Chapter 21: Where the island meets the water Frozen 30 September 2026 at 07:10 BST (game main at 085517d). Implemented, reviewed and merged work of the P3 stage (islands that sit in water): P3a waterline rim, swash and shallows, boat hull skirt; P3b rain rings on open water, CPU splashes retired; P3c cake-slice banks I; P3d cake-slice banks II. P3a to P3c were merged under the owner's overnight standing instruction, P3d on the owner's instruction on 30 September ("merge and update journal"). The P3 stage close has NOT run: no whole-stage bench, no protected-look check, no P3 sheet; P3-1, P3-2 and P3-3 are open. The plan's night 0.8x-of-lawn bank line is waived by an orchestrator ruling that the owner can reverse. Original source terminology is preserved; workflow run IDs and local paths are removed. Line numbers refer to the frozen copy of docs/visual-handoff.md taken at 07:10 on 30 September. Part codes: P3a waterline, P3b rain rings, P3c and P3d cake-slice banks. SOURCE: docs/visual-handoff.md (frozen copy, lines as of 30 September 2026, 07:10) 80: **Beautify P3a Waterline rim, swash and shallows; boat hull skirt (30 September 2026):** Opus 5.5 on this Mac. The water now reads the opaque depth under each fragment (`shaders/water.gdshader`): the ray's run below the surface gives the depth of the submerged face it meets and how many render px the fragment lies from that face's waterline, from the projection alone (no screen derivatives). From it: a rim of 1-2 render px that breathes with a slow lap (`RIM_PX`), coloured `mix(sky_ref, RIM_PALE, 0.45) + mirror * 0.35` at 0.75; a one-px swash line broken by a world-cell hash 3-5 px out (`SWASH_PX`, 0.4); shallows that lift toward turquoise by day or darken at night by `shallow_lift`; Glade's rim is a soft white line (`waterline_foam`). The shallows and swash fade out by `WATERLINE_REACH` 0.2 m, inside every submerged face (cut and quay sides 0.30 m, the new boat skirt 0.2 m); over open sea nothing is drawn. The dead shoreline foam (Defect 3) is deleted; the shore mask keeps G for the old shore band and R for the streaks. `WorldView.shallow_lift` (in `sync_profile`) is day +0.3, golden hour +0.2, night -0.15, piecewise linear in the hour's windows between the presets' 0.12/0.66/1.0, plus the look's RENDER `shallow_bias` (canal 0, Glade 0.1). New RENDER keys `shallow_bias` and `waterline_foam` (canal 0 and 0). Boats get a hull skirt from y -0.2 to 0 in the hull colour (`scripts/architecture.gd`); beached boats are not lifted (owner D13 and the P3a ruling): the tip is now `WorldView.beached_transform` (`BEACHED_TIP` -0.12, unchanged). One GL Compatibility trap found in tuning: the pixel viewport flips `PROJECTION_MATRIX[1][1]`, so the px scale takes `abs()` (without it the rim filled the whole submerged face, `a/look/t1-quay.png`). Branch `beautify/P3a` from `main` 5b163ad. Snapshot `build/backup-feel/21-P3/`; at part start the live tree matched it in scripts, shaders, tests, data, scenes and tools. Scratch tree `artifacts/beautify/21-P3/a/pre-tree/`; evidence `artifacts/beautify/21-P3/a/`; backups `a/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture and the suite (`a/cap-power.txt`, `a/suite/suite_all.out`). 81: - **Files:** `shaders/water.gdshader`, `scripts/world_view.gd`, `scripts/look_profile.gd`, `scripts/architecture.gd`, `tests/test_water.gd`, `tests/test_look.gd`, `tests/test_scene.gd`. `mirror.gdshaderinc` needed no change (`linear_depth` is right for both projections). 82: - **Tests:** test-first `a/P3a-testfirst-test_water.log(.out)` 149 checks, 18 failures; `-test_look` 462, 11; `-test_scene` 630, 7. After (`a/r1-*`): test_water **149, 0** (was 126), test_look **524, 0** (was 451), test_scene **633, 0** (was 619). New: `waterline_checks` (uniforms at neutral defaults; reach inside the 0.30 m face and the 0.2 m skirt; rim 1-2 px, swash 3-5 px; the depth read before and outside the mirror block, the rim mixed after it; panel-depth gate; per-fragment ray under perspective; no `fwidth`/`dFdx`; the ramps reach 0 at 0.2, 0.30 m; a mirror of the px span: 3.35 px at Harbour's camera, 6.8 at the quarter's; shore mask R and G kept), `shallow_checks` (RENDER keys; the lift at all ten anchors, the three presets exact, Glade +0.1 and 0.4 by day, between hours between neighbouring anchors), `water_wired` in every editor look state (start, Glade, undo, redo, by index; 13:00 = 0.3, back at night -0.15), `boat_skirt_checks` (bottom at -0.2 for all 3 variants, skirt in the hull colour, the four lights where they were, a beached boat stored at ground height, the raised tip's skirt 0-5 cm above the ground, the lowered half inside the terrain). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/a/suite` (a/b/v): **40 suites, 19,620 checks, 0 failures** (19,510 + 110). `--verify` 2880x1800 exit 0 in 31 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures; the known flaky check did not fire. 1440x900 not run (the inherited stall). 