, 2/3) | 9.068 → 9.254 (+0.126, 2/3) | 31: - The first 19:30 sitting ran while load rose to 4.2, and its dense round 1 was 15.69 against 17.25. So 19:30 was run again (`close/bench-r2.sh`, `bench-r2-t19.5.out`, `bench-r2/summary.txt`; 12:37-12:46, load 3.5-4.1). The repeat gave dense p50 16.555 → 16.536 (−0.019, 1/3) and p95 18.106 → 18.132 (+0.012, 2/3), and Harbour p50 7.937 → 8.006 (−0.002, 1/3) and p95 9.130 → 9.123 (−0.007, 1/3). That is within noise. 32: - Draw calls are identical: dense 2583 (19:30: 3433), Harbour 907 (19:30: 822). 33: - The Harbour p95 rounds scatter by 0.4-1.5 ms on both sides. The Harbour p50 at 13:00 is +0.039 ms (0.6%) in 3/3 rounds, the same lean as P8d fix 2's +0.037 at 01:00. More drops fall inside the frame for the same particle count. 34: - **Owner sheet** `artifacts/beautify/26-P8-II/close/sheet/` (`close/sheet.sh`, regions `close/sheet-regions.json`, crop picker `close/pick.js`). Before is on the left and after on the right. The close-ups are shown in colour and greyscale. 35: - Pages: 36: - `p8ii-01a-drops-close-up-lane.png` 37: - `p8ii-01b-drops-close-up-field.png` 38: - `p8ii-02-lane-night.png` 39: - `p8ii-03-quarter-night.png` 40: - `p8ii-04-harbour-night.png` 41: - `p8ii-05-existing-scenes.png` 42: - `p8ii-06-what-differs.png` 43: - The recording frames are frame 150 of P8d's clips. Lane and quarter come from `26-P8/d/fix2/rec/`. Harbour comes from `26-P8/d/fix3/rec/`, which ran on the current file. The stills come from this close. 44: - `sheet.md` has one plain paragraph for P8d with its acceptance numbers, then what existing scenes will see, the bench and the suite, and the D28 question. 45: - I looked at every page. Page 05's glade-rain thumbnails seem to show a dimmer lamp pool after, but at full size the four captures match (`close/look/glade-rain-4.png`), and the frame change, 2.48% / 1.50%, is inside noise (2.27% / 1.55%). 46: - **Not done / open:** 47: - The owner has not seen the sheet. **D28** (the veil: P8e/P8f, lamp-lit drops, or stop) is the gate before P8-IV; P8-III can run meanwhile. D27 and P8b's questions are still answered only by orchestrator rulings. 48: - From P8d: the rain is not pre-warmed (`ShaderWarmup` skips CPUParticles3D), and the perspective and photo-mode lenses were not captured. 49: - The P8 recording sheet of every clip is left for the P8 close. This sub-stage made no new recordings, and P8d's frames were already moved to the Trash by the part, apart from the kept `f0150.png` frames and clips. 50: - Not explained: the consistent water and cloud changes in 6 views, and the capped after-2 sitting. 51: - No 1440x900 `--verify`. 52: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 53: - Scratch trees: `close/pre-tree` was moved to the macOS Trash with `mv` as `beautify-26-P8-II-close-pre-tree-124946`, once its numbers were recorded. P8d's part trees (`26-P8/d/pre-tree`, `26-P8/d/fix2/d2-tree`) are outside `26-P8-II/` and were left in place. 54: - Platforms: Mac run; Windows not run. 56: **Beautify P8d Rain drops one art pixel wide at every zoom (6 October 2026):** Opus 5.5 on this Mac (plan section "P8d"; report polish backlog "rain in art pixels" (drops); step `26-P8`, sub-stage P8-II, with its own bench so the D28 veil decision does not block). Each rain drop is now a rectangle of whole art pixels, 1 wide and 6, 7 or 8 long, at every zoom and lens, and the rain volume follows the zoom, so a frame shows the same drops at the same art-pixel speed whatever the zoom; review fix 2 (below) raises the band the rain covers over the scene's tallest content, so the upper storeys have rain in front of them at lane and quarter zooms. The old drop was a 0.028 x 0.85 m billboard: a 1 px light core with a dark presentation-outline column each side (3 columns), 10 art px long at Harbour and 12-18 in the quarter, flickering in and out of the pixel grid as it drifted. D28 (the veil) is untouched: that is P8e/P8f. 57: - **Code:** 58: - `shaders/rain.gdshader` (new): the vertex stage projects the particle's origin, snaps it to the viewport's pixel grid and writes `POSITION` for a `WIDTH_PX` 1 by `LENGTH_MIN` 6 + (0-2) pixel rectangle. Length and alpha (`ALPHA_MIN` 0.16 to `ALPHA_MAX` 0.34) come from `drop_hash(INSTANCE_ID)`, so a drop keeps them while it falls. A dead particle (CPUParticles3D writes it a zero transform) draws nothing. Colour: the old drop's (0.66, 0.72, 0.95) mixed `HAZE_TINT` 0.35 toward the hour's haze. Unshaded, `blend_mix`, `depth_draw_never`, `cull_disabled`; the distance fog still applies, as before. 59: - `scripts/world_view.gd`: `_build_rain` uses the shader (the quad's size is only its bounds); `sync_profile` sends `haze_colour`; `update_camera` calls `_place_rain`, which sets the node transform, box, direction and speed from the new static `rain_volume(target, yaw, pitch, size, aspect)`. The whole old Harbour system (box height, spawn height, speed 16-20 m/s) is scaled by size / `RAIN_SIZE_REF` 38, turned to the view's yaw, and set toward the camera so every ray of the frame is inside the box from the water up to `RAIN_COVER_TOP` 20 m x scale. The world fall direction stays `RAIN_FALL` (-0.10, -1, 0.05). Photo mode's free camera keeps the old fixed 35 x 12 x 35 box over the target. The amount stays the rain setting's (`rain x 2600`, 909 at 0.35); the zoom never writes it, because setting `CPUParticles3D.amount` kills every drop at once. 60: - **Deviation from the plan's numbers:** the box is not an axis-aligned "about 1.15 x camera.size". The half-extents are 37.1 x 12 x 61.9 m at Harbour (0.98 / 1.63 x size across / toward the camera, 13.7 m toward it), 20.6 x 6.6 x 58.7 in the quarter (2.79 x size deep at pitch 0.315) and 15.6 x 5.1 x 36.0 at lane. A Monte-Carlo model of the particles (`d/model.js`, `d/model.out`, flat ground, no buildings) gives the plan's 1.15 x size box centred on the target a bottom row at 1% of the mean row density in the quarter and 28% at lane. Alpha does not change by zoom band: the scaled volume keeps the drop count per frame within 12% across the three zooms in the first sitting only (field max/min 1.115); after review fix 2 the whole-frame field max/min is 1.177 (`d/fix2/drops.tsv`, 237.1/201.5; unfixed code 1.165), and only the like-for-like meadow-cell figure (1.11) is within 12% (below), so no compensation was needed. 61: - **Files:** `shaders/rain.gdshader` (new), `scripts/world_view.gd`, `tests/test_wet.gd`, `tests/test_camera.gd`. Backups: `artifacts/beautify/26-P8/d/backup/`. Snapshot `build/backup-feel/26-P8-d/` (= the P8-I close, 7f11b43): the live tree equalled it at part start (diff -rq on scripts, shaders, tests, data, scenes, tools and `project.godot`: empty). Pre-part tree: `artifacts/beautify/26-P8/d/pre-tree`. No new `class_name`. 62: - **Tests:** 63: - `test_wet` 220 → 266 (+46): the shader contract (render modes, constants, `VIEWPORT_SIZE` snap, `INSTANCE_ID` hash, dead-particle guard, the fragment); a GDScript twin of `drop_hash` over 4,000 slots (alpha 0.160-0.340, lengths 6/7/8 about a third each); `rain_volume` for Harbour, quarter, lane and pitches 0.25 and 1.25 (all 567 samples (81 rays of the frame x 7 heights) lie inside the box from the water to the top of the band; the drops fall through the band; world fall = `RAIN_FALL`; speed scales with the zoom; at water height the lane volume is the Harbour one x 16/38); the view (shader material, amount 909, lifetime 1.7, haze at 13:00, 19:30 and 01:00, the free-camera box, dry stops it). 64: - `test_camera` 224 → 244 (+20): in the editor's SubViewport at the three zooms, the rain stands where `rain_volume` puts it; a twin of the vertex stage agrees with `Camera3D.unproject_position` (worst under 0.01 px), and every drop covers whole pixels 1 x 6-8, one art pixel (`art_pixel`) wide; the zoom path never writes `.amount` or calls `restart(`, and a zoom keeps the amount and emission. 65: - The headless dummy renderer does not simulate CPUParticles3D (`capture_aabb` stays empty; `d/probe/probe_aabb*.gd`), so the drop-position half of "no restart on zoom" runs only with a renderer: the test checks it when particles exist, and the windowed probe `d/probe/probe_restart.gd` (`probe_restart.out`) shows `capture_aabb` identical across a zoom 38 → 16 → 38, the drops inside the lane box 2.5 s later, and `amount + 1` emptying it. 66: - Test-first logs: `d/d-testfirst-test_wet.log(.out)` (parse failure on the missing `WorldView.rain_volume` and constants), `d/d-testfirst-test_camera.log(.out)` (2 failures). Affected suites (`d/suites/`): `test_wet` 266, `test_camera` 244, `test_water` 410, `test_warmup` 37, 0 failures. 67: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh d/suite`, a/b/v, 08:05-08:12, AC, Low Power Mode 0, `d/suite-power.txt`): **42 suites, 22,621 checks, 0 failures**. `--verify` 2880x1800 exited 0 on try 1 in 32 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. The umbrella checks are in that suite, unchanged. 1440x900 was not run (the inherited stall). 68: - **Changed expectations:** none in the part itself (only new checks; `test_water`'s amount `int(0.35*2600)` holds). Review fix 2 changes two: `test_wet` "the view's rain stands where rain_volume puts it" and `test_camera` "the rain stands where rain_volume puts it for this view", old: `rain_volume(target, yaw, pitch, size, aspect)` → new: the same with the scene's ceiling (`world.rain_ceiling`) as a sixth argument. **Expected journey count deltas:** none (fix 2 included: it moves only the particle box). Integer counts and draw calls are identical before→after in all ten still views and both noise pairs (`d/metrics/counts-*.txt`). 69: - **Recordings** (`d/rec3.sh`, `tools/art_record.sh --keep-frames`, 10 s, 30 fps, 1440x900, light clock running; frames measured with `d/dropscan.js`, then moved to the Trash; AC, Low Power Mode 0, `d/rec3/power.txt`; table `d/metrics-drops-rec3.tsv` from `d/droptable.js`). Views: harbour-night, quarter-night and lane-night (before-1, after, before-2), and `d/fixture/rainfield.json` (a flat 64 m meadow at night, rain 0.35, no buildings) at the Harbour, quarter and lane cameras (before, after). `dropscan.js`: background = per-pixel median of the 12 neighbouring frames; drop pixels differ by at least 6 luma; streaks are components at most 4 wide and at least 5 tall; 32 sampled frames per clip (pairs n, n+1 for tracking). "Stable" width counts only drop pixels whose background varies by at most 4 luma, so water glitter beside a drop is not booked as width. 70: - Looked at: `rec3/lane-night-after/contact.png`, `rec2/field-h-before/contact.png`, and the frame crops `look/hn-before-1-f150-crop.png`, `hn-after-f150-crop.png`, `field-h-before-f150-crop.png`, `field-h-before-f150-153.png` (old drops at fixed scale, a few per frame, different drops each frame). 71: - Earlier sittings `d/rec/` and `d/rec2/` (same code) are kept but superseded: their scanner took streaks at most 2 wide, which dropped the old 3-column drops (before counts 0.2-14 per frame there). 72: - **Stills** (`d/caps.sh`; before-1 / after / before-2 / after-2, one sitting; journey set + lane-night, lane-day, harbour-golden and `field-night`; AC, `d/cap-power.txt`). Every still has its own drop positions (CPU particles on the real clock). Frame >8 luma (`d/metrics/frame-table.txt`): before→after legs 0.18-7.08% against noise legs 0.69-6.75%. Some views sit above both noise legs in one cross leg (harbour-golden 4.17% / 4.17% against 0.90% and 4.76%; quarter-night 4.62% / 4.29% against 2.23% and 3.16%) and inside them in the other (1.26%, 2.09%): the TIME phase and drop positions, as in the P8-I close. Pixel profile (`d/profile.js` → `d/metrics/drop-profile.txt`, field-night): the old drop is +55 luma on a 51 meadow with a dark column each side (42/43 and 31-43); the new drop is +27 on 48 with neighbours unchanged. Sheet `d/sheet/` (`d/sheet.sh`, before left, after right, 3x): pages 1 to 6 (harbour-night, lane-night, quarter-night, field-night, harbour-day, harbour-golden), all looked at. 73: - **Acceptance, measured (after against before, `metrics-drops-rec3.tsv`):** 74: - **Every drop exactly 1 art px wide at three zooms:** met. On the field, stable drop rows exactly 1 px: 99.2% / 99.1% / 99.1% (Harbour / quarter / lane zoom), and 92.5% / 97.7% / 96.5% of streaks are 1 px in every row (the rest: drops crossing and neighbouring flicker). Before on the field: 23-40% of stable rows, and 96-99% of the old streaks are 3 columns. Stock views: 98.6% / 97.8% / 96.1% stable rows (before 54-65%). By construction (shader and tests) a drop cannot be wider. 75: - **6-8 art px long:** 98.6% / 98.9% / 98.8% of 1-wide streaks on the field (median 7); 95.5-97.8% in the stock views. Before: median 10 at Harbour, 18 in the quarter and **24 at lane** (corrected by the second review: `d/review2/scan.js`, `d/review2/table-rec.tsv`, lane-night and field-l before; this part's `dropscan.js` gave 10 at lane because it split the old long faint lane streaks). After fix 2 (`d/fix2/drops.tsv`): field 98.7% / 98.6% / 98.8%, median 7; stock views 97.0-97.6%. 76: - **Density constant (after review fix 2, below):** across zooms on the field, 237.1 / 208.7 / 201.5 streaks a frame at 38 / 21.05 / 16 (max/min 1.177; the unfixed code in the same sitting 1.165) and 204.5 at size 10. At size 95 the whole frame counts 334.7, but that is the fixture, not the density (review fix 3): the 64 m meadow fills only the middle of the frame and the drops register more often over water and sky. Like for like, cells over the meadow: 24.9 a frame at 95 against 24.2 / 23.2 / 22.4 / 22.7 at 38 / 21.05 / 16 / 10, max/min **1.11** over 10-95 (`d/fix3/field.tsv`; 1.04 in fix 3's own sitting). Down the stock frames, top band ÷ bottom band: lane-night 0.88 / 1.04 → **0.93**, quarter-night 0.77 / 0.76 → **0.88**; harbour-night **0.83**, the mean of eight after clips over three sittings and a trial (one clip 0.74-0.91), against 0.84 for seven before clips (review fix 3). The numbers that follow on this line are the part's first measurement, before fix 2: not met down the frame in the stock views (review fix 1 below). On the field, 232.6 / 208.7 / 208.6 streaks a frame at the three zooms (max/min 1.115). Before: 98.4 / 80.4 / 61.4 (1.60). Stock scenes after: harbour-night 214.8, quarter-night 230.2, lane-night 116.1 (before 103.9 / 97.6 / 48.6, noise pair within 1.4 a frame); at lane the houses fill the frame and hide or swamp the faint drops. Within a stock frame the drops now thicken toward the bottom. Top band ÷ bottom band of ten row bands, before (two legs) → after: lane-night 1.02 / 1.13 → **0.44**, quarter-night 0.76 / 0.76 → **0.61**, harbour-night 0.89 / 0.86 → **0.62** (`d/fix1/bands.tsv`). 77: - **No bare strips at lane zoom:** met. After review fix 2 every one of the ten bands has more drops a frame than both before legs in all three stock views (lane 10.0-13.9 against at most 6.6; `d/fix2/bands.tsv`), and the 3x3 grid min/mean is lane-night 0.733 / 0.757 → **0.864**, quarter-night 0.418 / 0.413 → **0.820**, harbour-night 0.766 / 0.775 → **0.755** (the unfixed code 0.568, 0.366, 0.582). Before fix 2 (the rest of this line) the stock views were less even than before. The field's 3x3 grid min/mean is 0.955 at lane and 0.93 at the quarter camera (before 0.738 and 0.427, the bottom row bare); Harbour 0.86 (before 0.555). Stock views, grid min/mean, before-1 / before-2 → after: `metrics-drops-rec3.tsv` harbour-night 0.785 / 0.676 → **0.642**, quarter-night 0.438 / 0.456 → **0.358**, lane-night 0.775 / 0.676 → **0.579**, each after below both before legs; `d/fix1/bands.tsv` (new sitting) 0.773 / 0.804 → **0.624**, 0.419 / 0.419 → 0.424, 0.744 / 0.802 → **0.531**. Every one of the ten bands has at least as many drops a frame as both before legs in all three views (lane top band 5.3 / 5.1 → 6.8, bottom 5.2 / 4.5 → 15.3). 78: - **No crawl:** a drop reappears in the next frame in the same column ±1, 1-40 rows lower: 94.6% / 95.3% / 95.9% on the field, 82-89% in the stock views (after fix 2: 94.8% / 95.1% / 96.2% and 85-88%). Before: 32% / 42% / 14% and 13-35% by this part's scanner; **at lane zoom the second review's scanner gives 61% (lane-night) and 69% (field-l)** (`d/review2/table-rec.tsv`, crawlMatched), the corrected before figures: the old lane drop did reappear, as a 24 px streak, and this part's scanner split it. It keeps its length on reappearing in 94-97% (before 59-75%). Fall: 6-7 art px a frame at every zoom (before 6, 12 and 14-15). 79: - **No particle restart on zoom; default amount unchanged; alpha 0.16-0.34 tinted from the haze; umbrella unchanged:** the windowed probe and tests above; amount 909 at 0.35. 80: - **Bench** (`d/bench.sh`, `bench/summary.txt` from `d/bench-summary.js`; pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, rain 0.45, `--motion-live`; dense and Harbour at 01:00 and 13:00; every run on try 1, 491 raw samples; AC, Low Power Mode 0, `bench/power.txt`; 07:57-08:04). Raw medians pre → live (paired median, rounds live slower): 81: - 01:00: dense p50 13.894 → 13.905 (+0.011, 2/3), p95 15.672 → 15.663 (−0.009, 1/3); Harbour p50 6.621 → 6.607 (−0.009, 0/3), p95 7.237 → 7.296 (+0.059, 2/3). 82: - 13:00: dense p50 13.932 → 13.902 (+0.019, 2/3), p95 15.725 → 15.805 (+0.009, 2/3); Harbour p50 6.240 → 6.256 (+0.023, 3/3), p95 7.988 → 7.846 (−0.198, 0/3). 83: - Draw calls identical (2583, 907). Within noise. 84: - **Open for the owner (look, with D28):** the new drops have about half the old contrast (+27 against +55 luma on the meadow; at night the haze tint darkens them further) and no dark rim, and there are more of them per frame: 2.07-2.41 times in the first sitting's stock views (`d/metrics-drops-rec3.tsv`), but about 1.7 times at harbour-night (177.0 against 101.4/102.8) and 1.8 times at quarter-night (183.6 against 100.9/102.1) on the shipped code (`d/fix2/drops.tsv`; on the field, 2.6 times at quarter and 3.4 times at lane zoom). In stills they read as a fainter, finer veil (`d/sheet/`). `HAZE_TINT` and the alpha band are one-line constants. Down the stock frames the rain was heavier at the bottom and thinner at the top than before (review fix 1); review fix 2 raised the band over the tallest content, and lane and quarter are now at least as even as before, and Harbour is as even as before (0.83 against 0.84, means over several clips; review fix 3). The owner sheet `d/sheet/sheet.md` carries these numbers with D28. 85: - **Review:** filled by the orchestrator (measuring part: second review required). 86: - **Review fix 1 (6 October 2026, Opus 5.5 on this Mac; evidence `d/fix1/`).** The second review found one must-fix item. The "Density constant" and "No bare strips" lines used only the flat-field fixture. They left out that drop density across the frame became less even in all three stock views, which the part's own `metrics-drops-rec3.tsv` already showed. The reviewer's scanner (`d/review2/scan.js`, `d/review2/rows-abs.tsv`, recordings `d/review2/rec/`) gave lane-night after 1.51x the mean in the bottom band and 0.62x in the top band; before it was 0.87-1.17. 87: - **Fixed:** documentation only, with no code change. The two acceptance lines above now give the stock-view numbers. The owner question carries them, and `d/sheet/sheet.md` (new) puts them before the owner with D28. At part start the live tree differed from `build/backup-feel/26-P8-d/` only in the part's four files plus `rain.gdshader.uid`, so nothing was unexpected. 88: - **Re-measured** (`d/fix1/rec.sh`: `tools/art_record.sh --keep-frames`, the rec3 recipe; before-1 from the pre-tree, after from live, then before-2 from the pre-tree; 09:05-09:14; AC power, Low Power Mode 0, `d/fix1/rec/power.txt`). The scanner is `d/fix1/dropscan-bands.js`, which is `d/dropscan.js` plus ten row bands, and the table is `d/fix1/bands.tsv` from `d/fix1/bands-table.js`. A first run lost its frames to a wrong `require` path and is kept as `d/fix1/rec-failed-scan/`. The lane-night bands over the mean, top to bottom, are 0.61 0.72 0.74 1.00 1.13 0.82 0.93 1.33 1.32 1.39 after, against 1.05 0.82 0.92 1.02 1.12 0.90 0.93 1.24 0.97 1.02 before. This matches the reviewer's result. 89: - **Cause:** the volume scales with the zoom. `RAIN_COVER_TOP` x k is 8.4 m at lane (k = 16/38) and 11.1 m in the quarter. The spawn slab's top is 11.4 m and 15.0 m over the target. The old fixed box spawned up to 27 m over the target at every zoom, which is about 30 m in the world. The top of the lane and quarter frames shows towers and upper storeys above those heights, so little rain falls in front of them. This shows in `d/fix1/lane-night-top-bottom-2x.png` and `quarter-night-top-bottom-2x.png` (frame 150; top third above, bottom third below; before on the left, after on the right; both looked at): the old long streaks over the lane roof terrace and sky are nearly gone after, while the lawn and quay below carry dense fine drops. At Harbour (k = 1) the heights are unchanged. There the heavier bottom bands (1.23-1.24x after, 0.77-1.03x before) are probably the box reaching 13.7 m further toward the camera, but this has not been proven. 90: - **Full suite** (`d/fix1/suite.sh`: `VERIFY_RES=2880x1800 tools/suite_all.sh d/fix1/suite all`; 09:14-09:22; AC power, Low Power Mode 0, `d/fix1/suite-power.txt`): **42 suites, 22,621 checks, 0 failures**. That includes `test_wet` 266 and `test_camera` 244. `--verify` at 2880x1800 exited 0 on try 1 in 33 s, with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures. 1440x900 was not run (the inherited stall). The changed expectations are still none, and the expected journey count deltas are still none. No bench was run, since no code changed. 91: - **Review fix 2 (6 October 2026, Opus 5.5 on this Mac; evidence `d/fix2/`).** The orchestrator ruled review fix 1's documentation-only answer not enough: the rain density had to be made even down the stock night frames (the band stopped at 8.4 m over the lane target and 11.1 m over the quarter's, below the towers' roofs; top band ÷ bottom band 0.44 / 0.61 / 0.62), without touching contrast, alpha or the particle amount, and the wrong lane "before" numbers corrected (done above: old median length 24 and crawl 61-69% at lane). At the start the live tree equalled `build/backup-feel/26-P8-d2/` (diff -rq on scripts, shaders, tests, data, scenes, tools and `project.godot`: empty), so nothing was unexpected. Backups `d/fix2/backup/`; the unfixed code as a scratch tree `d/fix2/d2-tree` (from `26-P8-d2`, `.tools` linked to the live one). 92: - **Code** (`scripts/world_view.gd`): `rain_volume` takes an optional sixth argument, `ceiling` (a world height, default `-INF`). The covered band is `RAIN_COVER_TOP` x scale + `RAIN_CEILING_REACH` 2.0 x (ceiling − target height, if positive): it rises 2 m for every metre the tallest content stands over the target. The spawn slab keeps its zoom-scaled foot (3 m x scale, so the drops still reach the water) and its top rises with the band (27/20 of it). Speed, slant, width and amount are unchanged; with nothing over the target the volume is exactly the zoom-scaled one, so the flat field fixture is untouched. The new static `rain_ceiling_of(sky, crowns)` is the highest cell of `Composition.sky_map` (buildings and walls with roofs, over the ground) or tree crown; `_refresh_foliage` stores it in `rain_ceiling` on each full foliage pass (a sync, or a spread release's later pass, which then re-places the rain) and `_place_rain` passes it. The stock scenes' ceiling is 15.67 m (Harbour scene) and 15.75 m (artwork), so the band is 33.8 m over the lane target, 39.4 m over the quarter's and 45.3 m at Harbour (was 8.4, 11.1 and 20). 93: - **How the rule was chosen** (trial clips, after only, `d/fix2/trial-*`, same recipe): the reviewer's floor ("max(band, tallest content)") leaves the content at the band's thin top, so it was tried as content + band (`trial-B`: lane 0.90, quarter 0.65, Harbour 0.72), content + 2 x band (`trial-C40`: 0.98 / 0.83 / 0.99, but a flat field with the target on the ground would jump to twice the band), then the continuous rule at reach 2 (`trial-R2`: 0.91 / 0.81 / 0.91) and 3 (`trial-R3`: 1.17 / 0.71 / 0.95, with fewer drops a frame: quarter 116). Reach 2 was kept. The trials show the single top ÷ bottom band ratio moves by about ±0.1 between sittings of the same code (Harbour 0.91 in `trial-R2`, 0.74 in the measured sitting). 94: - **Tests** (test first): `test_wet` 266 → 328 (+62: `rain_ceiling_checks`, the rule at five views and two content heights, every ray of the frame inside the box up to the raised band, the slab's foot and top, the speed, slant and width unchanged, the flat-field identity, `rain_ceiling_of` on the courtyard; and in `rain_view_checks` the view reads the ceiling and spawns over it); `test_camera` 244 → 247 (+3: at the Harbour, quarter and lane zooms the drops are spawned over the Harbour scene's 15.67 m tower by at least the zoom's own band). Failing logs on the unfixed code: `d/fix2/testfirst-test_wet.log(.out)` and `testfirst-test_camera.log(.out)` (first wording, with a `RAIN_CEILING_CLEAR` constant), and `testfirst2-*.log(.out)` (the final tests run in `d/fix2/d2-tree`): 2 and 3 failures. Affected suites on the fix: `live-test_wet` 328 and `live-test_camera` 247, 0 failures. 95: - **Full suite** (`d/fix2/suite.sh`, `VERIFY_RES=2880x1800 tools/suite_all.sh d/fix2/suite all`; 10:11-10:19; AC, Low Power Mode 0, `d/fix2/suite-power.txt`): **42 suites, 22,686 checks, 0 failures** (`test_wet` 328, `test_camera` 247, `test_water` 410, `test_warmup` 37). `--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). 96: - **Recordings** (`d/fix2/rec.sh` and `rec-z.sh`, one sitting 09:49-10:11, AC, Low Power Mode 0, load 2.9-4.4, `d/fix2/rec/power.txt`; the fix1 recipe and scanner `d/fix1/dropscan-bands.js`; tables `d/fix2/bands.tsv` and `d/fix2/drops.tsv` from `d/fix2/table.js`). Stock views lane-night, quarter-night and harbour-night: before-1 (pre-tree, the P8-I close), d2 (the unfixed P8d), after, before-2 (pre-tree). Field fixture: after at sizes 38, 21.05, 16, 10 and 95; d2 at 38, 21.05, 16, 10 and 95; pre-tree at 95. Frame 150 of every clip is kept (`rec//f0150.png`); looked at: lane-night, quarter-night and harbour-night after (drops in front of the upper storeys, the tower tops and the sky). 97: - **Acceptance, measured (after against before-1 / before-2, the unfixed code in brackets):** 98: - 1 art px wide and 6-8 long on the field: stable rows exactly 1 px 99.33% / 99.41% / 99.04% at Harbour / quarter / lane zoom (P8d entry 99.2 / 99.1 / 99.1; unfixed 99.26 / 99.29 / 99.23); 6-8 long 98.7% / 98.6% / 98.8%, median 7 (entry 98.6 / 98.9 / 98.8). No regression. 99: - Field density across zooms: 38 / 21.05 / 16 max/min **1.177** (unfixed 1.165 in this sitting; 1.115 in the P8d entry's sitting); with size 10 (204.5) still 1.177. Size 95 gives 334.7 a frame over the whole frame (unfixed 329.2; the fix's rule leaves a field's volume identical, tested), so the whole-frame 10-95 ratio is 1.66. Corrected by review fix 3: that figure compares different backgrounds, not densities. At 95 the 64 m meadow fills only the middle of the frame (`fix2/rec/field-z95-after/f0150.png`), and the cells over water and sky count 31.1-43.6 a frame against 24.9 for the centre cell over the meadow (unfixed 26.3; `grid_3x3_per_frame` in `fix2/rec/field-*/drops.json`). Like for like over the meadow (95: the centre cell; 38: the five non-corner cells; 21.05, 16, 10: all nine), 24.9 / 24.2 / 23.2 / 22.4 / 22.7 a frame, max/min **1.11** (unfixed 1.19), inside 1.25. 100: - Stock top band ÷ bottom band (target 0.8): lane-night 0.88 / 1.04 → **0.93** (0.43), quarter-night 0.77 / 0.76 → **0.88** (0.56), harbour-night 0.80 / 0.84 → 0.74 in this one clip (0.66). Corrected by review fix 3: one 10 s clip is not the figure; the ratio moves about ±0.1 between clips, and nothing forces it. The five after clips before fix 3's sitting (0.74, 0.91, 0.80, 0.80, 0.86) average 0.82. Over the eight after clips of the same code (this sitting 0.74, `trial-R2` 0.91, the reviewer's re-check 0.80 / 0.80 / 0.86, fix 3's sitting 0.87 / 0.84 / 0.85) the mean is **0.83**, which meets 0.8, against 0.84 for seven before clips (0.78-0.90; `d/fix3/bands.tsv`). Its top band is the sky over the painted skyline, the thinnest band in every leg before and after (0.68-0.69 of the mean before, 0.75 after); its bottom band, water just in front of the camera, was thin before (0.81-0.87 of the mean) because the old box did not reach there, and is now 1.02. The fix does change Harbour's sky rows (the earlier sentence here said it did not; corrected by review fix 3): the band over the target goes from 20 to 45.3 m and the spawn top from 27 to 61 m, and the top band falls from 16.8 drops a frame (unfixed) to 13.3 here (fix 3's sitting: 16.7 → 14.2-15.0; before 6.7-7.5), which fits the same 909 drops spread over a taller slab (not isolated). `trial-R3` (reach 3) gave Harbour 0.95 but quarter 0.71 and far fewer drops. 101: - No bare strips: every band of all three views has more drops a frame than both before legs (above); grid min/mean lane 0.864, quarter 0.820, Harbour 0.755 (Harbour just under its before legs 0.766 / 0.775). 102: - Drops a frame, after (before legs; unfixed): lane-night 115.9 (49.9, 48.3; 110.8), quarter-night 183.6 (100.9, 102.1; 236.4), harbour-night 177.0 (101.4, 102.8; 212.7). The amount stays 909. 103: - **Bench** (`d/fix2/bench.sh`, the P8d recipe: pre-tree (P8-I close) against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, rain 0.45, `--motion-live`; 10:19-10:26; every run on try 1; AC, Low Power Mode 0, load 2.3-3.4, `d/fix2/bench/power.txt`; `d/fix2/bench/summary.txt`). Raw medians pre → live (paired median, rounds live slower): 01:00 dense p50 13.902 → 13.905 (+0.006, 2/3), p95 15.686 → 15.725 (+0.021, 2/3); Harbour p50 6.619 → 6.653 (+0.037, 3/3), p95 7.771 → 8.298 (−0.012, 1/3; pre rounds spread 7.19-8.46). 13:00 dense p50 13.612 → 13.603 (+0.023, 2/3), p95 15.365 → 15.370 (+0.044, 2/3); Harbour p50 6.271 → 6.284 (+0.013, 2/3), p95 7.064 → 7.144 (+0.119, 3/3; pre rounds spread 7.02-7.63). Draw calls identical (2583, 907). Within noise, with one lean: Harbour p50 at 01:00 is 0.6% slower in all three rounds (the P8d bench had +0.023 ms, 3/3, at 13:00); the particle count is unchanged, so more drops land inside the frame, a few pixels each. 104: - **Not done in fix 2:** stills and journey counts were not re-captured (the fix moves only the particle box; draw calls are identical in the bench); Windows not run. (Fix 2 also listed the Harbour band ratio and the size-95 field count here; review fix 3 shows both meet their targets when measured like for like.) 105: - **Review fix 3 (6 October 2026, Opus 5.5 on this Mac; evidence `d/fix3/`).** The orchestrator ruled two documentation findings must-fix: (1) the size-95 field figure 1.66 was put down to P8d's density, but it comes from the fixture (at 95 the meadow fills only the middle of the frame, and drops register more over water and sky), so the entry must give the like-for-like figure; (2) the Harbour band ratio 0.74 "misses 0.8" came from one 10 s clip, and "the fix does not change the sky rows at Harbour" was wrong. Documentation only, no code change. At the start the live tree differed from `build/backup-feel/26-P8-d2/` only in fix 2's files (`scripts/world_view.gd`, `tests/test_wet.gd`, `tests/test_camera.gd`) and from `26-P8-d` only in P8d's (those plus `shaders/rain.gdshader`): nothing unexpected. One note: `world_view.gd` was last written at 10:20, after fix 2's recordings (09:49-10:11) and suite (10:11-10:19); no copy of the 10:11 state exists to diff, so fix 3's sitting and suite, both on the current file, are the check. 106: - **Changed:** the "Density constant" and owner lines above, fix 2's field-density and Harbour lines (the old figures kept, marked "corrected by review fix 3"), the not-done lists, and `d/sheet/sheet.md` (backups `d/fix3/visual-handoff.md.pre-fix3`, `d/fix3/sheet.md.pre-fix3`). 107: - **Re-measured** (`d/fix3/rec.sh h` then `f`, one sitting 10:42-10:53; the fix1/fix2 recipe, `tools/art_record.sh --keep-frames`, and scanner `d/fix1/dropscan-bands.js`; AC, 100%, Low Power Mode 0, load 2.5-3.4, `d/fix3/rec/power.txt`; tables `d/fix3/bands.tsv` and `d/fix3/field.tsv` from `d/fix3/table.js`, which also reads fix 2's sitting, the reviewer's re-check `d/review/fix2-recheck/rec/` (10:30-10:39, AC, Low Power Mode 0) and `fix2/trial-R2`): 108: - Harbour top band ÷ bottom band, pre-tree and live alternating: before 0.78 / 0.81 / 0.86, after 0.87 / 0.84 / 0.85, unfixed (d2) 0.63. Over all sittings: after 8 clips mean **0.834** (0.74-0.91), before 7 clips 0.838 (0.78-0.90), d2 2 clips 0.65. Target 0.8: met on the mean, and Harbour is as even as before. Top band a frame: before 6.7-7.5, d2 16.7-16.8, after 13.3-15.6; drops a frame after 172-181 (before 100-103). 109: - Field, cells over the meadow: fix 2's sitting 24.9 (95, centre) / 24.2 (38) / 23.2 (21.05) / 22.4 (16) / 22.7 (10), max/min **1.113** (d2 1.190); fix 3's sitting 23.8 (95) / 23.3 (38) / 22.9 (16), max/min **1.038** (d2 at 95: 22.3). The cells over water and sky at 95: 29.7-44.7. Whole-frame counts at 95 repeat (334.0 after, 331.2 d2). Target 1.25: met like for like. 110: - Looked at: `d/fix3/rec/harbour-night-after-1/f0150.png` (drops over the sky above the skyline, the towers and the water), `d/fix3/rec/field-z95-after/f0150.png`, `d/fix2/rec/field-z95-after/f0150.png` (the meadow a diamond in the middle third, water and sky round it), `d/fix2/rec/field-h-after`, `field-q-after`, `field-l-after`, `field-z10-after` `f0150.png` (meadow over the whole frame at 21.05, 16 and 10, apart from two small sky corners at 21.05; at 38 the four corner cells show water and sky). 111: - **Full suite** (`d/fix3/suite.sh` a, b, v; `VERIFY_RES=2880x1800 tools/suite_all.sh d/fix3/suite`; 10:53-11:01; AC, Low Power Mode 0, `d/fix3/suite-power.txt`): **42 suites, 22,686 checks, 0 failures** (as fix 2; `test_wet` 328, `test_camera` 247). `--verify` 2880x1800 exit 0 on try 1 in 33 s, `PIXELPUNK_INTEGRATION []`; the `test_creature_reactions` rerun passed. 1440x900 not run (the inherited stall). 112: - **Changed expectations:** none. Journey count deltas: none (no code change). No bench: no code change since fix 2's bench. 113: - **Not done:** 114: - Even density down the stock frames: done by review fix 2; Harbour's 0.83 is a mean over clips, single clips range 0.74-0.91 (review fix 3). 115: - Alpha by zoom band (the density holds without it, above). 116: - The rain is not pre-warmed: `ShaderWarmup.entries_of` walks MeshInstance3D and MultiMeshInstance3D, not CPUParticles3D, so neither the old nor the new drop material is warmed (its header lists rain). Not in this part's files. 117: - Perspective and photo-mode lenses: covered by the pixel snap in the shader, but no capture was taken; the free camera keeps the old box. 118: - The report's "about 60% of drops never rasterise" was not re-derived; measured visible old streaks are 98 a frame on the field at Harbour zoom against 233 now. 119: - Recordings at 1600x1000 (the Movie Maker limit noted in P8a). 120: - **Platforms:** Mac run; Windows not run. 122: **Beautify P8-I (parts a, b, c of P8) stage close (6 October 2026):** the stage-end overhead of sub-stage P8-I "Motion" (step `26-P8-I` in the plan's overhead table), run by Opus 5.5 on this Mac on the final P8-I code: `main` 1821f51, clean apart from the orchestrator's `artifacts/beautify/beautify-stage.js` (`close/head.txt`). The parts are P8a, P8b and P8c; P8a2 was not run (D17). No game code changed. Evidence is in `artifacts/beautify/26-P8-I/close/`. 