Chapter 18: Sweeping up the small faults Frozen 28 September 2026. Implemented, reviewed and merged work; decisions by the owner (the project's creator) are in the second source. Original source terminology is preserved; workflow run IDs and local paths are removed. Line numbers refer to the frozen copies. Part codes: P0a measurement kit, P0b wet-leaf glints, P0c island bank and lawn, P0d roof pepper, P0e mirror seam and stars in rain. SOURCE: docs/visual-handoff.md (frozen copy, lines as of 28 September 2026) 295: **Beautify P0 stage close (26 September 2026):** the stage-end overhead of P0 "Defect sweep" (step `18-P0`), run by Opus 5.5 on this Mac on the final P0 code (`main` 478fccd; the only uncommitted file was the orchestrator's `artifacts/beautify/beautify-stage.js`). No game code changed. Evidence: `artifacts/beautify/18-P0/close/`. Before side: the pre-stage scratch tree from `build/backup-feel/18-P0/` (checked equal to tag `beautify-18-P0` on scripts, shaders, data and scenes). Power: mains, Low Power Mode 0 throughout (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). 296: - **Full suite** (`close/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/18-P0/close/suite`, 03:02-03:08): **40 suites, 18,447 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; the `test_creature_reactions` rerun 5,549 checks, 0 failures. The known flaky check ("The re-plan is applied after the edits") did not fire. Same count as P0e's `e/suite3`. 1440x900 not run (the inherited stall). 297: - **Journey** (`artifacts/feel-journey/18-P0/`, the six views through `tools/art_capture.sh --set journey`, no pkill, all first try; `close/journey.sh`). Against `17-owner` (`18-P0/counts-vs-17-owner.txt`), only `draw_calls` and `rendered_objects` change: harbour-day and -night 925 → 892, harbour-lane 755 → 735, quarter-day and -night 905 → 914, styles-day 496 → 484. Harbour, lane and quarter are the parts' declared net deltas (P0c 925 → 891, 755 → 734, 905 → 914; P0d +1 on Harbour and lane, +9 on styles). The pre-stage tree already draws styles at 523 (identical to the P0a baseline), so styles moved 496 → 523 between `17-owner` and the P0 snapshot, outside P0; P0 takes it 523 → 484 (P0c −48, P0d +9). Re-run: every view was captured again in the art set (`close/after/`), with identical integer counts on all six (`close/metrics/counts.txt`). By eye: darker brown banks, no waterline beads, softer water under the island; nothing unexpected. `notes.json` entry `18-P0` added (backup `notes.json.bak-before-18-P0`); `index.html` and `journey-standalone.html` rebuilt. 298: - **Art capture and protected looks** (`close/caps.sh`; one sitting 03:09-03:19, in order `pre-1`, `after`, `pre-2`, `after-2`; all 24 views first try: the kit's 21 baseline views plus harbour-lane, quarter-r35 and quarter-r60). Metrics in `close/metrics/` (`ab-pre1-after`, noise pairs `ab-pre1-pre2`, `ab-after-after2`, `ab-baseline-*`, `acceptance.txt`, `protected.md`, `stars.txt`, `counts.txt`). The pre-stage tree's counts equal the P0a baseline on all 21 shared views. Whole-frame changed >8 in the pre pair 0.03-3.9%, in the after pair 0.03-3.2%. Protected-look sheet `close/protected/protected-1.png`, `-2.png` (before = `pre-1`): 299: - Canal mirror columns (quarter-night `mirror_columns`): p50 36.3 → 36.0, changed >8 19.5% against the noise pairs' 7.2% and 2.8%. That matches P0e's own 18.6% (far hits past 60% of the reach fade a little darker, as intended). By eye at 2x, the orange, cyan and pink streaks are all there. Quarter-2100: 20.7% (noise 3.9%, 3.5%), p50 37.7 → 38.6. 300: - HDR neon cores: fish sign 2.2% (noise 1.9%, 0.9%), bar sign 0.0%, lane sign 0.9% (0.6%, 1.5%); medians equal within 0.3 luma. Unchanged. 301: - 05:45 and 20:45 skies: zenith, horizon band and right sky within the noise pairs (for example 20:45 horizon band 1.3% against 0.9% and 1.5%); colours equal. **Stars:** the scene rain is 0.45, so D5's fade takes most of the faint twilight stars. `art.bspecks4` in `sky_right`, pre-1 → after: 05:45 11 → 4, 20:45 10 → 1, 21:00 12 → 2, 06:10 9 → 2, 22:00 13 → 14, quarter-night 14 → 11, dry 15 → 14, rain 0.35 16 → 13, 0.6 13 → 2. This is D5 as specified. It shows most at dawn and dusk, so the owner sheet shows it (page 7). 302: - Glade/Canal separation: Glade lawn unchanged (`#637f19`, p50 119.5, 0.00% changed); canal lawn `#466e26`, p50 97.3 → 98.7. Banks: Glade 125 → 79 (`#6b4a2a`), canal 137 → 85 (`#635342`). Both are darker now, and Glade's is warmer. 303: - Approved tree style (grove-day): crowns 1.1% changed >8 (noise 0.2%, 0.9%), conifers 1.4% (0.2%, 0.7%), colours and medians equal. Tree feet 2.4% (the flat lawn and tufts around them, P0c). 304: - **Whole-stage A/B bench** (`close/bench.sh`, `close/bench.out`, `close/bench/summary.txt`, `bench/power.txt`; 3 interleaved rounds, pre-stage tree against live, `--disable-vsync` before `--path`, 1920x1080, 13:00, rain 0.45; mains, Low Power Mode 0, battery 99% charging at start and end; runs report vsync "enabled", as in the part benches): 305: - dense.json: mean 14.36 / 14.43 / 14.38 → 14.80 / 14.75 / 14.80 ms, **median 14.38 → 14.80 (+0.42 ms, +2.9%)**, slower in every round with no overlap. p95 medians 16.02 → 15.58 (noisy). Draw calls 2,125 → 2,140. 306: - The parts measured P0b +0.7% (13:00), P0d +1.2% and P0e about 0; P0c had no bench, and its +15 draw calls on dense (terrain surfaces, moss) are the likely rest. Not isolated further. 307: - Harbour: 8.32 → 8.34 ms, p95 8.32 → 8.27, draw calls 936 → 902. Display-paced at 8.33 ms, so a weak A/B. 308: - **Owner sheet** `artifacts/beautify/18-P0/close/sheet/` (before left, after right, colour and greyscale where values matter; crops enlarged 2-3x by `close/crops.js`, pages by `close/sheet.sh`): `p0-1-overview.png`, `p0-2-P0b-wet-leaf-glints.png`, `p0-3-P0c-island-bank.png`, `p0-4-P0c-lawn-tufts.png`, `p0-5-P0d-roofs.png`, `p0-6-P0e-mirror-seam.png`, `p0-7-P0e-stars-in-rain.png`, `p0-8-existing-scenes-1.png`, `p0-9-existing-scenes-2.png`, and `sheet.md` (one paragraph per part with the numbers, then bench and suite). 309: - **Stage acceptance re-read on today's code** (`close/metrics/acceptance.txt`, pre-1 → after, both noise pairs agree): 310: - Harbour slate specks 30 → 0, cones 20 → 0, lane cone 12 → 4. 311: - Waterline black art pixels 20-68 → 0 on Harbour day, night and golden, Glade and lane. 312: - Bank/lawn: 0.87 by day, 0.86 at night, 1.47 at golden; Glade bank 79 against lawn 120. 313: - Seam region above luma 130: day 18.3 → 3.4%, golden 23.4 → 11.0%. Corner: day 42.3 → 6.9%, golden 41.6 → 12.8%. 314: - Quarter-golden plate crown `#353e27` → `#1e2f19`. 315: - **Not done / open:** 316: - P0c's night bank line still fails as recorded: bank minus open water 4.1, minus water below 3.2, target 8. It is deferred to the re-measure after P1a. 317: - For the owner, from the P0 sheet: P0d's far-focus deck (a dark flat roof where tiles are hidden), the Glade lip flaps (lighter than the lawn), and the twilight stars lost at the scene rain. 318: - In the grove, the flat lawn shows soft broad light and dark patches where the quilt was (sheet page 8). It is not a cell pattern (P0c's separated steps), but the owner may want to judge it. 319: - P0e's seam fade at night is re-checked after P1a. ... 332: **Beautify P0e Mirror seam and stars in rain (25 September 2026):** Opus 5.5, in a cloud session (Linux, no GPU), implemented report Defect 6 (the mirror seam) and the stars item of Defect 9, with owner decisions D4 (no foam change) and D5 (a gentler star fade). Code and headless tests in the cloud; **captures and bench: done on the Mac** (26 September 2026, "Mac half" below). The stepped seam is gone. The bright sky patch in the bottom-left corner (x 95-195, y 945-1000) of quarter-day and quarter-golden, which stopped the first Mac half and survived the first look-review fix, is closed by the escalation fix on the orchestrator's order (26 September, "Escalation: the closest pass" below): a miss now also takes the soft floor by how close it passed in front of geometry standing above the mirror. Share of corner pixels above luma 130: day 42.95 / 42.25% -> 6.00 / 6.00%, golden 41.35 / 40.98% -> 11.42 / 11.75%. The other small pale patches were soft hits held pale by `MIRROR_SOFT_BODY` 0.3; the first look-review fix set it to 0.7 ("Look review and fix" below). Branch `beautify/P0e`, from `beautify/P0d`. Files: `shaders/mirror.gdshaderinc`, `shaders/water.gdshader`, `shaders/puddle.gdshader`, `shaders/backdrop.gdshader`, `scripts/main.gd` and `scripts/world_view.gd` (the `--mirror=debug` diagnostic), `tests/test_wet.gd`, `tests/test_water.gd`, `docs/lighting-and-texture-plan.md` (the "shoreline done" correction the plan asks for), `docs/visual-handoff.md` (this entry), `docs/beautify-cloud-handoff.md`, and on the Mac `tools/art_regions.json` (two view aliases), and the escalation fix on the Mac in `mirror.gdshaderinc`, `water.gdshader`, `puddle.gdshader` and `tests/test_wet.gd`. Evidence is under `artifacts/beautify/18-P0/cloud/e/` (git-ignored, on the cloud container only) and, for the Mac half, `artifacts/beautify/18-P0/e/`. 333: - **What changed.** 334: - **The mirror seam (Defect 6), in the shared `mirror.gdshaderinc`.** 335: - Both marches (`mirror_march`, `mirror_march_view`) keep their `bool` return (`test_camera:67` pins `bool mirror_march_view(`) and gain two last out parameters, `bool exhausted` and `float gap`. 336: - `exhausted` is true when the ray ran out of steps inside the frame while its last sample was still geometry (`last_scene < backdrop_depth`). That covers the case the report found: under the town, a ray still in front of a building face fell back to the full `sky_ref`, a bright sky patch with a stepped edge along the town's 5 m contour. `hit_uv` and `travel` then hold that last sample, and `gap` is how many view metres in front of it the ray passed. 