83: - **Changed expectations (old -> new):** test_water P0e "the foam threshold is unchanged; P3a replaces the foam" (`water.contains("float foam = smoothstep(0.72, 0.94, shoreline)")`) -> "the dead shoreline foam is gone" (neither that smoothstep nor `foam * 0.18` in the shader). No check deleted. 84: - **Captures** (`a/caps.sh`, `tools/art_capture.sh`, one sitting 01:41-01:52 in the order `pre-1` (scratch tree), `after-1`, `pre-2`, `after-2`; 17 views each, all first try, plus `moff-beached`/`moff-quarter` (`--mirror=off`) in pre-1, after-1 and pre-2): harbour-day, -night, -golden, harbour-lane, glade-day, quarter-night, quarter-day, photo-golden, `photo-dof0` (photo lens at `--dof=0`), styles-day, the authored beached-boat scene `a/beached/beached.json` (`a/beached/make_beached.gd`: meadow island, sand strip, one boat beached by painting land under it, two afloat, a stone quay, stepping stones) at day, night and Glade, a 3-frame orbit on harbour-night (`orbit-a` yaw 0.70, harbour-night 0.78, `orbit-c` 0.86), and `creat-day`/`creat-night` (Harbour with `--creatures`). Crops looked at: `a/look/hd-cut.png`, `hn-cut.png` (Harbour cut, before|after), `t2-quay.png`, `t2-beached.png`, `t2-qn.png`, `t2-hn.png`, `glade.png`, `beached-glade.png`, `beached-boat.png`, `photo-water.png`, `orbit.png`, `ducks.png`. 85: - **Integer counts and draw calls identical on every view** (`a/metrics/counts.txt`); the skirt adds vertices to the boat mesh only. 86: - Whole frame, changed >8 (`a/metrics/summary-frame.txt`), A/B set 1 / set 2 against noise pre / after: harbour-day 2.12/1.25 (1.44/0.85), harbour-night 2.01/1.86 (1.41/1.27), quarter-night 3.13/2.79 (2.38/2.51), styles-day 1.23/0.90 (0.47/0.14), beached-day 1.59/1.63 (0.78/0.63), glade-day 0.80/0.78 (0.69/0.40). The change is the rim line; everything else is within the pairs. 87: - **Acceptance** (`a/rim.js`: on the render grid, a rim pixel is one brightened by more than 12 luma after -> before; `a/metrics/rim-set1.txt`, `rim-set2.txt`, `skyline-band.txt`): 88: - **1-2 render-px rim along cuts, quays, hulls, stepping stones:** met on every pixel-size-2 view. Share of vertical rim runs 1-2 px long, set 1 / set 2: quarter-night 91.9/92.8%, quarter-day 94.3/93.5, styles-day 98.9/98.6, beached-day 94.3/95.3, beached-night 93.9/93.1, harbour-day 91.5/81.9, harbour-night 83.5/86.1, harbour-golden 84.2/87.5, harbour-lane 79.6/75.5 (its noise pair has as many long runs: rain streaks), orbit 84.4/84.7 and 84.1/86.7. By eye a clean 1-2 px line on the cut, quays, both boats and the stepping stones. Not met as a count on the photo views: the photo lens blurs the line into long runs of slightly brighter pixels (photo-dof0 1-2 px 17.8%); by eye a thin soft line (`photo-water.png`). Glade (pixel size 1) draws the soft white line about 3 render px wide, with the broken swash beside it at close zoom (beached-glade: runs of 3 grid units dominate); intended, a look call for the P3 sheet. On a sandy beach slope the rim widens to 2-4 px (`hd-cut.png`), where a shallow bed lies within 1-2 px for a longer stretch. 89: - **None over open sea:** rim pixels in the open-sea rects are within noise: harbour-day 0.87%/0.25% against 0.77/0.72, harbour-night 0.15/0.08 against 0.11/0.12, glade-day 0.07/0.06 against 0.22/0.22, orbits and creatures within their pairs except creat-day set 1 (0.39 against 0.18; set 2 0.12 against 0.23). No rim at the skyline ring's feet: the band under the horizon has as many brightened pixels as the noise pair, with no row structure (harbour-golden per-row counts, `skyline-band.txt`). 90: - **Shallows and swash inside the submerged face, no hard outer line:** by construction (both reach 0 by 0.2 m) and measured: the mean luma change 4-12 render px past the end of each rim run is within ±0.9 of 0 on every pixel-size-2 view (e.g. quarter-night -0.30..0.13, beached-day 0.00..0.55, harbour-night -0.27..0.35) except harbour-lane (up to +1.46, where its noise twin reads up to +2.79). The swash shows as +3 to +9 at 2-3 px past the rim on the closer cameras (quarter, lane, beached); at Harbour's zoom the whole 0.30 m face spans only about 3.3 render px, so the swash has faded before it would draw (none there). The pixel just past the rim reads 5-12 luma darker: at night the -0.15 shallows; by day it is the presentation outline inking the new line (a temporary debug output of `shallow_lift` and `shallow` on harbour-day showed the lift positive and the shallows present there; the debug capture was not kept). 91: - **Rim with the mirror off or missing:** `moff-beached` 95.3% and `moff-quarter` 91.2% of runs 1-2 px; the Harbour rim facing the sky (mirror misses) is drawn the same. 92: - **No crawl across the orbit:** the same 1-2 px line along the same cut in all three yaws (`orbit.png`); run shares 84-87% in each. 93: - **Boats:** the afloat hulls now take the rim (`t2-quay.png`, `t2-qn.png`). The beached boat sits deeper in the sand with the skirt closing the gap under its raised end; nothing hangs below the sand (`beached-boat.png`). Lights and streaks unchanged (test and identical `reflection_streaks` counts). Ducks keep their own ring (`ducks.png`). 