123: - **Before side:** the pre-stage scratch tree `close/pre-tree`, from `build/backup-feel/26-P8/`. It equals `main` ee8379a (the P7f cost trim) on scripts, shaders, tests, data, scenes, tools and `project.godot`. Only the three `.ps1` tools differ (in line endings), plus a `.DS_Store` (`close/diff-pretree-vs-ee8379a.txt`). P8 adds no `class_name`, so the tree's class cache was not refreshed. 124: - **Power:** every run was on AC, 100% charged, with Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). Load average was 1.7-4.1. 125: - **Full suite** (`close/suite.sh` a, b, v; 05:33-05:41): **42 suites, 22,555 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). 126: - **Journey** (`artifacts/feel-journey/26-P8-I/`; six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): 127: - Counts against 25-P7 (`26-P8-I/counts-vs-25-P7.txt`) are identical in all six views, draw calls included. The only change is the one P8c declared: `pool_energy_ratio` 1.0000000065 → 1 in the day views. The stage capture re-ran the six views with identical counts (`close/metrics/counts.txt`), so no single-view re-run was needed. 128: - Frame >8 luma against 25-P7 is 0.85-7.35% (`26-P8-I/frame-vs-25-P7.txt`). By eye (`close/look/j-1.png`, `j-2.png`, looked at): no visible difference. The night life does not show in stills because every automated run holds the light clock. 129: - `notes.json` has a `26-P8-I` entry (`close/journey-note.js`, backup `notes.json.bak-before-26-P8-I`). `index.html` and `journey-standalone.html` (108.42 MB) were rebuilt. 130: - **Art capture and protected looks** (`close/caps.sh`; one sitting 05:43-06:20, every view on try 1). There are three groups: 131: - g1: the P7 close's 58 views; 132: - g2: its 30 P7 views and close-ups; 133: - g3: 8 P8 views: `h1725`, `h1730`, `q1823` and `q1908` (the stopped-clock band hours); `quarter-fog`, `bar-fog` and `signs-fog` (the P8c fixtures at fog 0.5); `quarter-0545-fog`. 135: That is 96 views per set, captured in the order pre-1, after, pre-2, after-2. Metrics are in `close/metrics/` (`close/metrics.sh`; regions `close/regions.json` are the P7 close's): `ab-*`, the noise pairs, `frame-*`, `frame-table.txt` (`close/frame-table.js`), `counts.txt`, `protected.md` (`close/prot.js`), `consistent.txt` (`close/consistent.js`, maps in `close/look/consistent/`) and `elapsed-ms.txt`, plus `ab-p7close-*` and `frame-p7close-*` against `25-P7/close/after` (the pre-trim P7 code). Protected sheets: `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P7-close.png`, `noise-pre.png` and `noise-after.png`. Page 1 was looked at. 136: - **Counts:** identical in every fixed view in both pairs, apart from `pool_energy_ratio` → 1. The orbit views vary inside one tree, as at P6 and P7. `qn-orbit` has `light_admits` 27 → 26 (evicts and reclaims 11 → 10) in both pairs while each tree's own pair agrees. That is consistent but small and not explained. It is probably P8c's pool bookkeeping reading the eased energy in float64. 137: - **The P7 close capture against today's pre-1:** counts are identical in every fixed view, so the P7f trim changed no count. 138: - **Systematic frame changes:** in 39 of 96 views, both before→after legs are above both noise legs (`frame-table.txt`). The pixels that change in all four cross pairs and in neither noise pair cover 0.5-2.5% of a frame (`consistent.txt`). They are rain rings on open water, water reflections, the far neon columns and cloud dither, all `TIME`-driven. One example is chemistry-day by day: water rings only (`look/consistent/chemistry-day.png`). The likely cause is that the live captures happen later after start: `elapsed_ms` is +47 ms median, and later in both runs in 86 of 96 views (`elapsed-ms.txt`), so `TIME` sits at another phase. This is not proven. 139: - **Real still changes:** P8b's stopped-clock fix (h1725 and h1730 Fish Market sign region mean luma 70.2 → 68.4 and 70.5 → 68.7, the off step) and P5b's glint dashes, now held at light clock 0 (D17). 140: - **Canal mirror columns:** within noise. p50 43.9 → 43.8 and 52.1 → 52.3. Changed >8 is 25.2% / 6.5% against noise of 19.2% / 4.5% (quarter-night), and 10.4% / 20.5% against 2.1% / 16.6% (quarter-21:00). 141: - **HDR neon cores:** within noise. Fish sign p50 66.3 = 66.3; bar sign 70.6 = 70.6 with 0% changed; lane sign 140.5 → 140.2 (0.9% / 0.3% against noise of 0.8% / 0.2%); the quarter tower window 96.0 = 96.0. 142: - **05:45 and 20:45 skies:** within noise (p50 equal to ±0.3). 143: - **Glade/Canal separation:** Glade's night zenith `#001d45`, horizon, water, lawn and bank are identical. The Harbour night zenith changes 22.7% in pair 1 and 0.24% in pair 2; the pre noise pair is 22.7%, so pre-1 is the odd capture (cloud dither). 144: - **Approved tree style:** unchanged. Grove crowns 55.1 → 55.0 (55.1 = 55.1 in pair 2), conifers 55.6 = 55.6, quarter undersides 27.7 = 27.7, golden grove 48.5 = 48.5. 145: - **Whole-stage A/B bench** (`close/bench.sh`; `bench-t1.out`, `-t13.out`, `-t19.5.out`; `bench/summary.txt` from `close/bench-summary.js`). Setup: pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync` before `--path`, 1920x1080, rain 0.45, `--motion-live` on both sides (so the live light clock runs; the pre-tree ignores the flag). 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): 147: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 148: |---|---|---|---|---|---|---| 149: | Dense | 14.23 → 14.26 (+0.03, 2/3) | 16.13 → 16.11 (−0.00, 1/3) | 13.93 → 13.92 (+0.02, 2/3) | 15.79 → 15.87 (+0.04, 3/3) | 15.95 → 15.97 (+0.00, 2/3) | 17.73 → 17.75 (+0.01, 2/3) | 150: | Harbour | 6.54 → 6.61 (+0.05, 3/3) | 8.20 → 8.19 (+0.35, 2/3) | 6.16 → 6.16 (+0.01, 2/3) | 7.29 → 7.17 (−0.28, 1/3) | 6.38 → 6.41 (+0.03, 2/3) | 8.17 → 7.84 (−0.17, 0/3) | 152: - Draw calls are identical: dense 2583 (19:30: 3433), Harbour 907 (19:30: 822). 153: - The night Harbour p50 +0.05 ms (3/3) matches P8b's +0.021 and P8c's +0.052. The Harbour p95 rounds scatter by 0.7-0.9 ms on both sides, as in P8b and P8c. 154: - **Owner sheet** `artifacts/beautify/26-P8-I/close/sheet/` (`close/frames.sh` pulls frames from the parts' `clip.mp4` recordings; `close/crops.js`, `close/sheet.sh`, regions `close/sheet-regions.json`). Before is on the left and after on the right; colour and greyscale where values matter. 155: - Pages: 156: - `p8-01-overview-stills.png`; 157: - `p8-02-P8b-faulty-tube.png` (pink turret bulb at clip frame 108, before, after, and Reduced); 158: - `p8-03-P8b-breathing-pools.png` (P8b's breathe map, after only); 159: - `p8-04-P8b-stopped-clock.png`; 160: - `p8-05-P8c-sign-spill.png` (sign 141 fixture, low step at clip frame 154, on step against low step, and Reduced); 161: - `p8-06-existing-scenes-changes.png` (consistent-change maps); 162: - `p8-07-existing-scenes.png`. 163: - `sheet.md` has one plain paragraph per part with its numbers, then existing scenes, the bench and the suite, and lists the owner questions: D27, breathe lamp pools only, and the faulty pick on lettered signs. 164: - Pages 2-6 were looked at; 1 and 7 were made but not opened. Page titles are short because the sheet tool does not wrap titles; the explanation is in `sheet.md`. 165: - **Not done / open:** 166: - The owner has not seen the sheet. D27 and P8b's two questions are unanswered (orchestrator rulings after P8b). 167: - The P8 recording sheet of every clip is left for the P8 close: this sub-stage made no new recordings and reused the parts' clips. 168: - From P8c: the reflection streaks of a faulty sign stay steady (`reflections.gdshader` has no night life), and the C3a CPU threshold still rounds half-centimetres in float64. 169: - Still not explained: the `qn-orbit` light-admit change and the `TIME`-phase shift in the live captures. 170: - No 1440x900 `--verify`. 171: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 172: - Scratch trees: `close/pre-tree` was moved to the macOS Trash with `mv` once its numbers were recorded, as `beautify-26-P8-I-close-pre-tree-064411`. The part trees `artifacts/beautify/26-P8/a/`, `b/` and `c/pre-tree` lie outside `26-P8-I/` and were left in place. 173: - **Corrected after the handoff check (6 October 2026):** the P8a, P8b and P8c entries below and `sheet.md` were reworded to match their files: stills against noise legs (P8a harbour-day, P8b lane-night/quarter-2045/quarter-day, P8c several views and after-3); the P8c sign 141 pixel counts, the sign 166 block table and the lamplighter per-pixel claim (console-only or no saved script); the sign 141 rect range, fix-pass glass and terrace values; P8c bench, acceptance, after-3 and `sign-fog-final` as pre-review code (not re-benched); P8b breathe levels and level counts; the P8a suite window. No numbers were re-measured. 174: - Platforms: Mac run; Windows not run (P6e, the Windows kit, is still deferred by the owner). 176: **Beautify P8c Night life on OmniLights, sign spill and fog (6 October 2026):** Opus 5.5 on this Mac (plan section "P8c"; report polish backlog "night lights breathe" (lights); step `26-P8`, last part of sub-stage P8-I). The real lights now take the night life P8b gave the glass and pools. Every OmniLight of the pool breathes with its pools, and the light of a faulty cyan or pink tube steps with the tube, so its spill on walls and wet paving and the fog glow it feeds blink with the glass. The factor goes on after the `move_toward` easing, so the steps are crisp and the dusk fade keeps its timing. Defaults D27 and D17 as in P8b (no owner answers yet); P8b's open questions (breathe pools only, the faulty pick, the BAR sign) are untouched. 177: - **Code:** 178: - `Dusk.life_seed(seed, origin)` is the seed the shaders form (`fract(UV2.x + hash of the instance origin)`). Since the review fix it hashes the origin with the shaders' float32 rounding (`Dusk.shader_cell`, `Dusk.shader_hash01`); off a half-centimetre that is also the s of the C3a threshold. `Dusk.light_life(seed, t, tube, amounts, motion)` is the breathe, times the stutter for a tube's light. It is exactly 1 with no amounts, by day and under Off. 179: - `WorldView._update_lighting`: each pool light keeps its eased energy (`entry.eased`, the same `move_toward` at `delta*3`). The node gets `eased * light_life` on the running light clock with `life_raw` (zero under a held clock, so every automated still is unchanged). `feed_fog_lamps` still sends the node energy, so the fog glow follows. 180: - The pool's bookkeeping reads `eased_energy()`: eviction, reclaim, lit-lost and the delivered share. A dip never evicts a light early. 181: - `light_seed` and `light_tube` are worked out when a light takes a pool slot (or is kept by `_retain_lights`), never in the light pass over every source. A first form that stored them on every source in the pass cost +2.4 ms `light_sync_ms` on dense (+3.8%, 3/3; `bench-v1.out`), and a one-hash form still +0.85 ms (`bench-v2.out`); both were reverted. 182: - A tube's light is a `neon` sign or a `bulb` whose colour is `Architecture.CYAN` or `PINK` (the glass in the `cyan`/`pink` flicker materials). A recessed window's cyan or pink pendant stutters on the GPU (P8b), but its window light does not. 183: - **Files:** `scripts/dusk.gd`, `scripts/world_view.gd`, `tests/test_dusk.gd`. Backups: `artifacts/beautify/26-P8/c/backup/`. Snapshot `build/backup-feel/26-P8-c/` was missing at part start and was made with `tools/feel_snapshot.sh save 26-P8-c`. It equals `26-P8-b2` and the live tree (`main` 56f1901) on scripts, shaders, tests, data, scenes and tools; nothing unexpected. Pre-part tree: `artifacts/beautify/26-P8/c/pre-tree` (from `26-P8-b2`). 184: - **Tests:** 185: - `test_dusk` 409 → 442 (+33; `light_life_checks`, `light_life_view_checks`). The checks cover: 186: - the light's seed equals the threshold's s and the glass's seed (Harbour: every tube light's seed is drawn by its cyan or pink glass); 187: - the tube rule, and pool lights carrying their source's seed and flag; 188: - energy = eased × `light_life` after the easing; 189: - a forced faulty light steps to `FLICKER_LOW` and back in one frame, and the fog's fed energy follows both ways; 190: - `lamp_power` is `Dusk.level` whatever the clock; 191: - eviction and reclaim read the eased energy, and the light pass is unchanged; 192: - held clock, Off and day are exactly the eased energy; Reduced gives no stutter and half the breathe; Glade gives half the breathe and no stutter; 193: - through 18:36-19:45 the eased course is identical with the clock running and held. 194: - Test-first log: `c/c-testfirst-test_dusk.log(.out)`, 5 gate failures (the `light_seed` gate came later). 195: - Affected suites (`c/suites/`): `test_lighting` 189 and `test_warmup` 37, 0 failures. 196: - **Full suite** (`c/suite/suite_all.out`; a, b, v; 04:48-04:56; AC, Low Power Mode 0, `c/suite-power.txt`): **42 suites, 22,549 checks, 0 failures**. `--verify` at 2880x1800 exited 0 on try 1 in 32 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). 197: - **Changed expectations:** none (only new checks). `performance.json` `pool_energy_ratio` reads exactly 1 at rest instead of 1.0000000065 (by day): the delivered share now sums the eased energy in float64, not the node's float32. **Expected journey count deltas:** none. Integer counts and draw calls are identical before→after and in both noise pairs (`c/metrics/counts-*.txt`). 198: - **Stills** (`c/caps.sh`; before-1 / after / before-2 / after-2, one sitting; journey set + lane-night + `quarter-fog` (quarter, fog 0.5) + `bar-fog` (first fixture, fog 0.5); AC, `cap-power.txt`; plus `after-3` on the pre-review code (before the float32 seed fix), `c/fin.sh`, 04:45-04:48): 199: - Frame >8 luma before→after against the noise legs (`metrics/frame-*.txt`): several views sit slightly or clearly above both legs. harbour-day 5.271% / 5.277% against 0.702% (before) and 4.558% (after); harbour-night 4.119% (before-2→after) against at most 3.921%; lane-night 2.294% against 2.102%; quarter-day 3.712% against 3.668%; styles-day 3.148% against 2.857%. The others are inside, e.g. quarter-night 7.0% / 7.4% against 10.9% (before) and 6.4% (after). The changed pixels are clock-driven (rain rings, water reflections; below). The P8-I close's note that the after captures start about 47 ms later (another `TIME` phase) is the likely cause; it is not proven. 200: - Pixels changed in all four cross pairs and in neither noise pair (`consistent.txt`; `look/consistent-harbour-day.png` and `-quarter-night.png` looked at): 0.009-0.115%, and 0.53% on harbour-day. They show water glitter, rain, smoke and the far neon columns, all `TIME`-driven. Lights hold under a held clock. 201: - `after-3` (pre-review code) has identical counts. Against after, harbour-day 5.187% and harbour-night 7.681% are above both noise legs and quarter-day 3.677% is just above (3.668%); against after-2, harbour-night 4.444% is above (3.921%). The others are inside (`frame-after-after3.txt`, `frame-after2-after3.txt`). 202: - **Not covered (review finding, queued):** `shaders/reflections.gdshader` has no night life, so the paving and canal reflection streaks of a faulty sign stay steady while its tube, spill and fog glow blink. Not in the plan's P8c file list. Also, the lamplighter filmstrip below ran under `--motion=off` (night life exactly 1 by construction), so the unchanged dusk timing with the night life running rests on the headless test only; there is no Full before clip. 203: - **Recordings** (`c/rec.sh`, `tools/art_record.sh`, 1440x900, 30 fps, light clock running; AC, `rec/power.txt`; frames measured, then moved to the Trash). Looked at: the `bar-fog-after` and `lamplighter-full-after` contact sheets and the `quarter-fog-after` diff sheet. 204: - **The demonstration fixture.** No stock scene has a faulty sign light: the quarter's two faulty glass seeds are an antenna bulb (no light) and a window pendant. `c/fixture/quarter-faulty-signs.json` is the quarter with neon 141 moved 6 cm along its wall and neon 166 13 cm, each 3 mm off the wall, so their seeds are faulty (`probe_fixture2.gd`, `probe_verify.gd`). Both lights are in the 16-light pool at the recording camera. 205: - **Sign 141 spill** (`sign-fog-before/after`, 20 s, fog 0.5; `stepdiff.js`, `blink.js`, `local.js`; `look/stepdiff-141-crop.png`, `look/sign141-steps.png`, looked at): the pink light on the pergola terrace under the sign steps with the tube only in the after clip. 206: - `blink.js` printed its pixel counts to the console only (1,440 px falling in both bursts, mean −14.4 luma against −0.28 before; 0 px on two burst-free windows). No file holds them and the frames were trashed (only the lossy `clip.mp4` remains), so they cannot be re-derived as measured. `look/blink-141.tsv` keeps only the per-frame mean of that pixel set: 15.95-42.35 after, 39.27-42.35 before. 207: - The rect's mean goes 35 → 26.70-27.96 on the twin's low steps (7.976 and 17.883 s; the first burst reads low, low, low then on, the second low/on/low/low) and back on the on step (`metrics/rec-141-spill.tsv`). 208: - The `sign-fog-final` clip (`c/fin.sh`, pre-review code) repeats it (`metrics/rec-final-141-spill.tsv`). 209: - Under Reduced (`sign-fog-reduced`) the rect's range across the burst is 1.83 luma against 9.06 at Full (`metrics/rec-141-reduced.tsv`). 210: - **Sign 166:** its glass stutters, but its light shows no measurable spill. `stepdiff.js` printed its block table to the console only; no text output was saved, only `look/stepdiff-166.png`, and the frames were trashed, so the block counts cannot be re-derived. 211: - **Lamplighter** (quarter from 16:48 on a 120 s day, fog 0.5, `--motion=off`, pre-tree vs live, all 300 frames; `lampcmp.js` → `metrics/lamplighter-off-cross.tsv`; `look/lamplighter-off-diffmap.png`, looked at, shows rain streaks): a median 1.37% of pixels per frame differ by more than 8 (1.11-1.62% over the 300 frames). That per-frame share is all the saved script computes. The earlier per-pixel statement (no pixel differs in more than half the frames, so the fade's timing is unchanged) and the diff map are not produced by any saved script, so they are not backed by a file. `lamplighter-full-after` shows the evening lighting up at Full. 212: - **quarter-fog before/after** (10 s): motion medians 9.48/3.13 → 11.55/3.02 (changed / >8 %). The diff sheet also shows the inherited ring-edged sky change at seconds 4 and 9. 213: - **Acceptance, measured:** 214: - "Spill, wet paving and fog glow blink with the tube": spill met on sign 141's terrace (rect mean 35 → 26.70-27.96 on the low steps, `metrics/rec-141-spill.tsv`; the −14.4 luma, 0-false-pixel count was console-only, see above). Wet paving: that terrace is in the rain; open paving was not separated. The fog glow follows the node energy (tests and `probe_run.gd`: fed 1.8 → 0.216 at 7.843 s); no glow-only pixel set was separated in the clip. 215: - "`light_sync_ms` within noise": met on the pre-review code (before the float32 seed fix; not re-benched; `bench.out`, 04:40-04:45, rounds 1-3): dense 64.262 → 64.200, Harbour 4.831 → 4.856 (inside the pre spread 4.815-4.861), quarter with `--creatures` 6.372 → 6.354. 216: - **Bench** (`c/bench.sh`, `bench/summary.txt`; pre-tree against live, interleaved, round 0 discarded, `--motion-live`, 01:00, 1920x1080, `--disable-vsync`, rain 0.45; dense, Harbour, and the quarter with `--creatures` and fog 0.5; AC, Low Power Mode 0, `bench/power.txt`; 04:40-04:45). Pre-review code (before the float32 seed fix; not re-benched), raw medians: 217: - dense p50 14.265 → 14.263, p95 16.173 → 16.126; 218: - Harbour p50 6.572 → 6.624 (+0.052, 3/3); 219: - quarter p50 7.240 → 7.280 (+0.040, 3/3), p95 8.939 → 8.847; 220: - the light factor costs about 0.03 ms a frame for 16 lights (`probe_lights.gd`); 221: - Harbour p95 7.152 → 8.103 here, but −1.30 and −0.72 in the two earlier sittings (`bench-v1.out`, `bench-v2.out`): noise, as in P8b; 222: - draw calls identical (2583, 907, 1043). 223: - **Review and fix pass (6 October 2026, about 05:00-05:24, AC, Low Power Mode 0; `c/fix/`).** One must-fix finding. `light_seed` hashed the instance origin with the CPU's float64 `floor(x*100+0.5)`, but the glass and pool shaders round the same float32 origin in float32. On an exact half-centimetre the two disagree (z 7.055 → 705 on the CPU, 706 on the GPU; 1,648 of 4,000 coordinates k/100+0.005; reviewer's `c/review/halfcm.js`). Such a light took a different life from its own glass and pools: a faulty pick that disagrees, and a breathe out of phase with its pool. The test's glass parity used the CPU hash on both sides, so it could not catch it. 224: - **Fixed:** `Dusk.life_seed` now hashes with `Dusk.shader_hash01`: the coordinate, ×100 and +0.5 each rounded to float32 through a `PackedFloat32Array`, then `hash32` as before. `Dusk.hash01`, the C3a thresholds and P8b's glass are unchanged. 225: - **Tests** (`test_dusk` 442 → 448, +6, all 0 failures; test-first log `c/fix/testfirst-test_dusk.log(.out)`, 3 failures on the reviewed code; after the fix `c/fix/test_dusk.log.out`, `test_lighting` 189 and `test_warmup` 37, 0 failures). The test now emulates the shader's float32 rounding itself (`shader_cell`, `shader_hash01`) for the glass side. Its cells are pinned to the reviewer's emulation (4.455 → 446, 7.055 → 706, 8.655 → 866, 13.355 → 1336, 0.005 → 1). New checks: a light's seed on each of the 1,648 disagreeing half-centimetres (in x, y and z) uses the shader rounding; the BAR origin gets the shader hash while `hash01` is unchanged; a Harbour neon sign moved onto x = k/100+0.005 keeps every tube light's seed among its glass's seeds. 226: - **Survey** (`c/fix/probe_halfcm.gd`, `.log.out`): in the stock quarter, 27 of 216 light sources stand where the roundings differ, 7 of them tube lights (neon 99, 124 BAR, 125, 130, 139, 141, 166). Before the fix all 27 breathed out of phase with their pools. No stock tube light's faulty pick flipped, so no stock sign blinked apart from its glass. Harbour has none. 227: - **Recording** (`c/fix/rec.sh`, 20 s, night, fog 0.5, 1440x900; AC, `c/fix/rec/power.txt`). The fixture is `c/fix/quarter-halfcm-141.json`: sign 141 nudged −2 cm along x, kept on its stock z 8.655. There the shader seed is faulty (bursts 7.24 and 16.805 s) and the old CPU seed is steady. Before is the reviewed code (`c/fix/prefix-tree`), after is the fix. Frames were measured, then moved to the Trash. 228: - Terrace spill rect (370,255)-(475,355), `c/fix/rec-halfcm-terrace.tsv`: before 44.6-45.9 through both bursts, no step. After 45.6 → 39.0 at 7.24 s and back at 7.840 (44.90), and 45.5 → 39.0 at 16.805 (still down at 16.872; the fixer wrote 16.87). 229: - `stepdiff.js` (`c/fix/stepdiff-halfcm.txt`, `look-stepdiff-halfcm.png`, looked at): 18 16x16 blocks drop more than 5 luma extra in the after clip, against 0 in the null windows and 0 the other way round (`stepdiff-halfcm-rev.txt`). The blocks are all on the pergola terrace. 230: - Crop sheet `c/fix/look-halfcm-steps.png` (looked at): the pink terrace glow is gone on the low steps only in the after row. 231: - Sign 141's lettered glass shows no separable dip in either clip: `rec-halfcm-glass.tsv` (`glass.js`) has a single rect, (624,680)-(648,704), whose mean runs 68.71-87.63 across both burst windows, drifting alike in both clips (before and after within 3.9 luma at every sample). The range is wide; the file does not show the glyph saturating. The glass side rests on the shader-rounding emulation and the reviewer's BAR observation, not on these pixels. 232: - **Cost** (`c/fix/probe_cost.gd`): a seed now costs 1.22 µs against 0.48 µs. It runs only when a light takes a pool slot, so at most 16 per take, under 12 µs. `light_sync_ms` was not re-benched. Stills were not re-captured: under a held clock the factor is exactly 1 (tested), so the seed cannot reach a still. 233: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh c/fix/suite`, a/b/v, 05:16-05:24, AC, `c/fix/suite-power.txt`): **42 suites, 22,555 checks, 0 failures**. `--verify` 2880x1800 exited 0 on try 1 in 33 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures. 1440x900 was not run (the inherited stall). 234: - **Still open (not this part):** `Dusk.hash01`, which sets the C3a CPU threshold (`source.dusk`, the light's dusk turn), still rounds in float64. A half-centimetre light can therefore turn on at a different moment from its glass. Fixing it changes C3a thresholds, so it belongs to its own part. The GPU rounding is modelled, not read back from the GPU. Fused multiply-add would give the same cells for the quarter's origins. 235: - **Review:** filled by the orchestrator. 236: - **Not done:** 237: - A visible spill demonstration in a stock scene: no stock sign light is faulty. 238: - The second fixture try (neon 161 on the turret): its light lights nothing visible there (`look/lights-pool-crop.png`: the turret takes no pool light). 239: - Fog glow and wet paving not separated in the pixels. 240: - The probes `probe_lights.gd`, `probe_match.gd`, `probe_parity.gd`, `probe_fixture*.gd` and `probe_verify.gd` read `source.life` from the first code form and no longer run as written. 241: - The C3a CPU threshold's float64 rounding on half-centimetres (see the fix pass). 242: - **Platforms:** Mac run; Windows not run. 244: **Beautify P8b Night life in the shaders: faulty tubes, breathing pools, stopped-clock fix (6 October 2026):** Opus 5.5 on this Mac (plan section "P8b"; report polish backlog "night lights breathe" (shaders) and the known fault "a clock stopped mid-band holds a dim sign"; step `26-P8`, sub-stage P8-I). At night the light pools breathe and about 1 cyan or pink tube in 10 is a faulty tube that stutters. A sign whose hour has stopped inside its dusk band now settles fully lit or off instead of holding dim. Plan default D27 applies (the owner answers after the part): faulty tubes are on by default, at most 3 flashes a second, and off under Reduced and Off. D17 applies: Reduced and Off gate only the new terms and P5b's glints. 245: - **`Dusk.night_life(seed, t, kind, amount, motion)`** (`scripts/dusk.gd`) is the CPU twin of `dusk_night_life()` (`shaders/dusk.gdshaderinc`). Both hash whole step numbers with C3a's `dusk_hash32`, so the CPU and GPU agree. `t` is P8a's `light_time`, which `WorldView` now wraps at `LIFE_PERIOD` 1380 s. That period holds a whole number of breathe steps (2484), stutter steps (6900), faulty cycles (120) and P5b glint periods (1150 at Full, 575 at Reduced), so the wrap has no seam (P8a review note 1). 246: - **Breathe (pools):** value noise at `BREATHE_HZ` 1.8, quantised to 3 levels: x0.89, x1, x1.11 (`BREATHE_AMP` 0.11). It rides in the pool's `share`, so the dry pool, its wet sheen and its glints breathe together. Lantern glass stays steady. 247: - **Faulty tubes (`cyan`/`pink`, the `flicker` materials):** a seed hash under `FAULTY_SHARE` 0.10 makes a tube faulty. Bursts come once per 11.5 s cycle, starting somewhere in the cycle's first 4.5 s (so 7-16 s apart). Each burst is 4 steps of 0.2 s; the first step is always `FLICKER_LOW` and the rest are hashed. A 1 s window can hold at most 5 changes (2.5 flashes). 248: - **Amounts (`Dusk.life_amounts`, set by `WorldView.apply_night_life` as `life_breathe` on `pool` and `life_faulty` on `cyan`/`pink`):** 249: - The breathe follows the lamps (smoothstep 0.15-0.5): nothing by day, full at golden hour and night. 250: - The stutter needs `smoothstep(0.8, 1, progress) x windows`: 0 at golden hour, 0.0095 at 20:45 and 05:45, 1 at night. 251: - Glade gets the breathe at 0.5 and no faulty tubes. 252: - Under a held light clock (every automated still), both amounts are 0, so stills rest at x1. 253: - `motion_scale`: Reduced halves the breathe and stops the tubes; Off stills both. 254: - **Stopped-clock fix (`Dusk.settle`, `dusk_settle`):** `WorldView._settle_dusk` counts frame seconds since the evening's progress last changed. After `SETTLE_SECONDS` 0.5 (or one frame under a held clock), `dusk_settled` on `cyan`/`pink` shows the band step's draw as lit (1) or off (the floor), never `FLICKER_LOW`. A moving hour (running clock, slider drag) sputters through `Dusk.flicker`, unchanged. 255: - **P5b glints (D17; P8a's carried item):** they re-roll on `light_time x motion_scale` instead of `TIME`, and the epoch wraps at the period. Reduced re-rolls at half rate; Off and a held clock hold them still. 256: - **Files:** `scripts/dusk.gd`, `scripts/world_view.gd`, `shaders/dusk.gdshaderinc`, `shaders/neon_body.gdshaderinc`, `shaders/light_pool_body.gdshaderinc`, `tests/test_dusk.gd`, `tests/test_wet.gd`. Backups: `artifacts/beautify/26-P8/b/backup/`. Snapshot `build/backup-feel/26-P8-b/` was missing at part start, so it was made with `tools/feel_snapshot.sh save 26-P8-b`. It equals `26-P8-a` and the live tree (`main` bb8cc70) on scripts, shaders, tests, data, scenes and tools; nothing unexpected. Pre-part tree: `artifacts/beautify/26-P8/b/pre-tree`. 257: - **Tests:** 258: - `test_dusk` 352 → 409 (+57; `night_life_checks`, `night_life_shader_checks`, `night_life_view_checks`). They cover: 259: - constant parity (period, breathe, faulty, stutter, salts, `FLICKER_LOW`); 260: - about 1 in 10 faulty over 1,000 seeds; 261: - steady non-faulty tubes; 262: - 7-16 s gaps, at most 3 flashes in any second, exactly 3 breathe levels and the step rate; 263: - periodicity at `LIFE_PERIOD`; 264: - nothing by day, the breathe only at golden hour, Glade half with no tubes, Reduced and Off; 265: - `settle` returns only 1 or 0; 266: - the view's uniforms, rest under a held clock, the wrap, and settling after 0.5 s but not while the hour moves. 267: - `test_wet` 218 → 220 (+2, one changed). 268: - Test-first logs: `b/b-testfirst-test_dusk.log(.out)` (16 gate failures) and `b/b-testfirst-test_wet.log(.out)` (3 failures). 269: - **Full suite** (`b/suite/suite_all.out`; a, b, v; 03:19-03:27; AC, Low Power Mode 0, `suite-power.txt`): **42 suites, 22,516 checks, 0 failures**. `--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). 270: - **Changed expectations:** `test_wet` "P5b: the only TIME terms are behind the wetness gate" → "P8b: the pool has no TIME; the glints' only clock is light_time x motion_scale, behind the wetness gate". `glint_port` gains an `epochs` argument (default 0 = the old behaviour). **Expected journey count deltas:** none. Counts were identical before→after and in both noise pairs, including draw calls (`b/metrics/counts-*.txt`). 271: - **Stills** (`b/caps.sh`, `caps2.sh`; before-1 / after / before-2 / after-2, one sitting, 10 views: the journey set plus quarter-golden, quarter-2045, lane-night and quarter-dry; AC, `cap-power.txt`): 272: - **Pixels changed by more than 8 luma, against the noise legs (`metrics/frame-*.txt`):** 273: - harbour-night: 11.6% / 12.4% before→after, against 5.7% before/before and 12.0% after/after. The water rings and cloud dither run on `TIME` at the capture moment. 274: - Slightly above both noise legs (before-1→after): lane-night 3.595% (before-2→after 3.077%) against 3.046% / 2.338%; quarter-2045 3.862% against 3.53% / 1.596%; quarter-day 3.098% against 2.975% / 1.07%. Their before-2→after legs (except lane-night's) are inside. 275: - Every other view: inside its noise legs. 276: - **Pixels changed in all four cross pairs and in neither noise pair (`consistent.js`, `metrics/consistent.txt`, pictures `b/look/consistent-*.png`, looked at):** 277: - quarter-night 0.69%, lane-night 0.77%, harbour-night 0.30%: P5b's glint dashes, now held at `light_time` 0 instead of `TIME`; 278: - quarter-day 4 px, styles-day 0, quarter-dry 0.03%. 279: - **The fix on the picture (`b/caps4.sh`; harbour at 17:25, where the twin puts the Fish Market sign on a dim band step; `metrics/fishmarket-band.txt`, `look/crop-h1725-fishmarket.png`, looked at):** the letters measure mean luma 196.4 (dim, both pre runs) → 184.1 (off, both live runs). The 17:30 hour gives 197.3 → 184.9. The quarter tries at 18:23 and 19:08 (`caps3.sh`) changed no sign: the BAR sign predicted dim there did not change on the GPU. Not resolved; see "Not done". 280: - **Recordings** (`b/rec.sh`, `tools/art_record.sh`, 1440x900, 30 fps, light clock running; AC, `rec/power.txt`). Contact and diff sheets of `quarter-night-before`, `lane-night-after` and `quarter-dry-after` were looked at. 281: - quarter-night before/after: 20 s each. 282: - lane-night after: 10 s. 283: - quarter-night Reduced: 20 s. 284: - quarter-dry before/after: 10 s each. 285: - Kept frames were measured, then moved to the Trash. 286: - Motion medians (changed / >8 %): quarter-night 10.10/3.47 → 10.61/3.40; quarter-dry 2.90/0.43 → 3.41/0.43; Reduced 10.25/3.39. 287: - The once-per-5-s shifted frame in the contact sheets also appears in the pre-tree's clip (P8a's note): not this part's. 288: - **Acceptance, measured:** 289: - **Faulty tubes:** the quarter has 23 distinct tube seeds, 2 of them faulty (`b/probe_life.gd`, `probe_screen.gd`). Both are bulbs, not lettered signs: a pink turret bulb and one cyan festoon bulb. Harbour has 6 seeds and 0 faulty. 290: - Pink bulb in the recording (`timeline.js`, `metrics/rec-after-faulty.tsv`): dips to luma 131 from 242 at clock 6.458-6.858, 7.058-7.258 and 17.292-17.692 s. The twin gives 6.484/6.884/7.084/7.284/17.320/17.720 (`probe_bursts.gd`): the same within one frame (hash parity on the GPU). 291: - Gap 10.8 s (target 7-16). At most 4 changes in 0.8 s: 2 flashes in any second (target ≤ 3). 292: - The cyan bulb dips at 11.058-11.258 and 11.658-11.858 (twin 11.078/11.678). 293: - Under Reduced, the pink bulb holds at 242.1 in all 600 frames (no stutter). 294: - CPU over 1,000 seeds (`probe_life.log`): 94 faulty, gaps 7.15-15.89 s, worst 2.0 flashes per second. 295: - **Pools breathe** (`breathe.js`, `metrics/breathe-quarter-dry.txt`, map `look/breathe-map-quarter-dry.png`, looked at): 4.41% of the frame is static in the before clip and steps in the after clip. 296: - 32,513 pixels take 3 levels and 24,657 take 2; 9,889 more take 4-6 (4,777 / 3,240 / 1,872) and 965 more than 6, where pools overlap. 297: - Steps: 1.30 per second median, 2.90 max (twin, `probe_life.log.out`: 3 levels at 29/42/29% for x0.89/x1/x1.11, i.e. 689,080 / 1,006,271 / 704,649 of 2,400,000 samples; 1.60 per second median). 298: - Luma range: median 2, p90 4. The x0.89/x1.11 is the shader constant, held by parity, not read off the pixels (the pool's own share of a pixel is not separable). 299: - **Lantern glass steady:** the `warm` material takes no stutter or breathe (test). 300: - **Bench** (`b/bench.sh`, `bench/summary.txt`; pre-tree against live, interleaved, round 0 discarded, `--motion-live` on both, 01:00, 1920x1080, `--disable-vsync`, rain 0.45; AC, Low Power Mode 0, `bench/power.txt`). Raw medians: 301: - dense p50 13.854 → 13.835, p95 15.581 → 15.598; 302: - Harbour p50 6.573 → 6.594 (+0.021, 3/3), p95 median 7.192 → 8.04 (+0.85 ms, +11.8%; live slower in 2 of 3 rounds). Rounds: pre 7.086/7.192/8.094, live 8.6/7.287/8.04. The live median is inside the pre spread, but one live round (8.6) is above the pre maximum. P8b has no cost line, the p50 delta is +0.021 ms and draw calls are unchanged, so the p95 move is recorded, not judged. 303: - Draw calls identical (dense 2583, Harbour 907). 304: - **Review:** filled by the orchestrator. 305: - **Not done:** 306: - Breathing window glow: the breathe applies to every pool, as the plan says, so the window pools on walls breathe as well as the lamp pools; the map shows mostly the wall halos. An owner question: lamp pools only (fixture order 1) is a one-line gate. 307: - The faulty pick lands on bulbs: in the quarter it picks bulbs rather than lettered signs. No salt was tuned to choose a sign. 308: - BAR sign mismatch: the quarter's BAR sign did not show the predicted dim step at 18:23. The Fish Market sign proves the fix; the BAR mismatch is not explained. 