337: - A ray that leaves the frame, goes behind the camera, or ends over the painted panel is not exhausted and keeps the sky, exactly as before. There was no separate screen-edge fade in the code; the report's "apply the screen-edge fade only to rays that leave the frame" is met by leaving that path as it was. 338: - The stride doubles after `MIRROR_STRIDE_AFTER` 48 steps, so the water's 72 steps reach 96 base steps (`mirror_reach(step_len, steps)`). A puddle's 48 steps are unchanged. 339: - Every hit now fades toward the soft floor `mix(sky, body, MIRROR_SOFT_BODY 0.3)` (0.7 since the Mac look review, 26 September; the puddle uses its `base`) over `MIRROR_SOFT_FROM` 0.6 to 1 of `travel / mirror_reach`. The existing metre fade (`MIRROR_TRAVEL`) is kept inside it. At the Harbour camera the reach is about 24 m, so hits start softening from about 14.6 m, before the metre fade's 15.3 m; because the step is in viewport pixels, the reach in metres also depends on the window size. 340: - A soft hit takes the floor by `mirror_soft_weight(gap)` = `1 - smoothstep(MIRROR_SOFT_GAP 0.5 m, 4.0 m, gap)`: all of it for a ray about to meet a wall (which equals a hit at the full reach, so the two meet without a step), none for a ray passing well above low ground, which reflects sky as before. The water blends its reflectivity from the miss value (`0.85 · fresnel`) to the hit value by the same weight, reads no screen for a soft hit, and still adds the panel's lights below it. The gap values are first values; the Mac's debug render is where they are set. 341: - **Review fix.** As first committed (`c6572fb`), a soft hit was yes or no: all floor wherever the last sample was geometry. The reviewer's `--mirror=debug` classification on llvmpipe (class only, not a look) found 70,944 soft-hit pixels (4.4% of the frame) of open Harbour water in front of the island, with a hard edge against sky that traced the island's outline pushed down the screen by the reach. The step there is about −0.073·(sky − body) by day and +0.069 at night. The weight above makes it continuous. 342: - **Stars in rain (Defect 9, D5).** `backdrop.gdshader`: `(1 - rain*0.8)` (72% at the default rain 0.35) becomes `(1 - smoothstep(STAR_RAIN_FADE.x, STAR_RAIN_FADE.y, rain))` with one named pair `const vec2 STAR_RAIN_FADE = vec2(0.10, 0.60)`: full when dry, 0.50 at rain 0.35, 0 from 0.6. Changing the pair is a one-line change. 343: - **Foam (D4): no change.** The foam threshold in `water.gdshader` is untouched, and a test pins it until P3a. `docs/lighting-and-texture-plan.md` now says, next to the second pass's "soften the shoreline", that the foam has never been able to appear (report Defect 3) and is left to P3a. 344: - **Hit/miss diagnostic (for the Mac's capture).** `--mirror=debug` sets `WorldView.debug_mirror_view`, which sets `mirror_debug` on the water and puddle materials: a hit paints green, a soft hit from red to yellow by its weight, a sky fallback red. It is off by default and never saved, so play is unchanged. ... 362: - **Seam line, quarter-day and quarter-golden (region `mirror_seam`, x150-850 y880-1000): the stepped seam is gone; a small bright sky patch remains in the bottom-left corner. Partial; not a tuning problem (below).** Share of region pixels above luma 130: day 18.3% -> 6.5% (noise pair 18.29/18.32, 6.54/6.76), golden 23.4% -> 16.1%; region mean day 101.3 -> 96.9, golden 103.5 -> 101.3. Changed >8: day 40.3% (noise 8.5% pre, 1.4% after), golden 35.8% (noise 9.4%, 10.3%). Classes in the region (quarter-day, Mac): 80,112 hits, 1,584 soft at 1/2 or more, 300 below, 1,740 sky (cloud recheck: 1,604, 324, 1,700). By eye at 4x (`e/look/qd-seam-left4.png`, `qd-seam-right4.png`, `qd-bl-context.png`, `qg-bl-context.png`): the long pale stepped band across the bottom is replaced by the town's reflection fading out; left over are a pale patch with a stepped top at x 95-195, y 945-1000 (in the region from x 150), and two small ones at x 265-300 and x 712-735 near y 990. The sky-class pixels there read luma 138 against 95 for the hits around them, the same as before. **Corrected by the look review (below):** only the corner patch is sky-class (or weight 0). The x 250-350 patch is mostly soft hits at weight 1/2 or more (1,428 pixels, luma 130 by day, 139 golden, against 324 sky-class), kept pale because the soft floor at `MIRROR_SOFT_BODY` 0.3 is 70% sky, and the hits near it that are past 60% of the reach fade toward that same pale floor (mean 120 by day, not 95). The first pass tried only `MIRROR_SOFT_GAP`; `MIRROR_SOFT_BODY` had not been tried. 363: - **Why `MIRROR_SOFT_GAP` stays 0.5-4 m.** A trial diagnostic on a scratch copy (`e/trial-tree/`, not shipped: exhausted rays binned by gap, sky misses blue; `e/cap/gapbins-1/`, sheet pages 5-6, `e/look/gapbins-qd-bl.png`) shows the corner patch is 648 rays that do not run out in front of geometry (they leave the frame or end over the panel, no constant reaches them) and 956 exhausted rays 8-16 m in front of their last sample; only 52 are at 4-8 m. Raising the far value to 8 m (the reviewer's suggestion) would change 52 pixels there and put 24,676 pixels of Harbour's open water (4-8 m) onto a partial floor; reaching 16 m would add 41,708 more, and the gap field jumps between those bands along the island's mirrored outline, so either would make the hard open-water edge the review fixed. The residual patch needs a different treatment (a design change, not tuning), so it is reported, not fixed here. 364: - **Stars in rain (D5).** `art.bspecks4` in `sky_right` (pre-1, pre-2 -> after-1, after-2): dry 14, 14 -> 14, 14; rain 0.35: 15, 15 -> 12, 12; rain 0.45: 12, 13 -> 10, 11; rain 0.6: 12, 12 -> 4, 0; day 0, 0 -> 1, 0. At rain a few of these specks are raindrops seen end-on (the 4 at 0.6 in after-1 are not at star cells; `e/look/sky-r60.png`), so strength is also measured at the 13-14 star cells found in each tree's dry frame (`e/stars.js`, `e/metrics/stars.txt`): excess over the local sky, as a ratio to dry, before -> after: rain 0.35 0.69, 0.67 -> **0.49, 0.50** (target about half); 0.45 0.62, 0.59 -> 0.23, 0.24; 0.6 0.49, 0.48 -> **0.03, 0.04** (target gone; 1 of 13-14 cells above 8 luma, a rain streak). The formula predicts 0.72 -> 0.50, 0.64 -> 0.22, 0.52 -> 0. **Met; `STAR_RAIN_FADE` stays (0.10, 0.60).** ... 373: - **Look review and fix (26 September 2026).** The orchestrator's look review found two things. (1) The seam line still fails in the bottom-left corner: x 95-195, y 945-1000 mean luma 121.2 -> 121.0 by day, 120.9 -> 120.4 golden, share above luma 130 42.8% -> 41.9%. Its rays are sky-class or weight 0 (`gapbins-1` shows the top of the patch as rays that did not run out of steps), so no constant reaches them. **The stop condition stands; not changed here.** It needs an order for a design change (for example, count a ray as exhausted when any of its last few samples was geometry, not only the last one) or a waiver. (2) The entry blamed the other patches on sky-class pixels; they were soft hits held pale by the floor (corrected above). Fix: a `MIRROR_SOFT_BODY` trial on `e/trial-tree/` (differs from the live tree only under `--mirror=debug`), one sitting 01:51-01:56, mains, LPM 0 (`e/body/power.txt`): pre-part tree twice, live 0.3, trials 0.5, 0.6, 0.7, 1.0 (`e/body/bodycaps.sh`, `e/cap/body-*`; numbers `e/body/measure.txt`, `abtab.txt`, `harbour-edge.txt`; stacked 2x crops `e/body/look/q{day,golden}-seam-x0-450.png`, `-x450-900.png`, rows pre, 0.3, 0.5, 0.7, 1.0). Region x 150-850, y 880-1000, share above luma 130 (day / golden): 0.3 6.71 / 16.11%, 0.5 4.90 / 13.47%, 0.6 4.80 / 13.03%, 0.7 4.47 / 12.28%, 1.0 3.63 / 10.21%. The corner stayed at 40.4-40.7% for every value. Harbour-day's soft blob (the roof tank's reflection) darkens with the body: soft/sky |step| 8.8 at 0.3, 10.7 at 0.5, 12.0 at 0.7, 13.0 at 1.0. At 4x it shows at 0.5-0.7 as a soft darker oval with a blurred top (`e/body/look/hd-tank-all.png`), and at 1x it cannot be seen (`hd-1x.png`). **Chosen: 0.7**, the top of the review's range. 1.0 would take the sky out of the floor entirely and is outside that range. `test_wet` updated first: 131 checks, 1 failure against 0.3 (`e/body/testfirst/`). After: `test_wet` 131, `test_water` 126, `test_camera` 191, all 0 failures (`e/body/tests/`). 374: - **Re-captured against the same pre-part tree** (`e/body/recap.sh`, one sitting 01:59-02:03, mains, LPM 0): `e/cap/pre-3`, `pre-4` (pre-part tree), `after-3`, `after-4` (live, 0.7), all 11 views, first try. Draw calls and every integer count identical on all 11 views (`e/body/counts-pre3-after3.txt`). Region above luma 130, pre -> after (noise pairs): day 18.38 / 18.23% -> **4.72 / 4.69%** (was 6.5% at 0.3), golden 23.24 / 23.55% -> **12.18 / 11.90%** (was 16.1%). The x 250-350, y 960-1000 patch: soft hits (weight 1/2 or more) day 146.5 -> 116.4, hits 138.8 -> 108.9 (at 0.3: 129.9 and 120.3); golden soft 143.5 -> 121.5, hits 135.9 -> 112.8 (0.3: 139.3, 126.6); share above luma 130 day 71.4% -> 13.3% (0.3: 33.3%), golden 66.5% -> 19.5% (0.3: 51.5%). The corner is unchanged: day 42.95% -> 40.47%, golden 41.16% -> 40.11% (`e/body/measure-after3.txt`). By eye at 2x (`e/sheet/p0e-7-softbody-0.3-vs-0.7.png`, `p0e-8-seam-final-zoom.png`): the pale patches right of the corner and near x 450-600 now fade into the reflection; the corner patch keeps its pale stepped top. ... 378: - **Escalation: the closest pass (26 September 2026, on the Mac).** The second look review (`e/review/recheck2/`) found the corner unchanged within noise after the `MIRROR_SOFT_BODY` fix (share above luma 130: day 42.95 / 42.33% before, 40.47 / 41.09% after; golden 41.16 / 41.89%, 40.11 / 40.62%) and kept the stop condition. The orchestrator then ordered a fix. This is a design change, not tuning. 