94: - **Review fix (floating lanterns, 30 September 02:05-02:21):** the review found that the plan's "ducks and floating lanterns with `--creatures`" capture had not been taken, although the afloat float disc (`Architecture.water_stand`, reaching 0.11 m under the surface) now sits in the depth the rim reads. Taken now: `a/lantern/make_lantern.gd` authors `a/lantern/lantern.json` (the Harbour starter with five afloat lanterns on the water by the beach corner where the ducks swim, all supported, camera size 14 on that corner), and `a/lantern/sitting.sh` captures `lantern-day` and `lantern-night` with `--creatures` in one sitting 02:12-02:13 in the order `pre-1` (scratch tree), `after-1`, `pre-2`, `after-2`, all first try (power AC, 100%, Low Power Mode 0, `a/lantern/*/power.txt`). Integer counts and draw calls identical on both views and in the noise pair (`a/lantern/metrics/counts-*.txt`; creatures 26 by day, 30 at night on both trees). Whole frame >8, A/B set 1 / set 2 (`ab1`, `ab2`): day 2.74/3.32% against noise 2.18 (pre pair, `noise-pre`) / 2.30 (after pair, `noise-after`), night 2.81/3.60 against 2.58/3.05. Day sets 1 and 2 and night set 2 are above both noise pairs; night set 1 lies between them. The A/B mean dluma (+0.35/+0.37 day, +0.23/+0.21 night, against +0.03/+0.02 in the pre pair and +0.05/+0.01 in the after pair) is a brightening, as the intended rim on the shore, the hull and the pillars in these frames would explain. `rim.js` on the frame: 90.3/92.2% (day) and 89.1/91.6% (night) of rim runs 1-2 px (`rim-set1.txt`, `rim-set2.txt`). Crops looked at: `a/look/lantern-day.png`, `lantern-night.png`, `lantern-day-b.png`, `lantern-night-b.png` (pre-1|after-1, 3x), `lantern-disc.png`, `lantern-disc-b.png` (6x, day and night), `lantern-noise.png` (after-1|after-2 at night). What they show: the float discs draw **no visible rim** and no artefact (no halo, dark ring or line); five lanterns were placed and four discs are in frame: the boxes at (527,517), (675,600), (862,678) and (1203,565) hold the lanterns at (6.4,16.8), (9.4,16.9), (12.6,16.4) and (14.0,11.0); the fifth, (12.4,13.4), is at the moored boat, whose deck and hull hide its foot and disc (`look/lantern-hidden.png`), so it has no box. Brightened render px per box, first three boxes (`a/lantern/disc.js`, `metrics/disc.txt`, now labelled by set and lantern): set 1 day 5/11/5 against noise 6/13/5, night 5/10/8 against 7/13/12; set 2 day 5/22/3 against 18/15/0, night 9/18/28 against 14/27/14. The set-2 counts over noise: day (9.4,16.9) 22 against 15 is scattered water flecks above the disc, with no line round it (`look/lantern-disc2-set2.png`); night (12.6,16.4) 28 against 14 has 18 samples in one column, a 2 px pale vertical line over the water in after-2 only, which is a rain streak (`look/lantern-streak.png`); without that column the count is 10. The fourth box, (14.0,11.0) (37 and 31 in set 1, 41 and 40 in set 2, against noise 4-5), is the rim on the boathouse pillar beside it, not the disc (`lantern-disc-b.png`). The shore, the moored boat's hull and the boathouse pillars in the same frames take the rim. The ducks keep their own rings, unchanged, with no rim round them. Why the disc draws none was not investigated (outside this fix); the acceptance line does not list floats, so it is a look note for the P3 sheet. No code changed in this fix: full suite re-run `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/a/suite-fix` (a/b/v, 02:14-02:21, power `a/lantern/suite-power.txt`): **40 suites, 19,620 checks, 0 failures**, `--verify` 2880x1800 exit 0 in 31 s, `PIXELPUNK_INTEGRATION []`, the creature_reactions rerun 0 failures. **Recheck fix (numbers only, 30 September 02:22-02:32):** the recheck found that the whole-frame >8 figures above had set 2's A/B printed as the pre noise pair (swapped), that the fifth lantern and the set-2 disc counts were not reported, that the set-2 blocks in `disc.txt` carried the set-1 labels, and that a stray `a/lantern/metrics/.txt` held only art_metrics' usage text. Fixed: figures corrected from the files on disk; `disc.js` labels each block by set and lantern and `disc.txt` was regenerated from the same captures (old copy `a/review/disc-before-relabel.txt`); the stray file moved to `a/review/lantern-metrics-stray-usage.txt`. No code or capture changed. Suite re-run `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/a/suite-fix2` (`a/suite-fix2.sh`, 02:25-02:31, AC 100%, Low Power Mode 0, `suite-fix2/power.txt`): **40 suites, 19,620 checks, 0 failures**, `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`. 95: - **Bench:** measured in P3b's 3-way bench (plan). Count probe only: draw calls and every integer count identical (`a/metrics/counts.txt`). The water adds one depth read and about 20 ALU lines per fragment. 