309: - Left to P8c: OmniLights, sign spill and fog. 310: - **Platforms:** Mac run; Windows not run. 312: **Beautify P8a Ambient-motion preference, motion clock, recording recipe (6 October 2026):** Opus 5.5 on this Mac (plan section "P8a"; report quick win 15; step `26-P8`, sub-stage P8-I). Players can set **Ambient motion: Full, Reduced or Off** in Scenery (below Creatures). The setting feeds three new shader globals that later P8 parts read. Nothing moves differently yet: no shader reads the globals, so every picture is unchanged. Owner default D17 applies (no P8 answers yet; the owner may change it): the setting will gate only the new motion terms and P5b's glints, so P8a2 is not run. 313: - **Preference:** `UserPrefs` stores `"scenery"/"ambient_motion"` (`full`, `reduced`, `off`) in `user://preferences.cfg`. A missing, unknown or non-string value reads as full. `motion_scale` is 1, 0.5 or 0. An interactive launch reads the file (`motion_choice`). An automated run never reads it; it takes `--motion=full|reduced|off` and defaults to full. Like the Creatures switch, changing it is no edit: it is not in the document, history or save, and the save stays version 2. 314: - **Globals** (`project.godot` `[shader_globals]`, written only by `WorldView.sync_motion_globals`, and only when a value changes): 315: - `motion_scale`. 316: - `light_time`: seconds, advanced by the frame delta, so it keeps running when the day clock is stopped. Every automated run holds it at `LIGHT_CAPTURE_TIME` 0 (`light_hold`, like `creature_hold`). `--motion-live` releases the hold. 317: - `art_pixel` = `view_camera.size` / SubViewport height. 318: - **Recording recipe:** `tools/art_record.sh --out --view [--seconds 10] [--fps 30]`. One command does all of this: 319: - runs Movie Maker (`--write-movie` PNG, `--fixed-fps 30`) on a `--smoke --record= --motion-live` run (new in `main.gd`: photo mode from frame 1, no stills, a 90-frame lead, then a clean quit, writing `record.txt`); 320: - keeps the last s x fps frames; 321: - writes `clip.mp4` (H.264 CRF 12, yuv420p), `contact.png` (1 frame per second at art resolution), `diff.png` (that frame minus the next, x4) and `motion.tsv` (share of changed pixels per pair, with the median); 322: - moves the frames and the .wav to the Trash (`/usr/bin/trash`). 323: - `art_capture.sh` gained `--def-of `. 324: - **Recordings are 1440x900, not 1600x1000.** Movie Maker on this Mac records at the project window size: a 1600x1000 trial came out resampled to 1440x900 with smeared art pixels (`a/trial/harbour-night-2s/`; its frames were looked at and then deleted). At 1440x900, 100% of 2x2 blocks are uniform (720x450 art pixels; `trial/harbour-night-2s-1440`). The mp4 round trip changes 0.6% of pixels by more than 8. 325: - **Files:** `scripts/user_prefs.gd`, `scripts/world_view.gd`, `scripts/main.gd`, `scripts/editor_ui.gd`, `project.godot`, `tests/test_ui.gd`, `tools/art_capture.sh`, `tools/art_record.sh` (new). Backups: `artifacts/beautify/26-P8/a/backup/`. Snapshot: `build/backup-feel/26-P8/`. The live tree equalled it at part start (diff -rq empty). Pre-part tree: `artifacts/beautify/26-P8/a/pre-tree`. 326: - **Tests:** `test_ui` `p8a_motion_checks`, 345 → 397 checks (+52), 0 failures. The checks cover: 327: - fallbacks, scales and `motion_choice`; 328: - global registration and the setter source; 329: - light clock hold per mode, `--motion-live` and advancing; 330: - `art_pixel` within 1% at sizes 16, 38 and 60 and after a SubViewport resize; 331: - the Scenery option (Full/Reduced/Off), the preference written only when interactive, no edit, no undo effect, no document key. 333: The test-first log (`a/a-testfirst-test_ui.log`) is a parse failure on the missing `WorldView.light_hold_for`. Affected suites (`a/suites/`): `test_warmup` 37, `test_creatures` 167, `test_camera` 224, `test_owner_0924` 87, all 0 failures. **Full suite** (`a/suite/suite_all.out`; a, b, v; 02:24:56-02:32:34): **42 suites, 22,457 checks, 0 failures**. `--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 verify was not run (the inherited stall). 334: - **Changed expectations:** none (only new checks). **Expected journey count deltas:** none. 335: - **Captures** (`artifacts/beautify/26-P8/a/`, `caps.sh`; journey set, pre-tree ×2 and live, one sitting; AC, 100%, Low Power Mode 0, `cap-power.txt`): 336: - `before-1/`, `after/`, `before-2/` (`/courtyard.png`). Counts are identical before→after and in the noise pair, including draw calls (`metrics/counts-*.txt`). 337: - Frame >8 luma before-1→after is 0.35-4.4% against noise of 0.31-5.3%. before-2→after is 0.56-5.4% (`metrics/frame-noise-vs-after.txt`). The file has one noise pair per view (before-1 vs before-2). Harbour-day's before-1→after 4.424% is above it (3.196%); its 1.444% is the other cross pair (before-2→after), not a noise leg. Harbour-lane (2.027% / 2.026% against 1.892%) and styles-day (0.353% / 0.564% against 0.312%) are also slightly above it in both legs. 338: - `after-scenery/scenery-panel/editor.png` shows the panel; the new row is below the fold. 339: - **Baselines for every later P8 part** (`a/rec/harbour-night/`, `a/rec/grove-day/`): 10.000 s, 300 frames, 1440x900, light clock running (`light_time` 12.96 at the end). 340: - harbour-night: median changed per sampled frame pair is 9.9%, >8 is 2.2% (rain, water, cloud dither). `motion.tsv` samples 10 pairs, one per second (frame a vs a+1), not all 299. The diff sheet `rec/harbour-night/diff.png` also shows a large ring-edged change in the top-left sky at seconds 0 and 5 only: existing behaviour (P8a adds no shader reads), so later P8 parts must not book it as their own regression. 341: - grove-day: 0.11% / 0.03% (the crown sway only). 342: - Their contact and diff sheets were looked at. 343: - **Measurements:** the part's step table has no bench. P8a adds one global write per frame and no draw call (counts identical). No bench was run. 344: - **Review:** fresh-context review, no must-fix. Findings kept as notes: `light_time` is never wrapped and reaches shaders as a float32 (about 1 ms spacing at 4 h, 7.8 ms at 24 h), so P8b wraps it at a period its sin terms divide evenly; `art_pixel` uses `view_camera.size` over the SubViewport height, which does not describe the free-camera reach path (P9a reuses it and should know). **Not done:** P8a2 (D17 default); a stills sheet of the option (covered by tests); recording at 1600x1000 (Movie Maker limit above). Platforms: Mac run; Windows not run (`art_record.sh` is Mac-only, like the other kit scripts). SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 6 October 2026, 16:17 BST, game main at 58c7827) 228: ## Orchestrator rulings after P8b (6 October 2026) 230: P8b review: no must-fix. Three items are taste or record questions, not defects, and go on the P8-I sub-stage sheet for the owner with D27: 231: 1. Breathe: every pool breathes (plan line), so window halos on walls step about 1.3 times a second while ground lamp pools barely register (breathing pixels luma range median 2). Question: breathe lamp pools only (one-line gate) or a stronger amplitude. Nothing changed now. 232: 2. Faulty picks: the quarter's two faulty tubes are a pink turret bulb and a cyan festoon bulb; no lettered sign is faulty in any stock scene (Harbour has none). Hash salt not tuned. Question: key the pick to lettered signs. 233: 3. Quarter BAR sign shows no dim step at 18:23 or 19:08: probably before its band (the probe reads LightingProfile.at_hour, which may differ from the scene's lit profile). Recorded, not fixed. 234: Also listed for the sheet: P5b glints now run on light_time x motion_scale (D17), so existing night and golden stills change 0.2-0.8% of the frame (glint dashes frozen at light_time 0 in automation). 236: ## P8-I approved (owner, 6 October 2026, in chat) 238: "All approved." Given after the P8-I sheet (`artifacts/beautify/26-P8-I/close/sheet/`, main 7f11b43) was sent. Read by the orchestrator as approval of P8a, P8b and P8c as built, and of the defaults in force (D17 Reduced and Off gate only new motion terms and P5b glints, no P8a2; D27 faulty tubes on at no more than 3 flashes per second, off under Reduced; Ambient motion default Full). The owner gave no change to the sheet's open questions, so they stay as built: every pool breathes (window halos included, no lamp-pool-only gate, amplitude as is); the faulty pick stays on the hash (no tuning toward lettered signs); faulty-sign reflection streaks do not blink; the dusk-turn threshold rounding (Dusk.hash01, float64) is not fixed, and would be its own part if the owner asks. These can be reopened by the owner at any time. 240: ## P8-II approved (owner, 6 October 2026, in chat) 242: "P8-II approved, move on to P8-III." Given while the P8-II stage close (journey, handoff check, suite, bench, sheet under `artifacts/beautify/26-P8-II/close/`) was still running, so before the owner could see the finished sheet. Read by the orchestrator as approval of P8d as built (one art pixel drops, rain box following view, zoom and the tallest content, contrast as is: about half the old drops' brightness, no dark rim) and as taking the D28 default for the later veil (P8e and P8f: the veil as the report designs it) with no change. P8-III starts (P8g1, P8g2, P8h, P8i1, P8i2, P8l, P8j, P8k) once the P8-II close has finished and no Godot is running. Reopenable by the owner when the sheet is seen. 244: P8-II approval repeated (owner, 6 October 2026, in chat): "P8-II approved, move on", after the P8-II sheet (`artifacts/beautify/26-P8-II/close/sheet/`, main 4d50efa) had been sent. The earlier approval is therefore confirmed on the finished sheet; nothing changes (rain contrast as built, D28 default stands, P8-III already running). === Appended 7 October 2026 === Frozen 7 October 2026 at 09:01 BST (game main at fa7f38d, clean). Implemented, reviewed and merged parts of sub-stage P8-III (one wind and the other motion parts): P8g1 wind field and whole-pixel crown shift (two fix passes), P8g2 bending trunk and limbs, P8h cloth for flags, pennants and valances (review fix: start-up report), P8i1 laundry, bunting and smoke gusts, P8i2 rain slant and water cat's-paws, P8j boat bob and more swarms, P8l cloud shadows (second review and fix), P8k placement dust and splash puffs (second review, wording fix) and the 7 October fix pass (cloth reach read at the start of the flap step; puffs respect the Water switch). The P8-III stage close and its owner sheet had NOT been done at the freeze, so the stage is still in progress. Owner answers: P8g1, P8g2 and P8h approved in chat on 6 October; P8i1 to P8k and the fix pass accepted under the owner's standing approval until 08:00 on 7 October, with open items for the P8-III sheet. P8-IV (the rain veil) had not started. Part codes: P8g1/P8g2 wind in the trees, P8h cloth, P8i1 washing, bunting and smoke, P8i2 rain and water in the gusts, P8j boats and swarms, P8l cloud shadows, P8k puffs; D17 motion setting scope, D25 release cost line, D29 whole-pixel crown shift. Line numbers refer to the frozen copy. Moves "to ~/.Trash/" read "to the macOS Trash"; the workflow run ID is replaced by . SOURCE: docs/visual-handoff.md (frozen copy = HEAD fa7f38d, 7 October 2026, 09:01 BST), lines 5-267 5: **Beautify P8k Placement dust and splash puffs (7 October 2026):** Opus 5.5 on this Mac (plan section "P8k"; report polish backlog "Placement dust and splash puffs on release" and its lens finding "Nothing puffs when a part lands or is removed"; step `26-P8`, sub-stage P8-III, last part; a measuring part, second review required). When a release sets a part down on the ground, a small ring of dust puffs (3-6 art px each) rises round its foot in the ground's colour; on water it is a splash of drops and one pale ring spreading on the surface (1-2 art px thick while small, a blocky band 3-6 art px thick as it grows: see "Second review" below). A removed part puffs when its sink ends. Undo and redo puff too, at REPLAY rate. One pooled burst per release; nothing on a drag step, a stroke's dab, a settled or forced sync, with build animations off (so no automated still or bench ever puffs) or under Ambient motion Off; Reduced fires half the puffs. P3b's CPU splash squares are not brought back. Snapshot: `build/backup-feel/26-P8-k/` (= `main` 1156239; the live tree matched it at part start, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/k/pre-tree`. Evidence: `artifacts/beautify/26-P8/k/`; file backups in `k/backup/`. 6: - **Code:** 7: - `scripts/placement_puffs.gd` (new, `class_name PlacementPuffs`, editor import `k/import.sh`): a pool of `POOL 4` slots made with the view, each a one-shot CPUParticles3D for the dust or drops (explosiveness 1, `PUFFS 12`, `LIFETIME 0.55 s`, outward `0.6-1.0 m/s` along each point's normal with `RISE 0.25 m/s` up, `DAMPING 2.5`, no gravity; hard-stepped fade 0.9/0.6/0.3) and a one-particle flat ring (PlaneMesh, grows `RING_GROW 0.4 → 0.85` of the part's reach + 0.3 m, fade 0.6/0.4/0.2). Statics: `count_for(motion)` (Full 12, Reduced 6, Off 0, D17), `size_range(art)` = 3-6 × art_pixel, `colour_of(finish, ambient)` = `Architecture.GROUND_COLORS[finish-1]` lerped `AMBIENT_MIX 0.4` to the profile ambient, `dim_of(windows)` = lerp(1, `NIGHT_DIM 0.5`, windows) (like the smoke's tone), `splash_colour(ambient)`, `ring_points`, `circle`. `fire(spots, water, colour, rate, motion, art)` takes the next slot in turn (the oldest is reused), thins to `MAX_POINTS 32`, sets `speed_scale = rate`, and the slot hides itself `LIFETIME/rate + 60 ms` later (`tick`), so an idle pool draws nothing. The dust material is ChimneySmoke's three-ring puff texture with `billboard_keep_scale` (without it the first film drew every puff a metre wide: `k/look/old/try1-place-tree.png`); the ring texture is clear to 0.78, opaque to 0.92 (the lens's 0.6-0.9 read as a broad white band, `k/look/old/try3-delete-boat.png`). 8: - `scripts/world_view.gd`: `puffs` made beside the smoke in `_ready`; `_puff_release(fire)` at the end of `sync_scene` compares each top-level part's XZ with the last full sync's (`puff_seen`, id → [pos, object]); new or moved parts (no parent, no `points` param) are set down, missing ones removed. A height-only change, a dab (`whole_next`), a settled or forced sync and any sync with animations off only record (animations off forget the record, `puff_based`). `_puff_fire` builds one burst: a building's or boat's footprint, else a ring (`TREE_RADIUS 0.7` for tree, blossom and conifer, just outside the foot collar; `PROP_RADIUS 0.35`), at `ground_height` or `WATER_Y`; a splash when most parts stand in water; the colour from the first land part's finish, × `dim_of(profile.windows)`. Removals wait in `puff_due` until `build_now() + REMOVE/sync_rate` (`_tick_puffs` from `tick_build`; `build_until` extended). `set_animate(false)` drops waiting removals and hides the pool. `sync_profile` keeps `puff_ambient` and `puff_windows`. 9: - `scripts/chimney_smoke.gd`: the puff texture builder moved into `static func puff_texture()` (same texture, smoke unchanged). `scripts/shader_warmup.gd`: `entries_of` adds the dust and ring materials as a particle system draws them ("multi", colours and custom data). 10: - Deviations: puff colour × `dim_of(windows)` is not in the plan (unshaded puffs glowed pale on the 01:00 lawn, `k/look/old/live-night-delete-tree.png`). Drawn parts with points (walls, pipes, fences, bridges) and mounted parts never puff. "A3 adapt paths": an adaptation that moves another part aside in the same release joins that release's one burst at the release's rate; only undo and redo run at REPLAY (`sync_rate`). 11: - **Tests:** test-first logs `k/testfirst-test_build_anim.log.out` (461 checks, 1 failure: no `placement_puffs.gd`) and `k/testfirst-test_warmup.log.out` (43 checks, 2 failures). `test_build_anim` 460 → 510 (`p8k_puff_checks`: pool of 4; 10-14 puffs at Full, fewer at Reduced, none under Off; 0.55 s; 3-6 art px; colour per finish; dim by day 1 and at full windows 0.5; the view's pool idle and hidden at init; settled and forced syncs puff nothing; one puff for a placed tree, round its foot at ground height, the lawn's colour, one burst at rate 1, no ring on land; an unchanged release, drag steps (`sync_live`), a height-only change and a dab puff nothing; the release after a drag puffs once where let go; two parts in one release, one burst round both; a boat on water splashes with its ring at the water line; a removal puffs only when the sink ends, where the part stood; redo at REPLAY rate (`speed_scale` 2.5) and an undo's removal at REMOVE/REPLAY; slots reused in turn; a finished burst hides its slot; Reduced fires 6; Off and build animations off fire nothing and drop a waiting removal; automation builds settled; nothing saved). `test_warmup` 41 → 43 (the pool exists at init; both materials warmed as particles draw them). Affected suites on the final code: `k/after3-test_build_anim.log.out`, `k/after3-test_warmup.log.out` (0 failures); `test_release_spread` 130/0 (`k/after2-…`). Full suite on the final code (`k/suite/suite_all.out`, 04:41-04:49, AC 100%, Low Power Mode 0, `k/suite-power.txt`): **44 suites, 23,392 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Final code times 04:22:03-04:22:10 (`k/final-code-times.txt`), before the final films, benches, start-up runs, still sets 3/4 and the suite. 12: - **Changed expectations:** none of the existing assertions changed. One of my own new checks changed while the part ran: the view's colour check went from `colour_of(2, ambient)` to `dimmed(colour_of(2, ambient), windows)` with the night dim (and a new check of `dim_of`). 13: - **Captures** (`k/film.sh` with `k/film/puffs.json`: the A0 drag harness with `--animate`, Harbour, camera 0.78,0.56,15 at (15.5, 0.6, 3.5), 1600x1000, rain 0; a tree placed on the east lawn and deleted, a boat placed on open water and deleted, the delete undone; frames at 0-900 ms; final films 04:22-04:25, AC, Low Power Mode 0, `k/film/power.txt`; before from `pre-tree`). Measured with `k/an/puffs.js` (changed art pixels > 24 summed RGB in each step's crop, their blobs and colours; `k/an/out/*.txt`, `totals.txt`), then every frame moved to the Trash (the macOS Trash; sheets, performance.json and scans kept). Looked at: `k/look/live-day-delete-tree.png` (a ring of pale green-grey puffs from 267 ms, fading by 750), `k/look/live-day-delete-boat.png` (drops and a ring growing and fading; thin at 267-400 ms, a stair-stepped band of blocks at 600-750 ms), `k/look/live-night-delete-tree.png` (darker dust at 01:00), `k/look/live-off-delete-tree.png` (no puffs), `k/look/pre-day-delete-tree.png`. 14: - **Acceptance, against the target:** 15: - **"A 3-6 art-px puff in the ground's colour":** land removal (delete-tree, live-day against pre-day): 10 puffs at 267-600 ms (12 fired; some overlap), blob width median 4-5 art px, max 6 (**target 3-6: met**). Colour: puff pixels #79936b against the lawn under them #537325 by day; #665f43 against #53521a at 01:00. **Same hue family, but lighter and greyer than the lawn** (about +36 luma by day): the plan's formula lerps the finish 40% toward the pale day ambient and the puff is unshaded while the lawn is lit. Left for the owner (AMBIENT_MIX and NIGHT_DIM are the two constants). 16: - **"Release `step_total` within 2 ms of before"** (D25, Harbour animated release; `k/bench.sh`, interleaved, rounds 1-3, AC, Low Power Mode 0, `k/bench/power.txt`, 04:26-04:36). Release probe (`k/release-summary.js`, `bench/release-summary.txt`; the live tree fires a puff on each of the 9 house moves): **Harbour animated 31.69 → 31.79 ms (+0.10, 3/3)**; Harbour off 25.96 → 25.99 (+0.03); dense on 110.84 → 111.53 (+0.69); dense off −0.28; dense-bare on −0.01, off +0.06; drained slices and worst slice unchanged. Drag harness with `--animate` (`k/drag-summary.js`, `bench/*-summary.txt`): **the first puff of each kind in a launch without the warm-up hitches**: place-tree 183.96 → 260.60 (+76.6), place-boat 46.45 → 113.90 (+67.5), the first use of the dust and the ring materials (a warm placement first, `bench/puffs-warm.json`: those steps carry the same +77.4 / +66.8, and the measured place-tree after them is +0.29, place-boat +0.61). **With `--warmup`, as a person's launch draws the warm-up: place-tree −0.69, place-boat +0.54, delete +0.11-0.19, undo −1.88 (`bench/wu-summary.txt`): met.** Automated runs without `--warmup` (and any headless or `--warmup=off` launch) still pay the first-use compile; reported, not fixed. 17: - **"Never on a drag step or preview":** tests. **"Nothing with build animations off or in automation":** still sets 3/4 below; tests. **"Off under Off":** `live-off` against `pre-day`: 0 changed pixels in the removal window, 7,238 changed art pixels over all steps against the noise pair `live-day2`/`live-day` 13,152. **Reduced:** 6 puffs (`live-reduced`). 18: - **Stills** (`k/caps.sh`, six views, sets 1/2 at 04:15 on the code before the night dim, sets 3/4 at 04:49-04:52 on the final code, AC, Low Power Mode 0, `k/cap-power.txt`; `k/metrics/`): every integer count identical, draw calls included, in all four sets. Pixels > 8 summed RGB, before-3 → after-3 / before-4 → after-4 against before-3/before-4 and after-3/after-4: harbour-day 0.84/1.10 against 1.19/0.72, harbour-night 1.09/0.88 against 0.81/0.62, lane-day 1.87/2.19 against 1.87/2.01, quarter-day 0.41/3.37 against 0.39/3.50, quarter-night 0.61/1.21 against 0.91/0.90, styles-day 0.22/0.47 against 0.18/0.50: within noise. Looked at `k/after-1/harbour-night/courtyard.png`. 19: - **Start-up** (`k/startup.sh`, `k/startup.out`, 04:36-04:41, same power; two new materials): medians of rounds 1-3: entries 142 → 144, pieces 306 → 310, ready_ms 6999 → 6991, warm_ms 4878.2 → 4872.4, all_ms 11451.1 → 11722.6 (+272, two more idle rounds). 20: - **Counts and draw calls:** identical on stills (an idle pool is hidden). While a burst shows, its emitters draw: 1 draw call for dust, 2 on water. 21: - **Expected journey count deltas:** none (journeys build settled; no puff fires). 22: - Review: filled by the orchestrator. 23: - **Second review (must-fix, wording only):** the entry called the splash "one thin pixel ring"; it is thin only for the first part of its life. The ring is one 16x16 texture (`PlacementPuffs.ring_texture`, nearest filter) scaled over the whole ring, so each texel grows with it: by 600-750 ms the band is 3-6 art px thick with stair-step blocks up to about 15x9 art px. Re-measured on the trashed final-code frames (the macOS Trash, delete-boat, live-day against pre-day, art grid 800x500; `k/fix2/ringwidth.js`, outputs `k/fix2/ring-live-pre.txt`, noise pair `ring-noise.txt` (live-day2/live-day), `ring-off.txt`). Thickness where the ring crosses its centre row and column (median / max art px): 200 ms 1/1; 267 ms 1-2 / 3-6; 333 ms 2 / 3-6; 400 ms 1 / 3-5; 500 ms 2 / 4; 600 ms 3-6 / 5-9; 750 ms 4 / 7. Ring pixels that survive a 1-px erosion: 0, 116, 115, 156, 242, 407, 620 (200-750 ms) against at most 21 in the noise pair and 0 under Off. The reviewer's own scan (`k/review2/ring.js`, re-run: run median 2 / max 6 at 267 ms, 1 / 6 at 400, 3 / 9 at 600, 4 / 11 at 750) agrees. Blobs at 750 ms (`k/review2/scan.js` re-run, `k/fix2/scan-live-pre.txt`): four ring arcs 32x30, 32x29, 32x24, 33x23 art px and separate 15x9 blocks. Looked at `k/look/live-day-delete-boat.png`: the 600 and 750 ms tiles show the blocky band. Fixed: this entry's wording (summary, captures, owner sheet). No code changed, so no suite, film or bench was re-run: the code files are unchanged since 04:22:10 (`k/final-code-times.txt`), before the full suite at 04:41-04:49; no test reads this document. The look (a finer ring texture or a thinner ring) is the owner's call and not changed. 24: - **Orchestrator fix pass (7 October 2026):** two must-fix findings, both fixed; evidence `artifacts/beautify/26-P8/k/fixpass/` (file backups `fixpass/backup/`). Snapshot `build/backup-feel/26-P8-fix/` (= `main` adff4c9; the live tree matched it, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected); pre tree `fixpass/pre-tree`. Asked: (1) P8k review finding 1: puffs on water ignored the Water switch, so with Water off a part set on a cell that is not land threw white drops and a ring at `WATER_Y` over the dry bed; (2) P8i1 review finding 2: cloth and line pieces read the gust's reach at t, so a 1 → 2 or 2 → 1 change could land inside a +n or −n flap step and leave a pose on screen for one frame. Changed: (1) `WorldView._puff_fire`: `water = not state.land_at(p) and bool(state.environment.water)`, so with Water off the existing dust path runs (ground colour of the cell's finish, at `ground_height`, no ring). (2) `wind_cloth_pixels` (`shaders/wind.gdshaderinc`) and its twin `WorldView.cloth_pixels` read n (`wind_line_pixels` / `wind_crown_pixels`) at t0, the first moment of the piece's own flap step: t0 = (floor(t × motion / CLOTH_STEP + ph) − ph) × CLOTH_STEP / motion with ph = 4 × the piece's phase (new twin `cloth_step_start`). The orchestrator's formula floor(t × motion / CLOTH_STEP) × CLOTH_STEP / motion leaves out the phase; without it a change could still fall inside a phased step, so the phase is included. Motion ≤ 0 returns 0 first (Off as before). Gust field, timing, amplitude rule and vertex counts untouched; a reach change now waits for the next step boundary (up to 0.25 s, 0.5 s under Reduced). 25: - **Tests** (test first: `fixpass/testfirst-test_cloth.log.out` 92 checks, 6 failures; `testfirst-test_build_anim.log.out` 515, 3; after: `after1-*.log.out` 0 failures): `test_cloth` 84 → 92 (`hold_checks`: include text reads the reach at t0; the twin over 300 s at 1/16-step sampling, modes 0 and 2, Full and Reduced, default and full rain, 3 anchors: every pose holds ≥ one whole step, shortest before 0.0156 s (Full) / 0.0312 s (Reduced), now ≥ 0.234 / 0.469; reach changes crossed inside gusts; Off still in modes 0-2; the pose's reach is n at t0). `test_build_anim` 510 → 515 (Water off: the cell is not land, one puff, dust not splash, no ring, at ground height). Full suite (`fixpass/suite/suite_all.out`, 05:15-05:23, AC 100%, Low Power Mode 0, `fixpass/suite-power.txt`): **44 suites, 23,405 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,583/0. 1440x900 `--verify` not run (inherited stall). Final code times 05:13:40-05:14:55 (`fixpass/final-code-times.txt`), before the suite and every capture and bench. 26: - **Changed expectations** (old → new): `test_cloth` P8h "its reach is the crown's n" compares with `crown_pixels` at t → at the flap step's start; P8h "Reduced never moves the free edge further than Full" likewise; P8i1 "a line piece flaps 0, +n, 0, −n with n its own reach" compares with `line_pixels` at t → at the step's start; P8i1's include text check `wind_line_pixels(anchor,t)` → `(anchor,t0)`. My own new text check first looked for `+4.0*` inside the floor; the code names the phase `ph`, so it now checks `+ph)` and the t0 line (changed before the after-run, not to pass a behaviour check). 27: - **Clips** (`fixpass/rec.sh`, 05:24-05:45, AC, Low Power Mode 0, `fixpass/rec/power.txt`; 20 s, 15 fps, 1440x900; still = `fixpass/still-tree`, the pre tree with `wind_cloth_pixels` returning 0, a measuring harness, never shipped; cameras as P8h/P8i1: `q` quarter laundry, `hb` Harbour bunting, `st` styles laundry, `pz` quarter flag at rain 1). Pose holds (`an/holds.js` per piece and `an/pixhold.js` per art pixel against the still clip; `an/out/holds-rows.txt`, `pixhold.txt`; step = 3.75 frames at Full, 7.5 under Reduced): quarter laundry Full piece hold_min 1 → **3** (hist {1:3, 3:2} → {3:5}), per pixel 75 pixels with a 1-frame hold → 0; Reduced 3 → **7** ({3:2, 4:2, 7:2} → {7:6}). Harbour bunting, art pixels whose shortest hold is under a step: 656 → 21 of 1,270 at Full (1-frame 403 → 13), 476 → 15 of 932 under Reduced. Styles laundry: 151 → 41 of 274 at Full (1-frame 102 → 35), 82 → 19 of 168 under Reduced. The rain-1 flag (`an/edge.js`, P8h's edge scanner; `an/out/pz-moving-runs.txt`): moving pose runs pre 1 1 2 4 3 4 3 4 3 4 → after 3 4 3 4 3 4 3 4; Reduced 7 1 6 7 3 → 7 1 6 7 7 (the 1 6 split is the tip column's threshold edge, the same in the pre clip; the free-edge shift holds 7 frames). Looked at `fixpass/look/q-pre-f328-332.png` (pre: the garment's hem steps out for frame 330 alone) and `q-after-f328-332.png` (after: it holds 330-331). 28: - **Not fully met:** acceptance line 1 asks every pose to hold a whole step in every clip. The quarter clips and the flag meet it. Harbour and styles still have some art pixels with holds shorter than a step, measured but not explained: 21 and 41 at Full, 15 and 19 under Reduced (`pixhold.txt` lists the first 12 of each with frames; e.g. Harbour (175,246) at 19.6 s, styles (312,215) at 18.8 s). Neighbouring garments and pennants overlap there, and each keeps its own phase, so one pixel can change when either piece steps; the twin test shows no pose shorter than a step. Telling the two apart would need a per-piece render, which was not done. 29: - **Timing, amplitude and Off** (`an/rows.js` against the still clip): first-to-last moving light_time quarter 17.53-22.00 → 17.67-22.47, Harbour 16.33-22.93 → 16.73-23.07, styles 18.27-22.87 both (moves quantised to step boundaries); largest per-frame shift still 1 or 2 at Full (quarter {1:23, 2:44} → {1:29, 2:42}) and 1 under Reduced. Off: `q-off` against `q-off-pre` 0 of 300 frames differ (whole frame); noise pair `q-after`/`q-after2` 0 of 300 (`an/out/pairs.txt`). Frames to the Trash after measuring (the macOS Trash). 30: - **Stills and counts** (`fixpass/caps.sh`, harbour-day, quarter-day, styles-day, before/after sets 1-2, 06:05-06:07, AC, Low Power Mode 0, `fixpass/cap-power.txt`; `fixpass/metrics/`): every integer count identical, draw calls included. Pixels > 8 before → after 5.05/3.63 (harbour-day), 0.46/1.03 (quarter-day), 3.74/0.40 (styles-day) against noise pairs 1.81/0.77, 1.27/0.53, 1.05/3.23. The Harbour excess is rain drops and rain rings on the water, as in its noise pair (`fixpass/look/hd-b1-a1-diff.png` against `hd-b1-b2-diff.png`, looked at); nothing on a building or cloth (all cloth rests at light_time 0). 31: - **Bench** (`fixpass/bench.sh`, `bench-r012.out`, 06:10-06:12, AC, Low Power Mode 0, `fixpass/bench/power.txt`; pre tree against live, interleaved, round 0 discarded, rounds 1-2 only (time), 1920x1080, vsync off, 01:00, rain 0.45, `--light-pin=19.0`): p50 Harbour 6.676/6.682 → 6.697/6.697, quarter 7.021/7.108 → 7.067/7.099; draw calls 916 and 1043 both. No dense run, no start-up run (no new materials). 32: - **Expected journey count deltas:** none. 33: - **For the owner sheet:** dust reads pale green-grey, lighter than the lawn (the plan's 40% ambient lerp on an unshaded puff); the night dim is mine. The splash ring is about the boat's reach + 0.3 m across at the end, with white drops. It starts 1-2 art px thick, but its 16-texel texture stretches as it grows, so by 600-750 ms it is a blocky band 3-6 art px thick (up to 9 at the sides) with stair-step blocks up to about 15x9 art px (`k/look/live-day-delete-boat.png`, last two ring tiles). A finer ring texture or a thinner ring is the owner's choice. Puffs at the default Harbour zoom are 1-2 screen-art pixels smaller than at this lane zoom (sized by art_pixel, so they stay 3-6 art px). Without the warm-up, the first puff costs about 70 ms (shader compile). 34: - **Not done:** no Glade, perspective, photo-mode or building (house) placement film (house footprints are tested headless only); the drag bench is Harbour only, the release probe covers dense; second review done (one must-fix, wording; above). Walls, paths and other drawn parts do not puff. 35: - Platforms: Mac run; Windows not run. 36: 37: **Beautify P8l Cloud shadows drifting across the island (7 October 2026):** Opus 5.5 on this Mac (plan section "P8l" (moved from P4g); report polish backlog "Cloud shadows crossing the island by day (a HeightFog-style quad, 2-step noise)" and its lens finding; step `26-P8`, sub-stage P8-III; a measuring part, second review required). By day and at golden hour soft cloud shadows lie on the town, the walls and the water and drift downwind along the one wind. Each is two flat steps (a half-strength penumbra round a core), no dither. They fade out with the sun's light in the evening and are not drawn from the lit evening (20:45) through the night to dawn. Glade draws them at half strength. Every automated still shows one fixed pattern (the light clock is held at 0); Off holds the pattern still, Reduced drifts at half speed. Snapshot: `build/backup-feel/26-P8-l/` (= `main` a894099; the live tree matched it at part start, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/l/pre-tree`. Evidence: `artifacts/beautify/26-P8/l/`; file backups in `l/backup/`. 38: - **Code:** 39: - `scripts/cloud_shadow.gd` (new, `class_name CloudShadow`, cloned from HeightFog): a camera-pinned full-screen quad, render priority 115 (after ContactShade 110, before HeightFog 120). Constants `ALPHA 0.12` (core; penumbra half), `SPEED 1.5 m/s`, `CELL 23 m`, `CELLS 45` (the pattern repeats every 1035 m along the wind; SPEED × the 1380 s light-clock wrap is 2 periods at Full and 1 at Reduced, so the wrap has no seam), `OCTAVE 0.35`, `PENUMBRA 0.58`, `CORE 0.64` (37% of the ground shaded, twin probe `l/probe/tune.js`), `HEIGHT 40 m`, `SUN_FLOOR 0.35`. `shade(profile, hour, heading, look)` sets `strength` and the true sun's direction and hides the node at 0. Static twin: `sun_up` (SunPath.handover while the sun is the key, 0 while the moon is), `strength_of` = ALPHA × look × min(sun_up, 1 − HeightFog.night), `drift`, `deck_of` (from the point toward the sun up to the deck), `pattern_of`, `cover` (two octaves of value noise, periodic along the wind), `level` (0, 0.5, 1). 40: - `shaders/cloud_shadow.gdshader` (new): the fog's ray rebuild (depth, or the water plane; perspective path kept). It discards the painted sky (panel depth and `SCREEN_UV.y < SKY_HORIZON`, the rows water.gdshader leaves to the backdrop; the fog's test alone shaded the sky in the first trial, `l/look/try-harbour-golden-shade.png`) and anything past the water's edge. Since the second review it also leaves the water's first art row under the horizon line unshaded (`CLOUD_HORIZON_GUARD 1.25` rows; twin `CloudShadow.SKY_HORIZON`, `HORIZON_GUARD`, `horizon_clear`), see "Second review" below. The pattern drifts by `CLOUD_SPEED × light_time × motion_scale` along `WIND_DIR` (includes `wind.gdshaderinc` and `sky.gdshaderinc`). Output `ALBEDO = 0`, `ALPHA = strength × level`, premultiplied. 41: - `scripts/look_profile.gd`: RENDER `cloud_shadow` canal 1.0, Glade 0.5. `scripts/world_view.gd`: `var cloud: CloudShadow`, made beside the fog, `cloud.shade(...)` in `sync_profile` after `fog.tint`, `backdrop_depth` set with the fog's. No saved key, no switch. ShaderWarmup picks the quad up as a node of the view. 42: - `tools/test.sh`, `tools/Test.ps1`: the new suite `test_cloud_shadow`. 