379: - **Diagnosis** (trial diagnostic on the scratch copy `e/trial2-tree/`, not shipped; `e/cap/near-dbg-1`, `near-dbg-2`, `e/near/nearbins.js`, `e/near/qd-near2-bl.png`): each missing ray also recorded its closest pass in front of geometry standing above the water. Of the corner's 5,492 water pixels by day, 2,116 are hits. Every one of the 3,376 misses passed within 2 m of geometry: under 0.25 m 696, 0.25-0.5 m 662, 0.5-1 m 1,152, 1-1.5 m 722, 1.5-2 m 144. The values grade smoothly from the top of the patch to the frame bottom. These rays skim the low pier's bushes and trees, which the 3 and 6 px stride steps over, and then end over the panel or run out high above the ground. That is the pale stepped top. The gap at the last sample (8-16 m, weight 0) never saw them. Harbour-day's misses sit in a band about 160 px deep under the island: under 2 m 57,148 pixels (8% of the water), 2-4 m 50,800. 380: - **Change** (shared `mirror.gdshaderinc`, both marches, water and puddles): both marches gain a last out parameter, `out float near`: the smallest gap (view metres) in front of any sample whose surface stands more than `MIRROR_NEAR_ABOVE` 0.2 m above the mirror's own plane. The value is 1e6 when there is none. The orthographic march computes that height as `rise * (travel - gap)`, where `rise = sqrt(0.5 * (1 - depth_per_metre))` is the sine of the pitch. The perspective march uses the plane through `start` with normal `normalize(dir_view - normalize(start))`. So the canal bed under the water and the paving around a puddle never count. A miss takes the soft floor by `soft_w = max(exhausted ? mirror_soft_weight(gap) : 0, mirror_near_weight(near))`, where `mirror_near_weight` is `1 - smoothstep(MIRROR_NEAR_GAP 0.5 m, 3.0 m, near)`. It still reads no screen, and with no weight the result is the sky, as before. In `--mirror=debug` a closest-pass miss shows as a soft hit. Hits, the stride, the reach fade and `MIRROR_SOFT_GAP` are unchanged. The `bool` returns are kept (`test_camera:67`). Physically, a rippled mirror reflects a cone of rays, so a ray that grazes an edge is partly a hit. 381: - **Window trials** (`e/near/trials.sh`, `trialC.sh`; `e/cap/near-A`, `near-B`, `near-C`; crops `e/near/quarter-day-bl-trials.png`, `quarter-golden-bl-trials.png`, `hd-trials.png`, `hd-edge-trials.png`, `hd-corner-col.png`). Corner share above 130, day / golden: A (0.25, 2.0) 10.33 / 17.16%; **B (0.5, 3.0) 5.93 / 11.60%**; C (reuse `MIRROR_SOFT_GAP` 0.5-4) 6.00 / 11.49%. At 3x, B and C remove the pale stepped top; A leaves a faint patch. C darkened more of Harbour: 67,136 soft pixels at weight 1/2 or more, against 640 before (A 35,056). Its corner result is no better, so **B was chosen**. 382: - **Tests first:** `test_wet` was updated before the code. Its log against the old shaders shows 145 checks, 20 failures (`e/near/testfirst/`). New checks pin the out parameter and its start value, `MIRROR_NEAR_GAP` (0.5, 3.0), the weight function, `MIRROR_NEAR_ABOVE` in (0, 0.5], both above-plane rules, the combined weight and the debug class in water and puddle. A numeric check covers pitches 20-60 degrees: `depth_per_metre` = cos 2x, and a sample stands `rise * (travel - gap)` above the mirror. After the change (`e/near/tests/`): `test_wet` 145, `test_water` 126, `test_camera` 191, all 0 failures. 383: - **Re-captured against the same pre-part tree** (`e/near/recap.sh`; one sitting 02:28-02:34; mains, Low Power Mode 0, battery 80%, `e/near/power.txt`). Sets: `e/cap/pre-5` and `pre-6` (pre-part tree), `after-5` and `after-6` (live), `dbg-2` (live `--mirror=debug`). All 11 views and 6 debug views came out first try. Draw calls and every integer count are identical on all 11 views (`e/near/counts-pre5-after5.txt`; for example quarter-day 914, harbour 892). 384: - **Corner** x 95-195, y 945-1000, share above luma 130 (`e/near/measure.txt`): day 42.95 / 42.25% -> **6.00 / 6.00%**, mean 121.2 -> 103.5; golden 41.35 / 40.98% -> **11.42 / 11.75%**, mean 120.6 -> 108.0. With 0.7 alone it was 40.47% and 40.11%. The golden residual is mostly the cream sky reflection in hits at the left edge of the box (x 83-117), which is also in the pre-part tree. 385: - **Plan region** `mirror_seam` x 150-850, y 880-1000: day 18.31 / 18.52% -> 3.28 / 3.49% (0.7 alone: 4.7%); golden 23.21 / 23.41% -> 11.06 / 11.20% (12.2%). 386: - **Patch at x 250-350:** day 11.5%, golden 17.8% (13.3% and 19.5% with 0.7 alone). 387: - **Debug classes** (`e/sheet/p0e-13-hitmiss-closest-pass.png`, `e/near/look/qd-bl-dbg.png`): the corner's red sky class is now yellow soft hits, weight about 1, with one orange streak lower down. 388: - **By eye**, at 3x (`e/near/look/quarter-day-bl-3x.png`, `quarter-golden-bl-3x.png`; rows pre-part, 0.7 alone, after): the pale stepped top is gone. The corner reads as the reflection fading into slightly lighter water. A faint vertical edge stays at x 95, where the hits of the tree reflection end. ... 390: - **Harbour-day** frame 3.82% (0.72%, 0.78%); across sittings, 0.7 alone -> after 5.54% (noise 2.22%). The change is a soft darker band under the island's reflection, about 60-100 px deep, fading to sky. At 1x it reads as the island's reflection softening, with no hard edge (`e/near/look/hd-1x.png`). At 4x (`hd-corner-4x.png`), a faint vertical tone step extends down from the island's left corner, where the reflection's own vertical edge already was: about 5 luma over 8-column means (`e/near/vstep.js`; 1.9 before). ... 394: - **Bench** (`e/near/bench.sh`, `e/near/bench.out`, `e/near/bench/summary.txt`, `power.txt`): 3 interleaved rounds, pre-part tree against live, `--disable-vsync`, 1920x1080, 13:00, rain 0.45; mains, Low Power Mode 0 throughout; runs report vsync "enabled", as in the first bench. dense.json mean median 14.78 -> 14.70 ms, p95 median 16.67 -> 16.48. Harbour 8.22 -> 8.18 ms, p95 8.33 -> 7.57. Draw calls 2,140 and 902 on both sides. The added compare per step has no measurable cost. 395: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/18-P0/e/suite3`, parts a, b, v, 02:36-02:45; mains, LPM 0, `e/suite3-power.txt`): **40 suites, 18,447 checks, 0 failures**; 18,447 = 18,433 + the 14 new `test_wet` checks. `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun: 5,549 checks, 0 failures. No Mac-only failures. 1440x900 not run (the inherited stall). ... 402: - **Still open:** nothing on the seam line after the escalation fix; the look review of the escalation is the orchestrator's. Harbour's open water under the island changed (a soft darker band, above), which the entry's earlier "open water unchanged" claims no longer cover. The seam fade is re-checked at night after P1a. 403: 404: - **Review:** code-only, Opus 5.5 in a fresh context: 3 must-fix. (1) The yes/no soft hit made a new hard edge in open water: fixed with the gap weight. (2) The entry's "open water unchanged": corrected above. (3) A vacuous reach check that read no shader: now computed from the shaders' constants. Optional findings taken: soft hits keep the panel's lights and read no screen; the reflectivity blend; the reach-fade size dependence noted; the tests described as text checks. Recheck (same reviewer, `f861410`): all three closed. The Harbour classification now shows 916 pixels above weight 1/8 in open water (was 70,944 at full weight) and only 78 neighbouring pairs, of 1.26 M, that jump by more than 0.5; the one soft area left is the reflection of a roof water tank, fading down its rows. The gap sign and units were verified in both marches. No new must-fix. The quarter bottom-left note above is the reviewer's. 405: - **Mac review:** Opus 5.5 look review in a fresh context [run ID removed] confirmed the Mac half's stop: 2 must-fix. (1) The bright corner patch at the bottom left of quarter-day and quarter-golden was unchanged; (2) the pale x250-350 patch was soft hits on a 70%-sky floor, and `MIRROR_SOFT_BODY` was untried. Fix pass 1: `MIRROR_SOFT_BODY` 0.3 -> 0.7 (item 2 closed). Fix pass 2 (the workflow's escalation round, which took the design change the review named before the orchestrator had ruled; the orchestrator reviewed it afterwards and endorses it): the closest-pass weight (`MIRROR_NEAR_ABOVE`, `MIRROR_NEAR_GAP`), corner above luma 130 43% -> 6% by day and 41% -> 11.5% at golden, recheck closed with the reviewer's own captures. Consequence the owner sees on the P0 sheet: a soft darker band in Harbour's daytime open water under the island's reflection (3.8% of pixels change by more than 8, noise 0.7%), which reads as the island's reflection shadow; the orchestrator accepts it. Not captured, left to the stage close: the protected 05:45 and 20:45 skies (the star fade changes them at the scene rain). Merged to `main` on 26 September 2026. ... 407: **Beautify P0d Roof pepper and per-tile value (25 September 2026):** Opus 5.5, in a cloud session (Linux, no GPU), implemented report Defect 5. Code and headless tests in the cloud; **captures and bench: done on the Mac** (26 September 2026). The first Mac half stopped on the look (the `plain` deck); after the look review the deck, the dormer prism and the tiles' side faces were fixed on the Mac (see "Look review fix" below). Branch `beautify/P0d`, from `beautify/P0c`. Files: `scripts/architecture.gd` (`profile_roof`), `scripts/masonry.gd`, `scripts/mesh_batch.gd` (`gable`, `quiet`), `scripts/openings.gd` (dormer gable), `scripts/build_anim.gd` and `scripts/build_audio.gd` (quiet pieces, review fix), `shaders/surface_body.gdshaderinc`, `tests/test_cluster_walls.gd`, `tools/art_regions.json` (three lane regions, Mac), `tools/probe_shaders.gd` and `tools/probe_surfaces.gd` (new cloud probes), `docs/visual-handoff.md` (this entry), `docs/beautify-cloud-handoff.md`. Evidence is under `artifacts/beautify/18-P0/cloud/d/` (git-ignored, on the cloud container only). 408: - **What changed.** 409: - **The deck.