96: - **Journey:** expected count deltas zero on all six views; look change: a 1-2 px waterline rim on every Harbour and quarter view and styles-day. 97: - **Review:** one Opus 5.5 review in a fresh context and two fix passes ([run ID removed]). Round 1: one must-fix (the plan's floating-lantern captures were never taken; taken in the fix, `a/lantern/`) and two notes (the open-sea wording, and the Journal line, which the orchestrator edited on 30 September: the edition 3 deploy status). Recheck 1: one must-fix (two swapped whole-frame >8 figures in the lantern bullet, corrected). Recheck 2: nothing open. The orchestrator looked at `a/look/hd-cut.png`, `hn-cut.png` and `glade.png`: a thin lit rim along the cut, white on sand, lavender at night. Merged to `main` under the owner's overnight standing instruction. 98: - **Not done / open:** the rim at night is a strong lavender line (the report's colour and 0.75); a look question for the P3 sheet, not tuned here. Swash is absent at Harbour's zoom (the face is too thin in px). The afloat lantern's float disc draws no visible rim (review fix above); cause not investigated, a look note for the P3 sheet. The beach slope's wider rim and Glade's 3-px line are left as drawn. The night shallows may undo a little of P1d's quay lift; P3d re-measures it (plan risk). The `docs/visual-handoff.md` "Journal (blog)" line was already edited in the working tree by another session before this entry; left as found. Scratch tree `a/pre-tree` left for the orchestrator. Platforms: Mac run; Windows not run. 55: **Beautify P3b Rain rings on open water; CPU splashes retired (30 September 2026):** Opus 5.5 on this Mac. Rain now drops pixel rings on the open water: the puddle's `rain_ring()` moved into a new include, `shaders/rain_ring.gdshaderinc` (after `sky.gdshaderinc`; it keeps `RING_PERIOD` 0.9 and `RING_SHIVER` 0.004 and takes cell, radius and spawn share as arguments). The puddle calls it with its own 0.7 / 0.25 / 0.4 and its own `max(fwidth)` pixel, line for line as before. The water calls it on 0.8 m cells, gone at 0.3 m, spawn `0.45 * clamp(rain * 2.5, 0, 1)` (none at rain 0, full from 0.4), with the render px on the plane from the projection (`px_m / sin_pitch`; the water still has no screen derivatives), and not at all when `rain` is 0. The ring line is `max(color, mix(sky_ref, mirror, 0.5) * 1.6 + 0.04)` at 0.55 over open water (after the review fix, 1.0 where the mirror hits the town: `ring_top`, `RING_OPACITY_MIRROR`) over the finished water and under the P3a rim; its push is added to the mirror wobble. The invisible rain fleck (`0.004 + rain * 0.003`) is gone. The CPU splash emitter (`WorldView.splashes`: untextured 0.34 m planes, `rain * 520`) is retired, which also removes the splash ticks over the painted sky (report quick win 7, Defect 9). Branch `beautify/P3b` from `main` 1b86dd0. Snapshot `build/backup-feel/21-P3-b/`; at part start the live tree matched it in scripts, shaders, tests, data, scenes and tools (nothing unexpected). Scratch tree `artifacts/beautify/21-P3/b/pre-tree/`; evidence `artifacts/beautify/21-P3/b/`; backups `b/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture, bench round and the suite (`b/cap-power.txt`, `b/bench/power.txt`, `b/suite/suite_all.out`). 56: - **Files:** `shaders/rain_ring.gdshaderinc` (new, with its editor `.uid`), `shaders/puddle.gdshader`, `shaders/water.gdshader`, `scripts/world_view.gd`, `tests/test_wet.gd`, `tests/test_water.gd`, `docs/visual-handoff.md` (this entry). 57: - **Tests:** test-first `b/P3b-testfirst-test_wet.log(.out)` 171 checks, 14 failures; `-test_water` 171, 13 failures. After (`b/r1-*`, `b/r2-*`): test_wet **171, 0** (was 145), test_water **171, 0** (was 149). Two of my new test_water checks were corrected after the first run, not the code: the fleck check now ignores comments (it had matched this part's own comment), and the "no splash planes" check matches `PlaneMesh` exactly (Godot's `QuadMesh` extends `PlaneMesh`, so the rain and smoke quads had counted). New: `ring_include_checks` in test_wet (the include exists; both shaders include it after the sky and keep no private copy; signature; no texture read and no TIME, SCREEN_UV, VIEWPORT_SIZE or derivatives in the include; RING_PERIOD/RING_SHIVER there and RING_CELL/RADIUS/SPAWN not; the ring body's six key lines unchanged; the puddle's call and its draw and shiver unchanged; a GDScript port: no ring at spawn 0, rings at the water's rain-0.35 share, none beyond its radius), `rain_ring_checks` and `splash_checks` in test_water (constants, spawn at rain 0/0.1/0.35/0.4/1, the call with the projection pixel and the dry guard, found before the mirror, `+ ring.yz` in the wobble, the ring mix after the reflection and before the rim at 0.55, no fleck; no `splashes` in the source or on the node, no flat-plane CPU particles, rain still falls at 0.35 x 2600, water rain 0.35 then 0). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/b/suite` (a/b/v): **40 suites, 19,668 checks, 0 failures** (19,620 + 48). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 0 failures, and the known flaky check did not fire. 1440x900 not run (the inherited stall). 