43: - Deviations: D17 is honoured by the drift alone. The pattern is a new term, but a still pattern under Off (the plan: "frozen ... scaled by motion_scale") is drawn rather than none. The drift speed is constant, not × wind_strength: a speed that followed the weather would jump the pattern whenever the rain changed, because the offset is light_time × speed. The deck is read toward the sun (not straight up), so a wall's shadow follows the building's own. 44: - **Tests:** test-first logs `l/testfirst-test_cloud_shadow.log.out` (11 checks, 11 failures: no script, shader or RENDER key) and `l/testfirst-test_warmup.log.out` (41 checks, 1 failure); no script errors. `test_cloud_shadow` (new, 71 checks): RENDER keys; ALPHA in 0.10-0.14; the priority sits between the contact shade and the fog; the wrap holds whole periods; shader text (premultiplied quad, depth, the wind include, `light_time*motion_scale`, black, `ALPHA=strength*`, no TIME, the sky discard; constants equal the twin's); strength (Day ALPHA, golden 0.10-0.14, zero at night and at 20:45, 21:00, 21:30, 22:00, 00:00, 03:00, 05:45 and 06:10, Glade half, minute by minute never above ALPHA × handover, zero while the moon is the key or the windows are lit, no pop); the pattern (no drift at t 0 or Off, Reduced half; two steps only; 20-50% coverage with both steps; seamless across the wrap at Full and Reduced; it rides downwind at SPEED; Off equals the still); no blips (every level holds at least 1/15 s, median at least 1 s, one band at a time, at Full and Reduced over 40 s × 60 points); the deck look-up; a real view (`world.cloud`, a camera child, priorities, ALPHA by day, the sun's direction, hidden at night, golden, Glade half, backdrop depth like the fog, no environment key, saved scene unchanged). `test_warmup` 40 → 41 (the quad's material is warmed). Full suite on the final code (`l/suite/suite_all.out`, 03:12-03:20, AC 100%, Low Power Mode 0, `l/suite-power.txt`): **44 suites, 23,334 checks, 0 failures**. `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`. creature_reactions rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Final code times 02:40-02:47 (`l/final-code-times.txt`), before every capture, clip, bench, start-up run and the suite. **Fix pass (second review):** test-first log `l/fix-testfirst-test_cloud_shadow.log.out` (72 checks, 1 failure: no `horizon_clear`); `test_cloud_shadow` 71 → 77 (`horizon_checks`: the twin's SKY_HORIZON is sky.gdshaderinc's, the shader's guard is the twin's, the shader text holds the guard; at art heights 300, 450, 500, 540, 600, 900 and 1000 rows the first water row is unshaded, the water from two rows down is shaded and at most 2 rows are clear; the near water is shaded), `l/fix-test_cloud_shadow.log.out`. Full suite on the fixed code (`l/fix/suite/suite_all.out`, 03:49-03:58, AC 100%, Low Power Mode 0, `l/fix/suite-power.txt`): **44 suites, 23,340 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,583/0. Fixed code times 03:40:30-03:40:51 (`l/final-code-times.txt`), before the fix captures and the suite; no code edited after. 45: - **Changed expectations:** none of the existing assertions changed. Two of my own new checks changed while the part ran. (1) "No pop: under 0.01 a minute" failed at 0.0140 (06:16) on the first code. It became "never faster than the key light's own fade (ALPHA × the largest change of SunPath.handover in a minute)". The code changed with it, for two reasons: sun_up was not 0 while the moon was the key (05:50-06:10), and the product of the two fades (0.0194 a minute) became their lesser. (2) The hold check was set before the code, from the twin probe (`l/probe/tune.js`: grazing band tips give holds down to 0.12 s), as "at least 1/15 s, median at least 1 s", in place of my first "at least 0.3 s". 46: - **Captures** (`l/caps.sh`, sets 1-2 interleaved before/after, 02:48-02:52, AC, Low Power Mode 0, `cap-power.txt`). Views: harbour-day, harbour-golden, harbour-night, quarter-day, quarter-night, glade-day, lane-day, styles-day, quarter-2045, quarter-0545, plus `harbour-2015` (Harbour camera, `--time=20.25`). Each set reports fail=1 only because `--views` named the `--def` view; it was captured. Measured with `l/an/shade.js` (alpha = 1 − after/before luma per art pixel; penumbra 0.03-0.09, core ≥ 0.09), `metrics/shade.txt`, `metrics/noise-*.txt`, `metrics/counts-*.txt`. Looked at: `look/shade-1-*.png` (shade maps), `look/try2-hg-crop.png` (golden: the right lawn and tower under a core with a stepped edge), `look/try2-hd-crop.png` (day: far water shaded), `look/hd-live-shade-90-135-180-225.png` (the drift), `look/q2045-b1-a1.png` (20:45: rain rings and drops only), `look/off-px-229-143.png`. 47: - **Acceptance, against the target:** 48: - **"Soft 2-step shadows at alpha 0.10-0.14 by day and golden":** met as the shader's alpha (ALPHA 0.12 core, 0.06 penumbra, tested), not as a uniform darkening in pixels. Restated after the second review with a noise-free measure (`l/an/stable-alpha.js`, `l/fix/stable-alpha.txt`): only pixels identical within both noise pairs (b1 = b2 and a1 = a2, so rain and TIME drop out), before summed RGB ≥ 120, penumbra and core split at each histogram's valley, the core also split by the before pixel's brightness (sum ≥ 480 is mostly the bright far water just under the horizon, `l/an/alpha-where.js`). Final-code sets 3/4 (sets 1/2 agree within 0.004): **water reads about 0.09**: bright core harbour-day 0.093 (p10 0.088, p90 0.102), quarter-day 0.093, styles-day 0.098, harbour-golden 0.085. **Ground, walls and roofs read 0.11-0.13**: harbour-day 0.114, quarter-day 0.118, styles-day 0.121, harbour-golden 0.128, lane-day 0.130. Penumbra medians 0.054-0.070. Share of stable measured pixels in the core: harbour-day 9.4%, golden 24%, styles 13%, quarter-day 2.2%, lane-day 1.2% (the quarter town and lane lie mostly in a gap of the still's pattern). Between the two before sittings every class holds ≤ 0.12% of pixels. The first-pass figures (core 0.104-0.134, `l/an/shade.js`) counted every pixel at ≥ 0.09 as core, rain included, and overstated how dark the core reads. 49: - **"Off at night":** night and twilight counts are identical (the quad is hidden). Pixels >8 before → after (set 1/set 2) against the noise pairs (before/before, after/after): harbour-night 6.56/0.75 against 5.97/1.52; quarter-night 7.01/0.69 against 6.08/1.41; quarter-0545 0.57/0.99 against 0.54/1.01; quarter-2045 2.85/2.90 against 1.69/1.79. 20:45 is a little above the pairs, but its differences are rain rings, drops and canal ripples only (`look/q2045-b1-a1.png`), and the quad is not drawn there. The protected mirror, neon cores and 05:45/20:45 skies are untouched. Final code (sets 3/4, `l/fix/night.txt`, % of pixels > 8 summed RGB, before → after against before/before and after/after): harbour-night 1.14/1.05 against 7.55/7.06, quarter-night 11.56/1.04 against 10.66/1.45, quarter-0545 2.59/0.66 against 4.02/2.98, quarter-2045 2.81/1.23 against 4.50/3.83. **The first pass's claim that by day and golden the painted sky was untouched was wrong** (second review): see "Second review" below for what changed the sky's bottom row and what holds now. 50: - **"Halved in Glade":** the shader strength is exactly 0.06 (test). In pictures (noise-free measure) the Glade's shaded pixels form one class at median 0.041 (p10 0.030, p90 0.057, 10% of pixels) against canal's ground core 0.11-0.13 and water 0.09, about 0.35-0.45× (through AgX at exposure 1.5). 51: - **Dusk blend:** harbour at 20:15: penumbra-class darkening median 0.041 (strength 0.042 at windows 0.78), on 17% of pixels. The strength reaches 0 at 20:43; dawn fades in 06:10-06:30 with the key light. 52: - **Freeze (two identical runs):** after-1 against after-2 on day views changes no more than the before pair does: harbour-day 3.14 against 7.09, golden 1.27 against 6.56, quarter-day 3.11 against 3.46, glade 0.47 against 0.76, styles 0.42 against 0.22. 53: - **Drift and no shimmer** (`l/rec.sh`, harbour-day 10 s clips, 15 fps, light clock 6 → 16 s, rain 0, smoke off, 02:53-03:00, AC, Low Power Mode 0, `rec/power.txt`; `l/an/drift.js`, `an/out/drift.txt`; on the 222,052 art pixels that never change in the pre clip). The noise pair `hd-live`/`hd-live2` differs in 0 of 150 frames. In `hd-live`, 37,488 pixels change 56,244 times. Blips (a change undone within 2 frames) number 80 (0.14% of changes); reversals within 4 frames, 543 (1.0%), are grazing band tips like those the twin predicts. The shaded share is 11.3 → 13.0% over 10 s. The patches move right and up the frame (the map sheet). The edge-speed scanner `an/onset.js` (4.8 px/s) is quantised and unreliable, so the speed rests on the twin test. Reduced: 28,425 changes (0.51×), 13 blips. Off (`hd-off` against `hd-off-pre`): the shaded share is 8.9% in every second, but 19 pixels change 70 times (0 in `hd-off-pre`). They are stone pixels beside the water that sit 2-6 levels lighter than before and flicker (`look/off-px-229-143.png`). The probable cause is the presentation pass reacting to the TIME-driven water beside a shaded edge; it was not investigated further. 54: - **Counts:** draw_calls and rendered_objects +1 on every view where the shadows are drawn: harbour-day 906 → 907, harbour-golden 828 → 829, harbour-2015 826 → 827, quarter-day 1023 → 1024, glade-day 905 → 906, lane-day 748 → 749, styles-day 500 → 501. Night, 05:45 and 20:45 are identical. 55: - **Bench** (`l/bench.sh`, `bench-r01.out`, `bench-r23.out`, `bench-summary.txt` ("mid" is the pre tree), 03:02-03:07, AC, Low Power Mode 0, `bench/power.txt`; pre tree against live, interleaved, round 0 discarded, 1920x1080, vsync off, **13:00** (the shadows are hidden at night), rain 0.45, `--light-pin=10`). p50: dense 14.204 → 14.268 (paired +0.015, slower 3/3), Harbour 6.419 → 6.471 (+0.052, 3/3), quarter 6.849 → 6.871 (+0.046, 2/3). p95: dense +0.086 (3/3), Harbour +0.086 (2/3), quarter +0.008. The plan's stop (dense p95 over about +0.5 ms) is not hit. Draw calls 2586 → 2587, 916 → 917, 1043 → 1044. Start-up (`l/startup.sh`, `startup.out`, 03:07-03:11, same power; new shader and material): entries 141 → 142, pieces 304 → 306; medians of rounds 1-3: ready_ms 6961 → 6963, warm_ms 4866.1 → 4868.6, all_ms 11339.0 → 11471.7 (+133). 56: - **Expected journey count deltas:** draw_calls and rendered_objects +1 on harbour-day, harbour-lane, quarter-day and styles-day; harbour-night and quarter-night identical. Pixels: the day views change (shadows in the still's fixed pattern); night views change by no more than noise. 57: - **Second review and fix pass (7 October, 03:39-03:59):** two must-fix findings. 58: - (1) **The sky's bottom art row was lightened at golden hour.** The presentation outline (`presentation.gdshader`, edge = smoothstep(0.14, 0.32, brightest − here)) darkens the painted sky's last row against the brighter water under it; shading that first water row weakened the outline, so the sky row got lighter wherever the pattern lay on the horizon water (review: harbour-golden rows y 228-229 at 1600x1000 brighter on 878/872 of 1600 stable pixels by +33 summed RGB; harbour-2015 233 pixels at +2.6). Fixed in the shader: the water's first art row under the horizon line is never shaded (`on_water && SCREEN_UV.y < SKY_HORIZON + CLOUD_HORIZON_GUARD / VIEWPORT_SIZE.y`, guard 1.25 rows, so exactly the first water row at 500 art rows and at most two at any height). Re-captured in one sitting (`l/caps.sh 3 4`, before from `pre-tree`, after on the fixed code, 03:41-03:46, AC, Low Power Mode 0, `cap-power.txt`) and re-measured with the review's scan (`l/an/horizon-sets.js`): harbour-golden y 228-229 brighter on 16 pixels at +4.4 (was 878 at +33.4), harbour-2015 on 21 at +1.5 (was 233 at +2.6); per-row runs (`l/an/rowruns.js`) show ±1-2 only, apart from one art pixel at x 542 (+28). In all sky rows (y < 230, `l/an/outline-lift.js`, `l/fix/outline-lift.txt`) the pixels brightened by > 6 next to a darkened pixel (an outline weakened by shading) fell from 1,762 to 8 on harbour-golden (two art pixels where the sky borders a shaded, brighter tower: x 542 and 678) and from 56 to 4 on harbour-2015; between the two before sittings 0 and 8. Not touched: a few hundred other sky pixels by day (harbour-day 124-204, styles-day 178-282, quarter-day 126-191 in sets 1/2 and 3/4; styles-day 156 between the before sittings, 0 on harbour-day) differ before → after by 7-25 summed RGB as single art pixels on the painted cloud bands' step edges, not beside any shading; the quad discards those rows, and the probable cause (not verified) is the backdrop's TIME-driven cloud drift (`backdrop.gdshader` `TIME * 0.010`) caught at a different moment by the after tree. Looked at: `l/fix/hg-horizon-left-b1-a1.png` (first pass) and `l/fix/hg-horizon-left-b3-a3.png` (fixed: the sky's dark bottom row is as before; the first water row stays pale above the shaded water), `l/fix/hg-horizon-b3-a3.png`. 59: - (2) **The core alpha was reported from a noisy measure** (rain counted as core, a fixed 0.09 cut). Restated with the noise-free measure above: water about 0.09, ground 0.11-0.13; the acceptance holds as the shader's alpha. 60: - Not re-run after the fix: the drift clips, the bench and the start-up report (the fix adds one compare-and-discard on the water's top row in the same shader; draw calls are unchanged, sets 3/4 against 1/2: +1 on the day views, identical at night). 61: - Review: filled by the orchestrator. 62: - **For the owner sheet:** every still shows the same pattern. At t 0 the Harbour island's left half and the quarter-day town are unshaded and the far water carries most of it; harbour-golden puts a core over the right lawn and the tall towers. The quad darkens everything under a patch the same: shaded faces, lamp pools and lit windows by day too (the lens's choice: no per-material sun mask). Shadows lie on the water as well. The drift is 1.5 m/s whatever the weather. Glade reads about 0.4× canal in pixels. Shaded water reads lighter than shaded ground (about 0.09 against 0.11-0.13). Under a shadow on the horizon water the first water row stays unshaded, a one-art-row pale line under the horizon (the price of keeping the sky's edge still; `l/fix/hg-horizon-left-b3-a3.png`). Where the sky borders a shaded tower that was brighter than it, the sky pixel beside it is lightened by the outline (two art pixels on harbour-golden). The 19 flickering stone pixels under Off are noted above. 63: - **Not done:** no Glade, quarter, perspective or photo-mode clip; no running-day-clock clip through the dusk fade (tests and the 20:15 still cover it); no night bench (nothing is drawn). No second review yet (the orchestrator runs it). Recording frames were moved to the Trash after measuring (clips, contact sheets and motion.tsv kept), so a re-scan needs new recordings. Windows cost unmeasured. 64: - Platforms: Mac run; Windows not run. 65: 66: **Beautify P8j Boats bob in whole art pixels; more fireflies and moths (7 October 2026):** Opus 5.5 on this Mac (plan section "P8j"; report polish backlog "Boats bob in whole art pixels at lane zooms" and "More fireflies and moths at night", with their lens findings; step `26-P8`, sub-stage P8-III). Boats afloat now heave, roll and pitch on the light clock in whole art pixels, so at lane zooms (and the quarter's own zoom) the hull steps up and down and rocks by single pixels, never crawling. At night the town gets up to 3 firefly swarms and 5 moth swarms instead of 2 and 2. Off keeps boats still and the night at 2 and 2; Reduced halves the bob and adds one moth swarm. Snapshot: `build/backup-feel/26-P8-j/` (= `main` f23d4f8; the live tree matched it at part start, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/j/pre-tree`. Evidence: `artifacts/beautify/26-P8/j/`; file backups in `j/backup/`. 67: - **Code:** 68: - `scripts/world_view.gd`: constants `BOB_HEAVE 0.035 m`, `BOB_ROLL 2°`, `BOB_PITCH 0.8°`, `BOB_RATES` = 2π × (373, 242, 198) / 1380 rad/s (1.698, 1.102, 0.902: whole cycles in light_time's wrap, so seamless), `BOB_RAIN (0.7, 0.6)`, `BOB_GLADE 0.6`, `BOB_TILT_ART 0.05`, `BOB_ROLL_ARM 1.55 m`, `BOB_PITCH_ARM 1.75 m`, `BOB_STEP 0.125 s`. Statics `bob_amount(rain, glade, motion)`, `bob_held(s, reach)` (the crowns' hold rule: top = round(reach), n held while |sin| ≥ n/(top+1)), `bob_steps(seed, t, amount, art, tilt)` (heave, roll, pitch in whole steps, the clock read at the start of each BOB_STEP, phases from `wind_hash(seed)`), `bob_transform(at, steps, art, up_y)` (heave straight up by n × art / the camera's up.y, so exactly n art pixels up the frame; roll and pitch by n × art / arm about the waterline; no step = `at` exactly). `bob_art(origin)` (= `crown_art`, so perspective uses the boat's depth). `_bob_boats()` runs in `_process` after `_lean_rain` and at the end of `_sync_decorations` (a rebuild never draws a frame at rest), writes a slot only when its pose changed (`bob_drawn`), skips boats in `deco_anim`/`deco_dying` (the build animation's) and boats a drag step moved (`bob_dragged`, set in `sync_live`'s decoration sync, cleared by `sync_scene`); amount 0 under `light_hold` (every still) and under Off. `_sync_decorations` lists afloat boats (kind boat, no parent, no land under it, not dying) in `bob_boats`. `sync_environment` sets `bob_rain` and `bob_glade`; `set_motion` and creature creation pass `motion_scale` to `Creatures.swarm_motion`. 69: - Design: boats already draw from their own small decoration MultiMeshes (key `boat__`, 1-3 instances each), so no new MultiMesh, no large buffer re-uploaded and **no draw-call change**. Lamps, emitters and reflection streaks stay put (the lens: a 1 px mismatch is fine). 70: - `scripts/creatures.gd`: `FIREFLY_MAX 3`, `MOTH_MAX 5` counted **outside MAX** (MAX 30 is now the cap for creatures with bodies, whose shares sum to 26); `FIREFLY_BASE 2`, `MOTH_BASE 2`; `swarm_caps(motion)` = base + floor(extra × motion) (Full 3/5, Reduced 2/3, Off 2/2, D17: only the new swarms are gated); `swarm_motion` member; the plan context and `signature()` carry the caps (a new setting re-plans within the view's 1 s check); `_admit` counts bodies against MAX and each swarm species (leavers included) against its own cap, so the fx multimesh (sized from the caps in `world_view.gd`) always has room. Site ranking is unchanged, so the 2 and 2 of before stay where they were and the new ones are added after them (the lens's "prefer lamps near water and plazas" is not applied). 71: - Deviation, found by measurement: with BOB_STEP 1/15 s (first form) the three independent motions changed one 15 fps frame apart, so a combined pose showed for a single frame (`rec-old/`, intermediate code). BOB_STEP 0.125 s puts every change on one 8 Hz grid; every combined pose holds at least one step. 72: - **Tests:** test-first logs `j/testfirst-test_creatures.log.out` (170 checks, 3 failures: no `bob_steps`/`_bob_boats`) and `j/testfirst-test_creature_reactions.log.out` (5,555 checks, 4 failures), no script errors. `test_creatures` 170 → 220 (`bob_twin_checks`: constants, rates whole cycles in 1380 s, amount ×(0.7+0.6 rain), Glade ×0.6, Reduced half, Off 0; at 60 fps over the whole wrap at 0.03 m and 0.02 m in full rain every component is whole, moves ≤ 1 px a frame, changes only at a step start and every pose holds ≥ one step; 1 and 2 px heave at the lane zoom; seamless; Harbour zoom 0.084 m: still at the default rain, 0/±1 (never 2) in full rain; roll and pitch only at ≤ 0.05 m; seeds out of step. `bob_projection_checks`: through a real orthographic camera a heave of n moves every hull point exactly n art px up, none sideways; a roll step swings the top about 1 px. `bob_view_checks`: afloat boat drawn at its pose, at once after a rebuild; beached untouched; Off and the automation hold exactly at rest; a building boat left to the build animation; a drag-moved boat at rest until the release, then bobbing; nothing saved). `test_creature_reactions` 5,555 → 5,583 (`swarm_checks`: quarter night 3 and 5 (target 3 and 4-5), none by day, Harbour gains swarms, Off 2/2, Reduced 2/3, the old swarms keep their sites, a setting change re-plans and the extra swarms leave, fx multimesh sized from the caps). Full suite on the final code (`j/suite/suite_all.out`, 02:20-02:28, AC 100%, Low Power Mode 0, `j/suite-power.txt`): **43 suites, 23,262 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Final code times 01:44-02:01 (`j/final-code-times.txt`), before the final clips, final stills, bench and suite. 73: - **Changed expectations (old → new):** `test_creature_reactions` budget: "The species shares fit in 30" (all six ≤ MAX) → "The species with bodies fit in 30" plus "FIREFLY_MAX 3 and MOTH_MAX 4-5, outside MAX"; "at most 30 creatures" per scene → at most 30 with bodies and the swarms within FIREFLY_MAX + MOTH_MAX; cap check while perches are reassigned: "never more than 30 creatures" → never more than 30 with bodies, and fireflies ≤ 3, moths ≤ 5 (leavers included); "each lamp's moths are spread apart (closest > 0.08 m at 0.35 s)" → closest ≤ 0.08 m in at most 16 of 160 instants over 4 s and mean > 0.2 m. The old snapshot failed on a new swarm (lamp:18, 0.065 m) because a lamp's moths turn both ways and cross: every swarm, the old two included, dips below 0.08 m in 1-6% of instants, mean 0.29-0.35 m (`j/probe/moths.txt`). `test_creatures` view check: "Within the 30 creature budget" (all spots) → creatures with bodies only (Harbour's default is night, so swarms are present). My own twin check moved from "every heave pose holds ≥ 0.2 s" (1/15 s steps, failed at 0.133 s) to the step-grid form above. 74: - **Captures** (`j/rec.sh`, `tools/art_record.sh --keep-frames`, 1440x900, 15 fps, light clock 6 → 16 s; before from `pre-tree`; final-code clips 02:01-02:06, AC, Low Power Mode 0, `rec/power.txt`; frames measured with `an/pairs.js` and `an/bob.js`, then moved to the Trash). Boats at a lane zoom (Harbour, camera 0.78,0.56,13.5,14,0.6,7.5, art 0.03 m, rain 0): with water (`bz-*`) and with the water switched off (`dz-*`, scene `j/probe/harbour-dry.json`, so the hulls stand on a still background); quarter-night with `--creatures`, rain 0 (`qc-*`). Looked at: `look/dz-boat-pre-live-frames.png` (frames 97, 98, 105, 114, before|after: the hull, cabin and lamp post step and tip by single pixels and read as one rigid boat), `look/bz-diffmap-150-step15.png` and `look/dz-diffmap-150-step15.png` (intermediate code: differences only on the three hulls and, with water, their reflections), `look/qc-pre-qc-live-120.png` (the quarter's boat bobs at its own zoom, 0.047 m; two new moth specks), `look/qc-moths-pre-live-120.png` (a new moth swarm at a lamp; small, a few pixels). 75: - **Acceptance, against the target:** 76: - **"Boats step 0 or 1 art px with no shimmer":** in the dry clip every frame-to-frame change of a hull falls on a twin pose change, and every twin change shows: boat 134 20 changes, boat 135 20, boat 136 22, each at the twin's frame or one frame earlier (`an/bob-dry-live-changes.txt` against `probe/twin_times.txt`), nothing changes between them. Holds median 6-8 frames (0.4-0.53 s); shortest 2 frames for boats 134 and 135, 1 frame once for boat 136 (two steps 0.125 s apart landing on adjacent 15 fps frames; 7-8 frames at 60 fps). Each motion moves ≤ 1 px per change (twin over the wrap; at 0.03 m and rain 0 heave, roll and pitch each reach ±1). Noise pairs `dz-live`/`dz-live2` and `bz-live`/`bz-live2`: 0 of 150 frames. 77: - **"Exact rest under the hold or Off":** `dz-off` vs `dz-pre` and `bz-off` vs `bz-off-pre`: 0 of 150 frames; stills below. Reduced (`dz-red`): 7, 11 and 9 pose changes against 20-22 at Full, holds median 15-28 frames, roll only (heave reach under half a pixel). 78: - **"3 firefly and 4-5 moth swarms on night quarter scenes":** quarter-night `--creatures` performance.json creatures 30 (pre, and live Off) → 34 (live Full) → 31 (Reduced); with 26 bodies that is 3 + 5 at Full; test: 3 and 5. `qc-off` vs `qc-off-pre`: 0 of 60 frames; noise pair `qc-pre`/`qc-pre2` 0 of 60. 79: - **"Drag and build of a boat unaffected":** in `bob_view_checks` (headless; no drag-harness film taken). 80: - **Stills** (`j/caps.sh` sets 1-2 at 01:47-01:50 on the intermediate code, `j/caps2.sh` sets 3-4 on the final code at 02:11-02:13, AC, Low Power Mode 0, `cap-power.txt`; `metrics/noise.txt`, `metrics/noise-table2.txt`): pixels changed > 8, final code, before-3 → after-3 / before-4 → after-4 against the noise pairs before-3/before-4 and after-3/after-4: harbour-day 1.01/0.79 against 0.88/0.90, harbour-night 0.66/1.32 against 0.64/1.34, lane-day 1.67/1.74 against 1.77/1.68, quarter-day 2.72/1.22 against 2.32/0.49, quarter-night 9.25/1.49 against 9.11/1.05 (before-3 is the odd take: before-4 → after-3 0.78), styles-day 0.21/0.33 against 0.71/0.46. Nothing moves at the hold. 81: - **Counts** (`metrics/counts-b1-a1.txt` to `counts-b4-a4.txt`): every integer key identical on every view, draw calls included. No new material, shader or surface (no start-up report taken). 82: - **Bench** (`j/bench.sh`, `bench-r01.out`, `bench-r23.out`, `bench-summary.txt` ("mid" is the pre tree), 02:14-02:19, AC, Low Power Mode 0, `bench/power.txt`; i2's form: interleaved, round 0 discarded, 1920x1080, vsync off, 01:00, rain 0.45, `--light-pin=19.5`): p50 dense 14.505 → 14.540 (paired +0.059, slower 2/3), Harbour 6.790 → 6.812 (+0.012, 2/3), quarter 7.165 → 7.174 (−0.004, 1/3); p95 dense +0.006, Harbour −0.847 (pre-tree spikes), quarter −0.094. Draw calls 2586 / 916 / 1043 both. Boats bobbing with the clock running (`j/bench-bob.sh`, `bench-bob-summary.txt`, Harbour at the lane zoom, rain 0, no pin, 02:19-02:20, same power): p50 5.193 → 5.207 (+0.006, 3/3), p95 5.978 → 6.067 (+0.079, 3/3), draw calls 578 both. `--churn` was not used: it only marks the lights dirty (`main.gd` automation_churn), never re-syncs decorations, so it does not measure the bob. 83: - **Expected journey count deltas:** none on stills (counts identical, boats at rest under the hold). A view captured with `--creatures` at night gains swarms (quarter-night `creatures` 30 → 34). 84: - Review: filled by the orchestrator. 85: - **For the owner sheet:** the bob reads at lane zooms and on the quarter (its 0.047 m zoom tips the boat too); at the default Harbour zoom boats stay still in dry and default rain and step 1 px only in heavy rain (the report's "about 0.5 px at the Harbour zoom", rounded). Roll is the most visible motion (the cabin top moves 1 px); a boat's pose changes about twice a second at Full. Swarms are added at the lamps and shores already ranked next, not re-ranked toward water and plazas. Moths are small (a few pixels) at quarter zoom. 86: - **Not done:** no Glade, perspective or photo-mode clip (the art at depth is `crown_art`, tested only through the twin); no drag-harness filmstrip of dragging or building a boat (headless view checks only); no full-rain boat clip (drops would swamp the diff; the twin covers 2 px). Recording frames were moved to the Trash after measuring (clips, contact sheets, motion.tsv and the analysis outputs kept); `rec-old/` holds the intermediate-code clips. 87: - Platforms: Mac run; Windows not run. 88: 89: **Beautify P8i2 One wind V: rain slant and water cat's-paws (7 October 2026):** Opus 5.5 on this Mac (plan section "P8i2"; report polish backlog "one wind": "Rain: the fall direction takes dir × 0.12 × (1 + g) ... Water cat's-paws: where g(wp.xz) is high, water.gdshader lowers the dash threshold from 0.86 to 0.78 and multiplies the mirror wobble by 1.8"; step `26-P8`, sub-stage P8-III, fifth part). The rain and the water now move in the one wind's gusts: while a gust crosses the view, new drops fall slanted (1 + g) times as far sideways, and each drop leans as it is drawn, its foot 1 px (default rain) or 2 px (heavy rain) downwind of its top as a one-pixel stair; ruffled patches (more ripple dashes, a harder-shivering mirror) ride the gust front across the water. Between gusts, at `light_time` 0 (every still) and under Off, rain and water are P8i1's; Reduced halves the slant and the ruffle and leans drops at most 1 px. Snapshot: `build/backup-feel/26-P8-i2/` (= `main` 93a1db9; the live tree matched it at part start, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/i2/pre-tree`. Evidence: `artifacts/beautify/26-P8/i2/`; file backups in `i2/backup/`. 90: - **Code:** 91: - `shaders/wind.gdshaderinc`: `wind_rain_pixels(xz, t)` (the crowns' reach and holds at a drop's ground point with no hashed lag, since a drop moves), `wind_rain_lean(xz, t, right)` (signed whole px along the camera's right, toward the downwind side, × how much of the wind crosses the frame), `wind_front(xz, t)` (a point's place in the moving gust front's frame, offset per gust; seamless over the 1380 s wrap). Earlier functions untouched. 92: - `shaders/rain.gdshader`: includes the wind; `drop_lean` and `drop_length` flat varyings; the quad widens by |lean| and the fragment keeps one pixel a row on the stair `floor((row + 0.5) × (span + 1) / L)`, discarding the rest; the side and the top follow the signs of `PROJECTION_MATRIX[0][0]` and `[1][1]`. Lean 0 draws P8d's drop exactly. 93: - `shaders/water.gdshader`: includes the wind; `PAW_DASH 0.78`, `PAW_WOBBLE 1.8`, `PAW_FROM 0.5`, `PAW_PATCH 6 m`, `PAW_EDGE (0.42, 0.62)`; `cat_paw(xz)` = smoothstep(PAW_FROM, 1, gust) × a patch (`sky_noise(wind_front / PAW_PATCH)`) × `motion_scale`, on `light_time`; the dash threshold is `mix(0.86, PAW_DASH, paw)` and the swell-and-row wobble × `mix(1, PAW_WOBBLE, paw)` (the rain ring's push is not scaled). 94: - `scripts/world_view.gd`: `rain_fall(g)` (RAIN_FALL with its sideways part × (1 + g)), `rain_gust(t)` (the gust at `view_camera.target` × `motion_scale`), `rain_lean_gust`, `_lean_rain()` (called each frame in `_process` after the clock and `_lean_smoke`; sets only `rain.direction` in the box's frame), `_place_rain` keeps the lean (both paths); twins `rain_pixels`, `rain_lean`, `paw_front`, `paw_noise`, `paw` and the `PAW_` constants. P8d's `rain_volume` (6 arguments), amount and box are unchanged. 95: - Deviation: the report's 0.12 is P8d's own slant |RAIN_FALL.xz| = 0.112 rounded; 0.112 is kept so the fall at no gust is P8d's exactly. The slant is one value per frame (the gust at the target); the drawn lean is per drop (the field at the drop). 96: - **Tests:** test-first log `i2/testfirst-test_water.log.out` (428 checks, 16 failures, no script errors). `test_water` 410 → 462 (`p8i2_rain_checks`: rain_fall at g 0, 0.5, 1; the gust is the wind's at the target, half under Reduced, none under Off, 0 at t 0; at a full gust's slant every frame ray at Harbour, quarter and lane zoom is still fed from inside P8d's box; drop lean whole px 0-2, none within 40 m at t 0, 1 px at the default rain, 1 and 2 in full rain, Reduced ≤ 1, Off 0, rides the front, seamless, every reach held ≥ 0.4 s; the Harbour foot leans to the frame's left; no sideways lean with the wind along the view. `p8i2_drop_checks`: shader text and the stair twin for lengths 6-8 and leans −2..2. `p8i2_paw_checks`: shader text and constants, twin: none below PAW_FROM, Reduced half, Off none, none within 40 m at t 0, patches ride the front and repeat over the wrap, about half of a gust's crest ruffles in patches metres across (`probe/paws.txt`: 48%, 48%, 56% of the crest, mean patch 9-15 m). `p8i2_view_checks`: a real view leans only `rain.direction` (amount, box and place unchanged), keeps it through `_place_rain`, Reduced half, Off P8d's). Full suite on the final code (`i2/suite/suite_all.out`, 01:15-01:24, AC 100%, Low Power Mode 0, `i2/suite-power.txt`): **43 suites, 23,175 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Final code times 00:07-01:05 (`i2/final-code-times.txt`), before the final stills, bench, start-up and suite. The first full run (`i2/suite-first/`, 00:56-01:04) failed P3b's "ring's push is added to the mirror wobble" because an intermediate form split the wobble into two statements; reverted to the single expression (its clips are pixel-identical to the final code's: `qn-live3`/`qn-final` and `hds-live`/`hds-final` 0 of 300 frames). 97: - **Changed expectations:** none. No existing assertion changed. 98: - **Captures** (`i2/rec.sh`, `tools/art_record.sh --keep-frames`, 1440x900, 15 fps, light clock 6 → 26 s, before from `pre-tree`; 00:09-01:06, AC, Low Power Mode 0, `rec/power.txt`; frames measured with `an/pairs.js` and `an/lean.js`, then moved to the Trash). Water clips dry: quarter-night (`qn-*`), harbour-night and harbour-day with chimney smoke off (`hns-*`, `hds-*`, plus Off and Reduced against the pre tree's Off and Reduced); rain clips at Harbour night, rain 1 (`hr1-*`, with Off) and the default rain (`hrd-*`). Looked at: `look/qn-canal-pre-live-285.png` (19 s: the canal's warm reflection columns break up harder inside the paw patch), `look/qn-diffmap-285.png` (the patch on the canal and a band on the far water), `look/hds-paw-pre-live-270.png` and `look/hds-diffmap-270.png` (by day the patches are scattered extra dashes; faint by eye in a still), `look/hr1-drops-pre-live-255.png` (17 s, full rain: the drops lean as "/" stairs, feet to the left, i.e. downwind, as the crowns step; some drops outside the front stay straight), `look/qn-still-lowerleft-b1-a1-a2.png` (the lower-left calm mirror in stills, before and two afters: the same apart from rain-ring phase), `look/hd-b3-a3-diff.png` (still differences are rain rings, drops and smoke only). 99: - **Acceptance, against the target:** 100: - **"Ruffled patches cross the water with the gusts":** the noise pairs `qn-live`/`qn-live2` and `hns-live`/`hns-live2` differ in 0 of 300 frames. Against the pre tree, `qn` differs in 167 of 300 frames, from 14.2 s (bottom right, upwind) to 24.3 s (left edge), the front crossing the canal; `hns` 208 of 300 (12.8-25.9 s), `hds` 207 of 300 (`an/out/pair-*.txt`). Before the front arrives, 7 `qn` frames (6.1-13.9 s) differ by 1 isolated art pixel each, mirror samples flipping at a nearest-texel boundary: the same flips appear under Off (`hds-off-pre`/`hds-off`: 10 of 300 frames, 1 px each, same pixels at 8.87 and 9.47 s as Full), so they come from the recompiled shader's rounding, not from the wind. 101: - **"The calm mirror streaks at the quarter-night lower left unchanged, proved by image diffs against the pre-part tree":** in the lower left (art 0-300, 300-449; `an/out/pair-qn-lowerleft.txt`) 81 frames differ only while the gust crosses it (17.2-22.5 s) and 9 more by 1 px each (the rounding flips above); noise pair 0 of 300. In stills (light_time 0) nothing ruffles: the counts are identical and every view changes within its noise pairs (below). 102: - **Rain slant (the plan's test):** in the twin and the view check above. In pictures (`an/lean.js`: drop pixels that continue down-left minus down-right, over all continuations, per second): `hr1-pre` −0.006 to −0.012 throughout; `hr1-live` rises from −0.010 (10 s) to +0.020 (17-18 s) and back to −0.003 (25 s); `hr1-off` matches `hr1-off-pre` within 0.002 every second. Default rain (`hrd`): +0.003 to +0.004 over 15-18 s (fewer drops, 1 px leans). 103: - **Reduced and Off:** Off: the 1-px flips only (above). Reduced: 85,575 changed pixel-frames against the pre tree's Reduced, against 150,500 at Full (0.57). 104: - **Stills** (`i2/caps.sh` sets 1-2 at 00:43-00:46 and `i2/caps2.sh` sets 3-4 on the final code at 01:07-01:10, AC, Low Power Mode 0, `cap-power.txt`; `metrics/noise-table.txt`, `metrics/noise-table2.txt`): pixels changed > 8, final code, before-3 → after-3 / before-4 → after-4 against the noise pairs before-3/before-4 and after-3/after-4: harbour-day 4.34/1.52 against 0.72/2.57 (after-3 is the odd take: before-4 → after-3 4.22, before-3 → after-4 1.66; its differences are rain rings, drops and smoke, `look/hd-b3-a3-diff.png`), harbour-night 1.40/1.97 against 2.68/2.16, lane-day 2.45/1.89 against 2.18/2.40, quarter-day 0.43/1.09 against 2.13/2.75, quarter-night 8.97/4.30 against 4.23/9.01, styles-day 0.62/1.75 against 2.83/1.54. 105: - **Counts** (`metrics/counts-b3-a3.txt`, `counts-b4-a4.txt`, and sets 1-2): every integer key identical on every view, draw calls included. No new material or surface. 