** (Superseded on the Mac: the deck now draws in the `tile` material, not `plain`; see "Look review fix".) The tiles' 4 mm insets (`masonry.gd`, `tiles()`) show the roof face under them, which drew in `stone` at the facet row tone; its paving grout was inked by the presentation outline into black pepper dots. With detail on and tiles on (`Masonry.enabled`), a face `Masonry.tiled_face(face)` accepts (not infill, sloping enough to have a down direction: exactly the faces `tiles()` lays tiles on; `tiles()` now calls the same function) draws in `plain` at `Masonry.deck_tone(color)` = `color.lightened(DECK_LIGHTEN 0.075)`, the mean of the two row tones (`row_tone`: 0.05 and 0.10). Gable infill keeps `stone`. Far roofs (detail off) keep their plain facets at the row tones; a scene with bricks off keeps the old stone deck, since nothing covers it. The deck is still appended in `MODE_POP`, so the build pop is unchanged (tested through a keyed batch). It is recorded **quiet** (below), so the build sound is unchanged too. 410: - **Where the tile instances are not drawn** (review finding, disclosed, not decided here): the view hides a tile chunk beyond `Masonry.RANGE` (40 m from the focus) or over `Masonry.BUDGET` (30,000), and editor thumbnails draw no instances. There the roof now shows this flat plain deck at the tiles' mean tone between the batched cut pieces, where it showed the stone shader's pattern at the facet row tones before. `plain` also takes no rain darkening. No planned capture shows it: every tile chunk is visible in the journey and bench scenes. The alternative, if the owner wants the old fallback, is a deck with its own LOD state. The comments in `Masonry.lod` and `profile_roof` now say so. 411: - **The dormer gable.** (Superseded on the Mac: `tile`, not `plain`.) `MeshBatch.gable` takes an optional material (default `stone`, so every other caller is unchanged). `Openings.dormers` draws the prism in `plain` at `deck_tone(colors.roof)`, quiet, when tiles are on; with bricks off it keeps the old stone at the roof colour. The prism's two small end triangles (the dormer's front gable) take the deck tone too. 412: - **Quiet pieces (review fix).** `BuildAudio.reveal_events` hears every `plain` piece of a timber cluster (medieval, japanese) as a plank, so the plain deck and dormer prism added plank clicks: a lone medieval tower went from 2 plank events (126 pieces) to 10 (726). `MeshBatch.quiet` marks the pieces recorded while it is set (a fourth element in each `pieces` entry); `BuildAnim.diff` passes the flag on as a seventh element of each `reported` row; `reveal_events` skips quiet rows. The deck and the dormer prism are recorded quiet. One sound does change: in a cluster whose bricks are not instanced (japanese, modern), the P0c stone dormer prism made one brick-course click as the reveal crossed it; the quiet plain prism makes none (a japanese house with a dormer: 27 events at P0c, 26 now, pinned). Plank events are otherwise identical to P0c for lone medieval and japanese houses and towers and a medieval house with a dormer (`d/probe/audio-pre.txt`, `audio-post.txt`; the unfixed build: `audio-unfixed.txt`). 413: - **Tile shader (`surface_body`, kind 0).** A new `varying flat float tile_piece` is 1 on an instanced tile and 0 on batched geometry. It reads the model matrix's y length over the shorter of its x and z lengths against `TILE_PIECE_MAX` 0.75: a tile instance is `TILE_T` 0.035 m along y against at least 0.07 m along x, a ratio of at most 0.5, while a batched mesh is uniformly scaled (ratio 1), also while a removed cluster shrinks away (review fix: the first gate, on the y length alone, turned on for the last tenth of a second of that shrink). Only instanced tiles take: 414: - a per-tile value `1 + (hash(piece origin) - 0.5) * TILE_VALUE 0.10` (±5%, on top of the CPU's existing ±4% per-tile shade); 415: - a lit lower lip, the lowest `TILE_LIP` 0.16 of the tile, up to `TILE_LIP_LIGHT` 0.12 brighter; 416: - the overlap shadow, the top `TILE_SHADOW` 0.16 under the next row's lip, up to `TILE_SHADOW_DARK` 0.12 darker. Lip and shadow are mirror images, so a tile keeps its mean tone and whole tiles and batched cut pieces keep one mean (the plan's "no tone seam"). 417: - Batched cut pieces share one model origin, so they are gated off. The constants are first values, not tuned: tuning them needs pictures. ... 436: - **Mac half (26 September 2026): STOP, acceptance fails for a reason that is not tuning** (the record of the first Mac run; fixed by the "Look review fix" below). Opus 5.5 on the Mac, live tree on `beautify/P0d` (`80c0128`), pre-part tree `artifacts/beautify/18-P0/d/pre-tree/` from `build/backup-feel/18-P0-d/` (tag `beautify-18-P0-d`). Everything is under `artifacts/beautify/18-P0/d/`. Power for every set and suite: AC attached, 84%, Low Power Mode 0 (the owner's D1 state; `cap/*/capture-meta.txt`, `suites/mac1/power.txt`). ... 441: - harbour-day `roof_slate`: 29/29 -> 29/29. **Fails, and unchanged:** the same pixels (`look/roof-slate-hday-x4.png`). The report's cause (stone grout on the deck) is not what makes this roof's pepper: the deck is now plain and every speck is still there. 442: - harbour-day `cones`: 20/20 -> 7/4. **Fails.** The black pepper becomes pale cream flecks along every tile gap and a cream cone tip (region mean luma 49.5 -> 64.2; 13.4% of pixels brightened by more than 0.1; `look/cones-hday-a-x4.png`, `cones-hday-b-x4.png`). 443: - lane-day `roof_lane` 1 -> 1, `cone_lane` 12 -> 6. **By eye it fails:** the canal roof's tile gaps show as pale sky-blue stripes along every row, the cone's gaps as a cream grid, and the dormer's gable prism turns pale blue-grey (`dormer_gable_lane` mean luma 20.9 -> 62.7, p90 34 -> 128; `look/lane-dormer-roof-x2.png`, `lane-cone-x3.png`). The tile rows read now only as bright lines. 444: - Night, no-regression check: harbour-night `roof_slate` 0 -> 0, `cones` 0 -> 1 (cone mean luma 12.5 -> 22.3: the same pale gaps, dimmer). 445: - Far focus (the owner's question): with the tiles hidden the flat plain deck glares pale blue on the slate roofs and cream on the medieval roofs and cones, where the stone deck read as dark roofs (`look/far-focus-x2.png`; whole-frame change 29.7% against a 23.2% noise pair). 446: - Why: `plain` (`shaders/plain_body.gdshaderinc`) is the old StandardMaterial, `ROUGHNESS` 0.55 with `SPECULAR` 0.5, so an up-facing slope takes a strong sky specular. The tiles (`surface_body` kind 0) are rough 0.77-0.84 and do not. P0c's look review met the same thing (plain grass at 2-3x the lawn's luma). None of the `TILE_*` constants touches the deck, so **no tuning was done**; the fix is a design change (for example the deck in the `tile` material, which the batched cut pieces already use, so no new surface; and a separate cause for the slate roof's pepper, likely the tiles' own dark edges inked by the luma outline) and is left to the orchestrator. ... 451: - **Look review fix (Mac, 26 September 2026).** Opus 5.5 on the Mac. The orchestrator's look review found five must-fix items: (1) the `plain` deck's sky specular (`plain_body`: roughness 0.55, specular 0.5) turned every tile gap pale sky-blue on slate and cream on shingle, so "tile rows read at lane zoom" failed; (2) the harbour-day slate roof's 29 specks did not move, in a regular lattice pointing at the tile geometry; (3) the dormer prism rendered as a pale blue-grey wedge; (4) styles-day was misfiled as unchanged (above); (5) bench, full suite and `--verify` not run. Everything is under `artifacts/beautify/18-P0/d/` (`fix/`, `cap/pre-3`, `pre-4`, `tuned-1`, `tuned-2`, `metrics/lookfix/`, `bench/`, `suite/`, `sheet/lookfix/`). Backups of the edited files: `fix/backup-pre-lookfix/`. Power for every set: AC attached, 81-83%, Low Power Mode 0 (`fix/power-caps.txt`, `bench/power.txt`, `suite/power.txt`). 452: - **Cause of the slate pepper, found by a diagnostic capture.** A scratch copy of the live tree (`fix/exp-tree/`, made by `fix/mkexp.sh`; the live tree was never edited for a diagnostic) coloured the instanced tiles' non-top faces: every one of the 20 specks in the `roof_slate` rect 650,250,735,295 turned green, the tiles' own side faces (`fix/diag-slate-x8.png`, left the P0d capture, right the diagnostic). The side faces are shadowed and show through the 8 mm joints between tiles; the outline inks them. The deck was never the cause on this roof. The same side faces made most of the cone pepper. 453: - **Changes (all in the part's own code).** `surface_body.gdshaderinc` (vertex): an instanced tile (`tile_piece`) shades every face with its top normal (`NORMAL = vec3(0,1,0)`, `wn` from it), so a joint shows the side lit like the tile. `architecture.gd` `profile_roof` and `openings.gd` `dormers`: the deck and the dormer prism draw in the `tile` material (the cut pieces' material: `surface_body` kind 0, roughness 0.77, specular 0.25, rain-darkened; a batched piece, so the per-tile gate is off) at `deck_tone`, still quiet and in `MODE_POP`. Tuning: `TILE_LIP_LIGHT` and `TILE_SHADOW_DARK` 0.12 -> 0.24 (the entry's first values; at 0.12 the rows barely read on the dark canal and styles roofs, `fix/exp1-*` against `fix/exp2-*`); still equal, so mean-preserving. Comments in `masonry.gd`, `mesh_batch.gd`, `build_audio.gd` updated. `build_anim.gd` unchanged. ... 456: - **Acceptance, number against target** (0 dark specks, `art.hspecks`, pre -> after, both captures of each side agree): 457: - harbour-day `roof_slate` 29 -> **0**; `cones` 20 -> **0**. Region means 53.0 -> 54.3 and 49.5 -> 51.3 (the review's bright flecks gone: the P0d plain deck had 64.2 on the cones). By eye (`sheet/lookfix/p0d-fix-slate-harbour-day-x6.png`, `p0d-fix-cone-harbour-day-x5.png`): no black specks, no cream flecks or cream tip; the slate rows read as soft diagonal bands. 458: - harbour-night (no-regression): `roof_slate` 0 -> 0, `cones` 0 -> 0 (mean 12.6 -> 13.3). 