58: - **Changed expectations (old -> new):** review fix: test_water "the water calls the ring with its projection pixel" `rain_ring(wp.xz, TIME, px_m / sin_pitch, ...) : vec3(0.0)` -> `rain > 0.0 && ring_far > 0.0 ? rain_ring(wp.xz, TIME, ring_px, ...) * ring_far`, and the ring mix `max(color, mix(sky_ref, mirror, 0.5) * 1.6 + vec3(0.04)), ring.x * RING_OPACITY` -> `max(color, ring_top), ring.x * mix(RING_OPACITY, RING_OPACITY_MIRROR, ring_mirror)` (open water unchanged). Part: test_wet "vec3 rain_ring reads no texture" read the body from `puddle.gdshader` -> from `rain_ring.gdshaderinc` (the helper loop now checks `puddle_mask` alone in the puddle). `RING_PERIOD` and `RING_SHIVER` moved from the puddle to the include; the puddle's RING_CELL/RADIUS/SPAWN assertion (0.7/0.25/0.4) is unchanged. No check deleted. 59: - **Captures** (`b/caps.sh`, `tools/art_capture.sh`, one sitting 02:40-02:45 in the order `pre-1` (scratch tree), `after-1`, `pre-2`, `after-2`; 11 views each, all first try): harbour-night, -day, -golden (rain 0.35), `quarter-r35` (quarter-night at `--rain=0.35`), quarter-night (0.45: plaza puddles and the qn sky), quarter-dry, glade-day (both dry controls), harbour-lane, lane-night, photo-golden, styles-day (rain 0.1); plus `photo-dof0` in `b/cap/photo-p1`, `-a1`, `-p2`, `-a2` (02:54). A first photo-dof0 attempt ran all four on the live tree (a shell word-splitting slip) and is kept apart in `b/discarded/`, not used. Crops looked at: `b/look/t1-hn-water.png` (P3a after | P3b tune), `hd-water.png`, `styles-water.png`, `hn-reflections.png`, `hn-column-zoom.png`, `qr35-columns.png`, `lane-night.png`, `photo.png`, `photo-dof0.png`, `qn-sky-pre1.png`, `qn-sky-after1.png`, `sky-ticks.png`, `tick-zoom.png`. 60: - **Counts** (`b/metrics/counts-set1.txt`, `-set2.txt`, `counts-noise-pre.txt`): every rainy view with water loses exactly **one draw call and one rendered object** (harbour 897 -> 896, golden 822 -> 821, lane 738 -> 737, quarter 970 -> 969, photo 1280 -> 1279, styles 491 -> 490), in both sets; quarter-dry and glade-day are identical; the noise pair is identical everywhere. 61: - Whole frame, changed >8, A/B set 1 / set 2 against noise pre / after (`b/metrics/summary.txt`): harbour-night 4.37/3.84 (1.84/1.50), harbour-day 4.79/4.30 (1.31/0.90), harbour-golden 4.16/4.35 (0.97/1.96), styles-day 2.33/2.92 (0.37/0.85); quarter-night 4.32/3.88 (3.42/3.74), quarter-r35 2.88/3.80 (2.10/2.81); inside noise: harbour-lane 2.15/2.13 (1.95/1.92), lane-night 1.71/1.91 (1.75/1.68), photo-golden 2.36/2.34 (2.47/1.41), quarter-dry 0.65/0.48 (0.35/0.77), glade-day 0.29/0.86 (0.79/0.36). 62: - **Acceptance** (`b/rings.js`: on the render grid a ring px is brighter than the median of its 9x5 neighbourhood by more than 12 luma; groups 8-connected; `b/metrics/rings-T12.txt`, `rings-hist.txt`; `b/ticks.js`, `b/metrics/sky-ticks.txt`; region A/B against a region-level noise pair `b/metrics/regions.txt`): 63: - **2-3 px ring ellipses stipple open water at rain 0.35:** met after the review fix (below). As first booked it held on the Harbour only: on `quarter-r35`'s canal the rings were drawn but lost in the mirrored walls (review finding 2); the fix raises them there. Harbour-night open sea (20-420 x 700-950): ring-shaped groups (2-3 px tall, 3-9 wide) 3/3 before -> 45/46 after (sets 1/2), groups 31/28 -> 57/58 per 10k render px; by eye evenly scattered pale ellipses about 5-8 px wide and 2-4 tall, some fresh drops as a 1-2 px dot (`t1-hn-water.png`, `hd-water.png`). By day and at golden hour the rings are clear (ring px 7.4 -> 11.0% of harbour-day's open sea, 7.2 -> 10.6% golden); styles-day at rain 0.1 has fewer (6.95 -> 8.7%). At lane zoom they are faint (lane-night within noise, `lane-night.png`); in the photo lens soft ellipses (`photo-dof0.png`). 64: - **Catch the reflection columns:** partly. Rings take the water's own mirror (the town image) and shiver it: in the Harbour-night mirrored far bay (1000-1500 x 235-420) bright groups 38/36 -> 55/57 per 10k (ring-shaped groups only 8/9 -> 21/20; the second reviewer's stricter hollow count 0.4/0.4 -> 1.7/1.7), and rings show faintly inside the pillar reflections (`hn-column-zoom.png`). The orange and neon columns are the separate additive streak pass (`reflections.gdshader`), drawn over the water; the rings do not change them. In the quarter's `mirror_columns` the ring count is flat (ring px 22.6/22.5 -> 22.9/23.0, groups within the pair) and no ring is visible among the streaks (`qr35-columns.png`). Making the streaks take the rings would touch the streak shader (P4c's neon bay); not done here, a look question for the P3 sheet. 65: - **None at rain 0:** quarter-dry whole frame 0.65/0.48 against noise 0.35/0.77, ring count flat (groups 287/288 -> 291/280), draw calls identical; glade-day (dry) within noise, identical counts. 66: - **Puddles unchanged within noise:** quarter-night `plaza` >8 1.53/1.20 against noise 1.53/1.21, `plaza_px` 1.84/1.09 against 1.63/1.30; quarter-r35 plaza 1.05/1.45 against 1.49/0.83; `paving` identical to its pair (1.11/0.82 both); quarter-dry plaza 0.10/0.07 against 0.06/0.10. 