106: - **Bench** (`i2/bench.sh`, final code, `bench-r01.out`, `bench-r23.out`, `bench-summary.txt` (its "mid" is the pre tree), 01:10-01:16, AC, Low Power Mode 0, `bench/power.txt`): pre tree against live, interleaved, round 0 discarded, 1920x1080, vsync off, 01:00, rain 0.45, `--light-pin=19.5` (the first gust over the middle: paws and leaning drops drawn). p50: dense 14.481 → 14.456 (paired −0.025, slower 1/3), Harbour 6.738 → 6.746 (+0.004, 2/3), quarter 7.099 → 7.118 (+0.037, 3/3). p95: dense −0.053, Harbour −1.264 (pre-tree spikes), quarter −0.056. Draw calls 2586 / 916 / 1043 both. An earlier bench on the intermediate form (`bench-form2/`) agrees (p50 −0.015, +0.027, +0.040). Start-up (`i2/startup.sh`, `startup.out`, final code, 01:24-01:29, same power; no new materials, but water, rain and the include changed): ready_ms 7113 → 7149, warm_ms 4960.0 → 5010.4 (paired +100.6, −14.4, +50.4), entries 141 and pieces 304 unchanged. 107: - **Expected journey count deltas:** none (counts identical; at light_time 0 nothing ruffles or leans). Pixels: none beyond noise. 108: - Review: filled by the orchestrator. 109: - **For the owner sheet:** by day the cat's-paws are scattered extra ripple dashes, faint by eye in a still and clearer in motion; on the night canal they break the reflection columns up visibly. The report's 0.78 and ×1.8 are used as given; a stronger ruffle is a one-constant change (PAW_DASH). Drops lean as one-pixel stairs, 1 px at the default rain and 2 px in heavy rain, only inside the front, so a frame mixes leaning and straight drops while a gust crosses it. The slant itself is one value for the frame (the gust at the target). Glade water ruffles the same (the report's Glade ×0.6 is not applied; the wind has no Glade factor since P8g1). 110: - **Not done:** no Glade, perspective or photo-mode clips (the photo-mode free-eye path takes the lean in code, not filmed). No 13:00 bench. No second review (not marked for this part). Recording frames moved to the Trash after measuring (clips, contact sheets and motion.tsv kept), so a re-scan needs new recordings. 111: - Platforms: Mac run; Windows not run. 112: 113: **Beautify P8i1 One wind IV: laundry and bunting, smoke gusts (6 October 2026):** Opus 5.5 on this Mac (plan section "P8i1"; report polish backlog "one wind": "pennants and garments swing about the cord ±1 px, doubled in a gust", "smoke: gravity = drift(rain) × (1 + 0.8 g at the town centre)"; step `26-P8`, sub-stage P8-III, fourth part). Laundry garments and bunting pennants are now cloth (P8h's material, reused): in a gust their hems and tips swing sideways by whole art pixels while the cord stays put, ±1 as the gust comes and goes and ±2 (doubled) at its height, in every weather. Chimney smoke leans further downwind while a gust crosses the plumes. Both rest between gusts, as the crowns and flags do, so nothing moves at `light_time` 0 (stills). Off stills both new terms; Reduced halves them (1 px, half the flap rate; half the lean). Snapshot: `build/backup-feel/26-P8-i1/` (= `main` a30b91d; the live tree matched it at part start, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/i1/pre-tree`. Evidence: `artifacts/beautify/26-P8/i1/`; file backups in `i1/backup/`. 114: - **Code:** 115: - `shaders/wind.gdshaderinc`: `wind_line_pixels(anchor, t)`: the crowns' front, lag and holds (n while the gust's shape is ≥ n/(top+1)), with top = round(`CROWN_PIXELS` × `motion_scale`) (≥ 1 while wind × motion > 0) instead of the crown's wind × strength. So each gust runs 0 → 1 → 2 → 1 → 0 at Full; under Reduced it runs 0 → 1 → 0. `wind_cloth_pixels(anchor, seed, t, mode)` takes the line's n for mode 2 and the crown's otherwise. `wind_cloth_shift` passes `floor(uv.y)`. Modes 0 and 1 (flags, valances) are unchanged. `wind_crown_pixels` is untouched (its 4-line cycle centre is repeated in the new function rather than refactored, so P8g1's text checks stand). 116: - `scripts/mesh_batch.gd`: `CLOTH_LINE = 2`; `cloth_face()` (`_append_face` in cloth, stamped; plain under `cloth_plain`). 117: - `scripts/architecture.gd`: `laundry_line`: each garment (and a shirt's shoulder box, same stamp) is `cloth_box(..., CLOTH_LINE, anchor = its peg point)`, with hoist and hem from `_hang_span` (the lowest top corner and the highest bottom corner of the turned box, so the whole top edge weighs 0 and the whole hem 1). The random draws keep their order. `bunting`: both faces of each pennant are `cloth_face` (hoist = the lower cord corner's height, straight down to the tip; anchor = the middle of its cord edge). Cords, posts and string lights are unchanged. Composition (sites, P7f's no-line-to-a-boxed-window rule) is untouched. 118: - `scripts/chimney_smoke.gd`: `GUST_LEAN = 0.8`; `drift(rain, gust = 0)` multiplies the sideways drift by 1 + 0.8 g (no gust: the old vector exactly); `centre` (the mean of the chimney mouths), `rain_amount` and `gust_now`; `lean(g)` resets `gravity` when g changes. Puff counts unchanged (24 a chimney, 480 at most; `probe/links-compare.txt`: 168/72/48/24 → same). 119: - `scripts/world_view.gd`: twin `line_pixels`; `cloth_pixels(..., mode = 0)`; `cloth_offset` passes the mode; `smoke_gust(t)` = `wind_gust(smoke.centre, t)` × `motion_scale`; `_lean_smoke()` is called in `_process` after the light clock steps. 120: - **Tests:** test-first logs `i1/testfirst-test_cloth.log.out` (70 checks, 4 failures), `testfirst-test_links.log.out` (190, 21), `testfirst-test_dusk.log.out` (450, 1); no script errors. `test_cloth` 65 → 84 (line shader text; twin: whole pixels, rest between gusts, reach ≥ the crown's, 2 px in dry/default/full rain, both 1 and 2 in each weather, every gust in the wrap shaped 1, 2, 1, 0 (none otherwise), Reduced 1, Off 0, seamless wrap, flap ±n both ways, modes 0/1 unchanged, rest at t 0 within 40 m; a real orthographic camera: hem exactly n px sideways, cord 0). `test_links` 150 → 325 (`cloth_link_checks`: laundry and bunting at 3 spans × 3 seeds against `cloth_plain`: same vertex count, plain before = plain + cloth, the same sorted vertex/normal/colour stream; mode 2, weights 0 at the cord and 1 at hem/tip, one anchor per garment or pennant; the lane's laundry and Harbour's bunting draw one cloth surface with UV2). `test_dusk` 448 → 461 (`smoke_gust_checks`: old drift at no gust, 1.8× sideways at a full gust and no higher, the gust is the wind's at the chimneys' middle, 0 at t 0, Reduced half, Off the old drift exactly, `_process` leans after the clock; puff counts). Full suite (`i1/suite/suite_all.out`, 23:41-23:49, AC 100%, Low Power Mode 0, `i1/suite-power.txt`): **43 suites, 23,123 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Final code times 22:32-22:35 (`i1/final-code-times.txt`), before every capture, bench and the suite. 121: - **Changed expectations:** none of the existing assertions changed. My own new check "rest at `light_time` 0" first covered ±60 m and failed on the new code: beyond about 43 m along the wind a gust is already blowing at t 0, as it is for the crowns. It now covers ±40 m. Every stock scene's cloth anchor is within 37.5 m, and 0 of them move at t 0 (`probe/links-compare.txt`). 122: - **Geometry unchanged apart from the material** (`i1/probe/links.gd`, pre-tree against live, `probe/links-compare.txt`): in Harbour, courtyard, styles, the quarter, lane.json and dense, the links mesh has the same vertex total, plain before = plain + cloth now, and an identical md5 of the sorted (position, normal, colour) stream. Surfaces go up by 1 (cloth). 123: - **Captures** (`i1/rec.sh`, `tools/art_record.sh --keep-frames`, 1440x900, 15 fps, dry unless named, before from `pre-tree`; 22:39-23:24, AC, Low Power Mode 0, `rec/power.txt`; frames measured, then moved to the Trash). Clips: Harbour bunting zoom, smoke off (`hb-ns-*`, camera 0.78,0.56,14,7,6.5,9: after, after2, before, off, off-pre, red, red-pre); Harbour smoke at that zoom (`hb-s-*`, `hb-ns26-*`) and at full rain in the whole Harbour view (`s1-*`, `ns1-*`); quarter laundry zoom (`q-*`, 0.845,0.315,8,2.4,5.8,2.8); styles laundry zoom (`st-*`, 0.55,0.68,9,0,4.5,-1, smoke off). I looked at all of these: `look/hb-bunting-right-b-a-285.png` (the pennants' tips sit 1-2 px aside; the cord stays), `look/q-laundry-before-after-280-282-284.png` (the pink and pale garments' hems step sideways, top edges fixed), `look/st-laundry-before-after-282-290.png` (the sheet's hem steps 2 px; the diff map holds only garment edges and their shadow), and `look/smoke-rain1-before-after-off-t23.png` (sparse puffs; the lean is not visible by eye at this zoom). 124: - **Acceptance, against the target:** 125: - **"Laundry and bunting swing ±1 px about the cord, doubled in gusts"** (`an/rows.js`: the best whole-pixel sideways shift of each changed row run against the pre-tree clip; `an/out/*-rows.txt`): bunting (`hb-ns-after` vs `before`) moves only from 16.33 to 22.93 s, with per-frame largest shift 1 in 33 frames and 2 in 61, both signs, explained median 0.74. Quarter laundry moves 17.53-22.00 s: 1 in 23 frames, 2 in 44, both signs, explained 0.84. Styles laundry moves 18.27-22.87 s: 1 in 18, 2 in 33 and 3 in 8. The 3s are 1-pixel runs near the top of a garment, where the sheared side edge flips a pixel and a ±3 fit happens to match (`DEBUG=3`; the twin caps at 2). The twin predicted these windows (`probe/links-compare.txt`: first moves 16.4-19.1 s Harbour, 17.5-18.7 quarter, 16.2-20 styles). Pose holds per piece (`an/holds.js`): median 4 frames at Full (CLOTH_STEP 3.75), 7 under Reduced; shortest 1, where two pennants' diffs touch or a 1 → 2 step lands inside a flap. This reading of "doubled" is mine: ±1 on a gust's shoulders and ±2 at its height, with rest between gusts. The report's constant ±1 breeze would move laundry in every still and at all times, against the crowns' and flags' approved rest. **For the owner sheet.** 126: - **"`test_links` geometry unchanged apart from the material":** above, and in the suite. 127: - **"Smoke leans harder in gusts":** through the view's own per-frame path (`probe/lean.gd`, `probe/lean.txt`) the Harbour emitter's sideways drift goes 0.16 → 0.271 m/s² dry (×1.69 at the first gust's peak, 19.5 s), 0.27 → 0.457 at Harbour's rain 0.22, 0.66 → 1.116 in full rain. The lean is on 17.6-24.6 and 46.6-53.5 s, its height (gravity y) is unchanged, Reduced gives ×1.35 and Off ×1.00. In pictures (`an/smoke.js`, full rain, smoke = the clip minus the same tree's smoke-off clip, 1-px-wide drops eroded; `an/out/smoke1-summary.txt`) the mean smoke position along the downwind screen direction moves +10.6 / +10.6 art px from the calm (8-16 s) to the gust and its lag (20-25 s) in two after takes, against +3.1 / +3.2 in two before takes and −1.2 under Off. The dry zoom clips were too noisy to read (the plumes differ take to take by ±10-20 px, `an/out/smoke-lean.txt`). 128: - **"Still under Off":** `hb-ns-off` against `hb-ns-off-pre` (the whole frame): 1 frame of 300 differs, by 3 isolated art pixels at 6.87 s, outside any gust (`an/out/hb-rows.txt`). `q-off` against `q-off-pre`: 0 of 300. The noise pairs `hb-ns-after`/`after2` and `q-after`/`after2` differ in 0 of 300 frames. 129: - **How the shear reads on sloped edges** (P8h review item): a garment's side edges lean by 1-2 px stairs while the hem steps whole. `an/cloth.js` called the quarter's sideways regions monotone in 25/29, 19/31, 19/31 and 16/32 moving frames, and the styles regions in 59/59, 30/30, 23/35 and 12/30. The pennants are 4-8 art px triangles: their sloped sides restep and read as a flutter (looked at). The P8h profile scanner explained 0 there (it shifts whole rows over a textured background), so the pennants were read with `rows.js`. 130: - **Stills** (`i1/caps.sh`; before-1/after-1/before-2/after-2, 23:27-23:31, AC, Low Power Mode 0, `cap-power.txt`; quarter-night, quarter-day, harbour-day, harbour-night, styles-day, lane-day; `metrics/noise-table.txt`): pixels changed > 8, before → after against the noise pairs. Harbour-day 0.80/1.44 against 1.00/1.93; harbour-night 0.77/3.86 against 1.11/4.21; lane-day 2.02/2.05 against 1.85/2.06; quarter-night 0.67/1.97 against 1.15/1.29; styles-day 0.79/2.51 against 0.23/1.29. Quarter-day 3.48/3.81 against 0.43/0.86, all in the canal water (`metrics/quarter-day-split.txt`: rows ≥ 780 11-12%, the buildings 1.3-1.5%; `look/qd-b1-a1.png`). The same live code captured from a scratch folder differs from the live captures by as much, 3.46% (`metrics/livetree.txt`), and P8h's before-3 → after-3 had 2.97%. So this is a per-folder water effect, not this part (no anchor moves at t 0 and cloth at rest draws as plain). 131: - **Counts** (`metrics/counts-b1-a1.txt`, `counts-b2-a2.txt`): draw_calls and rendered_objects +3 on every view with laundry or bunting: harbour-day/night 903 → 906, quarter-day/night 1020 → 1023, lane-day 745 → 748, styles-day 497 → 500 (the links mesh's cloth surface in the main and shadow passes). Nothing else changed. 132: - **Bench** (`i1/bench.sh`, `bench-r01.out`, `bench-r23.out`, `bench-summary.txt` (its "mid" is the pre tree), 23:31-23:37, AC, Low Power Mode 0, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, 1920x1080, vsync off, 01:00, rain 0.45, `--light-pin=19.0` (laundry, bunting and the Harbour smoke in a gust). p50: dense 14.554 → 14.525 (paired +0.014, slower 2/3), Harbour 6.752 → 6.757 (−0.072, 1/3), quarter 7.111 → 7.114 (+0.030, 2/3). p95: dense +0.036, Harbour −0.108, quarter +0.171 (2/3). Draw calls 2583 → 2586, 913 → 916, 1040 → 1043. Start-up (`i1/startup.sh`, `startup.out`, `bench/startup-summary.txt`, 23:37-23:41, same power; there are no new materials, but the include changed): ready_ms 7084 → 7110, warm_ms 4937.5 → 4984.2 (paired +74.6, +46.7, +24.0), entries 141 and pieces 304 both unchanged. 133: - **Expected journey count deltas:** draw_calls and rendered_objects +3 on harbour-night, harbour-day, harbour-lane (lane-day is its scene and camera; measured there), quarter-night and quarter-day; styles-day +3. Pixels: none beyond noise (everything rests at `light_time` 0), apart from the quarter-day canal folder effect above. 134: - Review: filled by the orchestrator. 135: - **Orchestrator fix pass (7 October 2026):** the reach n of a line piece (and a flag) is now read at the start of its own flap step, not at t, so a 1 → 2 step no longer lands inside a flap. Quarter laundry hold_min 1 → 3 frames at Full ({1:3, 3:2} → {3:5}), 3 → 7 under Reduced; Harbour bunting 1-frame pixel holds 403 → 13, styles 102 → 35 (the rest is not explained). Details, tests and numbers in the P8k entry above (`artifacts/beautify/26-P8/k/fixpass/`). 136: - **For the owner sheet:** the reading of "doubled in gusts" above (no breeze swing between gusts); laundry and bunting reach 2 px in every gust while flags reach 1 px dry (laundry is the lighter cloth); the pennant flutter is a 1-2 px restep of 4-8 px triangles; the smoke lean is real but subtle (about 10 art px of mean plume position in full rain at Harbour zoom, not visible by eye in a still). 137: - **Not done:** no Glade, perspective or photo-mode clips (the twin covers the depth step, as P8h's). No 13:00 bench. No Reduced picture of the smoke (the probe covers it). The dry smoke clips do not show the lean above take-to-take noise. Recording frames were moved to the Trash after measuring, so a re-scan needs new recordings (clips, contact sheets and motion.tsv are kept). 138: - Platforms: Mac run; Windows not run. 139: 140: **Beautify P8h One wind III: cloth material, flags and pennants (6 October 2026):** Opus 5.5 on this Mac (plan section "P8h"; report polish backlog "one wind" (cloth): "flag and laundry free-edge shear"; step `26-P8`, sub-stage P8-III, third part). Flags, the canal tower's pennant and awning valances are now cloth: when a gust crosses the island (the same front, strength and holds that step the tree crowns), each one's free edge flaps by whole art pixels, 0, +n, 0, −n in 0.25 s steps, n = the crown's 1 px (dry and default rain) or 2 px (full rain), while the hoist (pole or rail) stays put, so the piece shears. Between gusts it rests, like the crowns. Off stills it; Reduced halves the wind (the crown rule, so 1 px) and the flap rate (0.5 s steps). Snapshot: `build/backup-feel/26-P8-h/` (= `main` 841006b; the live tree matched it at part start, `diff -rq` on scripts shaders tests data scenes tools empty, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/h/pre-tree`. Evidence: `artifacts/beautify/26-P8/h/`. 141: - **Code:** 142: - `shaders/cloth.gdshader`, `shaders/cloth_build.gdshader` (new): the plain body with `#define CLOTH` (the build variant adds `BUILD_REVEAL`). At rest cloth draws exactly as plain. 143: - `shaders/plain_body.gdshaderinc`: under `#ifdef CLOTH`, after `wp` and before `grown_hidden`, each vertex takes `wind_cloth_shift(anchor, UV, wp, mat3(MODEL_MATRIX), MAIN_CAM_INV_VIEW_MATRIX)`, the anchor read from UV2 through `MODEL_MATRIX` (a free-standing flag is a decoration MultiMesh instance). plain and plain_build never define CLOTH. 144: - `shaders/wind.gdshaderinc` (reusable for P8i1): `CLOTH_STEP = 0.25`, `wind_cloth_phase` (stamp fraction + anchor hash, so MultiMesh flags do not flap in unison), `wind_cloth_pixels` (n = `wind_crown_pixels(anchor)`, flap index on `light_time × motion_scale / CLOTH_STEP`; 4 steps divide the 1380 s wrap at Full and Reduced), `wind_cloth_axis`, `wind_cloth_shift` (weight × n × `wind_art_pixel` at the vertex's own depth). The axis is one screen axis, so the step is whole pixels: a valance swings along the camera's right; a flag flaps across its own length as the frame shows it (up and down when it lies across the view, sideways when it points toward the eye; a flap along its length would only stretch it). 145: - `scripts/mesh_batch.gd`: `CLOTH`, `CLOTH_FLAG`, `CLOTH_HANG`; `cloth_box()` and `stamp_cloth()` write UV = (free-end weight 0..1, mode + fraction: a flag's yaw, a valance's seed), UV2 = the anchor's ground point in the batch's outer space; unstamped cloth gets weight 0 (still) and an anchor; `build()` writes UV2 on the cloth surface. `cloth_plain` (default off) draws cloth as the plain boxes of before P8h, for fingerprint tests. 146: - `scripts/architecture.gd`: `materials()` registers `cloth`; the flag prop's two cloth boxes, the canal tower's pennant (`_roof_trim`) and the awning's ten valance strips (one anchor, so the strip moves as one) use `cloth_box`. Positions, sizes and colours unchanged. `scripts/build_anim.gd`: `cloth` in `MATERIALS` (so `build_materials`, the build clock and the warm-up take it). 147: - `scripts/world_view.gd`: the twin `CLOTH_STEP`, `cloth_phase`, `cloth_pixels`, `cloth_axis`, `cloth_offset`. 148: - Not cloth: bunting and laundry (P8i1), umbrellas. Glade: cloth moves whole art pixels there too (no unsnapped path). 149: - **Tests:** test-first logs `h/testfirst-test_cloth.log.out` (18 checks, 15 failures), `testfirst-test_build_anim.log.out` (1), `testfirst-test_warmup.log.out` (3), no script errors. New suite `tests/test_cloth.gd` (added to `tools/test.sh` and `tools/Test.ps1`): 65 checks (shader text and order, registration, twin: whole pixels ≤2, tied to the crown's n, rest between gusts, both signs, ±2 full rain / ±1 default and dry, step holds 0.25 s Full / 0.5 s Reduced, Off still, seamless wrap, neighbours' phases differ; offsets exactly n px on one axis through a real orthographic camera, hoist 0, half-way half, perspective at the vertex's depth, instance turn; geometry: flag 72 cloth + 36 plain + 252 painted metal = 360 as before, awning 360 + 360 = 720 as before, weights and anchors, canal pennant 36 vertices in its cluster, medieval finial not cloth, Harbour cloth surfaces carry UV2). `test_build_anim` 444 → 460 (`p8h_cloth_reveal_checks`: a new canal tower's pennant keyed and revealed with CUSTOM0, wears `cloth_build`, hands back, sinks and is freed on removal), `test_warmup` 37 → 40. Full suite (`h/suite/suite_all.out`, 21:49-21:57, AC 100%, Low Power Mode 0, `h/suite-power.txt`): **43 suites, 22,916 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). The first full run (`h/suite-first/`) failed test_dusk's P7b fingerprint (below), fixed before the final run. 150: - **Changed expectations:** `test_build_anim` "A2b: ... wrapper around its body": cloth's body is `plain_body.gdshaderinc` (old: `_body.gdshaderinc` for every animated material but pane_glass). `test_dusk` P7b "every stock building cluster ... draws the same geometry as before P7b": the clusters are now built with `cloth_plain = true` (old: as drawn), because the canal pennant moved from plain to cloth and changed 7 of 297 fingerprints (courtyard, castle, artwork, towers ×3, audit). 151: - **Vertex counts** (`h/probe/counts.gd`, pre-tree against live, `probe/counts-compare.txt`): every scene's cluster and decoration vertex totals identical (Harbour, courtyard, castle, styles, quarter); plain + cloth = the old plain. 152: - **Draw calls** (`h/metrics/counts-b3-a3.txt`, bench): Harbour 897 → 903 (+6: the flag and awning decoration meshes each gain a cloth surface), quarter 1011 → 1020 night/day, 988 → 997 dry (+9: the tower cluster and three awning meshes), styles-day and dense identical (no cloth). 153: - **Stills** (`h/caps.sh`; before-1/after-1/before-2/after-2 at 21:10, before-3/after-3 on the final code; quarter-night, quarter-day, quarter-dry, harbour-day, harbour-night, styles-day; `metrics/noise.txt`): at `light_time` 0 every cloth piece in these scenes rests (`probe/t0.txt`), so stills only move by noise: before-3 → after-3 > 8 change 1.35-3.29% against pair noise 0-2.5% (before) and 1.3-11.3% (after); the pennant and awning crops are identical at rest (`look/quarter-dry-b3-a3-pennant.png`, `look/qd-b1-a1-right.png`, looked at; the only red is a cyan sign and water). 154: - **Captures** (`h/rec.sh`, `tools/art_record.sh --keep-frames`, 1440x900, 15 fps, light clock 6 → 26 s, dry unless named; before from `pre-tree`; 21:16-21:31, AC, Low Power Mode 0, `rec/power.txt`; frames kept, 3.1 GB): quarter (`qd-*`), pennant zoom (`pz-*`, camera 0.845,0.315,8,5.3,13.5,9.6), Harbour Fish Market zoom (`hz-*`, 0.78,0.56,12,4.5,4.5,11), Off, Reduced, full rain. Looked at: `look/pz-flap-rest-p1-m1-p2-m2.png` (pennant at rest, +1, −1, +2, −2: the free end drops and lifts, the hoist stays), `look/hz-flag-282.png` (Fish Market flag sheared 1 px; its shadow moves too), `look/hz-awning-296.png` (valance's lower edge steps 1 px sideways), `look/qd-mid-awning.png`, `look/qd-273-left.png`, `look/film-after-strip.png` (A2 reveal). 155: - **Acceptance, against the target:** 156: - **"Flags shear ±1-2 px with gusts"** (`an/edge.js`, silhouette edges of the pennant against the pre-tree clip, `an/out/edge-pz-*.txt`): dry free edge −1/0/+1, rain 1 −2..+2, hoist 0 in every frame; moves only 18.2-21.3 s (dry) / 17.9-21.9 s (rain 1) of 6-26 s; poses held 3-4 frames (0.25 s). Harbour (`an/cloth.js`, `an/out/cloth-hz.txt`): the flag and its shadow move 18.3-21.7 s, the awning 19.3-22.7 s, as the twin predicts (`probe/t0.txt`: 18.5-21.7, 19.7-22.8); quarter (`cloth-qd.txt`): pennant 18.2-21.3 s, max shift 1, both signs; three awnings 14.5-17.7, 17.0-20.5, 18.2-21.7 s. The noise pair `qd-after`/`qd-after2` differs in 0 of 300 frames, and `qd-before` against `qd-after` only in frames 218-325 (14.5-21.7 s, the gusts) (`an/out/pairs.txt`). 157: - **"Still under Off"**: `pz-off` against `pz-before` 0 moving frames; in the four quarter cloth boxes `qd-off` changes in 3, 69, 0 and 0 frame pairs (stars, water and a sign behind them) against 3, 69, 2 and 0 in `qd-before` (cloth static) and 17, 84, 18 and 16 in `qd-after` (`an/out/still-qd.txt`). Reduced: ±1, holds 7-8 frames (0.5 s), 12 pose changes against 18. 158: - **"Draw-call delta reported"**: above. 159: - **A2 reveal and removal on cloth**: `film/film.sh` (quarter, select tower, delete, undo filmed with `--animate`): `look/film-after-strip.png` shows the tower rising and the pennant revealed by the sweep at 1,200 ms, whole at 1,600 ms; removal held by `test_build_anim`. 160: - **Bench** (`h/bench.sh`, `bench-r01.out`, `bench-r23.out`, `bench-summary.txt`, 21:35-21:41, AC, Low Power Mode 0, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, 1920x1080, vsync off, 01:00, rain 0.45, `--light-pin` with the cloth in a gust (Harbour and dense 20.3 s, quarter 19.5 s). p50: dense 14.500 → 14.496 (−0.014, 1/3 slower), Harbour 6.726 → 6.698 (+0.029 paired, 2/3), quarter 7.085 → 7.121 (+0.036, 3/3); p95 dense +0.124 (3/3), Harbour −0.040, quarter +0.063. Draw calls 2583 / 907 → 913 / 1031 → 1040. 161: - Start-up (`--startup-report`, `h/startup.sh`, `h/startup.out`, `bench/startup-summary.txt`; ordinary Harbour launch 1600x1000, pre-tree against live, interleaved, round 0 discarded, medians of rounds 1-3; 22:08-22:13, AC, Low Power Mode 0, `bench/power.txt`): ready_ms 6801 → 7013, warm_ms 4647.4 → 4879.3, all_ms 10992.5 → 11480.7; entries 137 → 141, rounds 138 → 142, pieces 295 → 304; idle_worst_ms 134.9 → 133.5. Paired warm_ms +217.1, +212.5, +235.4 (3/3 slower): cloth and cloth_build are four more warm entries, about 54 ms each (P7f: one entry, about +179 ms). The plan sets no start-up budget; recorded, not a stop. 162: - **Review fix (6 October 2026):** the review found the plan's risk table (`docs/beautify-plan.md` "Start-up warm-up grows ... new materials", P8h) asked for a `--startup-report` before and after, and it had not been run or listed as not done. Fixed: run and recorded above. No code changed in the fix. Because `scripts/mesh_batch.gd` carries a timestamp (21:59) later than the first final suite (21:57), the full suite was run again: `h/suite2/suite_all.out` (22:14-22:22, AC 100%, Low Power Mode 0, `h/suite2-power.txt`), **43 suites, 22,916 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549/0. 163: - **Expected journey count deltas:** draw_calls and rendered_objects +6 on the Harbour views with the Fish Market flag and awning in frame (measured on harbour-day and harbour-night), +9 on quarter-night and quarter-day; styles-day none. Pixels: none beyond noise (cloth rests at `light_time` 0). 164: - Review: filled by the orchestrator. 165: - **Orchestrator fix pass (7 October 2026):** flags, the pennant and valances read the crown's reach at the start of their own flap step (`wind_cloth_pixels` t0, twin `cloth_step_start`), so every pose holds a whole step. The rain-1 flag's moving pose runs went from 1 1 2 4 3 4 … to 3 4 3 4 … frames. Expectation changed: "its reach is the crown's n" and "Reduced never further than Full" now compare at the step's start. Details in the P8k entry above. 166: - **For the owner sheet:** cloth rests between gusts (the report's "±1 px, doubled in a gust" would flap all the time; this follows the crowns' approved rest); the flap is a 4-step pixel flap at 4 steps a second in gusts; a flag seen end-on flaps sideways; the valance swing is subtle (1 px on a 2-4 px strip). 167: - **Not done:** bunting, laundry and smoke (P8i1). No Glade, perspective or photo-mode recording (the twin and test cover the depth step). No 13:00 bench. Recording frames kept, not moved to the Trash. 168: - Platforms: Mac run; Windows not run. 169: 170: **Beautify P8g2 One wind II: bending trunk ring and limbs follow (6 October 2026):** Opus 5.5 on this Mac (plan section "P8g2"; report "rigid whole-pixel crown shifts with a bending trunk ring"; step `26-P8`, sub-stage P8-III, second part; D29 kept the P8g1 shift, so the part runs). A tree's trunk and limbs now follow its crown in the wind: when a gust shifts a crown by its whole art pixels, the limbs inside the crown and the top of the trunk move with it by the same pixels, and the bare trunk below bends (its 1-2 px jog sits in the lower trunk), so the crown no longer slides off its trunk. Same gust timing as P8g1 (the same `wind_crown_shift` call). Snapshot: `build/backup-feel/26-P8-g2/` (= `main` ceeff35, P8g1 as committed; the live tree matched it at part start, nothing unexpected). Pre-part scratch tree `artifacts/beautify/26-P8/g2/pre-tree`. Evidence: `artifacts/beautify/26-P8/g2/`. 171: - **Code:** 172: - `shaders/wind.gdshaderinc`: `wind_bend(y, bend)` = smoothstep(from, to, y) over the span a surface carries as CUSTOM1 = (from, to, 1, 0); −1 with no span. A function of height alone, so pieces that meet move together. 173: - `shaders/plain_body.gdshaderinc`: includes the wind; `uniform bool crown_bend = false` and `uniform float banding = 1.0` (forwarded from the leaf's). With `crown_bend && !grown`, each vertex takes `wind_crown_shift(MODEL_MATRIX[3].xyz, MAIN_CAM_INV_VIEW_MATRIX, banding > 0.5)` × `wind_bend(rest.y, CUSTOM1)`, after `wp`, before the hidden-fullness collapse. `rest` is the vertex's place on the grown tree before the sapling swell (`grown_rest`, new in `shaders/grown.gdshaderinc`; `grown_morph` unchanged), so a growing tree (T4b) bends by the same weights. Every other plain surface draws as before. 174: - `shaders/foliage_body.gdshaderinc` (crown block only): a crown whose (non-grown) surface carries a span takes `wind_bend` too (host trees); every other crown keeps weight 1 (rigid, as P8g1). FoliageLamps, the `COLOR.a` stamp and the P6 lamp term untouched; no `*ALBEDO` in `light()`. 175: - `scripts/world_view.gd`: `bend_materials` (settled and growth copies of plain, `crown_bend` on; `bend_copies`), `bend_weight` (twin), `crown_joint` (lowest leaf vertex or T4b leaf anchor), `BEND_SHARE = 0.6`. `_build_grown_mesh` gives tree groups the copies and the span (0, 0.6 × joint, 1, 0); `_build_host_mesh` gives hosts the settled copy and writes the host span on the plain and leaf surfaces; `set_material_parameter` forwards plain uniforms and the leaf's `banding` to the copies. 176: - `scripts/architecture.gd` `grown_mesh(..., bend)`: writes the span as CUSTOM1 on the common plain surface (limbs carry no hue); without it, output is unchanged. `scripts/mesh_batch.gd` `build(materials, custom, custom1)`. 177: - `scripts/tree_house.gd`: `BEND_SPAN = 1.0`; `draw_host` reports `bend` = (top of everything that carries or is a house: boughs, prongs, decks, collars and each hosted house's box; max(that + 1 m, lowest leaf)). A host tree is still up to its house and bends above it, leaves and bark by one weight of height, so the house and its boughs never part from the tree. 178: - First form (span to the joint itself) was recorded and set aside (`g2/rec-v1/`, not used for numbers): its upper trunk sat at fractional offsets under the crown; `BEND_SHARE` 0.6 makes the trunk move whole pixels where it goes into the crown. 179: - **Tests:** test-first logs `g2/testfirst-test_foliage.log` (9 failures), `testfirst-test_tree_growth.log` (1), `testfirst-test_tree_house.log` (3), `testfirst-test_foliage_merge.log` (3). Counts: `test_foliage` 437 → 465 (`bend_checks`: shader text and order, the twin, Harbour and grove tree meshes all carry (0, 0.6 × joint, 1, 0) with every leaf and anchor at weight 1, uniform forwarding), `test_tree_growth` 256 → 259 (joints of tree/blossom/conifer at 4.2 m and 10-14 m; `grown_mesh` span beside the morph), `test_tree_house` 161 → 167 (host span above deck, collars and house box; weight 0 on every carrying vertex), `test_foliage_merge` 84 → 87. Full suite (`g2/suite.sh`, `g2/suite-final/suite_all.out`, 20:33-20:41, AC, Low Power Mode 0): **42 suites, 22,832 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 5,549/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). 180: - **Changed expectations:** `test_foliage` P8g1 check: old `...VERTEX+=inverse(mat3(MODEL_MATRIX))*crown;` → new `...*crown*bend;`. `test_foliage_merge` "grown surfaces, and only they, wear the grown materials": a CUSTOM1 surface may now also wear the trees' bend copy (a tree's limbs carry their span as CUSTOM1); new check that they do. 181: - **Captures** (`g2/rec.sh`, `tools/art_record.sh --keep-frames`, 1440x900, 15 fps, dry `--rain=0`, light clock 6 → 28 s (30 s for the house); before from `pre-tree`, after live, Off live; 19:37-20:11, AC 100%, Low Power Mode 0, `g2/rec/power.txt`; frames kept, 6.3 GB): 182: - Noise pairs (`g2/an/pair.js`, `an/pairs.txt`): `g22-after` vs `g22-after-2` 0 of 330 frames differ; `g22-off` (live) vs `g22-off-pre` 0 of 330 (so at rest the new code draws bit-identically). After vs Off differs only in frames 256-350 (the gust, light 17.06-23.33 s), exactly as before vs Off. 183: - Looked at: `look/g22-b9-after-off-before-285.png` (blossom: forks and trunk step with the crown after; before, the crown slides over still forks), `look/house-centre-after-off-before-300.png` and `house-ul-...` (host crown and trunk above the house move, the house stays; the red line on the gable is the crown's moving shadow, as before), `look/grow-crop-after-off-before-285.png` (`--tree-grow=0.5`: the sapling's trunk bends, nothing tears), `look/glade-crop-after-off-before-290.png` (Glade unsnapped, trunks follow), `look/persp-crop-after-off-285.png` (perspective, trunks follow). 184: - **Acceptance: "no 1-2 px gap at the trunk joint at three zooms"** (`g2/an/joint.js` v4, `an/joint-g{12,22,38}-{before,after}.txt/.json`, joints from `probe/joints.gd` → `probe/joints.json`): for each tree and each frame the twin says it is held shifted (≥ 2 frames from a step), the crown's best offset against Off and whether the bare trunk band (0.6 × joint to joint − 0.45 m, bark pixels) matches Off at no offset or at the crown's offset. Gap = crown offset − trunk offset. 185: 186: | Zoom | After: held frames with gap 0 | Before (P8g1): gap 1 / gap 0 / tie | 187: |---|---|---| 188: | 12 m | **243 / 243** (5 trees) | 194 / 49 / 0 | 189: | 22 m | **195 / 195** (4 trees) | 146 / 49 / 0 | 190: | 38 m | **195 / 195** (4 trees) | 98 / 0 / 97 (3-px trunks score equal) | 191: 192: Before's gap-0 frames are one tree (0,8) whose band shows 7-21 bark pixels crossed by its crown's shadow edge. Two trees behind evergreens and the small evergreens (band under 2 rows) are not measurable and skipped. Hosts, growth, Glade and perspective are judged by the looks above, not this scanner. 193: - **Stills** (`g2/before-1`, `after`, `before-2`, `after-2`; harbour-day, harbour-night, quarter-day, grove-day, glade-day; `metrics/`): integer counts, draw calls included, identical in every pair. Pixels changed > 8, before-1→after against the noise pairs (before-1→before-2 / after→after-2): grove-day 0.001% (0.001 / 0.001), glade-day 0.386 (0.449 / 0.347), harbour-day 6.05 (5.18 / 6.69), harbour-night 0.73 (1.06 / 3.07), quarter-day 0.37 (0.33 / 0.57): within noise (no crown is in a gust at `light_time` 0). 194: - **Bench** (`g2/bench.sh`, `bench-r0.out`, `bench-r123.out`, `bench-summary.txt` (its "mid" column is the pre tree), 20:23-20:30, AC, Low Power Mode 0, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, 1920x1080, vsync off, 01:00, rain 0.45, `--light-pin` at the busiest gust. p50: dense 14.444 → 14.507 (+0.066 ms, slower 3/3), Harbour 6.632 → 6.661 (+0.045, 3/3), forest 5.911 → 5.959 (+0.047, 3/3); p95 dense +0.089, Harbour −1.18 and forest −0.06 (scatter). Draw calls identical (2583 / 907 / 682). Not benched at 13:00 or under the perspective lens. 195: - **Expected journey count deltas:** none (stills unchanged; counts identical). 196: - Review: filled by the orchestrator. 197: - **For the owner sheet:** a host tree's crown now moves only above its house (below the house top it is still, rising to whole over 1 m), where P8g1 moved the whole host crown; the trunk's jog lands in the lower 60% of the bare trunk. 198: - **Not done:** no numeric joint scan for hosts, growth, Glade or perspective (the scanner assumes the orthographic camera and whole-pixel holds; judged from the looks). No Harbour or lane clips (P8g1 showed water and drops hide the crowns there; the grove at three zooms is the seam measure). Recording frames are kept, not moved to the Trash. 