459: - Tile rows at lane zoom: `roof_lane` 1 -> 0 specks, mean 28.3 -> 30.5 (P0d plain: 40.6); `cone_lane` 12 -> 4, mean 56.2 -> 57.3. By eye (`p0d-fix-lane-dormer-roof-x3.png`, `p0d-fix-cone-lane-x4.png`): no pale stripes; the canal roof's rows read as fine dark lines on the lit slope and tone steps on the shaded one, a little softer than the stone grout's lines before; the cone reads as separate shingles with per-tile value and rows. The four cone specks left are single dark joint pixels at cut-piece edges, not the target region's acceptance line (that is harbour-day and harbour-night). 460: - Dormer prism: `dormer_gable_lane` mean 20.9 -> 26.0, p90 34.2 -> 36.9 (P0d plain: 62.7 and 128; corrected at the stage close, the 25.2 -> 32.7 first written here was the mean blue channel). It reads as dark slate with the roof, no wedge. 461: - No tone seam between whole tiles and cut pieces: none visible on the cone crops or the slate roof crops. 462: - styles-day: whole-frame change 12.2% (>8: 0.76%) against noise pairs of 22.0% (0.85%) and 19.2% (0.73%): at noise. The gabled roof keeps no stripes; its rows read faintly (`p0d-fix-styles-roofs-x2.png`). quarter-day 10.1% (>8: 1.67%) against 9.1% (1.47%) and 8.8% (1.43%): at noise, a little above on >8. 463: - Far focus (the owner's question, not decided here): the deck, where the tile chunks are hidden, now reads as a dark flat roof at the tiles' tone, without the glare; it has lost the stone pattern's texture (`p0d-fix-far-focus-deck-x2.png`, pre left). 464: - **Draw calls** (`metrics/lookfix/counts-pre3-tuned1.txt`; both noise pairs identical): harbour-day and harbour-night 891 -> 892, lane-day 734 -> 735, styles-day 475 -> 484, harbour-far and quarter-day identical; `rendered_objects` up by the same. This differs from the declared delta (down by up to 1 per pass): a cluster mesh whose roof has no batched cut pieces now gains a `tile` surface for its deck (styles' gabled roofs), while Harbour's one stone-only deck surface still goes. Tile instance counts unchanged. Headless surfaces on the live tree (`fix/surfaces-lookfix.log`): Harbour 151, quarter 138, styles 119 (not comparable with the cloud's probe numbers, whose tile chunk visibility differs). 465: - **Bench** (`bench.sh`, the P0b form: dense.json, 1920x1080, `--disable-vsync` before `--path`, 13:00, rain 0.45, 3 interleaved rounds against the pre-part tree; `bench.out`, `bench/`; AC, Low Power Mode 0 at start and end): mean frame pre 14.58 / 14.56 / 14.56 ms, live 14.74 / 14.78 / 14.74; **medians 14.56 -> 14.74 ms (+0.18 ms, +1.2%)**, higher in every round. p95 medians 14.64 -> 16.67 ms (pre 16.53 / 14.64 / 13.50, live 16.67 / 15.23 / 16.67; p95 swings more than the change). Draw calls 2140 on both. The likely cost is the deck's fragments in the surface shader (three noise lookups) where `plain` was cheap. ... 469: - **Review:** code-only, Opus 5.5 in a fresh context: 3 must-fix. (1) The plain deck and dormer prism were heard as plank clicks in timber clusters: fixed with quiet pieces and pinned. (2) Roofs whose tiles are hidden now show a flat deck: disclosed above, comments updated, a Mac check added; the look decision is the owner's. (3) Two entry errors (which run missed which timing check; styles-day holds no cones or dormers): corrected. Optional findings taken: the scale-proof tile gate, the lone-tower surface note, the cones' scattering of instanced tiles, the dormer front gable check, the instanced-variant shader compile. Recheck (same reviewer, `923d5ca`): all three closed (every click event of lone medieval and japanese houses and towers and of medieval and canal houses with a dormer identical to P0c; the tile gate verified; 0 shader errors on the plain, instanced and build variants). One new minor finding, verified: the japanese dormer's lost course click, disclosed above and pinned. 470: - **Platforms:** Linux headless (cloud) and Mac run (suites, captures, bench, full suite and `--verify` at 2880x1800); Windows not run. 471: - **Mac review:** Opus 5.5 look review in a fresh context [run ID removed] confirmed the Mac half's stop: 5 must-fix. (1) The `plain` deck took sky specular, so tile gaps showed pale blue or cream stripes; (2) the harbour-day slate pepper (29 specks) did not move, because it came from the tiles' own shaded side faces, not the deck; (3) the dormer prism read as a pale wedge; (4) styles-day was mislabelled unchanged; (5) no bench or full suite. One fix pass: the deck and dormer prism in `tile` (quiet, same pop), instanced tile side faces shaded with the top normal, `TILE_LIP_LIGHT` and `TILE_SHADOW_DARK` 0.12 -> 0.24; specks 29 -> 0 and 20 -> 0 on harbour-day, bench +0.18 ms (+1.2%) on dense.json, full suite 18,392 checks, 0 failures. Recheck: all five closed, acceptance met. Owner decision on the P0 sheet: the far-focus deck where tile instances are hidden (now a dark flat roof). Draw calls rise by 1 on harbour and lane and 9 on styles-day (a `tile` deck surface on roofs without cut pieces). Merged to `main` on 26 September 2026. ... 473: **Beautify P0c Island bank and canal lawn (25 September 2026):** Opus 5.5, in a cloud session (Linux, no GPU), implemented report Defects 2 and 7 with owner decision D3 (waterline lobes to the top lip as overhanging grass; quay copings keep one moss tuft). Code and headless tests in the cloud; **captures and bench: done on the Mac** (25 September 2026, see "Mac half" below; no part bench, the stage bench covers it). **Stop condition hit on the night line** (below; it needs an orchestrator or owner decision). After the Mac look review, the lip grass was redrawn on the Mac on 26 September 2026 ("Look review fix" below). Branch `beautify/P0c`, from tag `beautify-18-P0-c`. Files: `scripts/architecture.gd`, `scripts/look_profile.gd`, `tests/test_look.gd`, `tests/test_nature.gd`, `tools/run_tests_linux.sh` and `tools/suite_linux.sh` (new, the Linux runner), `docs/visual-handoff.md` (this entry), `docs/beautify-cloud-handoff.md`, and on the Mac `docs/beautify-plan.md` (the open night line). Evidence is under `artifacts/beautify/18-P0/cloud/` (git-ignored, on the cloud container only; the numbers that matter are copied here). 474: - **What changed.** 475: - **Bank (Defect 2).** The plain cut side of meadow, sand and an earth path along the water is `Architecture.BANK` 43382f (was an inline 605650, lighter than the lawn). The quay face is unchanged (`STONE.darkened(0.2)` in `stone`). 476: - **Waterline lobes (D3).** The three 36574b leaf lobes that sat at y 0.23-0.29 on every third cell's cut and quay sides are gone. 477: - A meadow cell's cut sides carry them at the top lip instead: `Architecture.bank_lip_grass(state, x, z, feet)`, three flaps a side in `LIP_GRASS` 5f8a55. Each flap is one sheet (`LIP_FLAP` 0.24 along the side, scaled 0.9-1.1, by 0.16 out), drooping `LIP_DROOP` 0.35 rad outward, drawn in the lawn's own `grass` material (`ground.gdshader`) with `MeshBatch.append_quad`. Its centre is `LIP_OVERHANG` 0.08 m out over the cut and 2 cm below the lip, so its inner end meets the lip. The flaps sit at t 0.2-0.8 along the side (jittered, so they never reach the neighbouring cell). As the cloud first wrote them, these were flattened `leaf` lobes (`LIP_LOBE` 0.24 x 0.09 x 0.24, 0.05 m out, 1.5 cm above the lip). The Mac look review found them inked to near-black at Harbour zoom, so they were changed on the Mac ("Look review fix" below). One side in three grows, picked per side by `bank_side_grows(x, z, side)` (a non-linear `cell_hash`; see the review fix). No lip grass on a side onto stepping stones. 478: - Like the tufts (B4a), a lip lobe is not drawn where a ground footprint covers its foot (`LIP_FOOT` 0.06 m in from the lip), and `hidden_tufts` reports those feet, so the chunk fingerprint in `WorldView.sync_scene` rebuilds the ground when a house moves over or off the lip. 479: - A quay side keeps one `MOSS_GREEN` cushion (`COPING_MOSS` 0.2 x 0.06 x 0.1, long along the side) on its coping, on the sides `bank_side_grows` picks, at t 0.2-0.8 along the side by `cell_hash`. It is not hidden under footprints, as the coping itself is not. 480: - Sand and an earth path's cut keep no lobes (grass on sand or a path's lip would be wrong). This is the part's decision on "lobes can overhang a path": lip grass is only on meadow cells and only over water. 481: - **Flat meadow (Defect 7).** `Architecture.ground_color` drops the per-mark lightness step on the meadow in canal too (Glade already had). Every other finish keeps it in both looks. 482: - **Tufts (Defect 7).** `meadow_tufts` returns one `{pos, index}` per blade (index = height step 0-2) instead of a bare `Vector3`, because a canal tuft is no longer a row along x at (0.25, 0.4): 483: - `LookProfile.tuft_rows(environment, mark, cell)` gives each row as `[ox, oz, dx, dz]`. Which cells are tufted is still the mark's (every fourth cell). Canal: `canal_tuft_row(cell)`, from three 10-bit fields of `Architecture.cell_hash(x, z, 1)`: the middle blade in the cell's middle 0.56 m square and the direction over half a turn (Harbour's 155 tufted cells: 145 start offsets at 0.01 m and 32 directions at the test's 0.1 rad binning). Glade: its two x-aligned rows from the old offsets, **byte-identical** (tested blade by blade on Harbour and styles). 484: - Canal blades are `MEADOW.darkened(CANAL_TUFT_SHADE 0.15)`; only the tallest blade of each tuft keeps 789269, on its top 40% (`TUFT_TIP` 0.6, `LookProfile.tuft_tip`). Glade has no tip (tip = body), so Glade's tuft drawing is unchanged. 485: - `LookProfile.tuft_origins` is replaced by `tuft_rows`; `TUFT_ORIGINS` keeps only the Glade offsets. 486: - **Fringe (the plan's open question).** The part decided that the canal path fringe follows the tufts: `LookProfile.fringe_color` is the canal tuft body (each blade keeps its ±0.06 lightness step), so 789269 is left only on the tuft tips. Glade's fringe keeps `FRINGE_GRASS` 789269, unchanged. ... 