67: - **No splash planes or sky ticks:** the emitter is gone (tests; one draw call fewer). Sky ticks (a run of 2+ render px brighter than the px above and below by more than 6 luma) over the quarter's sky (1200-1600 x 0-180) and the Harbour's (0-1600 x 0-150): before, 5 (set 1) and 6 (set 2) ticks that come and go (three of them, the second reviewer found, are cloud edges rather than splashes, and more splash ticks lie outside the box scanned; the conclusion of none after holds), e.g. (1312,40) in quarter-night pre-1 (`tick-zoom.png`), (1390,8) in quarter-r35 pre-2 and (944,130) in harbour-night pre-1 (`sky-ticks.png`); after, none in either set. Two ticks present in every capture of both trees, (1540,172) on the quarter (also in quarter-dry) and (768,136) in harbour-golden, are fixed features of the skyline, not splashes. 68: - **Bench (3-way, `b/bench/`, `bench.sh`, `analyse.js`, `summary.txt`):** trees `preP3` (21-P3, before P3a), `postP3a` (21-P3-b) and `now` (a copy of the live tree), `--disable-vsync`, 1920x1080, `--rain=0.45`, Harbour and dense at 01:00 and 13:00, round 0 discarded, 5 interleaved rounds (tree order rotated and reversed), 72 runs all first try, vsync disabled in every run. AC 100%, Low Power Mode 0; load average 2.3-4.8 (uptime 3.29 at part start; the Mac was busy, round spreads 0.02-0.09 ms except dense round 1). Raw p50 medians preP3 / postP3a / now: Harbour 13:00 6.61 / 6.64 / 6.61, Harbour night 6.65 / 6.65 / 6.66, dense 13:00 13.14 / 13.14 / 13.14, dense night 13.18 / 13.22 / 13.18 ms. Paired steps: **P3a** +0.02 (5/5), -0.00, +0.00, +0.06 (5/5) ms; **P3b** -0.02 (0/5), +0.02 (4/5), -0.02, -0.06 ms; P3a + P3b together +0.00, +0.01, -0.01, +0.00 ms. Both parts are within about 0.06 ms of nothing on every scene. Draw calls 907 -> 906 (Harbour), 2361 -> 2360 (dense). 69: - **Journey:** expected count deltas: `draw_calls` and `rendered_objects` -1 on every rainy journey view with water (harbour-night, harbour-day, harbour-lane, quarter-night, quarter-day, styles-day), 0 elsewhere. Look change: rings on the open water in the Harbour and styles views; the quarter views within noise outside the water. After the review fix below: rings also show on the quarter's canal (quarter-night, and quarter-day when it rains), and by day the rings over the mirrored town near the Harbour island are paler and brighter than over the open sea; the counts are unchanged by the fix. 70: - **Review fix (30 September 03:34-04:03; evidence `b/fix1/`):** the review had two must-fix findings. (1) Once the render px on the plane (`px_m / sin_pitch`) grew past the ring radius, the one-px line lit the whole drop disc, clipped to its 0.8 m cell. Grazing far water in a low perspective view, and the zoomed-out orthographic camera (size 95, pitch 0.25: 0.77 m px), filled with bright cell-sized dashes and brightened as a band. None of the part's captures covered those views. (2) The ring line was booked "met", but on `quarter-r35`'s canal the rings did not read. Measured on ring px from a debug capture of `ring.x` (`b/fix1/predict.js`), they lifted the canal by 10-15 luma, about the Harbour's lift, but the canal's mirrored walls vary by 33 luma (p10-p90) against the Harbour's 5. A floor lift over the water colour (`color + 0.03..0.18`) changed nothing, because the old target already sat above it (`b/fix1/exp/r1-*`, `r2-*`). Fixed in `water.gdshader` only (the include and the puddle are untouched): `ring_px = px_m / sin_pitch`; `ring_far = 1 - smoothstep(0.55 R, 1.1 R, ring_px)` (0.165 to 0.33 m; Harbour's pixel is 0.143 m, the quarter's 0.136, the lane's 0.078) scales line and push, and the ring is not computed at all once faded out. Where the town is in the water (`ring_mirror = max(step(0.75, mirror_hit), soft_w)`), the ring catches at least the sky at 1.6 x + 0.04 at opacity 1.0 (`RING_OPACITY_MIRROR`); over open water (a miss) the target and 0.55 are as before. Tests first: test_water `b/fix1-testfirst-test_water.log(.out)` 191 checks, 16 failures; after (`b/fix1-r1-*`) test_water **191, 0** (was 171), test_wet **171, 0**. New `ring_far_checks`: the constants and the line; the Harbour, quarter, lane and Glade pixels keep full rings; size 95 and a 2.4 m pixel draw none; 0.25 m fades; a GDScript port shows that the line lights every point of its disc at a 0.77 m pixel but not at 0.143 m. Also the `ring_mirror` and `ring_top` lines, opacity 1.0, and the opacity for five hit/soft cases. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/b/fix1/suite` (a/b/v, 03:55-04:03): **40 suites, 19,688 checks, 0 failures** (19,668 + 20). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 0 failures; 1440x900 not run (inherited stall). Captures (`b/fix1/caps.sh`, `sitting.sh`), one sitting 03:48-03:55 in the order `pre-1` (the pre-part scratch tree), `prefix-1` (`b/fix1/exp-tree`, a copy of the live tree holding the part's pre-fix water shader), `after-1`, `pre-2`, `after-2`. There are 15 views each, all first try: the part's 11 plus the reviewer's `persp-low` and `persp-low-day` (Harbour, perspective, camera 0.78,0.25,38, rain 0.35) and `zoomout-day`/`-night` (orthographic size 95, pitch 0.25, rain 0.35). Power AC, 100%, Low Power Mode 0; load 2.7-3.5 (`b/fix1/cap-power.txt`, `suite-power.txt`). Metrics are in `b/fix1/metrics/`: 71: - **Far band** (the review's `band.js` rows, x 0-400; `bands.txt`), pre / pre-fix / after: persp-low-day y235-300 221.3 / 225.9 / 221.2, y300-400 200.2 / 206.5 / 200.2; night 59.6 / 61.8 / 59.4-59.6 and 57.1 / 59.7 / 57.0-57.2; near water y850-1000 keeps its rings (day 132.1 / 134.3 / 134.3). zoomout-day y300-400 204.6 / 211.4 / 204.6 and y400-500 169.7 / 179.8 / 169.7; zoomout-night 56.3 / 59.6 / 56.3. So the band is back to the pre-tree to 0.2 luma. By eye the dashes are gone and the ripple is as before (`b/fix1/look/far-band-day.png`, `far-band-night.png`, `zoomout-day.png`; pre | pre-fix | after). 72: - **Quarter canal**: rain-minus-dry lift histogram (the review's method; `lift-hist.txt`) over [0,880,400,1000], share of render px lifted >16 luma: pre 2.81/3.54%, pre-fix 3.65%, after 6.16/5.38%; >32 luma: 1.68/2.20 -> 1.69 -> 3.54/3.15%. Front canal [560,880,1000,1000] >16: 3.06/3.69 -> 4.47 -> 7.07/7.26%. The review's hollow-ring detector (`rings2-T10.txt`) moves only a little on the busy canal: [0,900,250,1000] 0.0/3.2 -> 4.8 -> 9.6/6.4 per 10k, front 10.6/12.9 -> 9.8 -> 15.2/15.2. By eye, pale lavender-blue ellipses 6-10 px wide now read over the mirrored walls, and none did before (`b/fix1/look/q-canal.png`, `q-front.png`). 73: - **Harbour open sea unchanged by the fix** (a miss): hollow groups 8.0 pre-fix -> 7.2/7.2 (harbour-night), 13.6 -> 14.8/14.0 (day), 17.2 -> 16.0/14.4 (golden), styles 7.6 -> 8.0/7.2; by eye identical (`b/fix1/look/hn.png`). 74: - **Where the town is mirrored, the rings are stronger**: whole frame >8, pre-fix -> after: harbour-day 2.66% and harbour-golden 2.13%, against noise 0.61-0.78 and 1.07-1.26. There the rings over the town's reflection near the island are paler and brighter, and fresh drops read as white 2-3 px dashes (`b/fix1/look/hd-mirror.png`). The photo lens shows soft pale ellipses (`photo.png`). The other views are within or near their pairs (quarter-r35 2.55 against 1.77/2.41, quarter-night 3.83 against 2.70/3.24, lane-night 1.90 against 1.89/2.01). 75: - **Counts**: integer counts and draw calls identical pre-fix -> after on all 16 views (`counts-prefix-after1.txt`), and in the after noise pair. Puddles: the fix does not touch `puddle.gdshader` or the include. 76: - **Bench**: not re-run for the fix. It adds a smoothstep, a step and a mix per fragment, and skips the ring where it has faded; the 3-way bench above is the part's. 77: - **Review:** one Opus 5.5 review and a second, independent review (measuring part; [run ID removed]), and one fix pass. Round 1 found a must-fix (rings on grazing and zoomed-out water turned into cell-sized bright blocks) and the reflection-column clause only partly met; the second review found a must-fix (no visible rings on the quarter canal, where the plan measures this line). Both were fixed (`ring_far`, `ring_top`) and rechecked: nothing open. Not fixed, recorded: the ring gate `rain > 0.0` in the water shader against `rain > 0.01` for the falling rain, so at 0 < rain <= 0.01 rings stipple with no rain falling (cosmetic; one threshold to align later); the loosened no-flat-plane test matches only the exact PlaneMesh class; the Tests line's `r2-*` for test_wet does not exist (test_wet 171/0 is from `r1-*` and the suite). The bench was not re-run after the fix. The rings inside the neon and orange reflection columns are P4c's (a look question for the P3 sheet). Merged under the owner's overnight standing instruction. 78: - **Not done / open:** review fix: by day the rings over the mirrored town near the Harbour island are paler and brighter, and fresh drops there read as white dashes (`b/fix1/look/hd-mirror.png`), a look question for the P3 sheet. The experiment scratch tree `b/fix1/exp-tree` is left for the orchestrator. the reflection-column part of the acceptance is only met for the water's own mirror, not the additive streaks (above); for the P3 sheet. Fresh drops draw as a bright 1-2 px dot before the ring opens; left as drawn. Rings at lane zoom are faint. The pre-stage scratch tree built for the bench is `b/bench/trees/preP3/`; scratch trees left for the orchestrator. Platforms: Mac run; Windows not run. 34: **Beautify P3c Cake-slice banks I: the cut in its ground material, lip and strata (30 September 2026):** Opus 5.5 on this Mac. An island's cut (every meadow, sand and along-the-water earth-path side that `water_edge` calls "cut") is no longer one flat `BANK` quad in `plain`. It is drawn in its cell's own ground material (`GROUND_MATERIALS[finish-1]`: grass, sand or earth, already in the chunk) and ground colour, with UV.x the bank mark `Architecture.BANK_UV_MARK` (-8192) and UV.y the metres below the bank top. `ground.gdshader` branches on the mark (`bank_albedo`): top to bottom, a lip in the ground's own colour (meadow 0.10-0.18 m, ragged by 0.12 m column, a 