199: - Platforms: Mac run; Windows not run. 200: 201: **Beautify P8g1 One wind I: wind field and whole-pixel crown shift (6 October 2026):** Opus 5.5 on this Mac (plan section "P8g1"; report polish backlog "one wind" (foliage); step `26-P8`, sub-stage P8-III, first part). The town now has one wind field, and each tree's crown moves rigidly by 1 or 2 whole art pixels along the camera's right vector each time a gust crosses the island, every 20-40 s. Shrubs, ivy and cluster foliage keep their old sway. Under the perspective lens and in photo mode the step is whole pixels at each crown's own depth (second fix pass). The part had two orchestrator-directed fix passes (below). Snapshots: `build/backup-feel/26-P8-g1/` (= P8-II close code, the before for every number here) and `build/backup-feel/26-P8-g1b/` (the part before the first fix pass). The second fix pass backed up the files it changed in `g1/fix2/backup/` and built `g1/fix2/mid-tree` (the code before it) for same-sitting A/Bs. Evidence: `artifacts/beautify/26-P8/g1/`; the fix passes are in `g1/fix/` and `g1/fix2/`. 202: - **Code:** 203: - `shaders/wind.gdshaderinc` (new; P8h, P8i1, P8i2 and P8l will share it): the gust field on `light_time`. It has one gust per 30 s cycle, centred within ±5 s of the cycle's middle, so gusts are 20-40 s apart. Each gust rises over 2 s, falls over 5 s and has a hashed strength of 0.55-1. A front travels downwind at 6 m/s, so the phase comes from `MODEL_MATRIX[3]`, plus a hashed lag of ±0.6 s per tree. The wind blows along `RAIN_FALL` on the ground. There are 46 cycles × 30 s = 1380 s = `light_time`'s wrap, so the field has no seam. 204: - Crown shift on the banded path: `top = clamp(round(2 × wind × motion_scale × strength), 1, 2)` pixels while `wind × motion_scale` is above 0, else 0 (the 1-pixel floor is the fix pass's). A crown holds n pixels while the gust's shape is at least n/(top+1), so a step lasts at least about 3.5 s and never blinks. Glade is unsnapped at 0.04-0.08 m along the wind. It reads `motion_scale`, so Off stills it (D17: a new term); Reduced halves the strength, which with the floor means 1 px where Full would move 2 (full rain), and the same 1 px otherwise. 205: - `project.godot`: a new global, `wind_strength`. 206: - `scripts/world_view.gd`: 207: - `wind_strength = wind_of(rain) = 0.3 + 0.7 × rain`, written by `sync_environment` and never by the zoom. 208: - CPU twins: `wind_gust`, `crown_gust`, `crown_pixels`, `crown_offset`. 209: - `crown_materials` builds the trees' own copies of the settled, grown and growth leaf, with `crown_shift` on and `sway` 0. 210: - `_build_grown_mesh` gives them only to `TreeGrowth.KINDS` groups; `_build_host_mesh` gives the settled copy to host trees (tree houses; fix pass, one line). 211: - `set_material_parameter` forwards leaf uniforms to the copies. 212: - Test-only `--light-pin=` (`light_pin_for`, `light_pin`, read in `scripts/main.gd`; automated runs only, wrapped at 1380 s): pins the running light clock for a bench, night life stays live. Not a shipped option. 213: - Second fix pass: a fifth global, `art_pixel_slope` (`project.godot`, default 0): 2 tan(fov/2) / rows of the camera the frame is drawn with under the perspective lens, 0 under the orthographic camera, written by `sync_motion_globals` from `camera.fov`, so photo mode's own lens and free camera are followed. `wind_art_pixel(origin, cam)` in the include (twin `WorldView.crown_art`) is `art_pixel` when the slope is 0, else slope × the crown origin's depth in front of the main camera. `wind_crown_shift` now takes the main camera's whole inverse view matrix (`MAIN_CAM_INV_VIEW_MATRIX`; right vector, eye, forward) instead of its right vector. The orthographic path is unchanged pixel for pixel (below). 214: - `shaders/foliage_body.gdshaderinc` (included by `foliage.gdshader`; the plan named `foliage.gdshader` :15/:45): `uniform bool crown_shift = false`. The shift is applied after `wp` (so the leaf pattern, glints and P6 lamp term ride with the crown) and before `grown_hidden`. FoliageLamps, the `COLOR.a` stamp and the lamp list are untouched. No `*ALBEDO` in `light()`. 215: - **Deviation for the owner (D17):** tree crowns lose the old `TIME` vertex sway (0.018 m ≈ 0.24 art px at Harbour, which crawled sub-pixel every frame), at Full and Reduced as well as Off, Glade included. So under Off, trees are now fully still, where before they still swayed; at Full they rest between gusts. Shrubs, ivy and clusters keep the old ungated sway. The orchestrator accepted this (the old sway is the crawl D29 exists to remove) and puts it on the owner sheet. The Glade check is in the fix pass below. 216: - **Tests:** test-first logs `g1-testfirst-test_foliage.log` (6 failures plus a script error) and `g1-testfirst-test_foliage_merge.log` (3 failures); fix pass `fix/testfirst-test_foliage.log` (3 failures plus a script error on the missing `light_pin_for`) and `fix/testfirst-test_tree_house.log` (1 failure, taken with the host line reverted: the line was written before its test). 217: - Second fix pass, test first: `fix2/testfirst-test_foliage.log` (7 failures, no script error) and `fix2/testfirst-test_camera.log` (4 failures); after: `fix2/after-test_foliage.log` 437/0, `fix2/after-test_camera.log` 260/0. 218: - `test_foliage` 348 → 410 (first sitting) → 421 (fix pass, `wind_floor_checks`) → 437 (second fix pass, `perspective_crown_checks`: the include and project global, the twin against a real perspective `Camera3D` at 0.75, 1 and 1.3 × the target's depth at fov 24 and 36, worst under 0.01 px where the target-depth step missed by over 0.3 px, the view's slope under both projections); `test_foliage_merge` 81 → 84; `test_tree_house` 160 → 161; `test_camera` 247 → 260 (second fix pass, `p8g1_photo_crowns` in photo mode on entry by P, at a wider photo lens, from the free camera and after leaving: the slope follows the camera's own lens, n steps are n pixels at 0.7, 1 and 1.4 × the target's depth, the trees' leaf keeps the shift, the light clock keeps running). 219: - Final full suite (second fix pass; `fix2/suite.sh`, `fix2/suite-final/suite_all.out`, 18:47-18:56, AC 100%, Low Power Mode 0): **42 suites, 22,792 checks, 0 failures** (first fix pass: 22,763; first sitting: 22,751). `--verify` 2880x1800: exit 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). 220: - **Changed expectations:** `test_foliage_merge` "grown surfaces, and only they, wear the grown materials" now also counts the trees' grown copy as grown (old: only `grown_materials`). Second fix pass: `test_foliage` "the shift takes its phase from MODEL_MATRIX[3] ..." old `wind_crown_shift(MODEL_MATRIX[3].xyz,MAIN_CAM_INV_VIEW_MATRIX[0].xyz,banding>0.5)` → new `wind_crown_shift(MODEL_MATRIX[3].xyz,MAIN_CAM_INV_VIEW_MATRIX,banding>0.5)`. Everything else is added checks. My own first draft of the fix's `_process` text check read a brace block that GDScript has not, and failed on correct code; it was rewritten as a plain text match before the code was judged. 221: - **Stills** (first sitting: `caps.sh`, `metrics.sh`; before-1/after/before-2/after-2, 9 views, AC; automation holds `light_time` 0, where no crown in any of these scenes is in a gust under either form, `fix/an/still-t0.txt`): 222: - Counts, draw calls included, are identical in every pair. 223: - grove-day changes 0.54-0.55% in all four cross legs against noise of 0.013/0.001%: tree crowns at rest instead of at their `TIME` sway pose. `look/final-stills-sheet.png` shows no visible difference. 224: - The other views are inside their noise pair on 3 of 4 legs; the before-1 leg is the noisy capture: harbour-day before-1 → after 2.401% against noise 1.319/1.561%, quarter-day 2.307% against 1.74/0.743%, quarter-night 6.4% (`metrics/frame-table.txt`). Harbour has trees, so a crown-at-rest change there is expected. 225: - Fix pass: the 9 views again on the final code (`fix/after-3`, 16:58, AC). Counts identical to the first sitting's after. Across sittings the >8 change is 0.004% on grove-day and 0.7-5.9% elsewhere against same-sitting noise of 0.26-1.7% (`fix/an/stills-vs-first.txt`), the cross-sitting spread rule 1 warns of; I looked at `fix/look/harbour-day-after-vs-after3.png`: rain rings, drops and smoke, no tree crown. The fix cannot reach a still (no crown is in a gust at `light_time` 0). 226: - **Acceptance, against the target.** First-sitting numbers are marked (first); the rest are the final code, all from one sitting, 15:50-17:05, AC 100%, Low Power Mode 0 (`fix/rec/power.txt`), with before clips from the pre-part tree `pre-tree` (P8-II close) and the part's own scanner (`an/crawl.js`, `an/steps.js`, `an/same.js`, `an/wind.js` with the floor; `fix/an.sh`, `fix/an2.sh`, `fix/gusts.js`, `fix/vsoff.js`). 1440x900, 15 fps, the light clock from 0, so a clip covers `light_time` 6 to 6+seconds. 227: - **Rigid 1-2 px, no internal crawl:** dry grove, 22 m (70 s): 54 step frames against the twin's 53, 52 of them seen; the best integer shift explains 80.7% of step-pair change, interior 2.2% (`fix/an/gusts.txt`). 12 m (30 s): 24/24, 84.9%, interior 3.05%. 38 m cropped to the trees (`RECT=230,70,490,270`): 24/24, 85.7%, interior 2.6%. The 38 m number the first sitting quoted (85.3%/2.6%) is `an/out/steps-dry.txt` (first, cropped to the trees). `an/out-final/steps-g38-dry.txt` (48.7%, interior 9.2%, 84 step frames) is a whole-frame run whose water and cloud changes come in a 0.5 s rhythm the wind cannot make; `fix/an/steps-g38-dry-after-fullframe.txt` repeats it (48.7%). 228: - **Rest frames:** median changed art pixels in non-step pairs 0 at 22, 12 and 38 m against 43, 170 and 10 before (`fix/an/crawl-summary.txt`). Full against Off, dry, frame by frame: the frames differ only from 17.06 to 23.32 s, and by exactly 0 px in all 351 other frames (`fix/an/vsoff-dry.txt`). A second take of the 12 m clip matches the first pixel for pixel (`fix/an/same.txt`, 0%: the noise pair). 229: - **Gusts every 20-40 s:** twin over the whole 1380 s wrap for the grove (`an/out-fix/holds-grove.txt`): every tree moves on all 46 gusts at rain 0, 0.22, 0.35, 0.45 and 1, Full and Reduced; holds 3.47-3.53 s (4.33 s in full rain); gaps from one move's end to the next's start 23.7-35.3 s (23-34.7 s in full rain). Before the fix (first): rain 0 moved 20 times a tree with gaps up to 177.9 s, Reduced rain 0 never, Reduced rain 0.35 10 times with gaps up to 270.9 s. In the 70 s clips the gusts start at 17.06, 45.99 and 74.79 s (start gaps 28.9 and 28.8 s) at rain 0, 0.22 and 0.35 at Full and at Reduced rain 0 and 0.35; in the rain clips the scan takes only the pairs within a frame of the twin's steps (`PRED_ONLY=1`, `fix/an/gusts-pred.txt`): 52-53 of 53 seen, 72.8-76.4% explained (drops through the crowns take the rest), interior 3.5-4.7%. Reduced dry is the same clip as Full dry pixel for pixel (`same.txt`): both move 1 px, by the floor. 230: - **Off:** zero moves. The twin: 0 of 33,120 samples moved (rain 0, 0.35 and 1, `test_foliage`). Off dry clip (`fix/rec/g22-off-dry`): rest median 0, one rigid component in 449 pairs, and identical to Full dry outside the gust. 231: - FoliageLamps, the `COLOR.a` stamp and the lamp term: not touched by the fix (the fix pass changes, against `26-P8-g1b`: `wind.gdshaderinc` (comment and lines 78-79), `world_view.gd`, `main.gd`, the restored whitespace in `foliage_body.gdshaderinc`, `test_foliage.gd`, `test_tree_house.gd`). 232: - **Bench, the moving path** (fix pass; `fix/bench.sh`, `fix/bench-summary.js`, `fix/bench-summary.txt`, `fix/bench-t1.out`, `fix/bench-t13.out`; 16:46-17:01, AC 100%, Low Power Mode 0, load 2.7-3.1, `fix/bench/power.txt`). Three trees interleaved, round 0 discarded, rounds 1-3, 1920x1080, `--disable-vsync` before `--path`, rain 0.45, `--motion-live` with `--light-pin` at the moment the most trees are in a gust (`fix/pin.js`: dense 20.6 s, 44 of 100 trees shifted; Harbour 20.3 s, 9 of 12; forest 18.15 s, 43 of 100). Pre is the P8-II close plus only the pin lines (`bench-pre-tree`); forced is the live code with every crown shifted in every frame (`bench-forced-tree`, a harness, never shipped), so the whole scene is in a gust. Median raw p50 and p95, paired median delta and rounds slower: 233: 234: | Scene | 01:00 p50 live / forced | 01:00 p95 live / forced | 13:00 p50 live / forced | 13:00 p95 live / forced | 235: |---|---|---|---|---| 236: | Dense (pre 14.220 / 13.598) | +0.144 (3/3) / +0.136 (3/3) | +0.137 / +0.076 | +0.105 (3/3) / +0.085 (3/3) | +0.109 / +0.070 | 237: | Harbour (pre 6.534 / 6.175) | +0.065 (3/3) / +0.074 (3/3) | −0.509 / +0.281 | +0.068 (3/3) / +0.059 (3/3) | +0.769 / +0.291 | 238: | Forest (pre 5.694 / 5.252) | +0.170 (3/3) / +0.190 (3/3) | +0.180 / +0.410 | +0.164 (3/3) / +0.182 (3/3) | +0.179 / +0.379 | 239: 240: - Dense p50 is +0.14 ms at most with a gust moving (target 0.3) and Harbour +0.07 (target 0.5), forced or not, so no cheaper form is needed. The inverse matrix on moving crowns costs about nothing measurable (forced ≈ live); the cost is the per-vertex field on every tree leaf vertex, main and shadow passes. Harbour p95 scatters ±0.5-0.8 ms on both sides. Draw calls are identical (2583/907/682). 241: - The first sitting's bench (`bench-summary-final.txt`, clock from 0, no gust forced) gave dense +0.117/+0.077 and Harbour +0.452 (2/3, scattered)/+0.063 at 01:00/13:00; its Harbour 01:00 figure did not repeat here (+0.065, 3/3). 242: - **Fix pass (orchestrator, 6 October 2026):** the orchestrator ruled, on review, that: 243: 1. Dry weather and Reduced barely moved crowns (rain 0: 20 moves a tree a wrap, gaps up to 178 s; Reduced rain 0: none). Changed: the 1-pixel floor in `wind_crown_pixels` and its twin `crown_pixels` (`wind.gdshaderinc` line 79); numbers above. 244: 2. The bench never measured the moving path. Changed: the test-only `--light-pin`, the three-tree bench above. 245: 3. Glade and tree houses. Glade before (`fix/rec/glade-before`, old sway) against after (`glade-after`), same moment, motion live, same scanner (`fix/an/glade-window.txt`, `crawl-summary.txt`): in the first gust's window (16-24 s) the trees change a median of 23 art px a frame pair against 22 for the old sway; between gusts 0 against 22 (38 of 328 rest pairs change at all, against 318). So at Full the Glade trees visibly move as each gust passes (unsnapped, about 0.8 art px at 22 m dry) and are still between gusts: they read alive in gusts and calm between, less constantly alive than the old sway. I looked at `fix/look/glade-after-255-285.png`: edge speckle across the whole crown, 62% explained by a 1 px offset, nothing tears. Tree houses: host trees kept the shared leaf, so they swayed beside stepping neighbours. That is a one-line material change in `_build_host_mesh`, so it is fixed: on the carrier-fix tree-house demo (`artifacts/trees/carrierfix/demo.json`, camera 0,0.32,24,12,8,24; there is no tree-house starter, the four StarterDemos have no tree) the host crowns now step 1 px with the gusts (frames 240 → 290: a −1 px offset explains 3,847 of 4,447 changed art px of the centre host, `fix/look/h-centre-240-290m.png`, `house-after-278-d-1.png`, `house-after-331-d1.png`, looked at). The house, deck and boughs stay; the trunk and boughs under the crown do not follow (P8g2's joint, as for every tree). Rest median 156 against 230 before (water and ducks remain). 246: 4. Hygiene: the stray `if (grown &&grown_hidden(` is restored (and checked). The perspective lens (Feel D2, `--projection=perspective`) was recorded before and after (`fix/rec/g22-persp-*`): 24/24 steps, 77.0% explained, interior 3.7%, rest median 4 against 67 before (the shift was whole pixels only at the target's depth, so trees nearer or farther moved a fraction more or less: the lower explained share). Photo mode was not recorded. Both are closed by the second fix pass. 247: - One recording froze (`fix/rec/g22-rain035-frozen`: all 1,050 frames byte-identical although the run reported 15 fps); it was recorded again and the frozen one is kept for the record. 248: - **Second fix pass (orchestrator, 6 October 2026).** The orchestrator ruled four items must be fixed: photo mode untested; the harbour-day after clips at three zooms not analysed; the first fix pass's frames lost to the Trash; perspective 77.0% explained, below the 80% bar of the orthographic path. One sitting, 17:17-18:59, AC 100%, Low Power Mode 0, load 3.2-3.8 (`fix2/rec/power.txt`, `fix2/rerec/power.txt`, `fix2/suite-power.txt`, `fix2/bench/power.txt`). Every clip keeps all its frames in `/frames` (`fix2/rec` 11 GB, `fix2/rerec` 6.3 GB), and `fix2/frozen.js` checked each for frozen frames (`fix2/an/frozen-check.txt`). 249: 1. **Perspective (the code change above).** The same-sitting A/B (`fix2/rec/g22-persp-{before,mid,after}`, grove 22 m, `fix2/an/gusts-persp-photo.txt`): mid (the code before) reproduces 77.0%, interior 3.7% exactly (its `motion.tsv` equals `fix/rec/g22-persp-after`'s); after: 24/24 steps, **81.4% explained, interior 2.73%**, rest median 4 (before 66: `fix2/an/crawl-persp-photo.txt`). I looked at `fix2/look/g22-persp-after-288-289.png`: the tall tree moves one pixel left, green inside, red only on its silhouette and shadow edge. 250: 2. **Photo mode** (`--photo-lens`: the perspective lens at the default 24°, the pivot at the subject's depth and the gentle depth of field; `fix2/rec/g22-photo-*`): before (old sway) rest median 39; mid 24/24 steps, 80.2%, interior 2.8%; after 24/24, **84.8% explained, interior 2.56%, rest median 0** (`fix2/an/gusts-persp-photo.txt`). I looked at `fix2/look/g22-photo-after-288-289.png`: the crown steps whole under the blur, nothing tears. `test_camera` holds the slope to photo mode's own lens, its wider lens and the free camera, and the clock running there. 251: 3. **Harbour at three zooms** (the plan's captures; before from `pre-tree`, after on the live code, same sitting). The default-rain clips (`fix2/rec/harbour-day-*`, `harbour21-*`, `lane-day-*`, Harbour's rain 0.2) change by 3,000-8,000 art px a pair (water, drops, smoke), so read tree by tree in projected windows (`fix2/harbour.js`, `fix2/an/harbour-trees.txt`, windows checked on `fix2/look/harbour-day-boxes.png`): at the twin's step pairs 44.0% / 51.9% / 41.3% explained after against 27.6% / 32.0% / 29.3% before at 38 / 21 / 16 m, with the water and drops inside the windows taking the rest; rest medians 64 / 97 / 56 after against 64 / 96 / 68 before, the same: water and drops hide the crowns there. So the crowns were isolated in dry clips (`--rain=0`) against an Off clip of the same view, with the chimney smoke masked: two dry takes differ only in the smoke (`fix2/look/harbour21-dry-take-vs-take-200.png`), and Full against Off at rest only in the smoke (`harbour21-dry-full-vs-off-100.png`) (`fix2/hvo.js`, `fix2/an/harbour-vs-off.txt`): 252: 253: | Zoom | After: step pairs, explained, interior | After: rest pairs with any crown change | Before (old sway): pairs with crown change | 254: |---|---|---|---| 255: | 38 m (harbour-day) | 13, 78.1%, 1.3% | 86 of 436, median 0, max 29 px | 449 of 449 (median 14 px) | 256: | 21 m | 21, 84.8%, 2.4% | 90 of 428, median 0, max 29 px | 449 of 449 (430 of them ≥ 30 px) | 257: | 16 m (lane-day) | 6, 83.1%, 3.8% | 17 of 443, median 0, max 2 px | 449 of 449 (290 ≥ 30 px) | 258: 259: Each moves in one gust, 16.0-24.4 s at 38 and 21 m and 18.1-24.4 s at 16 m (a 30 s clip holds one; the twin covers the 20-40 s spacing). The 38 m share is below 80% because its crowns are small: the unexplained is silhouette edge (554 of 2,684 px), interior 1.3%. I looked at `fix2/look/harbour21-dry-after-239-240.png` (the small tree by the bridge steps 1 px left, green) and `lane-day-dry-after-291-292.png` (the blossom's crown steps whole; its trunk stays, P8g2's joint). Two lane-day dry recordings froze (1 distinct frame in 450; kept in `fix2/rec/frozen/`) and were recorded again. 260: 4. **The lost frames.** Restoring the Trash is the owner's; instead the first fix pass's cited clips were recorded again with their frames kept (`fix2/rerec/`, 15 clips; the perspective ones are `fix2/rec/g22-persp-before` and `-mid`). Recordings are deterministic outside rain and smoke: 14 clips have a `motion.tsv` byte-identical to the first fix pass's clip (`fix2/rerec/` `g22-dry-before/after`, `g22-reduced-dry`, `g12-dry-before`, `g22-off-dry`, `g38-dry-before/after`, `glade-before/after`, `house-before/after`; `fix2/rec/` `g12-dry-after`, `g22-persp-before`, `g22-persp-mid`), so their frames are the cited frames. The four rain clips (`g22-rain022`, `g22-rain035`, `g22-reduced035`, `g22-off`) differ only in the random drops; their scans repeat within the drop noise (`fix2/an-rerec.sh`, `fix2/an/rerec/`): rain 0.22 / 0.35 / Reduced 0.35 76.9% / 73.3% / 73.5% explained against 76.4% / 72.8% / 73.1%, gusts at 17.06, 45.99 and 74.79 s in all; dry 22 m 80.7%, interior 2.17%, 53 predicted / 54 seen, identical; Full against Off dry moved only 17.06-23.32 s and by 0 px in the other 351 frames, identical. 261: - The orthographic path is unchanged by the second fix pass: the 12 m dry grove on the code before it (`fix2/rec/g12-dry-mid`) and after it (`g12-dry-after`) match pixel for pixel in every frame (`an/same.js`: 0%), and both equal the first fix pass's clip by `motion.tsv`. No crown is in a gust at `light_time` 0, so no still can change; stills were not captured again. 262: - **Bench** (`fix2/bench.sh`, `fix2/bench-summary.txt`; mid-tree against live, interleaved, round 0 discarded, rounds 1-3, 1920x1080, vsync off, rain 0.45, `--light-pin` at the busiest gust, orthographic, 18:56-18:59, AC, Low Power Mode 0): dense p50 14.293 → 14.363 ms (paired median +0.002, slower 2/3), p95 −0.025; Harbour p50 6.596 → 6.588 (−0.004, 1/3), p95 −0.238. Draw calls identical (2583, 907). The perspective path was not benched (it adds one dot product per crown vertex). 263: - **Expected journey count deltas:** none. No crown is in a gust at `light_time` 0 in any journey or capture scene, so stills are unchanged by the shift; tree crowns at rest lose their sub-pixel sway pose (grove-day 0.55%). 264: - Review: filled by the orchestrator. 265: - **For the owner sheet:** the D17 deviation above; under the perspective lens and in photo mode the step is now whole pixels at each crown's depth (81.4% / 84.8% explained), and at Harbour 38 m the small crowns read 78.1% (silhouette edges, interior 1.3%). 266: - **Not done:** the first fix pass's original frames stay in the Trash (restoring them is the owner's; the second fix pass recorded the cited clips again, above). The perspective path was not benched. Photo mode was recorded at its default lens only (the wider lens and the free camera are held by `test_camera`, not recorded). Stills were not captured again (no crown is in a gust at `light_time` 0, and the orthographic path is pixel-identical). 267: - Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy = HEAD fa7f38d), lines 246-276 246: ## P8g1 approved (owner, 6 October 2026, in chat) 247: 248: "P8g1 approved, move on." Given while the orchestrator-directed fix pass on P8g1 (dry-weather and Reduced gust minimum of 1 art pixel, gust-forced bench, Glade and tree-house check, handoff corrections; run ) was still running. Read by the orchestrator as approval of P8g1 as it stands after that fix pass, including the deviation from D17 that tree crowns no longer have the old sub-pixel sway (they move only in whole-pixel gusts at Full and Reduced, and are fully still under Off, while shrubs, ivy and clusters keep their old sway). If the fix pass reports a result the owner would not accept (a gust-forced bench cost over +0.3 ms dense or +0.5 ms Harbour with no cheaper form, or crawl in the Glade), the orchestrator will bring that back to the owner instead of applying this approval. P8g2 starts once the fix pass has finished and P8g1 is committed. 249: 250: ## Orchestrator note after P8g2 (6 October 2026) 251: 252: P8g2 review: no must-fix. Left as recorded open items for the P8-III stage sheet, not fixed now: (1) on tree-house hosts the leaves in the 1 m band above the house take a fractional share of the whole-pixel shift, so the lower crown shears instead of stepping rigidly (hosts were judged by crops only; no numeric scan); the owner sees the change: a tree-house tree's crown now moves only above its house, where P8g1 moved the whole host crown; (2) host bark reaches full shift only at the lowest leaf when the house is low on a tall tree (no BEND_SHARE equivalent); (3) only 1 px gusts were scanned (dry clips), the 2 px case rests on construction; (4) no perspective before-clip, no Harbour or lane clips, no 13:00 bench; (5) latent: the host leaf path does not clamp wind_bend's -1 result like the plain path does, and a stale comment in grown.gdshaderinc. If the owner dislikes the host look, a directed pass fixes (1), (2) and (5) together. 253: 254: ## P8g2 approved (owner, 6 October 2026, in chat) 255: 256: "P8g2 approved, move on." Given after the P8g2 report (main 841006b) and the orchestrator note listing its open items. Read by the orchestrator as approval of P8g2 as built, with the open items left as built: a tree-house tree's crown moves only above its house and the leaves in the metre above it take a fractional shift (no directed pass), host bark with a low house on a tall tree, only 1 px gusts scanned, no perspective before-clip. They can be reopened by the owner at any time, and the P8-III stage sheet will show the tree-house look. P8h (cloth material) was already running. 257: 258: ## P8h approved, and a standing approval until morning (owner, 6 October 2026, in chat) 259: 260: "P8h approved, move on, Approve everything until morning. I need to use the laptop at 8am so dont start anything that won't finish before then." Read by the orchestrator as: (1) P8h approved as built (cloth material for flags, pennants and awning valances; the sloped pennant edge crawls 0/1 and cloth rests between gusts, both left as built); (2) until the owner is back on the laptop at 08:00 on 7 October 2026, every P8 part and fix result that clears review is treated as approved as built, with its open items left as built and listed for the owner sheet, so the pipeline does not wait for answers; (3) the standing approval covers plan approvals and owner-sheet questions only, not pushing to GitHub, deploying the journal, or any outward-facing action, which still need an explicit request; (4) the orchestrator starts no workflow it does not expect to finish before about 07:30 on 7 October: parts normally take 1.5 to 2 hours and up to about 3.5 hours with fix passes and a second review, so no new part is launched after about 04:15 (a part with a second review after about 03:45) and the P8-III stage close (about 1.5 hours) is launched only if it can finish by 07:30; whatever is not started is left for the morning with the resume point written at the top of the board. No Godot run is left running past 07:30. 261: 262: ## Orchestrator note after P8i2 (7 October 2026, under the standing approval) 263: 264: P8i2 review: no must-fix; approved as built under the standing approval until 08:00. Left for the owner sheet: (1) the cat's-paw strength follows the gust shape only, not wind_strength, so a dry calm day ruffles as hard as full rain (the report's own formula; rain lean does scale with wind_strength); (2) the daytime ruffle is faint (one constant, PAW_DASH, makes it stronger); (3) slant kept at P8d's 0.112 not the report's 0.12; Glade water ruffles the same (the report's x0.6 for Glade not applied); (4) latent: wind_front's per-gust offset puts the sky_noise lattice about 400 cells out, so the sin-based hash runs on arguments near 1e5 in float32 and the GPU pattern may not match the CPU twin or other GPUs (Windows not run); a smaller offset (fract-based or mod 16 cells) would fix it; (5) dashes at a patch's soft edge can be drawn partly (optional: evaluate paw at the dash cell centre); (6) rain falls about 1.6% slower vertically at full gust (direction is normalised), wording only. 265: 266: ## Orchestrator note after P8j (7 October 2026, under the standing approval) 267: 268: P8j review: no must-fix; approved as built under the standing approval until 08:00. Left for the owner sheet: boats reuse their existing per-kind decoration MultiMeshes instead of the plan's 'small MultiMesh of their own' (no draw calls added); the bob has a 0.125 s step (the first 1/15 s form showed a one-frame pose); under a G drag the original boat keeps bobbing under the ghost (as every decoration does today); drag-harness film, Glade, perspective and photo-mode clips not taken; under perspective the roll/pitch gate reads the global art_pixel; swarm sites are not re-ranked toward water and plazas; the moth-spacing test was loosened (at most 16 of 160 instants at or below 0.08 m; every swarm, old ones included, dips below 0.08 m in 1-6% of instants); extra swarms (3 fireflies, 5 moths) are gated by the motion preference (Full 3/5, Reduced 2/3, Off 2/2) and creatures at the night quarter go 30 to 34. 269: 270: ## Orchestrator note after P8l (7 October 2026, under the standing approval) 271: 272: P8l: the first review had no must-fix, the second review found two (the sky's bottom row was lightened at golden hour by the outline pass; the alpha figures were skewed by rain noise), both fixed and rechecked; approved as built under the standing approval until 08:00. Left for the owner sheet: under Off the pattern holds still instead of being hidden (the plan's 'hidden at night and when off' is read as night only; one line gates it); about 19 stone pixels beside the water flicker under Off (probably the outline pass meeting the moving water); Glade reads about 0.4x canal in pixels although the shader strength is exactly half, and its shadow falls on far water rather than the island; the stills (light clock held at 0) put little shadow on the town, so patches reach the towers only as they drift; the first art row of water under the horizon is left unshaded (a pale line) and two art pixels beside shaded towers are lightened; overlays drawn before priority 115 (selection grips, placement fill) are darkened by up to 12% under a passing patch by day; the CPU twin's hash differs from the GPU's in float32 (the same class as P8i2's open item 4); strength uses min(sun_up, 1 - night), not the product; perspective and photo mode not captured; Windows not run; start-up all_ms +133 ms and one more draw call by day. 273: 274: ## Orchestrator note after P8k (7 October 2026, under the standing approval) 275: 276: P8k: the first review had no must-fix; the second review found one (the handoff called the splash ring thin; it is a blocky 3-6 px band late in its life), fixed in wording and rechecked; approved as built under the standing approval until 08:00. Left for the owner sheet: puff colour is lighter and greyer than the lit lawn (+36 luma by day; AMBIENT_MIX and NIGHT_DIM are the constants); the first puff of each kind hitches (+77 ms tree, +67 ms boat) when run without --warmup because its shader compiles on first use (a player's warm-up covers it; start-up warm-up +272 ms, 144 entries); Water switched off: a part set on an unpainted cell would throw splash drops over the dry bed (one-line gate); building footprints use one ground height at the centre; the undo and redo checks fake the replay (no test through EditHistory); dense with animations +0.69 ms; no Glade, perspective, photo-mode or house-placement film; Windows not run. === Appended 8 October 2026 === Frozen 8 October 2026 at 03:38 BST (game main at 8e170d0, clean; the frozen handoff and owner-decisions.md equal HEAD). Covers the P8-III stage close (7 October) and the owner's P8-III approval, then sub-stage P8-IV: P8e rain veil (layered, depth-gated streaks) and P8f lit rain (as built, the second review's wording fix, the recheck's measured cause, and the third, orchestrator-directed fix pass: 300 CPU drops in a near box in front of the camera and the light floor read on the shown colour), and the P8-IV stage close. With P8-IV closed, all four P8 sub-stages are closed; P8e and P8f stand under the owner's standing approval until 08:00 on 8 October and the owner had not seen the P8-IV sheet at the freeze. The handoff entries' "Review: filled by the orchestrator" lines were still unfilled at the freeze. Part codes: P8e rain veil, P8f lit rain; D28 the rain veil default. Local paths and run IDs are replaced by placeholders; moves "to ~/.Trash/" read "to the macOS Trash". Line numbers refer to the frozen copy. SOURCE: docs/visual-handoff.md (frozen copy = HEAD 8e170d0, 8 October 2026, 03:38 BST), lines 5-211 5: **Beautify P8-IV (parts e, f of P8) stage close (8 October 2026):** this is the stage-end overhead of sub-stage P8-IV "Motion" (step `26-P8-IV` in the plan's overhead table). Opus 5.5 ran it on this Mac on the final P8-IV code: `main` 8ccf61a, clean (`close/head.txt`). The parts are P8e (rain veil) and P8f (lit rain, with its three fix passes). No game code changed. The evidence is in `artifacts/beautify/26-P8-IV/close/`. 6: - **Before side:** the pre-stage scratch tree `close/pre-tree`, built with `tools/feel_snapshot.sh tree 26-P8-III`. Every file changed since `523751a` that exists in it equals `523751a`, and it differs from live only in P8-IV's 13 files (`close/diff-pretree.txt`, `close/git-changed-since-523751a.txt`; `project.godot` equal). 7: - **Power:** every run on AC, 100% charged, Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`); load averages 1.6-3.5. 8: - **Full suite** (`close/suite.sh` a, b, v; 02:19-02:27): **44 suites, 23,554 checks, 0 failures** (`close/suite/suite_all.out`). `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,583 checks and 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 9: - **Journey** (`artifacts/feel-journey/26-P8-IV/`, six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): 10: - Counts against 26-P8-III (`26-P8-IV/counts-vs-26-P8-III.txt`): +1 draw call and rendered object on all six views (harbour-day 907 → 908, harbour-lane 749 → 750, harbour-night 906 → 907, quarter-day 1024 → 1025, quarter-night 1023 → 1024, styles-day 501 → 502). That is P8e's declared +1 on rainy views (the veil quad); P8f declared none. A single-view re-run of harbour-lane and quarter-night (`close/rerun/journey-live/`, `rerun/counts-journey-vs-rerun.txt`) gave identical counts. 11: - Frame >8 luma against 26-P8-III 2.1-9.5% (`26-P8-IV/frame-vs-26-P8-III.txt`); the within-tree re-run 0.3% (quarter-night) and 1.8% (harbour-lane). Looked at `close/look/journey-diff-*.png`: veil streaks in the sky, drops in new places, rain rings, reflections and cloud dither (TIME). 12: - `notes.json` has a `26-P8-IV` entry (`close/journey-note.js`, backup `notes.json.bak-before-26-P8-IV`); `index.html` and `journey-standalone.html` (109.88 MB) rebuilt. 13: - **Art capture and protected looks** (`close/caps.sh`, the P8-III close's 101 views in groups g1, g2, g3; 02:30-03:10; pre-1, after, pre-2, after-2, every view on try 1). Metrics `close/metrics/` (`metrics.sh`, `frame-table.js`, `prot.js` → `protected.md`, `elapsed.js`, `consistent.js` → `metrics/consistent.txt`; regions = the P8-III close's). Against the previous stage close `26-P8-III/close/after`: `ab-p8iiiclose-*`, `frame-p8iiiclose-*`. Sheets `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P8-III-close.png`, `noise-pre.png`, `noise-after.png`. 