507: - **Acceptance, number against target** (`metrics/accept-*.json`; `accept.js`): 508: 509: | Line | Before | After | Target | | 510: |---|---|---|---|---| 511: | `bank_over_lawn` harbour-day | 1.413 | 0.864 | ≤ 1 | pass | 512: | `bank_over_lawn` harbour-night | 1.732 | 0.863 | ≤ 1 | pass | 513: | `bank_minus_water_open` night | 38.6 | 4.1 | ≥ 8 | **fail** | 514: | `bank_minus_water_below` night | 25.7 | 3.3 (2.9 in `after-2`) | ≥ 8 | **fail** | 515: | `bank_over_lawn` harbour-golden | 2.448 | 1.475 | ≤ 1.6 | pass | 516: | glade-day bank / lawn p50 | 125.2 / 119.5 | 78.7 / 119.5 | bank ≤ lawn | pass | 517: | `waterline` `art.black` day / night / golden / glade | 24 / 33 / 29 / 68 | 0 / 0 / 0 / 0 | 0 | pass | 518: | `waterline` `art.specks4_pct` day / night / golden / glade | 1.47 / 0.25 / 1.72 / 0.35 | 0 / 0 / 0 / 0 | none isolated | pass | 519: | grove-day cell-to-cell step (below) p95 / p99 / max | 2.57 / 5.29 / 8.42 | 1.13 / 1.57 / 3.5 | under 1.5 | pass at p95; max is a shadow edge | 520: | grove-day `med5_steps_max` (kit, one column) | 8.9 (13 over 4) | 5.8 (2 over 4) | under 1.5 | reads tufts and gradients | 521: 522: The rendered bank is `#635342` by day, `#252528` at night, `#67472c` at golden (was `#93887c`, `#424856`, `#96775b`). 523: - **Grove steps, separated** (`lawnsteps.js`, `metrics/lawnsteps.txt`): 11 columns x = 300-500 over y 600-960, a 7x7 art-pixel median (removes tufts, keeps straight cell edges), then per art pixel the jump in excess of the local slope (the cell-to-cell step) and the local slope itself (the in-cell gradient). Step p95 2.57 → 1.13, p99 5.29 → 1.57, max 8.42 → 3.5, steps over 4: 50 → 0. In-cell gradient p95 1.42 → 1.07 luma per art pixel. The 3.5 at (300,746) is the soft edge between a shaded blue-green band and a sunlit band, not a cell edge (`look/grove-300-746.png`, x8). Without the 2D median (tufts left in) the largest jumps are tufts (19 luma). 524: - **By eye** (x3 nearest-neighbour crops, before left, after right, `look/`; copies in `sheet/`): the grove lawn's diamond quilt is gone, only smooth light-and-shade bands remain, and the tufts are scattered with no rows (`grove-lawn.png`); the same on the harbour's right lawn (`lawn-right-harbour-day.png`). No tuft lattice on grove-day or harbour-day. The black beads at the waterline are gone in all four views (`bank-harbour-*.png`, `bank-glade-day.png`). The lip grass reads as three small flattened green lobes hanging over the top edge of the cut at lane zoom (`lip-harbour-lane.png`) and as a broken dark green fringe along the lip at harbour zoom. That second reading was not measured, and the look review showed it wrong: at harbour zoom the lobes were ink pixels (see "Look review fix"). I saw none over a path in these frames (by code and test it is only on meadow sides over water). At night the bank is a flat, neutral dark grey band (`#252528`) that separates from the water mostly by hue; the wet-dark earth look by day and golden is right. 525: - **Tuning trial: the night line cannot be met by `BANK`.** In a trial copy of the P0c code (`trial-tree/`, `trial.sh`, `trial//`; 43382f there reproduced the live numbers exactly) and then in the live tree: 526: 527: | `BANK` | day b/l | night b/l | night − open | night − below | golden b/l | 528: |---|---|---|---|---|---| 529: | 43382f (cloud) | 0.865 | 0.861 | 4.1 | 3.3 | 1.475 | 530: | 443b3b | 0.926 | 0.937 | 7.4 | 5.4 | 1.576 | 531: | 3f3b3d | 0.912 | 0.924 | 6.8 | 4.8 | 1.544 | 532: | 473c34 (trial) | 0.937 | 0.954 | 8.1 | 5.3 | 1.589 | 533: | 473c34 (live, `cap/tuned-1`, `-2`) | 0.937 | 0.954 | 8.1 | 5.3 | **1.609** | 534: | 403c46 | 0.933 | 0.961 | 8.4 | 5.4 | 1.597 | 535: | 4c4038 | 1.010 | 1.086 | 13.8 | 9.6 | 1.741 | 536: 537: A cooler hue barely helps: the night bank gains about as much as the golden bank whatever the hue. The open-water line (≥ 8) and the golden cap (≤ 1.6x lawn) meet in a window under 1 luma wide, and 473c34 fell out of it between two sittings (golden bank 81.2 in the trial, 81.9 live). The below-water line cannot pass while the night bank stays under the night lawn (43.2): `water_below` carries the bank's reflection and rises about 0.35 per luma of bank, so it needs a night bank of about 44.5 (from 43382f to 4c4038 the bank gained 9.7 and `water_below` 3.4), over the lawn. That is not a tuning problem, so it is a **stop condition**: `BANK` stays at the cloud's 43382f; the tuned code and test (473c34 with `OLD_CHUNKS` re-recorded) are kept in `backup/tuned-473c34/`, the probe that re-recorded it in `probe/old_chunks.gd` (it reproduced the committed record 46 of 46 with 43382f first). The plan already moves the night line to after P1a, which changes the night water and sky; the orchestrator or owner decides whether P1a should own it, or whether the line is changed to `_open` only. 538: 539: What the trial shows, and what it does not: within the trial sitting, both 473c34 (8.1 / 1.589) and 403c46 (8.4 / 1.597) met the open-water line and the golden cap. 473c34 missed the cap when it was run live (1.609), and 403c46 was never run live. So the runner's summary "no `BANK` value meets both" says more than the evidence does. The accurate statement is this: no value met the open line and the golden cap in a live sitting; the window between them is under 1 luma wide; and the below-water line needs a night bank of about 44.4, above the night lawn (43.2), which contradicts "bank no higher than lawn at night". The look review re-derived both night numbers from the PNGs (`review/measure-*.json`) and the trial table from `trial//measure.json`, and they match. **Decided (orchestrator, 26 September 2026): the night line is deferred to the re-measure after P1a**, which the plan already calls for; P1a recolours the night sky and water, so the water side of the line moves. P0c closes on its other lines. The re-measure after P1a reads both night margins (open water and water below) and reports them against the conflict above; if they still fail, the line goes to the owner with the P1 sheet (G1). Reversible by the owner. The plan's P0c section records the same open item. ... 544: - **Look review fix (Mac, 26 September 2026).** Opus 5.5 on the Mac, same live tree. All the evidence is under `artifacts/beautify/18-P0/c/lookfix/` and `cap/`, and the pre-fix files are in `lookfix/backup/`. Power for every look-fix capture set and the suite: **mains (AC attached, not charging), 86-94%, Low Power Mode 0**, which is the owner's D1 state (`cap//capture-meta.txt`, `lookfix/power-suite.txt`). The look review found three must-fix items. 545: 1. **The lip grass inked to near-black at Harbour zoom (fixed).** 546: - What the review found: each `leaf` lobe was 1-2 art pixels, one lit green pixel and one ink pixel. The foliage shader's rim step and shadow band darkened the lobe's underside, and that underside is seen because it overhangs the cut. The waterline beads had moved up to the lip. The plan's `waterline` region could not see this. 547: - Scorer: `lookfix/lipspecks.js`, the review's rule: in the lip strip, art pixels under luma 25.5 that are more than 15 darker than both horizontal neighbours, plus window pixels under luma 10. It reproduces the review's numbers except glade-day `after`, where it gives 7 and the review gave 4. 548: - Change: each lobe is now one sheet of grass in the lawn's own `grass` material, drooping over the cut. See "What changed" for the constants. ... 596: - **Acceptance, re-measured** (`metrics/accept-pre-11_pre-12_fix-5_fix-6.json`): every line is as in the table above, to the digit. Bank/lawn: day 0.865, night 0.861, golden 1.475, glade bank 78.7 against lawn 119.5. Night minus open water 4.1 and minus water below 3.3 (**still failing**, item 2). `waterline` `art.black` 0 in all four views. Grove cell-to-cell steps (`lookfix/lawnsteps.txt`, W=3): p95 2.57 → 1.13, p99 5.29 → 1.57, max 8.42 → 3.5. 597: - **Full suite** (`lookfix/suite/suite_all.out`, `VERIFY_RES=2880x1800`, live tree): **40 suites, 18,355 checks, 0 failures**. `--verify` at 2880x1800 gave `PIXELPUNK_INTEGRATION []` in 31 s, and the `test_creature_reactions` rerun passed. There was no Mac-only failure. 1440x900 was not run (inherited stall). 598: - **Sheet:** `sheet/p0c-lookfix-lip.png` (the kit's sheet generator, before `pre-11` | after `fix-5`, harbour night, golden and day, glade-day and lane, colour and greyscale). It uses a lip region that exists only in `lookfix/lip_regions.json`; `tools/art_regions.json` is unchanged. The x6 and x3 crops above are for reading on a phone. 599: - **Not done:** the night line (deferred to the re-measure after P1a, item 2); no bench (the stage bench covers it). 600: - **Platforms:** Linux headless (cloud) and Mac run; Windows not run. 601: - **Mac review:** Opus 5.5 look review in a fresh context [run ID removed]: 3 must-fix. (1) The leaf lip lobes inked to near-black at Harbour zoom by night and at golden: fixed with one-sheet flaps in `grass`, rechecked at noise level. (2) The night line: deferred to after P1a by the orchestrator (above). (3) The power record named the wrong set: corrected. Left for the owner (P0 sheet; P3c redoes the lip): in Glade the flaps read as bright lime squares above the lawn (130.9 against 120.5). Draw calls fall on every journey view but quarter (+9, moss). The orchestrator fixed the stale `bank_lip_grass` comment. Merged to `main` on 26 September 2026. ... 604: **Beautify P0b Wet-leaf glints instead of gloss plates (25 September 2026):** Opus 5.5 replaced the foliage wet gloss (one hard `step(0.94, N·H)` cap per crown in the light's colour, times `rain*0.2`) with sparse wet glints (report Defect 1). Two orchestrator-directed fix passes followed. The first came after the first review [run ID removed]: stable cell size, one-pixel flecks, a leaf-green tint and a night bench. The second came after the re-check review: per-instance cell sizing in perspective, and less crawl. This entry describes the final state after the second fix pass. Files: `shaders/foliage.gdshader`, `tests/test_foliage.gd`, `docs/visual-handoff.md` (this entry). The pre-part copy is `artifacts/beautify/18-P0/b/visual-handoff.md.pre-P0b`, the reviewed copy is `b/orch-pre-fix/visual-handoff.md`, and the copy after the first fix pass is `b/backup/visual-handoff.md.fix2-start`. Backups are in `b/backup/`: ... 