14: - **Counts** (`metrics/counts.txt`): +1 on the 86 rainy views, identical on the 15 dry or Glade views (amsterdam ×3, banks-glade, disc-gg-sun145, glade-1930/-2045/-day/-golden/-night, grove-1930, grove-day, quarter-dry, shops-day/-night), the same in both pairs. glade-rain (Glade with `--rain=0.45`) gets +1: a Glade scene with its rain raised by hand draws the veil. Within a tree only the orbit views vary, as before. 15: - **Frames:** 78 of 101 views above both noise legs (`frame-table.txt`). The live capture is later in both pairs in 89 of 101 views, median 108 ms (`elapsed-ms.txt`). The consistent maps (`look/consistent/*.png`) show red veil streaks in the sky, drops in new deterministic places (P8f's near box) and darker rain rings on the water (a later ring phase; the rings themselves look the same, `look/rings-q2100-pre1-after.png`). disc-gg-sun145 (Glade, dry) is above noise by a hair (0.30/0.35 against 0.26/0.30, 0.06% consistent): TIME phase. 16: - **Canal mirror columns:** p50 43.4 → 43.5 (43.5 → 43.4) quarter-night, 52.1 → 52.3 at 21:00. Changed >8 9.2%/17.2% and 17.0%/17.7% against noise 6.3%/0.5% and 2.2%/1.7%: above noise, from the veil over the far water and the drops (P8e declared the veil on far water), p50 within 0.2. 17: - **HDR neon cores:** fish 66.3 = 66.3, bar 70.4 → 70.6 (70.6 = 70.6), lane sign 140.0 → 140.4 (140.2 = 140.2), tower windows 96.0 = 96.0; changed shares at noise. 18: - **05:45 and 20:45 skies:** changed 2.2-7.5% against noise ≤0.9% (the veil streaks, declared by P8e and P8f); p50 within 0.7 (20:45 horizon 150.5 → 151.2, zenith 97.5 → 97.7; 05:45 horizon 113.6 → 113.9, zenith 56.6 = 56.6). 19: - **Glade/Canal separation:** Glade night zenith `#001d45`, horizon, water, lawn and bank 0% changed; Harbour night zenith p50 53.9 = 53.9. 20: - **Approved tree style:** grove-day and grove-1930 crowns, conifers, feet and lawn 0.00% changed (p50 55.3, 55.6, 59.4; golden conifers 42.9 = the P8-III close); quarter-night undersides 27.7 = 27.7. 21: - **Whole-stage A/B bench** (`close/bench.sh` = the P8-III close recipe; `bench-t1.out`, `-t13.out`, `-t19.5.out`, `bench/summary.txt`; 03:11-03:23): pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync` before `--path`, 1920x1080, rain 0.45, `--motion-live`, dense `artifacts/weather-pass9/dense.json` and Harbour; 491 samples each, every run on try 1. Raw medians pre → live (paired median; rounds live slower): 22: 23: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 24: |---|---|---|---|---|---|---| 25: | Dense | 14.496 → 14.509 (−0.003, 1/3) | 16.447 → 16.314 (−0.133, 0/3) | 14.244 → 14.282 (+0.038, 2/3) | 16.173 → 16.247 (+0.066, 3/3) | 16.322 → 16.345 (+0.044, 2/3) | 18.139 → 18.130 (−0.003, 1/3) | 26: | Harbour | 6.890 → 6.909 (+0.025, 3/3) | 7.412 → 8.130 (+0.560, 2/3) | 6.566 → 6.573 (+0.021, 3/3) | 7.938 → 7.240 (−0.698, 1/3) | 6.667 → 6.687 (−0.013, 1/3) | 8.277 → 7.482 (−0.795, 1/3) | 27: 28: - Draw calls +1 (dense 2586 → 2587, 2587 → 2588, 3438 → 3439; Harbour 916 → 917, 917 → 918, 829 → 830). The stage costs −0.01 to +0.04 ms p50. Harbour p95 spreads 7.2-8.8 ms in both trees with no direction (P8e saw the same); the plan's stop (dense p95 over about +0.5 ms) is not hit. 29: - **Owner sheet** `artifacts/beautify/26-P8-IV/close/sheet/` (`close/sheet.sh`, regions `close/sheet-regions.json`, `rainalone.js`, `crop.js`). Pages `p8iv-01-veil-at-night.png`, `-02-veil-twilight-and-day.png`, `-03-lit-rain-bar-sign.png`, `-04-lit-rain-lamps.png`, `-05-drops-by-day.png`, `-06-existing-scenes.png`, `-07-dry-and-glade.png`, `-08-what-differs.png`, and `sheet.md`. Stills: close pre-1 and after. Pages 03-04: P8f's fix-3 recordings (`26-P8/f/fix3/rec/qn-pre`, `qn-live`, `qn-off`); the rain alone is the largest |clip − off| over frames 150-165, ×4; their "before" is the P8e code, so the veil is in both. I looked at every page: the lit drops (pink at the BAR, amber by the lamps) show clearly in the rain-alone crops and only faintly in a plain frame (page 03, second pair). 30: - **Not done / open:** 31: - The owner has not seen the sheet. P8e and P8f stand under the standing approval until 08:00 on 8 October; their open items (P8e and P8f entries below, "For the owner sheet") are left as built. 32: - No new recordings were made; the lit-rain pages reuse P8f's clips. No perspective, photo-mode, Reduced or Off captures at this close. 33: - The P8 recording sheet of every clip and the P8 stage close (all four sub-stages) are still open, as is M5 (dropped by D8). 34: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 35: - Scratch trees: `close/pre-tree` was moved to the macOS Trash with `mv` as `beautify-26-P8-IV-close-pre-tree-033359`, once its numbers were recorded. The part trees `26-P8/e/pre-tree` and `26-P8/f/pre-tree` are outside `26-P8-IV/` and were left in place. 36: - Platforms: Mac run; Windows not run. 37: 38: **Beautify P8f Lit rain: the veil takes neon and lamp colour; CPU rain trimmed (7 October 2026):** Opus 5.5 on this Mac (plan section "P8f"; report polish backlog "rain in art pixels" (the lit veil) and its lens finding "Replace the sparse hard rain streaks with a layered, lit pixel rain veil": "take 4 taps of the screen copy at ±6 px and add any brightness above 0.35, ×1.2, in its own hue ... Then retire the CPU rain, or keep at most 150 near drops"; step `26-P8`, sub-stage P8-IV, second part; a measuring part, second review required; D28 default). From the lit evening on, a rain pixel (P8e's veil streaks and P8d's drops) near the neon, a lit window or a lamp takes that light's colour: it reads the screen copy 6 art px left, right, above and below. By day and at golden hour nothing changes. The CPU keeps at most 150 drops (P8d had 909 at the default rain 0.35 and 1170 at 0.45). **Third fix pass (8 October, orchestrator ruling): the CPU keeps at most 300 drops, born in a near box in front of the camera so every one is on the frame, and the floor is read on the shown colour; the acceptance line below is the fix-3 state.** Snapshot: `build/backup-feel/26-P8-f/` (missing at part start, so I saved it before the first edit; it then equalled `26-P8-e` and `26-P8-e-post` = `main` 2d1d4d7: `diff -rq` on scripts shaders tests data scenes tools and `project.godot` was empty, nothing unexpected in the code). `26-P8-f` was re-saved after the third fix pass and now holds the fix-3 code (equal to `26-P8-f-fix` and the live tree); the pre-part tree is `build/backup-feel/26-P8-f-pre/` (= `26-P8-e-post` = `f/pre-tree`). Two things I did not expect, neither of them code: `26-P8-e` now equals the post-P8e tree, although the P8e entry below said it equalled `26-P8-III`; and `artifacts/beautify/owner-decisions.md` carries the orchestrator's new sections, written during this part. The as-built (post-part) snapshot is `build/backup-feel/26-P8-f-asbuilt/`. The pre-part scratch tree is `artifacts/beautify/26-P8/f/pre-tree`; the pre-P8e tree for the bench is P8e's `artifacts/beautify/26-P8/e/pre-tree`, re-checked equal to `26-P8-III`. Evidence: `artifacts/beautify/26-P8/f/`; file backups in `f/backup/`. 39: - **Code:** 40: - `shaders/lit_rain.gdshaderinc` (new): `screen_texture` (hint_screen_texture, nearest), `lit_night` (0 by default), `LIT_REACH_PX` 6, `LIT_FLOOR` 0.35, `LIT_GAIN` 1.2. `lit_rain(uv, size)` sums, over the 4 taps, each tap's Rec. 709 luma above the floor in the tap's own hue, × gain × `lit_night`; at `lit_night` 0 the copy is not read. `lit_colour(own, lit)` = (own × (1 − w) + lit) / max(1, its largest channel), with w = min(lit's largest channel, 1). No TIME. 41: - `shaders/rain_veil.gdshader`: includes it and sets `ALBEDO = lit_colour(streak_colour.rgb, lit_rain(SCREEN_UV, size))`, only after the discard (on streak pixels). `ALPHA` is P8e's. 42: - `shaders/rain.gdshader`: the same on P8d's drop colour. `ALPHA` is P8d's. 43: - `scripts/rain_veil.gd`: `LIT_REACH`, `LIT_FLOOR`, `LIT_GAIN`; the twins `lit_of(taps, night)` and `lit_colour_of(own, lit)`; `tint()` sets `lit_night` = `HeightFog.night(profile)`. 44: - `scripts/world_view.gd`: `RAIN_NEAR_CAP` 150, and the amount is `mini(maxi(1, rain × 2600), 150)`. `sync_profile` sets the drops' `lit_night`. `debug_rain_draw` is a diagnostic: false draws neither drops nor veil. 45: - `scripts/main.gd` (test-only): `--rain-draw=off` sets that diagnostic, for the bench. 46: - No saved key, save version 2 untouched, draw calls unchanged. 47: - **Third fix pass (8 October 2026):** `scripts/world_view.gd`: `RAIN_NEAR_CAP` 150 -> **300**, new `RAIN_NEAR_FRONT` 2.0 and `static func rain_near_volume(target, yaw, pitch, size, aspect, shift)`: a box turned with the camera (`CameraRig.basis_of`), from 2 m in front of the eye back by `size` m, spanning the frame plus a drop's drift over its life at a full gust across, and the frame plus a drop's fall over its life up; `_place_rain` uses it under the orthographic view (following `camera.v_offset`), P8d's `rain_volume` under the perspective lens, the old fixed box in photo mode. `shaders/lit_rain.gdshaderinc`: `LIT_FILMIC_WHITE` and `lit_shown(c)` = sRGB(filmic(cubic(c))), the canal look's measured colour path (`f/fix2/an/model.js`); `lit_tap` takes the luma of `lit_shown(c)` and still adds c x (l − 0.35) / l. `scripts/rain_veil.gd`: the twin `shown_of`, `lit_of` on it, `LIT_FILMIC_WHITE`. File backups in `f/fix3/backup/`. 48: - **Deviations from the lens and plan, each with its reason (measured, `f/metrics/lit-qn-*`):** 49: - **The drops are lit too.** The plan names only the veil. At Quarter zoom P8e draws the veil only over the sky and the far water, so no veil stands in front of the BAR sign. 50: - **The 4 taps are summed, not averaged.** Round 1 averaged them (×0.25). In the quarter-night clip the lit rain then stayed blue: its delta hue was 192° next to pink sources (pre 199°) and 206° next to warm ones (pre 218°) (`lit-qn-*-r1.txt`). 51: - **The rain's own colour gives way to the light, and is scaled into range instead of clipped.** Round 2 added the light and clipped each channel. The light sources' cores are near-white (the BAR letters store 255,240,255; a lamp 255,255,146), so drops beside them went white and kept their blue cast. In the BAR box the delta hue was 238° (`lit-qn-bar-r2.txt`). 52: - **Lit only from the lit evening on** (`HeightFog.night`; 1 at 05:45, 20:45 and night, 0 at 19:30, 13:00 and golden hour). By day every daylit surface is over the 0.35 floor. 53: - **As built, the 150 drops kept P8d's whole volume** (no near-only box). Superseded by the third fix pass: 300 drops in a near box under the orthographic view (the plan's "keep at most 150 near drops" in number became 300 to meet the ruling's density; P8d's volume stays for the perspective lens and its checks stand on `rain_volume`). The floor on the shown colour is a second deviation from the lens's "brightness above 0.35" of the copy, made under the ruling. 54: - **Tests:** the test-first logs are `f/testfirst-test_wet.log(.out)` (427 checks, 18 failures) and `f/testfirst-test_water.log(.out)` (462, 1). The affected-suite run `f/after2-test_camera.log.out` (261, 1) found a third P8d amount check, and I inverted it too. Runs during the rounds: `f/after1-`, `after2-`, `after3-`, `after4-` and `after5-*`. `test_wet` 403 -> 439, with new checks on: the include (taps, floor, gain, sum, the day skip, no TIME); the twin against an independent port of the lens formula (pink and amber in their own hue, the floor, ×4 for four taps, nothing at night 0); `lit_colour` (own colour exactly when unlit, scaled not clipped, giving way by the light); both shaders' `ALBEDO` lines and unchanged `ALPHA`; the cap at rain 0.02, 0.05, 0.35, 0.45 and 1.0; `lit_night` at 01:00, 20:45, 13:00 and 19:30 on both materials; dry and Glade draw nothing; the saved scene unchanged. 55: - **Third fix pass:** test-first logs `f/fix3/testfirst-test_wet.log.out` (445 checks, 13 failures), `testfirst-test_water` (462, 1), `testfirst-test_camera` (261, 7). After: `f/fix3/after2-*` and, on the final code (one comment in `rain_veil.gd` changed after the full suite; `f/fix3/code-md5.txt` -> `final-code-md5.txt`), `f/fix3/final-test_wet` **475, 0**, `final-test_water` **462, 0**, `final-test_camera` **261, 0**. `test_wet` 439 -> 475: `near_rain_checks` (the box turns with the camera, starts past the near plane, stays ≥40 m in front of the target with a drop's fall, every frame ray crosses it at shift 0 and 0.7, it holds a full gust's drift and a life's fall, P8d's speeds; a seeded model of where the drops are at a random moment: on the quarter frame 117 near drops against 162 for P8d's 909 and 27 as built, must be at least half of P8d and twice the as-built at all three zooms; the view uses the near box, the perspective lens P8d's volume); the shown colour (an independent port matched to the measured path at grey 0.15, 0.35, 0.5, 0.8; the twin against it; a tap that shows at the floor, stored grey 0.278, adds nothing; stored 0.35 and 0.30 now light the rain); the real Glade check (item c, below). **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh f/fix3/suite` a/b/v, 01:45-01:54, AC 100%, Low Power Mode 0, `f/fix3/suite-power.txt`): **44 suites, 23,554 checks, 0 failures** (`f/fix3/suite/suite_all.out`); `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 5,583 checks, 0 failures. 1440x900 not run (inherited stall). 56: - **Full suite as built** (`VERIFY_RES=2880x1800 tools/suite_all.sh f/suite` a/b/v, 23:50-23:59, AC 100%, Low Power Mode 0, `f/suite-power.txt`): **44 suites, 23,518 checks, 0 failures** (`f/suite/suite_all.out`). 57: - `--verify` at 2880x1800: exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,583 checks and 0 failures; the known flaky check did not fire. The 1440x900 `--verify` was not run (inherited stall). 58: - The code did not change from the final captures to the suite (`f/caps-code-md5.txt` = `f/final-code-md5.txt`). 59: - **Changed expectations (old -> new), inverted deliberately (P8d's amount checks):** 60: - `tests/test_wet.gd` rain_view_checks: "P8d: the default amount is unchanged (rain × 2600 = 909)" -> capped at 150. 61: - `tests/test_water.gd`:428: "P3b: the rain itself still falls (rain × 2600 drops)" -> it still falls, 150 drops. 62: - `tests/test_camera.gd`:860: "P8d: the rain setting gives back its amount (909)" -> 150. 63: - **Third fix pass:** the three amount checks 150 -> **300** (`test_wet` `RAIN_CAP_WANT`, `test_water`:430, `tests/test_camera.gd`:864), and their messages and comments "was 910" -> "was 909". `test_wet` rain_view_checks "the view's rain stands where rain_volume puts it under the scene's ceiling" -> where `rain_near_volume` puts it; "the drops are spawned over the scene's tallest content" -> "the near box stands ≥40 m in front of the target". `test_camera` rain_zoom_checks, Harbour, quarter and lane: the same two (was rain_volume and "spawned over the ceiling by the zoom's 20 m band"). `test_wet` lit_rain_checks: "each tap's luma" -> the luma of `lit_shown`; "a tap exactly at the floor adds nothing" (stored 0.35) -> a tap that shows at the floor (stored 0.278); "by x1.2 its luma over 0.35" -> by its shown luma over 0.35. "Glade (rain 0) draws neither" (review item c: the rain was already 0 when the look was set, so it proved nothing) -> from a raining canal night (drops and veil drawn), `LookProfile.choose(..., "glade")` writes rain 0 and the view draws neither, `lit_night` 0. 64: - **Expected journey count deltas:** none (third fix pass: counts identical on all 12 views in both pairs, `f/fix3/metrics/counts-*.txt`). Draw calls and rendered objects are identical before and after on all 12 views, in both pairs (`f/metrics/counts-b1-a1.txt`, `counts-b2-a2.txt`). The pixels change on every rainy view: there are far fewer drops (by day too), and at night the rain is lit. 65: - **Captures after the third fix pass** (`f/fix3/caps3.sh`: before-1, after-1, before-2, after-2 interleaved in one sitting, 01:31-01:37 on 8 October, 1600x1000, AC 100%, Low Power Mode 0, `f/fix3/cap-power.txt`; 12 views including quarter-0545 and quarter-2045; before = `f/pre-tree` = `26-P8-f-pre`). The as-built sets are `f/before-*`, `f/after-*` (metrics in `f/metrics/`). 66: - **Frame share over 8 luma** (`f/fix3/metrics/frame-table.txt`, as b1>a1 / b2>a2 against the noise legs b1>b2 / a1>a2): 67: - glade-day 0.29/0.34 against 0.27/0.34; glade-night 0.45/0.37 against 0.49/0.34 (unchanged). 68: - quarter-day 0.62/1.07 against 0.99/0.45 (as built 3.70/2.91: with the near box the day frame has about as many drops as before P8f). 69: - quarter-night 0.52/0.46 against 0.39/0.26; harbour-night 1.25/1.09 against 1.26/0.97; lane-night 1.39/1.37 against 1.17/1.23. 70: - quarter-2045 2.13/0.94 against 1.37/0.36 and quarter-0545 1.05/0.79 against 0.55/0.29: over noise (the horizon, below). 71: - quarter-dry 0.28/0.37 against 0.14/0.04: see "Dry scene" below. 72: - Draw calls and every integer count identical on all 12 views in both pairs (`f/fix3/metrics/counts-b1-a1.txt`, `counts-b2-a2.txt`). 73: - **Protected looks** (`f/fix3/metrics/protected.txt`, from `ab-b1-a1`, `ab-b2-a2` against the noise pairs `ab-b1-b2`, `ab-a1-a2`): 74: - **The 05:45 and 20:45 horizon band changes above noise; its p50 and the zenith p50 do not.** 20:45 (quarter-2045): horizon_band changed 6.80% / 6.69% (>8: 0.54% / 0.54%) against noise 0.54% / 0.38% (>8: 0.35% / 0.35%), p50 151 -> 151.2 and 151.1 = 151.1; the zenith also changes above noise now, 7.23% / 7.41% (>8: 0.20% / 0.37%) against 0.45% / 0%, p50 97.7 = 97.7. 05:45 (quarter-0545): horizon_band 7.44% / 6.21% (>8: 1.23% / 0.05%) against 0.99% / 0.35%, p50 113.9 = 113.9; zenith 0.39% / 0.39% against 0.67% / 0% (at noise), p50 56.6 = 56.6. As built (second review's numbers) the 20:45 band was 6.65% (>8 0.59%) against noise 1.13% and 0.59%, and the 05:45 band 3.04% against 1.08% and 0.10%: the shown floor brings more of the twilight sky over 0.35, so the veil streaks over it take its own colour (the sky's hue, a few levels). Looked at in `f/fix3/look/b2-a2-protected.png`. 75: - bar_sign (quarter-night): b1>a1 p50 70.6 = 70.6, changed 12.4% (>8 0%), the same as its noise pair b1>b2. b2>a2 reads p50 70.6 -> 69.6, changed 64.1% (>8 2.29%), and so does the noise pair a1>a2: the a2 capture's sign body is a shade darker than in the other three captures (`f/fix3/look/a1-a2-bar.png`), a capture-state difference inside one tree, not this part. 76: - fish_sign (harbour-night) p50 66.3 = 66.3 in both pairs. Canal mirror columns p50: quarter-night 43.5 -> 43.6 and 43.4 -> 43.5; 20:45 57.5 -> 57.3 and 57.4 -> 57.3; 05:45 54.9 -> 54.8 and 55.0 -> 54.9 (noise legs 43.5 -> 43.4, 57.5 -> 57.4, 54.9 -> 55.0). 77: - **Dry scene (second review item d), checked:** ten quarter-dry captures of one sitting (the four above plus `f/fix3/dry-pre-1..3` and `dry-live-1..3`, interleaved), all 45 pairs (`f/fix3/an/dry.js` -> `f/fix3/metrics/dry-pairs.txt`). Mirror columns change 0.10% to 35.7% between two captures of the same tree (20 pairs) and 1.08% to 41.2% between the trees (25 pairs); the most alike pair of all is one from each tree (before-2 and dry-live-3: frame 0.15%, mirror columns 1.08%). Paving and tower windows are 0% in every pair. The capture's `elapsed_ms` gap does not explain it: a same-tree pair 1 ms apart differs on 24.6% of the mirror columns, another 7 ms apart on 0.10%. So the as-built entry's quarter-dry "over noise" came from a lucky noise pair: the water, mirror and horizon state at the moment of capture varies from run to run in either tree, by a source I did not find (not the elapsed time). In code nothing is drawn at rain 0 (tests). 78: - **Recordings** (`f/fix3/rec3.sh` -> `f/rec.sh`, `tools/art_record.sh --keep-frames`, quarter-night 6 s at 30 fps, 1440x900, 01:20-01:28, AC 100%, Low Power Mode 0, `f/fix3/rec/power.txt`): fresh `f/fix3/rec/qn-pre` (pre-tree), `qn-off` (live with `--rain-draw=off`), `qn-live` (live, cap 300, recorded again at 01:27 after the cap change; the cap-250 clip and its scans are in `f/fix3/cap250/`). Aligned (`f/fix3/metrics/align.txt`): live and pre against off 1.6-2.1% at offset 0 and 3.9-13% elsewhere; the new off clip equals the kept `f/rec/qn-off` exactly (0.00%); the copy clip `f/fix2/rec/qn-copy` aligns with the new off clip at offset 0 only. 79: - Three classings of the same rain pixels (clip minus off, on the art grid, still pixels only; `f/fix3/an/regions3.sh`): `an/litscan.js` (taps from the final frame, the entry's first table), `fix2/an/litscan2.js` (taps from the screen copy, floor on the stored luma: what the as-built shader saw) and `fix3/an/litscan3.js` (taps from the copy, floor on the shown luma: what the fix-3 shader sees). Outputs `f/fix3/metrics/lit-qn-regions-{final,copy,shown}.txt`, summary `f/fix3/metrics/lit-qn-summary.txt`; the as-built clip under the shown classing `lit-qn-asbuilt-shown.txt`. 80: - **Acceptance, against the target:** 81: - **Rain turns pink before the BAR sign and amber near lamps: MET after the third fix pass, for every light the screen copy holds.** Quarter-night clip, 180 frames: 82: - Density (all rain pixels, the three classings agree within 2%): island 300,300,1300,800 **83,142 against 128,160 before P8f = 1.54x fewer** (target: no worse than about 2x; as built 6.01x, 18,877 against 113,405, `f/metrics/owner-sheet-asbuilt.txt`; with 250 drops 2.16x, `f/fix3/cap250/`). BAR sign box 940,490,1010,640: **936 rain pixels** (987 under the shown classing) against 2,739 before (target: several hundred; as built 36). Other bands: left 1.48x, right 1.32x, bottom 1.40x, skyline 1.08x fewer. 83: - Colour, classed as the fix-3 shader reads (shown floor): island next to pink light **291°** (before P8f 217°, as built 229°) on 2,655 px, 39% pinkish, magenta +8.31 (+1.24); next to warm light **29°** (220°; as built 30°) on 5,482 px, 43% amber, warmth +24.50 (−11.75). BAR box **332°** (218°) on 273 px, 70% pinkish, magenta +15.12 (as built 339° on 16 px). Left quay-lamp band warm 32° on 1,943 px, 55% amber. Whole frame pink 310°, warm 31°. Island rain pixels next to copy light: 8,137 (as built 1,294; before 13,032). 84: - Classed by the final frame (the entry's first method, which counts light the copy cannot hold): island pink 270° (216°) on 4,037 px, magenta +5.62; warm 351° (222°) on 10,094 px, warmth +2.96 (−15.17), 24% amber. That light (the engine glow and pixel halo, the water's reflections, the toe grade) is about half of the island's light in the final frame; the shown floor reaches 52.3% of it against 30.7% as built (`f/fix2/decomp-island.txt`). The rest stays out of reach of any screen copy, as the ruling allows. 85: - **As built: why the island rain stayed blue (second fix pass, `f/fix2/`), kept for the record:** 86: - Classed by the copy itself, the island's rain does take the light. In the island region the rain next to pink light is **318°** (pre 218°), 68.6% pinkish, magenta +14.31. Next to warm light it is **30°** (pre 223°), warmth +27.07, 41% amber. The BAR box is **339°** on 16 px (`f/fix2/lit-qn-regions-copy.txt`, from `an/litscan2.js` with the copy clip `f/fix2/rec/qn-copy`). 87: - Only **30.7%** of the island pixels that the final frame shows over 0.35 are over the floor in the copy (`decomp-island.txt`). The 69.3-point drop from 100% splits as follows (the first three are light added after the copy, which the copy cannot hold): 88: - fog quad: 0.4 points (100 -> 99.6%); 89: - engine glow and the P5f pixel halo: 23.1 points (99.6 -> 76.5%); 90: - later passes, mainly the water's reflections and the post's toe grade: 24.2 points (76.5 -> 52.3%); 91: - the floor read on the stored colour, not the shown one, 21.6 points (52.3 -> 30.7%). The filmic tonemap brightens the mid-tones after the copy, so a stored luma of 0.35 shows at about 0.45 in grey. 92: - The whole frame gives 25.2%, the skyline 22.0% and the BAR box 44.4%. 93: - Few rain pixels sit next to copy light on the island: 70 pink-class and 870 warm-class over 180 frames, against 2,062 and 7,061 before (the 150-drop cap). 94: - How it was measured: scratch trees `f/fix2/diag-*`, built from `26-P8-f-asbuilt`; the live code is unchanged. In one tree the veil draws the screen copy last, with the presentation pass as the identity (`cap-copy2`, `rec/qn-copy`). In another it draws masks of the copy's luma (`cap-mask`). A third draws an albedo ramp (`cap-ramp`). 95: - The colour path model fits the ramp: shown = sRGB(filmic(cubic(albedo))), canal look, mean error 1-2 levels of 255 up to albedo 1 (`rampfit-day-le1.txt`). Inverting it gives the copy, and the inverted copy agrees with the mask on 99.96-99.99% of pixels. 96: - The copy clip is aligned in time with the off clip: its frame-to-frame changes overlap the off clip's at offset 0 (2.9-3.2%) and almost nowhere else (≤0.2%; `align-copy-off.txt`). 97: - Images looked at: `cap-mask`, `cap-copy`, `cap-copy2`, `cap-noglow` (all `quarter-night/courtyard.png`) and `off-f150.png`. In the mask the island's windows, blossoms and awnings mostly sit under the floor; the lamp heads and the neon pass it. 98: - The stills hardly show the lit rain: drops are random per still, and the veil appears in front of the island only on its sparse near layer. 99: - Looked at: `f/fix3/look/rain-bar-pre-asbuilt-fix3-f150-165.png` and `rain-island-pre-asbuilt-fix3-f150-165.png` (the rain alone, |clip − off| x4, frames 150-165; left before P8f, middle as built, right fix 3: the as-built rain is sparse, the fix-3 rain is near the pre density and has pink and orange drops by the sign and the lamps); `bar-pre-asbuilt-fix3-f150.png`, `island-pre-asbuilt-fix3-f210.png` (the clip frames, same order: single 1-px drops barely show at 1x). 100: - **Drops plus veil at most 0.5 ms at 1920x1080:** met. Live against live with `--rain-draw=off`, 01:00, rain 0.45, paired medians (`f/fix3/bench-summary.txt`): dense p50 **+0.061** (3/3), p95 +0.002; Harbour p50 **+0.133** (3/3), p95 +0.169 (2/3). Draw calls 2587 -> 2585 and 917 -> 915. 101: - **CPU particle time lower:** met on the drops' whole per-frame cost; the CPU share alone is not resolved (second review item e). `f/fix3/probe/run_cpu.sh` (the as-built `probe_cpu.gd`: each tree's drops node alone in a bare windowed scene, drops on and off in 400-frame phases, 3 rounds, run twice per tree; `cpu-*-1.out`, `cpu-*-2.out`; 01:44-01:46, AC, Low Power Mode 0). Two figures, both stated: 102: - Frame wall time, on minus off: live 300 drops **0.054 ms** (median of 6, 0.047-0.070); pre-part 1170 drops **0.107 ms** (0.089-0.125). As built 150 drops 0.046 ms and 1170 drops 0.26 ms, at a higher load: these six-round medians rest on `f/probe/summary.txt`, whose first run per label was kept only in the session transcript; the `f/probe/cpu-*.out` files hold only the last run (frame-wall medians live 0.0425, pre8f 0.199), so they cannot be fully re-derived from files. This is the drops' whole cost per frame with vsync off: the CPU simulation and buffer fill plus the GPU draw of their quads. It is lower than the pre-part tree's in this sitting and below P8d's 0.26 ms. 103: - `Performance.TIME_PROCESS`, on minus off (the main loop's process time, where CPUParticles3D simulates): live per round −0.74 to +0.13 ms (run medians −0.001 and 0.024), pre-part −2.04 to +4.26 ms (2.98 and −0.45). Whole 400-frame phases jump by 0.5-4 ms in either tree, so its sign flips; in the quiet rounds it reads 0.02-0.13 ms live and 0.07 ms pre-part. It cannot separate a CPU difference under 0.1 ms, so "CPU particle time lower" rests on the wall time and the drop count (300 against 1170 simulated). 104: - **Glade, dry, 05:45 and 20:45, neon cores:** Glade unchanged (frame share at noise, counts identical). Dry: see "Dry scene" above. 05:45 and 20:45: the horizon band's changed share is above noise, p50s unchanged (Protected looks). The neon cores show near 1 on either floor, so the lit rain beside them is as it was (BAR box 332° against 339° as built). 105: - **Flashing (WCAG 2.3.1, D17):** none. Over the fresh clips the mean luma of the BAR box, two lamp boxes and the island moves at most 0.09-0.49 a frame (pre 0.09-0.75), their brightest 10% at most 0.18-1.68 (pre 0.19-2.37), 0 flash pairs (`f/fix3/metrics/flash-qn.txt`). 106: - **Reduced and Off:** unchanged in code: the near box changes where drops are born, not how they move; the veil's drift and the drops' lean keep their motion terms (P8e, P8i2 tests pass). 107: - **Bench** (`f/bench.sh` with `BENCHDIR=fix3/bench HOUR=1 ROUNDS="0 1 2 3" TREES="pre8f live off"`, `f/fix3/bench-t1.out`, `f/fix3/bench-summary.txt`; round 0 discarded; 1920x1080, vsync off, rain 0.45, `--motion-live`, `--light-pin=20`; 01:38-01:44, every run on try 1, AC 100%, Low Power Mode 0, load 1.9-4.4, `f/fix3/bench/power.txt`): 108: - This part against the pre-part tree (live against 26-P8-e), paired medians: dense p50 −0.074 (0/3 slower), p95 −0.192 (1/3); Harbour p50 −0.058 (0/3), p95 −0.014 (1/3). Draw calls equal. 109: - Drops plus veil (live against off): above. The as-built sitting's numbers (23:30-23:46, 7 October, including 13:00 and 26-P8-III) are in `f/bench-summary.txt`; only 01:00 was re-run, the lit case. 110: - **For the owner sheet:** 111: - **What the orchestrator ruled** (`artifacts/beautify/owner-decisions.md`, "Orchestrator ruling on P8f lit rain", 8 October, under the standing approval): take "more rain in front of the island" and "read the floor on the shown colour" together; "stop lit rain at the skyline" rejected. Both are built: the CPU drops (300) stand in a near box in front of the camera, and the floor is read on the shown colour. 112: - Island rain is back near its pre-P8f density: 1.54x fewer rain pixels on the island (300,300,1300,800), 1.64x in the foreground (0,600,700,900), 1.07x in the sky (0,0,1440,200) (`f/fix3/metrics/lit-qn-summary.txt`, "owner" lines). As built it was 6.01x, 7.89x and 1.12x (`f/metrics/owner-sheet-asbuilt.txt`, reproduced from the kept as-built clips by `f/fix3/an/owner-asbuilt.sh`; the first figures, 6.0x, 7.9x and 1.1x, had no saved output). 113: - Rain before the BAR sign turns pink (332°; about 5.5 rain pixels a frame in the box, 1.5 of them next to pink light), and by lamps and windows amber (29-32°). 114: - What remains for the owner: lit rain is still subtle, 1-px drops at alpha 0.16-0.34, pale next to the near-white cores. Light added after the copy (the glow halos, the water's reflections) does not tint the rain. At light rain (below 0.115) the near box shows about 2.2x as many drops on the frame (the box model) as P8d did at the same setting, because the amount is still rain x 2600 up to the cap. The near drops are 2-60 m from the eye (before: up to about 150 m), so the distance fog (from 100 m) no longer touches them. All of these are taste calls with the current defaults kept. 115: - **Not done:** no harbour-night or lane clip; no light-rain clip; the perspective lens keeps P8d's whole volume (not measured since the cap); no photo-mode, Reduced or Off clip; Glade with its rain raised at night reads the floor through the canal look's filmic curve, not its own AgX (untested); no start-up timing; the Windows iGPU cost is unmeasured; `--verify` at 1440x900 not run (inherited stall). 116: - **Fix pass after the second review (8 October 2026):** one must-fix finding. The lit-rain numbers (316° and 45°) were whole-frame averages. About two thirds of their weight came from veil streaks over the skyline band, so they did not describe rain in front of the BAR sign or near the island's lamps. What I fixed: the acceptance line now gives the numbers per region, is marked partly met and triggers amendment 7 and an owner decision (above). No code changed: `diff -rq` of scripts shaders tests data scenes tools and `project.godot` against `build/backup-feel/26-P8-f-asbuilt/` was empty, and the six code files' md5 equal `f/final-code-md5.txt`. The scan used the kept clips `rec/qn-live`, `qn-pre` and `qn-off`, with no new recording. Full suite re-run (`VERIFY_RES=2880x1800 tools/suite_all.sh f/suite-fix` a/b/v, 00:27-00:35, AC 100%, Low Power Mode 0, `f/suite-fix-power.txt`): **44 suites, 23,518 checks, 0 failures** (`f/suite-fix/suite_all.out`); `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun had 5,583 checks and 0 failures. 1440x900 not run (inherited stall). 117: - **Second fix pass after the recheck (8 October 2026):** one must-fix finding, still open after the first fix pass. The acceptance line was partly met, and the reason the island's rain stays blue had not been looked into. What I did: 118: - I measured the cause (see "Why the island rain stays blue" under Acceptance). The lit rain works on every pixel where the screen copy holds light. Most of the island light the scan counted arrives after the copy: glow and halo, reflections and the grade, and the tonemap's brightening of the mid-tones. 119: - I updated the owner options with what a floor on the shown colour would reach (52.3% of the island's sources against 30.7%). 120: - I changed no code, because each remaining lever is a design change or an owner call: more rain on the island, a floor on the shown colour, or lit rain limited to the skyline. The acceptance line stays **PARTLY MET**. Amendment 7's one-tier-up re-run and the owner decision stay open. 121: - The live tree still equals `build/backup-feel/26-P8-f-asbuilt/` (`diff -rq` of scripts shaders tests data scenes tools and `project.godot` is empty). The file backups in `f/fix2/backup/` were not needed. 122: - Captures and the copy clip: 00:40-00:49, AC 100%, Low Power Mode 0 (`f/fix2/cap-power.txt`). The copy clip's frames (`f/fix2/rec/qn-copy/frames`, 140 MB) are kept for the second review and are to be moved to the Trash after it. 123: - Full suite re-run (`VERIFY_RES=2880x1800 tools/suite_all.sh f/suite-fix2` a/b/v, 00:55-01:04, AC 100%, Low Power Mode 0, `f/suite-fix2-power.txt`): **44 suites, 23,518 checks, 0 failures** (`f/suite-fix2/suite_all.out`). 124: - `--verify` at 2880x1800: exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,583 checks and 0 failures. The 1440x900 `--verify` was not run (inherited stall). 125: - No expectations changed, and no journey count deltas are expected. 126: - **Third fix pass (orchestrator-directed, 8 October 2026, 01:06-02:10; ruling "Orchestrator ruling on P8f lit rain" in `artifacts/beautify/owner-decisions.md`):** two must-fix items. 127: - Item 1, asked: put the capped drops where they are seen so the island keeps rain no worse than about 2x fewer than before P8f and the BAR box gets several hundred rain pixels, and read the floor on the shown colour; keep the cost, CPU, flashing, protected-look and Reduced/Off limits. Done: the near box and cap 300 (250 was tried first: island 2.16x fewer, `f/fix3/cap250/`), and `lit_shown`. Island 1.54x fewer (as built 6.01x), BAR box 936 rain pixels (as built 36), island rain next to pink light 291° and next to warm light 29° under the shader's own classing; drops plus veil +0.061 / +0.133 ms; drops' frame cost 0.054 ms against 0.107 pre-part; no flashing; Glade and the p50s unchanged; the 05:45 and 20:45 horizon changes above noise (stated). All under Acceptance. 128: - Item 2, asked: (a) state the 05:45 and 20:45 horizon change above noise and re-measure: under Protected looks; (b) save the output behind 6.0x, 7.9x and 1.1x: `f/metrics/owner-sheet-asbuilt.txt` (6.01x, 7.89x, 1.12x) and re-measured after the fix (1.54x, 1.64x, 1.07x); (c) the Glade check made real; (d) quarter-dry checked: run-to-run capture state in either tree, not elapsed time and not the code (Dry scene); (e) both CPU figures stated (Acceptance). 129: - The live tree before the pass equalled `26-P8-f-asbuilt` (`diff -rq` of scripts shaders tests data scenes tools and `project.godot` empty); nothing unexpected. Post-pass snapshot: `build/backup-feel/26-P8-f-fix/`. Handoff backup: `build/backup-feel/visual-handoff.md.pre-P8f-fix3`. The last Godot runs were the final tests `f/fix3/final-test_wet`, `final-test_water` and `final-test_camera`, written 02:03:39-02:03:52 on 8 October. 