612: - **What the orchestrator asked for in the second fix pass, and what changed.** The work is in `b/fix2/`. 613: 1. **Speckle in a perspective free-camera view** (must fix). Cells were sized from `NODE_POSITION_VIEW`. Street trees are one MultiMesh whose node sits at the world origin, so in `--photo-lens --free-camera=-3,4,4,3.14159,0.12` the cells came out far under an art pixel. The reviewer measured 14.4-14.9% of foliage cells brightened relative, with blobs of 530-640. 614: - Now `vertex()` writes `draw_w`, the clip w of `MODELVIEW_MATRIX[3]`, which includes the instance transform (so it is each tree's base), into a `varying flat float`. The whole crown gets one size. 615: - A fragment more than `GLINT_LOCAL` (0.5) octaves in depth from its draw's origin uses its own depth. This covers world-space terrain chunks, whose node is also at the origin. 616: - The comment that said "one per draw" is rewritten. 617: - pf-behind now: 1.35-2.57% relative, 0% absolute, largest blob 32-63. The streak-free control against itself gives 1.84%, blob 79. See the table. 618: 2. **Crawl at ordinary orbit speeds** (ruled must fix). I measured it with a new leaf-tracked metric (see Crawl) and found the cause: under an orbit, about 15-20% of one-pixel anchors vanished each frame. They vanished because the anchor depended on which triangle, and which shell, the pixel sampled. The glint is rebuilt: 619: - Each cell has a fixed anchor point, up to `GLINT_SPREAD` 0.25 of the cell from its middle, so the glints do not line up on the cell lattice. 620: - The glint lights the art pixel the anchor projects to and the pixel to its right: a 1x2 fleck, which the first review allowed. 621: - The leaf seen along the anchor's ray is read off the fragment's own plane, from the derivatives of wp. It must lie inside the cell by the pixel's footprint plus `GLINT_MARGIN` 0.1 art px. 622: - The highlight normal is the shaded normal carried to the anchor along its screen derivatives, plus the tilt. The flat face normal jumped between a lobe's triangles. 623: - The glint key drops the per-lobe shift, so a shell in front passing through the same cell keeps the glint. 624: - Constants: cells of 4-5.7 art px (`GLINT_PIXELS` 2 → 4), `GLINT_SHARE` 0.6 → 0.65, `GLINT_CONE` 0.97 → 0.95, `GLINT_SUN` 0.4 → 0.3. The lamp strength stays 0.5 and the tint stays `mix(leaf_hue, 1, 0.25)`. 625: - Tuning rounds are `b/fix2/e1`-`e13` and `f2`-`f7` (`.gdshader`, built by `mk.js`). The first candidate (`fix2-r1`, cone 0.92, sun 0.4) gave 3.5-3.7% relative in photo-golden, so it was thinned to `f5`, which is the final shader. 626: - Result: at 0.05 rad/s and on a pan the reviewer's plain r1 now beats the reviewed build. At 0.2 and 1.0 rad/s the plain r1 stays below it, although the leaf-tracked persistence matches or beats it. **This is the stop condition below.** ... 644: - **Acceptance (rain 0.45, streak-free, against the gloss-off control; two controls → range).** Targets: at most 3% on both readings, runs of at most 4, no plate, no 3×3 blobs. 645: 646: | View | Reviewed P0b: absolute / relative | Final: absolute / relative | Final: run, blob, 3×3 | 647: |---|---|---|---| 648: | quarter-golden | 1.94% / 2.61% | 0.40-0.44% / 0.60-0.85% | 2, 2, 0 | 649: | quarter-day | 0.90% / 1.66% | 0.32-0.41% / 0.57-1.08% | 2, 2, 0 | 650: | photo-golden | 0.25% / 4.83% | 0.21% / 1.81-1.83% | 2, 2, 0 (relative blob 37-38, 3×3 14-15*) | 651: | glade (rain 0.45) | 0.08% / 1.71% | 0.08% / 0.08-0.89% | 1, 1, 0 | 652: | quarter-night | 0% / 1.64% | 0.31-0.32% / 0.58-0.98% | 2, 2, 0 | 653: | harbour-night (0.35) | 0% / 0.91% | 0% / 0% | none | 654: | pf-behind (frame, review metric) | 0.02-0.03% / 2.4-4.2% (blob 155-164) | 0% / 1.35-2.57% | blob 32-63, 3×3 96-132* | 655: | persp-q (frame, review metric) | 0.44-0.47% / 0.91-1.07% | 0.16-0.23% / 0.36-0.59% | blob 14-16, 3×3 0 | 656: 657: \* Relative blobs and 3×3 blocks in photo-golden and pf-behind are blur halos: each 1×2 fleck is spread over a few pixels on near-black crowns, which raises their luma by 1-10. In the absolute set the blob is 2. 658: - For pf-behind, the streak-free control against itself gives 1.84%, blob 79, 3×3 78. The live build against itself gives 0.23%. Before this pass the reviewer measured 14.4-14.9% with blobs of 530-640, and the pre-part tree measures 23-25%. 659: - Noise pair, streak-free control twice (part metric): at most 0.09% absolute and 0.57% relative. 660: - Colour: in quarter-golden, the relative set's hue is 111-114° (green), against 95° (olive) for the reviewed build. Saturation goes 92-94% → 58-68%. ... 682: | Motion | Plain r1: final / reviewed | Leaf-tracked r1: final / reviewed | Glints per frame: final / reviewed | 683: |---|---|---|---| 684: | still | 99.5 / 99.9% | 99.5 / 99.9% | 108 / 302 | 685: | orbit 0.05 rad/s | **94.7 / 92.1%** | 94.2 / 91.8% | 108 / 287 | 686: | pan 0.5 m/s | **95.3 / 93.6%** | 95.0 / 93.5% | 107 / 305 | 687: | orbit 0.2 rad/s | 68.0 / 74.9% (r2 89.8 / 89.9%) | **93.6 / 90.5%** | 101 / 259 | 688: | orbit 1.0 rad/s | 7.8 / 19.8% (r2 21.0 / 32.4%) | 89.1 / 89.8% | 115 / 237 | 689: 690: - The first fix pass (reviewer's numbers) had plain r1 of 84.9, 89.4, 40.8 and 8.1%, and leaf-tracked 79.9% at 0.2. 691: - At 1.0 rad/s the leaf-tracked gap (0.7 points) is inside its sampling error (about ±0.9 over about 1,150 glints). 692: - **Bench** (`b/fix2/bench.sh ` → `b/fix2/bench-t22/`, `bench-t13/`, `bench-t22.out`, `-t13.out`). Setup: pre-part tree against live, interleaved over 3 rounds, 1920x1080, `--disable-vsync`, rain 0.45. Power: AC, charged, Low Power Mode 0 (`power.txt`). Read the means: p95 snaps to refresh multiples, as explained in the first fix pass. 693: - **Night, 22:00, lamps lit (dense has 16 lights):** 694: - dense: mean 15.89 / 15.83 / 15.84 → 15.88 / 15.90 / 15.81 ms (median +0.3%). p95 16.667 in all six runs. 2,397 draw calls both. 695: - forest: 8.56 / 8.78 / 8.75 → 8.70 / 8.83 / 9.04 ms (median +0.9%). 978 draw calls both. 696: - **Day, 13:00:** 697: - dense: 15.16 / 15.15 / 15.34 → 15.34 / 15.27 / 15.24 ms (median +0.7%). 698: - forest: 8.91 / 8.84 / 8.79 → 8.96 / 8.92 / 8.98 ms (median +1.4%, 0.12 ms, which is the size of the pre runs' own spread). 699: - Draw calls are unchanged (2,125 and 940). 700: - **Captures.** This sitting's sets are `b/cap/fix2-*`. The folders `b/before-/`, `b/after-/` and `b/control-/` still hold the first fix pass's sitting. 701: - **Stop condition (second fix pass), for the orchestrator to rule.** At 0.2 and 1.0 rad/s the reviewer's plain r1 (68.0 and 7.8%) stays below the reviewed build's (74.9 and 19.8%), so "at or above the reviewed build at every speed" is not met on that reading. 702: - Plain r1 looks for a glint in the same 3×3 window. At 1-3 art px of motion a frame, only a fleck wider than the motion stays inside that window. The reviewed build's glints were blobs of 2-12 cells. 703: - The first review's look ruling allows only 1-pixel or 1×2 flecks with no 3×3 blobs, so the two rulings conflict at those speeds. 704: - Following each leaf, the final glints persist as well as or better than the reviewed build's (93.6 against 90.5% at 0.2 rad/s; 89.1 against 89.8% at 1.0). 705: - I did not widen the flecks. 706: - **Review (orchestrator; [run IDs removed]):** the first fresh-context Opus 5.5 high review found no must-fix, but the orchestrator ruled four items must be fixed (crawl from `fwidth` cell sizing, olive dirt-like flecks, both the absolute and the relative 3% reading must hold, a night bench with lamps). Fix pass 1 did them; its recheck found a new must-fix (perspective cells sized from the shared MultiMesh origin gave dense speckle in the photo free camera, 14.4-14.9% relative) and fix pass 2 closed it (`draw_w` per instance; pf-behind 1.1-1.5% relative, 0% absolute). The second recheck (`b/review/r3/`) re-derived every number and found acceptance met. **Orchestrator ruling on the crawl stop condition:** plain r1 is the crawl measure at 0.05 rad/s and on pans (94.7 and 95.6%, above the reviewed build); leaf-tracked persistence is the measure at 0.2 rad/s and above (93.2-93.6% at 0.2 against 90.5%; at 1.0 rad/s 89.1-89.7% against 89.8-89.9% with `track2.js --S 16`, 85.6 against 86.9% at the default `--S 8`, which loses about a third of the leaf points). The flecks stay 1x2. Crawl item closed. The known flaky check did not fail. The second review (amendment 14) was not run for P0b (optional in the plan). ... 713: **Beautify P0a Measurement kit, runner and P0 baseline (25 September 2026):** Opus 5.5 built the measuring kit the beautify plan's acceptance lines are read with, checked it against the report's numbers, captured the P0 baseline and ran the pre-flight full suite. No game file changed: `diff -rq` of scripts, shaders, tests, data, scenes and `project.godot` against `build/backup-feel/18-P0/` is empty. The only difference under `tools/` is the four new files. Files: `tools/art_capture.sh`, `tools/art_metrics.js`, `tools/art_regions.json`, `tools/suite_all.sh` (all new), `docs/visual-handoff.md` (this entry; the previous copy is `artifacts/beautify/18-P0/a/visual-handoff.md.pre-P0a`). Snapshot: `build/backup-feel/18-P0/`. The scratch tree is `artifacts/beautify/18-P0/a/pre-tree/`. At part start the live tree matched the snapshot in all six folders. ... 734: - **Tests.** There are no new tests: no game code changed, so there is no test-first log. The kit's own check is `validate`: **43 of 43 report numbers reproduced, 16 of 16 on the P0a acceptance list**, on `cap/` (`artifacts/beautify/18-P0/a/validate.txt`, `.json`), and 43 of 43 again on the new baseline (`validate-on-baseline.txt`). 735: - Pre-flight full suite on the live tree, whose game files equal `18-P0`: **40 suites, 18,284 checks, 0 failures**. The `test_creature_reactions` rerun also passed, on AC power. Files: `artifacts/beautify/18-P0/a/preflight/suite_all.out` and `suites/`. ... 738: - **Inherited, not booked to any beautify part.** `--verify` at 1440x900 stalls on the 18-P0 tree: it exits 0 at `--quit-after` after about 300 s, with no `PIXELPUNK_INTEGRATION` line and 2 to 7 PNG captures (`preflight/verify2/`, run standalone). At 2880x1800 it passes: `PIXELPUNK_INTEGRATION []`, 52 PNG captures, 33 s (`preflight/verify-2880/`, and again through the runner in `suite_all.out`). 739: - Nothing under scripts, shaders, tests, data or scenes changed after the 11:44 2880x1800 pass in `artifacts/ui-scale-0925/r2/`. The earlier 1440x900 pass predates the tray autohide change, so the stall is at 1440x900 only. It is probably the tray change: an inference, not isolated. 740: - Until it is fixed, `suite_all.sh` spends about 5 min on the stalled size each run. `VERIFY_RES=2880x1800` skips it. 741: - The first runner verify's `verify.log` was overwritten by the 2880 rerun; `verify2/` keeps the stall evidence. 742: - Every run also logs the known `terrain_flowers.svg` missing-resource error, as the audit runs did. 743: - A background task was suggested for the stall. 744: - In the fix-pass run (14:18, same game files, AC power, fully charged) `--verify` at 1440x900 **passed**: `PIXELPUNK_INTEGRATION []`, 52 PNG captures, 21 s (`a/fix/suite/verify.log`, `integration/`). So the stall is intermittent, not a fixed failure at that size. The cause is not isolated. Keep the default `VERIFY_RES`, and skip 1440x900 only when it stalls, saying so in the entry. ... 750: - **Measurements.** 751: - The P0a acceptance numbers on `cap/` (report value in brackets): 752: 753: | Measurement | Kit | Report | 754: |---|---|---| 755: | quarter-night below 0.1 | 46.2% | 46.2 | 756: | front quay p50 | 3.7 | 4 | 757: | front quay below 3 | 47.6% | 48 | 758: | zenith p50 | 31.1 | 31 | 759: | horizon band | 88.9 | 89 | 760: | harbour bank, day / night / golden | 137 / 71.7 / 123.6 | 137 / 72 / 124 | 761: | lawn, day / night / golden | 102.7 / 45.7 / 54.1 | about 100 / 44-49 / 52 | 762: | grove steps, `med5_steps_max` | 8.9, with 13 steps over 4 | 5-11 | 763: 764: - Other numbers: 765: - The quay ratio is 0.444 (0.45): paving mean 31.1, `#142130` (31, `#14202d`). 766: - Specks at a 0.04 threshold, which is the lens's: 0.617 / 0.465% (0.60 / 0.46). 767: - The slate roof reads 2.45% (1.9%). The lens's region is not given. 768: - Harbour zenith reads 27.3 (30), for the same reason. 769: - Determinism: the four noise views were captured three times from the scratch tree, all on AC power: `a/noise-1` and `a/noise-2` in one sitting with the baseline (13:14-13:16), and `a/fix/noise-3` about 50 minutes later (14:04). The reviewer's `a/review/cap` (13:46) adds quarter-night and grove-day. Per-region values of every capture and the ranges are in `a/fix/noise/noise-regions.txt` (and `.json`); the pair diffs are `a/fix/noise/ab-*.txt`. 770: - Pixels differing by more than 8 on any channel, within one sitting (noise-1, noise-2, baseline): grove-day 0.02-0.04%, quarter-day 1.1-1.9%, harbour-night 1.6-2.9%, quarter-night 2.1-3.4%. 771: - Against the later sitting (noise-3): grove-day 0.26-0.30%, quarter-day 3.8-4.3%, harbour-night 4.9-5.6%, quarter-night 6.3-6.9%, about twice the same-sitting figure. The reviewer's capture gives quarter-night 6.3% and grove-day 0.15% against noise-1. 772: - The cause is the run length. The water, rain and star phase follows TIME, and In the first sitting `elapsed_ms` was 3,685-3,788 ms on the quarter views and 4,065-4,142 ms on harbour-night. In noise-3 it was 3,346-3,367 and 3,538 ms, and grove-day ran 2,362 ms against about 3,200. noise-3 and the reviewer's capture, whose run lengths are close, differ by only 3.1% on quarter-night. So check `elapsed_ms` in both `performance.json` files. When the two runs differ by more than about 300 ms, use the cross-sitting figure as the noise. 773: - Region p50s are not all stable, and the earlier claim here that every p50 moved by 0.1 or less was wrong. The p50 ranges (same sitting / all sittings) are: 774: - quarter-day: `boat` 1.2 / 1.3, `crowns` 0.1 / 1.0, `windows_tower` 0.2 / 0.7, `paving` 0.5 / 0.5, `railings` 0.2 / 0.5, `plaza` 0.4 / 0.4, `iron_rail` 0.4 / 0.4. 775: - quarter-night: `boat` 0.5 / 0.9, `iron_rail` 0.4 / 0.8, `sky_right` 0.5 / 0.5, `mirror_seam` 0.1 / 0.5, `back_row` 0.4 / 0.4, `roof_terraces` 0.3 / 0.4, `cabinets` 0.2 / 0.4. 776: - harbour-night: `boats` 0.4 / 0.7, `lawn_right` 0.2 / 0.5. 777: - These moved by 0.1 or less across all sittings: zenith, horizon band, skyline, bank, lawn, front quay, left pier and `pipes`. `land_wall` moved by up to 0.2. Among the derived values, `quay_over_paving` (a ratio of means) moved by up to 0.01 and quarter-day `skyline_minus_horizon` by 0.1; the others did not move. 778: - Speck counts also move: `sky_right.bspecks4` 12-15, quarter-day `crowns.bspecks4` 28-37, harbour-night `boats.bspecks4` 88-108. 779: - A part calls a value unchanged only when its change is inside that region's all-sittings range in `noise-regions.txt`, or inside a noise pair captured with the same run length. Water moves most: `mirror_columns` changed 14% at more than 8. ... SOURCE: artifacts/beautify/owner-decisions.md (frozen copy) 1: # Owner decisions for docs/beautify-plan.md (section 6) 2: 3: Answered by the owner in chat on 25 September 2026, while P0a ran. These override the plan's defaults where they differ. 4: 5: | # | Answer | Differs from the default? | Effect on the plan | 6: |---|---|---|---| 7: | D1 | Mains power with Low Power Mode off (the owner sets this up) | yes | Benches record the power state; never mix power states inside one A/B. The P0a baseline and pre-flight suite may have run on battery: note it, don't compare their timings with mains runs. The flaky creature check should stop failing on mains; if it still fails, record it as known. | 8: | D2 | Second Opus review in a fresh context (owner said Opus 5.5 high in every role) | no | as budgeted | 9: | D3 | Waterline lobes move to the top lip as overhanging grass; quay copings keep one moss tuft | no | P0c | 10: | D4 | **No foam stopgap now**: leave the foam to P3a | yes | P0e drops the foam item (Defect 3 stopgap). Do not change the foam threshold in `water.gdshader`. | 11: | D5 | **Gentler star fade in rain**, not the report's (gone by 0.2) | yes | P0e: stars dim with rain but some still show at the default rain 0.35 (about half strength there), gone by about 0.6. One named constant pair so it is a one-line change. | 12: | D6 | **Stop after P1c** for the G1 verdict | yes | P1 runs P1a, P1b, P1c, then the orchestrator sends the G1 sheet and waits. P1d-P1f run only after the verdict. | 13: | D7 | **A new `painted_metal` material**, not a flag | yes | P1f takes the material route: +30-45 min, a part bench, `test_warmup` coverage, and the extra tests the plan lists for that route (`test_build_anim`, `test_warmup`, `test_release_spread.gd:353`). | 14: | D8 | **Builds at M3 and M6 only** | yes | P1g (M1) is not run; P2's M2 export is dropped; M3 after P4 is built; M4 (P5) and M5 (P8) are dropped; M6 at the end of P9. | 15: | D9 | Leaded medieval windows; no "under 0.6 m transom only" rule | no | P2a | 16: | D10 | Fog and smoke on for starters, SampleScene and the new canvas; `artwork.json` unchanged | no | P2b | 17: | D11 | Canal contact shade, cost reported | no | P2d | 18: | D12 | Collar follows the ground finish | no | P2e | 19: | D13 | Beached boats: no lift | no | P3a | 20: | D14 | Beacons and ring neon steady until P8a; the skyline switch hides the 3D ring | no | P4a | 21: | D15 | Drop the frame tubes on marked blades; no `icon` param | no | P5a | 22: | D16 | Tray: fixed dock offset | no | P5e | 23: | D17 | Motion preference covers the new terms and P5b glints only; puffs off under Off | no | P8a (no P8a2) | 24: 25: Orchestrator ruling, 26 September 2026 (not an owner answer; the owner may reverse it): P0c's night bank line (bank at least water + 8 at night) conflicts with the part's other lines on 43382f (open 4.1, below 3.3). It is deferred to the re-measure after P1a that the plan already calls for; if it still fails there it goes to the owner with the G1 sheet. 26: 27: Orchestrator ruling, 26 September 2026 (reversible by the owner): P1b's line "quay p50 at least 8 at 21:00" (5.5 after P1b) is not pushed in P1b, because 8 at 21:00 would sit above both 20:45 (6.8) and full night (3.7) and flash. It is re-measured after P1c (night fills) and after P1d (quays lit by the water, which owns the quays), and goes to the owner with the G1 sheet if still short. 28: 29: D18 to D33 need evidence from the parts and are asked when they come up. The blocking stops the owner set: before P2, the P3 sheet before P4b, and the P9 live pan. D6 adds a stop after P1c. ... 55: ## G2 verdict and open P0 sheet questions (owner, 28 September 2026: "accept recommendations and move to P2 next") 56: 57: Answered before the G2 sheet was finished, on the orchestrator's list of questions. Any of these can be reversed when the owner sees the sheet. 58: 59: | # | Question | Verdict | 60: |---|---|---| ... 68: | P0-1 | Far roofs with hidden tiles show a dark flat deck | Accept | 69: | P0-2 | Glade lip flaps read as bright lime squares | P3c redoes the lip | 70: | P0-3 | Harbour daytime water: soft darker band under the island's reflection | Keep | 71: | P0-4 | Stars in rain (D5: half at 0.35, gone by 0.6) | Keep | 72: | Gate | Blocking stop before P2 | Passed: go to P2 | ...