130: - Review: filled by the orchestrator. 131: - Platforms: Mac run; Windows not run. 132: 133: **Beautify P8e Rain veil: layered, depth-gated pixel streaks (7 October 2026):** Opus 5.5 on this Mac (plan section "P8e"; report polish backlog "rain in art pixels" (the veil) and its lens finding "Replace the sparse hard rain streaks with a layered, lit pixel rain veil"; step `26-P8`, sub-stage P8-IV, first part; owner D28 default: the veil as the report designs it). Behind P8d's drops the rain now has a fine veil: three layers of faint one-art-pixel streaks (far 3-4 px, mid 6-8, near 12-16) falling at 150, 240 and 360 art px/s, each drawn only where the scene lies beyond its depth, so the far rain vanishes behind the nearer buildings. The pattern is counted from the projected world origin, so it pans with the world in whole pixels; it drifts sideways with the drops' own fall and leans 1-2 px in a gust like the drops (P8i2). Every still shows one fixed veil (light clock held at 0); Off holds it still, Reduced falls at half speed. By day it is drawn at 0.6 of its alpha; dry scenes and Glade (rain 0) draw nothing. The lit-rain taps and the CPU rain cap are P8f's, not here. Snapshot: `build/backup-feel/26-P8-e/` (made by me before the first edit because it was missing; it then equalled `26-P8-III` = `main` 523751a: `diff -rq` on scripts shaders tests data scenes tools and `project.godot` was empty, nothing unexpected). It has since been re-saved: `26-P8-e` now equals `26-P8-e-post` (the post-P8e code, 9 diffs from `26-P8-III`), so the pre-part tree is `build/backup-feel/26-P8-III/`. Post-part snapshot `build/backup-feel/26-P8-e-post/`. Pre-part scratch tree `artifacts/beautify/26-P8/e/pre-tree`. Evidence: `artifacts/beautify/26-P8/e/`; file backups in `e/backup/`. 134: - **Code:** 135: - `scripts/rain_veil.gd` (new, `class_name RainVeil`, cloned from HeightFog/CloudShadow): a camera-pinned full-screen quad, render priority 117 (after ContactShade 110 and CloudShadow 115, before HeightFog 120). Per layer (far, mid, near): `SPEED` 150/240/360 px/s, `CELL` 36/48/72 rows, `LENGTH` 3-4/6-8/12-16, `DENSITY` 0.42/0.03/0.014 (chance a column of a cell holds a streak, at `RAIN_REF` 0.35, in proportion to the rain, capped 0.9), `ALPHA` 0.08/0.11/0.16, `DEPTH` +25/0/-6 m from the target's view depth (x size / 38, as P8d's rain volume). `DAY_ALPHA` 0.6 (HeightFog.night), colour = haze lerped 0.45 toward (0.78, 0.84, 1.0). `configure(rain)`, `tint(profile)`, `frame(target_depth, size)`, `advance(dt, slope)`; static twin `density_of`, `alpha_of`, `colour_of`, `fall_of`, `period_of`, `depths_of`, `gate`, `slope_of`, `drift_of`, `veil_hash`, `streak_of`, `stair`, `layer_lit`. 136: - `shaders/rain_veil.gdshader` (new): the fog's depth rebuild (depth buffer or water plane; perspective path kept; the painted sky counts as beyond every layer); anchor = floor of the projected world origin; rows counted down the picture whatever the projection's flip; fall = floor(SPEED x `light_time` x `motion_scale`); cell index hashed mod `period_of` (SPEED x 690 / CELL: 2875, 3450, 3450), so the 1380 s wrap is seamless at Full and Reduced; a streak never uses its cell's first row (streaks in a column never touch). In a gust each streak takes `wind_rain_lean` at its head's ground point on its layer and draws P8d's one-pixel stair (up to two columns tried to the lean's side). Unshaded, `blend_mix`, `depth_test_disabled`, `fog_disabled`; no TIME. 137: - `scripts/world_view.gd`: `var veil`, made beside the cloud shadows; `veil.configure` in `sync_environment` (beside the drops' `emitting`, same 0.01 threshold), `veil.tint` in `sync_profile`, `veil.frame` in `update_camera`, `backdrop_depth` in `_apply_depths`, `_advance_veil()` each frame (the drift: the light clock's advance x `motion_scale` x `RainVeil.slope_of(camera.basis, rain_fall(rain_lean_gust))`, so it follows P8i2's gust slant and holds under the automation hold, a pin and Off). ShaderWarmup picks the quad up as a node of the view. 138: - `scripts/main.gd` (test-only): `--pan=` with `--record` slides the view along the frame's x through a recording (the crawl check). Not in the pre-tree, so there is no before pan clip. 139: - No saved key, no switch, save version 2 untouched. 140: - **Deviations from the lens design that D28's default adopts** (`artifacts/beautify-0925/lens-packets.json`, finding "Replace the sparse hard rain streaks with a layered, lit pixel rain veil"; checked against `scripts/rain_veil.gd`): the far layer is about 7x the lens's density per cell (0.42 of the columns per 36-row cell against about 6% per 40 rows; about 7.8x per row); alphas are 0.08/0.11/0.16 against 0.07/0.11/0.18; cells are 36/48/72 rows against 40 for all three; the lens's fixed shear of 1 px left per 5 rows is replaced by a sideways drift at the drops' own slope (about -0.128 px a row in the quarter-night clip) plus a lean only in gusts (P8i2). Also priority 117 and `blend_mix` where the lens had 110 and `blend_add`. Reason: the plan's acceptance of about 5% faint streak pixels, and consistency with P8d's and P8i2's drops. 141: - **Tests:** test-first logs `e/testfirst-test_wet.log(.out)` (347 checks, 19 failures: no script, shader or view veil; `layer_alpha` was called `alpha` in that log, renamed before the code), `e/testfirst-test_camera.log.out` (261, 1), `e/testfirst-test_warmup.log.out` (44, 1); no script errors. `test_wet` 328 -> 403: the twin (priority between 110/115 and 120; three layers, far shorter/slower/fainter/deeper; lengths and alpha bands; seamless wrap; fall on light_time x motion_scale, none at t 0 or Off, half Reduced; densities follow the rain; twin frame share 3.5-6.5% at 0.35 and x1.6-2.2 at 0.7; every streak one run of its length; a fall of n moves it n rows; flipped projection; drift in whole pixels; wrap; the stair; a lean neither adds nor removes pixels; depth gate; depths by zoom; day x0.6; colour; slope and drift), the shader contract (render modes, depth texture, the wind include, `light_time*motion_scale`, no TIME, `wind_rain_lean`, the projected world origin, constants equal the twin's) and the view (camera child, priorities, densities, day and night alpha, colour, layer depths, panel depth like the fog, drift held under the hold and Off and half under Reduced, hidden when dry and in Glade, saved scene unchanged). `test_camera` 260 -> 261 (the panel depth under the perspective lens), `test_warmup` 43 -> 44 (the quad is warmed). Two of my own new checks were corrected while running (`e/after1-*`, `e/after2-*`): the wind-include check read comment-stripped text ("res://" lost), and the drift check compared float32 values near 4096 exactly (now within 0.01 px). `e/after1-test_wet.log(.out)` also stops on a SCRIPT ERROR from a third bug in my checks ("Invalid type in function 'vec' ... Cannot convert argument 1 from Vector3 to Array" at `tests/test_wet.gd:1207`, veil_view_checks), gone in `after2`. **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh e/suite` a/b/v, 21:59-22:08, AC 100%, Low Power Mode 0, `e/suite-power.txt`): **44 suites, 23,482 checks, 0 failures** (`e/suite/suite_all.out`). `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. creature_reactions rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Code unchanged from the first capture to the suite (`e/final-code-md5.txt`). 142: - **Changed expectations:** none of the existing assertions changed (only new checks). **Expected journey count deltas:** draw_calls and rendered_objects **+1 on every view where it rains** (harbour-day, harbour-night, harbour-lane, quarter-day, quarter-night, styles-day); dry scenes (Glade, rain 0) identical. Pixels: every rainy view gains the faint veil, by day barely visible. 143: - **Captures** (`e/caps.sh 1 2`: before-1, after-1, before-2, after-2 interleaved, one sitting, 1600x1000, AC 100%, Low Power Mode 0, `e/cap-power.txt`; views harbour-night, quarter-night, lane-night, harbour-day, harbour-golden, quarter-day, glade-day, glade-night, quarter-dry, styles-day, quarter-2045). Counts (`e/metrics/counts-b1-a1.txt`, `-b2-a2.txt`, same deltas in both pairs): +1 draw call and rendered object on the eight rainy views (harbour-night 906 -> 907, quarter-night 1023 -> 1024, harbour-day 907 -> 908, quarter-day 1024 -> 1025, lane-night 748 -> 749, styles-day 501 -> 502, harbour-golden 829 -> 830, quarter-2045 1026 -> 1027); glade-day, glade-night and quarter-dry identical. Frame >8 luma (`e/metrics/frame-table.txt`): glade-day 0.38/0.40 against noise 0.51/0.28, glade-night 0.36/0.37 against 0.36/0.38, quarter-dry 0.29/0.05 against 0.07/0.18 (dry: unchanged); rainy views 1.8-9.8 (lowest lane-night b1>a1 1.777) against noise 0.3-8.7. Protected looks (`e/metrics/ab-b2-a2.txt` against `ab-b1-b2.txt`): fish and bar neon cores p50 within 0.1, canal mirror columns p50 43.4 = 43.4 (quarter-night) and 57.3 -> 57.5 (quarter-2045), zenith and horizon p50 within 0.2 except the quarter-2045 horizon band, 150.5 -> 151 (+0.5) (sky means +0.2-0.8, the most quarter-2045's zenith 98.5 -> 99.3: the veil). The protected 05:45 sky (added afterwards, 22:40, one capture per tree, pre-tree then live, same settings, AC 100%, Low Power Mode 0, `e/cap-0545-power.txt`; `e/before-0545`, `e/after-0545`, `e/metrics/ab-0545.txt`, `counts-0545.txt`): it rains there (the artwork scene's rain 0.45, lighting Night), zenith p50 56.6 = 56.6 (`#2b3a69` -> `#2c3b6a`, mean 58.0 -> 58.9), horizon p50 113.6 -> 113.8 (`#766791` -> `#766892`), frame >8 luma 1.45% (no noise pair at 05:45), draw calls 996 -> 997. Looked at: `e/look/sky-quarter-night-b1-a1.png`, `sky-harbour-night-b1-a1.png`, `sky-harbour-day-b2-a2.png`, `e/try-1/qn-crop-r.png`, `e/try-1/ln.png` (lane-night), `e/look/qn-after2-streakmap-tl.png`, `e/look/panoff-f200-202.png`. 144: - **Recordings** (`e/rec.sh`, `tools/art_record.sh --keep-frames`, 1440x900, AC, `e/rec/power.txt`; frames measured, then moved to the Trash; clips, contact sheets and per-pair tables kept): quarter-night 10 s pre and live at 30 fps; live under Reduced and Off (6 s); two pans (`--pan=13` under Off, `--pan=30` at Full, 6 s); harbour-night pre and live at 15 fps over light_time 6-22 s (a gust). Scanners `e/an/veildiff.js` (veil = live - pre at the same frame number, same TIME, 3-14 luma) and `e/an/lean.js`; summary `e/metrics/clips.txt`. (`e/an/veilscan.js` and `an/out-veilscan.txt` are superseded: its median background counted the clouds' dither.) 145: - **Acceptance, against the target:** 146: - **Faint streak pixels in the qn sky crop about 5% (today 1.2%, reference 5.2%), hard at most 1%:** met. `e/an/streaks.js` re-builds the lens's vertical-line detector on the art grid (a pixel at least 4 (faint) or 25 (hard) luma above both horizontal neighbours, with one above or below too); `e/metrics/streaks.txt`. Quarter-night, the report's top-left crop (0,0,300,170): before 1.56/1.16% faint, after **4.70/4.81%**; hard 0.23/0.16 -> **0.05/0.12%**. The kit's `sky_right` (1250,0,1600,175): before 0.17/1.25%, after **4.72/4.65%**, hard 0.02-0.18%. The audit capture gives 1.25% faint (all hard) on the same detector. About 0.2-1.3 points of every figure are the painted clouds' dither rows (in the before too, `look/qn-after2-streakmap-tl.png`), so the veil alone adds about 3.2-4.5 points; in the clips the veil covers 5.0-6.6% of the sky's art pixels at Full. Harbour-night (rain 0.35 against the quarter's 0.45): 0.2-0.4% -> 2.5-2.9% faint. By day (harbour-day sky, 1200,0,1600,180) 0.08 -> 0.24% faint in both pairs: the veil is nearly invisible by day. **The detector is not calibrated on the reference:** it reproduces today's 1.2% on the old quarter-night crop (audit capture 1.25%), but on `artreference.png` 0,0,300,180 it gives only 1.07/2.53/3.78% faint at `--art` 1/2/3, not the lens's 5.2% (`e/metrics/streaks-regions.txt`). So "about 5% on this detector" is not "equal to the reference", and the veil may be denser than the art reference. 147: - **Hidden behind nearer buildings:** met. Pixels brightened by 4+ luma, before-2 -> after-2 (`e/metrics/bright.txt`, `an/out-bright-qn.txt`): quarter-night sky crops +6.5% and +6.1%, the round tower (540,180,700,480) 0.18% and the right block's facade (900,280,1200,420) 0.11%, against the after noise pair 0.17% and 0.18%: nothing is drawn over the near buildings. The detector's faint share over those facades is unchanged too (4.17 -> 4.17%, 5.0 -> 4.8-5.0%). 148: - **No crawl on pan:** met. Under Off with a 13 art px/s pan, 97.8% (min 94.7%) of each frame's veil pixels are the previous frame's moved by one whole-pixel shift, 0 or -1 px (77 one-pixel steps in 179 frames = 0.43 px a frame, the pan's speed); the still Off clip gives 98.4% at shift 0 (the remainder is the drops). At Full the falling layers move 4-6 rows a frame (the far layer, 5 = 150/30) with a -1 px sideways step about one frame in two (the drops' slope -0.128 x 5); Reduced 1-4 rows. 149: - **Motion preference and flashing (D17, WCAG 2.3.1):** Off holds the veil still (98.4% unmoved), Reduced falls at half the rows a frame; the veil's share of the sky stays 5.0-6.6% frame to frame and the sky's mean luma changes at most 0.13 a frame at Full (pre 0.11; `rec/*/veildiff-sky.tsv`, `an/out-veilscan.txt`): nothing flashes. 150: - **Gust lean:** in the harbour-night clip the share of veil pixels on a stair step rises from 0.1-1.1% (6-15 s) to 4.3-5.0% (18-22 s), the gust P8i2's drops lean in. 151: - **Bench** (`e/bench.sh`, `bench-t1.out`, `bench-t13.out`, `bench-t13b.out`, `bench-summary.txt`; pre-tree against live, interleaved, round 0 discarded, 1920x1080, vsync off, rain 0.45, `--motion-live`, `--light-pin=20` (in a gust), every run on try 1, 491 samples; 21:50-21:59, AC 100%, Low Power Mode 0, `bench/power.txt`), against 26-P8-III: **01:00** dense p50 14.458 -> 14.491 (paired +0.042, 3/3), p95 +0.019 (2/3); Harbour p50 6.869 -> 6.932 (+0.057, 2/3), p95 +0.054 (2/3). **13:00** dense p50 14.249 -> 14.314 (+0.048, 3/3), p95 +0.062 (3/3); Harbour (7 rounds) p50 6.547 -> 6.620 (+0.088, 7/7), **p95 7.472 -> 8.196 (paired +0.622, 4/7)**: Harbour's p95 by day spreads in both trees. Pre, rounds 0-7: 7.208, 7.244, 8.038, 7.574, 7.472, 7.292, 7.912, 7.467 (2 of 8 in 7.9-8.4 ms). Live: 8.113, 8.215, 7.165, 8.196, 8.337, 8.403, 7.794, 7.442 (`bench-t13.out`, `bench-t13b.out`): only 5 of 8 are in the 7.9-8.4 band (8.11-8.40), one is at 7.79 and two are at 7.17 and 7.44. Raw mean 8.39 -> 8.53 ms. Not explained; the plan's stop (dense p95 over about +0.5 ms) is not hit. Draw calls +1 (2586 -> 2587, 916 -> 917; by day 2587 -> 2588, 917 -> 918). Start-up (`e/startup.sh`, `startup.out`, 22:08-22:13, same power, rounds 1-3 medians): ready_ms 7252 -> 7341, warm_ms 5077 -> 5154, all_ms 12038 -> 12145; entries 144 -> 145, pieces 310 -> 312. 152: - **For the owner sheet:** the veil by day is nearly invisible; Harbour at the default rain shows about half the quarter's veil (its rain is lower); the near water and foreground quays (nearer than every layer) get no veil, only drops; under Off the veil stands still over the scene rather than being hidden (the cloud shadows' precedent; one line hides it); P8d's drops are unchanged, so the near rain is still the drops. At Quarter zoom no building face or roof on the island shows veil: detector faint shares before -> after (`e/metrics/streaks-regions.txt`, quarter-night, both pairs) are unchanged on the tower (540,180,700,480) 4.17-4.19%, the right block (900,280,1200,420) 4.79-5.02%, the left terrace (0,500,300,700) 3.05-3.18% and the near water (0,880,500,1000) 2.43-2.72%, `mirror_seam` 3.48-3.68%. The veil appears only over the sky, the far water (0,200,280,280) 0.73-1.30% -> 5.09-5.29% and (1250,210,1600,300) 1.52-1.73% -> 5.83-6.04%, and a little of the back roofs (`roof_terraces`, which includes some sky, 2.88-3.32% -> 3.83-3.91%). The veil reads as sky rain only. 153: - **Not done:** no before pan clip (the pre-tree has no `--pan`); no perspective, photo-mode, lane or Glade-in-rain clip; the Harbour day p95 tail not investigated further; the lit rain and the CPU cap are P8f. Windows cost unmeasured. 154: - Review: filled by the orchestrator. 155: - Platforms: Mac run; Windows not run. 156: 157: **Beautify P8-III (parts g1, g2, h, i1, i2, l, j, k of P8) stage close (7 October 2026):** this is the stage-end overhead of sub-stage P8-III "Motion" (step `26-P8-III` in the plan's overhead table). Opus 5.5 ran it on this Mac on the final P8-III code: `main` fa7f38d, clean (`close/head.txt`), equal to `build/backup-feel/26-P8-III-post`. The parts are P8g1, P8g2, P8h, P8i1, P8i2, P8l, P8j and P8k, plus the P8 fix pass. No game code changed. The evidence is in `artifacts/beautify/26-P8-III/close/`. 158: - **Before side:** the pre-stage scratch tree `close/pre-tree`, built with `tools/feel_snapshot.sh tree 26-P8-II`. That snapshot equals `build/backup-feel/26-P8-g1`, which is the P8-II close code. Each of the 46 files changed since `2de5e65` that exists in the pre-tree equals `2de5e65` (`close/git-changed-since-2de5e65.txt`). The pre-tree differs from live only in P8-III's files and in `project.godot`, which gains the `art_pixel_slope` and `wind_strength` globals (`close/diff-pretree.txt`). 159: - **Power:** every run was on AC, 100% charged, with Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). Load average was 2.9-6.6 during the captures and 2.8-4.2 during the bench. 160: - **Full suite** (`close/suite.sh` a, b, v; 19:46-19:54): **44 suites, 23,405 checks, 0 failures** (`close/suite/suite_all.out`; the new suites are `test_cloth` 92 and `test_cloud_shadow` 77). `--verify` at 2880x1800 exited 0 on try 1 in 34 s, with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,583 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 161: - **Journey** (`artifacts/feel-journey/26-P8-III/`; six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): 162: - Counts against 26-P8-II (`26-P8-III/counts-vs-26-P8-II.txt`): draw calls and rendered objects are harbour-day +10, harbour-night +9, harbour-lane +8, quarter-day +13, quarter-night +12 and styles-day +4. That is the sum of the declared deltas: P8h +6 at Harbour and +9 in the quarter, P8i1 +3, and P8l +1 on day views. P8h did not declare harbour-lane; its +4 there is P8i1's before count (`26-P8/i1/before-1/lane-day`, 745) minus the 26-P8-II journey (741). A single-view re-run of harbour-lane and harbour-night (`close/rerun.sh`, `close/rerun/journey-live/`) gave identical counts, and the stage capture shows the same deltas in both pairs. 163: - Frame >8 luma against 26-P8-II is 3.7-20.8% (`26-P8-III/frame-vs-26-P8-II.txt`); the re-run against the journey within one tree is 0.7-1.7%. The day views change by the cloud shadows. At night the changes are rain rings, reflections, neon columns, cloud dither and smoke, all on `TIME` (`close/look/journey-diff-*.png`, looked at). 164: - `notes.json` has a `26-P8-III` entry (`close/journey-note.js`, backup `notes.json.bak-before-26-P8-III`). `index.html` and `journey-standalone.html` (109.39 MB) were rebuilt. 165: - **Art capture and protected looks** (`close/caps.sh`; 19:56-20:42, every view of every set on try 1). Groups g1 (58 views) and g2 (30) are the P8-II close's. Group g3 is P8-II's 11 views plus `harbour-2015` (P8l's dusk view) and `qn-creatures` (quarter night with `--creatures`). That makes 101 views per set, captured as pre-1, after, pre-2, after-2. 166: - The metrics are in `close/metrics/` (`metrics.sh`, `frame-table.js`, `prot.js`, `consistent.js`, `elapsed.js`; regions `close/regions.json` = the P8-II close's). Against the previous stage close, `26-P8-II/close/after`, they are `ab-p8iiclose-*` and `frame-p8iiclose-*`. 167: - The protected sheets are `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P8-II-close.png`, `noise-pre.png` and `noise-after.png`. I looked at page 4. 168: - **Counts:** both pairs show the same deltas, and they never change within a tree apart from the orbit views (`qn-orbit`, `quarter-orbit`, as at earlier closes). Night-only scenes without cloth are identical: banks-night, beached-night, sheet-night, shops-night and the three rain fields. qn-creatures goes from 30 to 34 creatures. The P8-II close's after against today's pre-1 differs only in the two orbit views. 169: - **Frame changes:** 86 of 101 views are above both noise legs (`frame-table.txt`). The live capture is later in both pairs in 100 of 101 views, with a median of 296 ms (`elapsed-ms.txt`). The final code starts later, so every `TIME` term (rings, reflections, neon columns, cloud dither) sits at another phase. That explains the night maps (`look/consistent/*.png`; harbour-night 2.9%, quarter-night 2.4%, glade-night 0.25%). By day and at golden hour, the consistent change is the cloud-shadow pattern: harbour-day 11.3%, harbour-golden 21.8%, grove-1930 37.1%, and amsterdam-day and -night 24.6% (that scene keeps its own daytime, so its "night" capture is lit by day). 170: - **Canal mirror columns:** p50 43.7 → 43.6 (43.6 → 43.5) in quarter-night and 52.2 → 52.6 (52.1 → 52.6) at 21:00. The changed shares are 60%/56% and 57%/37%, against a pre-tree noise pair of 58% and 53%. The pre-tree's own runs differ that much. 171: - **HDR neon cores:** within noise. Fish 66.3 = 66.3, bar 70.6 = 70.6, lane sign 140.1 → 140.4 (pre-2 140.2 → 140.3), tower windows 96.0 = 96.0. 172: - **05:45 and 20:45 skies:** p50 within ±0.3, and the hex is unchanged. 173: - **Glade/Canal separation:** Glade night zenith `#001d45` and horizon unchanged. Glade lawn and bank 0%, and Harbour night zenith `#2b337f`. The Harbour night horizon and water change 14-19% against a pre noise pair of 11-19% (rings and reflections). 174: - **Approved tree style:** grove-day crowns 55.1 → 55.3, conifers 55.6 = 55.6 and feet 59.5 → 59.4. The changed pixels match the pre-tree's own noise (6.1%), which comes from the sub-pixel sway that P8g1 removed; the final code's pair shows 0.00%. Quarter-night undersides are 27.7 = 27.7. **At golden hour the grove now lies partly under a cloud-shadow core:** conifers 44.9 → 42.9 (17.9% >8, noise 0.4%/0.0%), crowns 48.4 → 48.3 and the upper lawn 97.2 → 86.3. This is P8l's fixed still pattern (declared in P8l), shown to the owner on sheet pages 09a and 09b. 175: - **Whole-stage A/B bench** (`close/bench.sh`; `bench-t1.out`, `-t13.out`, `-t19.5.out`; `bench/summary.txt` from `close/bench-summary.js`; 20:49-21:01). The setup was the P8-II close recipe: pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync` before `--path`, 1920x1080, rain 0.45, `--motion-live`. The scenes were dense `artifacts/weather-pass9/dense.json` and Harbour, at 01:00, 13:00 and 19:30. Every run has 491 raw samples, vsync disabled, and finished on try 1. Raw p50 and p95 medians, pre → live (paired median; rounds where live was slower): 176: 177: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 178: |---|---|---|---|---|---|---| 179: | Dense | 14.225 → 14.483 (+0.206, 3/3) | 16.153 → 16.479 (+0.334, 3/3) | 13.949 → 14.249 (+0.294, 3/3) | 15.864 → 16.176 (+0.274, 3/3) | 16.920 → 17.306 (+0.302, 3/3) | 18.701 → 19.037 (+0.347, 2/3) | 180: | Harbour | 6.657 → 6.852 (+0.211, 3/3) | 7.199 → 7.462 (+0.271, 3/3) | 6.279 → 6.540 (+0.256, 3/3) | 7.222 → 7.358 (+0.136, 2/3) | 7.409 → 7.820 (+0.411, 2/3) | 9.018 → 9.251 (+0.210, 3/3) | 181: 182: - Draw calls go from 2583 to 2586 for dense at night and 2587 at 13:00 (3433 → 3438 at 19:30), and for Harbour from 907 to 916 or 917 (822 → 829 at 19:30). 183: - The stage costs about +0.2-0.3 ms p50 in both scenes at every hour, in 3/3 rounds. The 19:30 Harbour figure of +0.41 has one round at −0.20, and that sitting's dense pre rounds spread 16.0-17.3. The parts' own p50 deltas add up to about the same: dense about +0.25 (g1 +0.12, g2 +0.07, j +0.06, l +0.02, the rest within ±0.03). The plan's stop is dense p95 over about +0.5 ms, and it is not hit (+0.27 to +0.35). 184: - **Owner sheet** `artifacts/beautify/26-P8-III/close/sheet/` (`close/sheet.sh`, `sheet-src.sh` with the frame list `sheet/src/sources.txt`, regions `close/sheet-regions.json`, crop picker `close/pickabs.js`). Before is on the left and after is on the right. 185: - The motion pages use one frame from each part's recordings at a gust time. Before is that part's own pre-part tree; for the trees it is the P8-II code. The frames are the kept PNGs (g2, h) or decoded from `clip.mp4`. 186: - The still pages use this close's pre-1 and after. 187: - Pages: 188: - `p8iii-01-trees-in-a-gust.png` 189: - `p8iii-02-flag-and-banner.png` 190: - `p8iii-03-laundry-and-bunting.png` 191: - `p8iii-04-water-in-a-gust.png` 192: - `p8iii-05-cloud-shadows.png` 193: - `p8iii-06a-boats-bob.png` 194: - `p8iii-06b-fireflies-and-moths.png` 195: - `p8iii-07-puffs.png` 196: - `p8iii-08-existing-scenes.png` 197: - `p8iii-09a-approved-trees-golden.png` 198: - `p8iii-09b-grove-lawn-golden.png` 199: - `p8iii-10-what-differs.png` 200: - `sheet.md` has one plain paragraph per part with its acceptance numbers, then what existing scenes will see, the bench, the suite and the open questions. I looked at every page. Page 06b shows the swarms only faintly at this crop (the count 30 → 34 is the evidence). On page 08, harbour-golden's lamp-lit courtyard reads clearly darker after: a cloud-shadow core lies over the right lawn and towers, and it darkens lamp pools too, as P8l declared. 201: - **Not done / open:** 202: - The owner has not seen the sheet. **D28** (the veil) is still the gate before P8-IV. 203: - P8k's puff colour and splash ring are for the owner to judge (sheet page 07). 204: - Cloud shadows over the approved grove and over lamp pools at golden hour are new in stills (pages 05, 08, 09a, 09b). 205: - The P8 recording sheet of every clip is left for the P8 close. The part recordings' kept PNG frames under `26-P8/g2/rec/*/frames` and `26-P8/h/rec/*/frames` are still in place; this close only read them. 206: - No new recordings were made at this close. The motion pages compare each part's own before and after, so a single P8-II-against-final clip of a gust was not made. 207: - The perspective and photo-mode lenses were not captured. The orbit views' counts vary within a tree, as before. There is no 1440x900 `--verify`. 208: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 209: - Scratch trees: `close/pre-tree` was moved to the macOS Trash with `mv` as `beautify-26-P8-III-close-pre-tree-210853`, once its numbers were recorded. The part trees `26-P8//pre-tree` are outside `26-P8-III/` and were left in place. 210: - Platforms: Mac run; Windows not run. 211: SOURCE: artifacts/beautify/owner-decisions.md (frozen copy = HEAD 8e170d0), lines 278-304 278: ## P8-III stage close finished; P8-IV started on the owner's instruction (7 October 2026, evening) 279: 280: The owner said "Do the close and continue afterwards with the beautify plan". The P8-III close (main 523751a, snapshot `26-P8-III`, sheet `artifacts/beautify/26-P8-III/close/sheet/`, 12 pages sent in chat) finished: 44 suites, 23,405 checks, 0 failures; draw-call deltas match the parts' declarations; stage cost +0.2 to +0.4 ms p50, below the dense p95 stop line of about +0.5 ms. The handoff check's 12 findings were corrected in `docs/visual-handoff.md`. P8-IV (P8e, then P8f with a second review) starts under the D28 default (veil as the report designs it). The owner has NOT yet given a verdict on the P8-III sheet: cloud shadows at golden hour (pages 08, 09a, 09b), tree-house look, cloth, cat's-paws, puffs (dust colour, blocky splash ring) and the earlier open items are all still open for them, reopenable at any time. No standing approval is in force; the pipeline continues on the owner's instruction only. 281: 282: ## P8-III approved (owner, 7 October 2026, in chat) 283: 284: "P8-III approved, move on." Given after the P8-III sheet (`artifacts/beautify/26-P8-III/close/sheet/`, main 523751a) had been sent, while P8f was running. Read by the orchestrator as approval of the whole P8-III sub-stage as built: wind field and crowns (tree-house crown moves only above its house; leaves in the metre above shear), cloth, laundry and bunting, rain lean and cat's-paws, boats and swarms, placement puffs (dust colour and blocky splash ring left as built), cloud shadows (including the golden-hour shading of the approved grove and the Harbour courtyard), with every open item recorded on the sheet left as built. They stay reopenable by the owner at any time. P8-IV (P8e done, P8f running) continues; D28 default stands. 285: 286: ## Laptop needed again by 08:00 on 8 October 2026 (owner, in chat) 287: 288: "again ill need laptop by 8am". Read by the orchestrator as the same arrangement as the night of 6 to 7 October, except that no new blanket approval was given: (1) the pipeline keeps running (P8f now, then the P8-IV close) and results that clear review are carried forward as provisional, not approved, until the owner gives a verdict; they stay reopenable; (2) nothing outward-facing (push, journal deploy) happens without an explicit request; (3) no workflow is started that is not expected to finish by about 07:30 on 8 October: P8f is expected to end between 01:30 and 02:30, the P8-IV close (about 1.5 to 2 hours) is launched only if it can finish by 07:30; (4) no Godot is left running past 07:30; whatever is not started is left for the morning with a resume point on the board. 289: 290: ## Standing approval until 08:00 on 8 October 2026 (owner, in chat) 291: 292: "approve all unti morning". Same terms as the standing approval of 6 to 7 October: every P8 part and fix result that clears review is treated as approved as built, with its open items left as built and listed for the owner sheet, so the pipeline does not wait for answers. It covers plan approvals and owner-sheet questions only, not pushing to GitHub, deploying the journal or any other outward-facing action. No workflow is started that is not expected to finish by about 07:30 on 8 October, and no Godot run is left running past 07:30. Supersedes the "provisional" reading in the section above for P8f and the P8-IV close. 293: 294: ## Clean-up of scratch (owner asked "clean up the folder", chose "All old stages' scratch", 8 October 2026) 295: 296: 87 raw PNG frame folders (the `frames/` directories of recordings, about 28 GB; all under `artifacts/beautify/26-P8/` for parts a to l) were moved to the macOS Trash (folder `beautify-frames-0708`) (folder name = old path with `_` for `/`), plus the duplicate snapshots `26-P8-e-pre` and `26-P8-g1-commit1`. Kept: every clip, metric, log, sheet, look crop and still, and everything under `26-P8/e` and `26-P8/f`. Handoff paths into the moved `frames/` folders no longer resolve unless the owner puts them back; the analyses cite the `.out`/`.txt` files that were computed from them. Earlier stages (P0 to P6) had already been cleared of frames. 297: 298: ## Orchestrator ruling on P8f lit rain (8 October 2026, 01:15, under the standing approval) 299: 300: P8f as built (main 081f834, snapshot `26-P8-f`; the pre-part snapshot is `26-P8-f-pre`, the as-built copy `26-P8-f-asbuilt`) meets the cost and flashing lines but only partly meets "rain turns pink before the BAR sign and amber near lamps": the island is 6x drier, 15 pink rain pixels in the BAR box over 180 frames (0.08 a frame; 2,455 before) and the island rain stays blue because 69% of its light is not in the screen copy. The owner is not asked (standing approval). Ruling: take the options "more rain in front of the island" and "read the floor on the shown colour (reverse the tonemap)" together, as a directed fix pass, because the plan's own words are "keep at most 150 NEAR drops", which the whole-volume cap does not do. The fixer may: (1) change `RAIN_NEAR_CAP` and/or give the capped drops a near-only placement so the island and the BAR box keep a rain density near P8d's (island rain pixels no worse than about 2x fewer than pre, BAR box several hundred rain pixels over 180 frames, not 36); (2) read the floor on the shown colour in `shaders/lit_rain.gdshaderinc` (invert the post's tonemap) so the island's lights count. Limits: drops plus veil still at most 0.5 ms (1920x1080, dense and Harbour), CPU drop time still below P8d's, no flashing above 3 per second, Glade, dry and 05:45/20:45 skies as measured, Reduced/Off behaviour unchanged. If both cannot be met, the order of preference is the density first, then the colour, and the entry says what was left. "Stop lit rain at the skyline" is rejected as a fallback unless nothing else works. 301: 302: ## P8-IV close finished; P9 started (8 October 2026, 04:00, under the standing approval) 303: 304: P8-IV close (main 6f706a9 with code 8ccf61a, snapshot `26-P8-IV`, sheet `artifacts/beautify/26-P8-IV/close/sheet/`, 8 pages sent in chat): 44 suites, 23,554 checks, 0 failures; +1 draw call on rainy views (the veil), stage cost -0.01 to +0.04 ms p50. Treated as approved as built under the standing approval until 08:00; the owner has not seen the sheet. Open for the owner: veil reads as sky and far-water rain only at Quarter zoom, lit rain faint in plain frames, the 300-drop near box (about 2.2x P8d's drops at light rain, model estimate), 05:45/20:45 horizon changed shares above noise, rain-ring phase differences, Glade with rain turned up gets the veil. P8 is done. P9a (edge ink and rim, second review required; pre-part snapshot `27-P9` = code 8ccf61a) is launched with the rule that the whole part must finish by 06:45; ink and rim stay OFF by default (D30 is the owner's gate after P9b). === Appended 9 October 2026 === Frozen 9 October 2026 at 19:36 BST (game main at 8418d4e, clean; owner-decisions.md equals HEAD). Records the owner's approval of the P8-IV sheet on 8 October (with D30), which the chapter now reports. Six wording corrections from the fact-check of the 8 October edition were applied to the chapter text; their evidence is in the 8 October excerpt above (frozen handoff of 8 October): the draw-call deltas at the P8-III close are the sum of the parts' figures, P8h's harbour-lane figure being derived afterwards (frozen line 162); 97.8% is the central figure of the veil pan test, minimum 94.7%, shift 0 or -1 px (line 148); "stop lit rain at the skyline" was the rejected option (owner-decisions line 300); the fix-3 lit-rain figures come from a fresh quarter-night clip of the same view (line 78); the open Glade item is that a Glade with raised rain reads the floor through the canal look's filmic curve instead of its own AgX, untested (line 115); the P8-IV close twilight medians moved by at most 0.7 (20:45 horizon 150.5 to 151.2, line 18). The P8 stage close of all four sub-stages and the whole-stage recording sheet remain open (line 33). SOURCE: artifacts/beautify/owner-decisions.md (frozen copy = HEAD 8418d4e, 9 October 2026, 19:36 BST), lines 326-328 326: ## D30 answered and P8-IV approved (owner, 8 October 2026, in chat) 328: "ink off p8iv approved". Read by the orchestrator as: (1) D30: ink and rim stay OFF by default (keys `edge_ink`, `edge_rim` stay 0 in all anchors; the pass stays hidden and costs nothing; `--edge-ink` and `EDGE_PROPOSED` remain for tests and later use; no outline level chosen, the open ink items 2:1 dashes, cone ticks, crown specks and the night ink line are not pursued); (2) the P8-IV sheet is approved as built (rain veil, lit rain with the 300-drop near box, the open items on that sheet left as built). No owner decision from the beautify plan is pending now except the optional P10 polish backlog, journal publishing and pushing to GitHub, none of which has been asked for.