0.10 m drip on about one column in eight; earth 0.09 m; sand 0.12 m), a 0.09 m shadow line at soil x0.55, and soil in 0.22 m strata whose boundaries wander by 0.06 m, each stratum toned ±10% and alternately warmer and cooler; per-finish soils `BANK_SOIL_MEADOW` (= `Architecture.BANK`, now 8b796a), `_EARTH`, `_SAND`. The face coordinate is the masonry grid's `masonry_face_coord`, the footprint `length(fwidth(vec2(face.x, wp.y)))` (taken outside the branch). Glade's `CALM_LIFT` lifts the lip only, not the soil. At night the soil leans half-way to a cool grey of its own value and lifts by `BANK_NIGHT_LIFT` 0.36 (`smoothstep(0.66, 1, night_amount)`: none by day or at golden hour, about half at 05:45/20:45); `night_amount` is a new ground uniform (default 0), set from the hour's windows in `sync_profile` for grass, earth and sand. The lip flaps (P0c, D3) keep their geometry but take the lawn's own colour (`LIP_GRASS` 5f8a55 -> 70846a; owner P0-2: no more lime squares in Glade). Report Signature A land side (material, depth UVs, lip with drips, shadow line, strata) and the Defect 2 follow-on; owner G1 Q3 (night bank at least water + 8). Branch `beautify/P3c` from `main` 68db4d9. Snapshot `build/backup-feel/21-P3-c/`; at part start the live tree matched it in scripts, shaders, tests, data, scenes and tools (empty `c/diff-live-vs-21-P3-c.txt`; nothing unexpected). Scratch tree `artifacts/beautify/21-P3/c/pre-tree/`; evidence `artifacts/beautify/21-P3/c/`; backups `c/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set (13 of 13 in `c/cap-power.txt`; its header lines name the first view and the tree, pre or live, not the set name, a slip in `caps.sh`) and the suite (`c/suite/suite_all.out`). 35: - **Files:** `scripts/architecture.gd` (`BANK_UV_MARK`, `BANK` value, `LIP_GRASS`, the cut in `terrain_chunk`), `shaders/ground.gdshader` (bank branch, masonry grid include, `night_amount`), `scripts/world_view.gd` (`night_amount` on grass, earth, sand), `tests/test_chemistry.gd`, `tests/test_nature.gd`, `tests/test_look.gd`, `docs/visual-handoff.md` (this entry). 36: - **Tests:** test-first (`c/P3c-testfirst-*.log.out`): test_chemistry 394 checks, 34 failures; test_nature 282, 3; test_look 542, 18; all new or changed checks. After (`c/r3-*`): test_chemistry **394, 0** (was 351), test_nature **282, 0**, test_look **542, 0** (was 524). New `bank_face_checks` (test_chemistry): the mark constant; a meadow and a sand island mark exactly their cut sides (six vertices each) in grass/sand, in the cell's ground colour, depth 0 on the top edge and top-to-bed at the foot, nothing of the cut in plain; an earth path along the water marks its cut in earth; earth and cobbled step treads and earth and road shores carry no mark (the shore faces and tread boxes keep their UVs); Harbour and `chemistry.json` mark exactly their 141 and 293 cut sides; the shader's include, mark, depth, face coordinate, face footprint, `night_amount`, the twelve named constants, `BANK_SOIL_MEADOW == Architecture.BANK`, per-finish lip depths; the A2 terrain reveal (grass, earth, sand have no build variant; an island raised with the build animation on is built exactly as with it off, in settled materials, the cut surfaces marked). `bank_wired` (test_look): grass, earth and sand take the hour's `night_amount` at every editor look state. `shore_kind` checks (`test_chemistry.gd:348-349`) and `ground_calm` wiring unchanged. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/c/suite` (a/b/v, 04:33-04:39): **40 suites, 19,749 checks, 0 failures** (19,688 + 61). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 37: - **Changed expectations (old -> new):** test_nature `bank_checks`: "the cut side's colour is BANK 43382f" -> removed (BANK is now the meadow soil 8b796a, held to the shader by test_chemistry); "the lip grass is 5f8a55" -> "the lip grass is the lawn's own colour 70846a"; "the meadow island's cut draws in BANK (plain), never 605650" -> "draws in grass with the bank mark, none in plain"; "BANK is only on the upright cut faces" -> "the marked grass is only on upright faces, in the lawn colour"; "a sand island's cut is BANK too" -> "is in sand with the bank mark, not BANK in plain"; `lip_of` now tells a flap from the lawn and the cut by lying off the flat island's ground and facing up (the flaps share the lawn colour). `OLD_CHUNKS` re-recorded with `c/probe/old_chunks.gd` (the P0c probe; it reproduced the committed record first, `c/probe/old_chunks-pre.txt`; new `old_chunks-post.txt`): 32 of 46 chunks change hash, every chunk keeps its vertex count. No check deleted. 38: - **Captures** (`c/caps.sh`, `tools/art_capture.sh`, one sitting 04:25-04:31 in the order `pre-1` (scratch tree), `after-1`, `pre-2`, `after-2`; 14 views each, all first try): harbour-day, -night, -golden, harbour-lane, lane-night, glade-day, styles-day, grove-day, chemistry-day (earth steps, sand shore), quarter-night (journey counts), and a 3-frame orbit at night and by day (`orbit-a`/`-c` yaw 0.70/0.86 with harbour-night/-day as the middle frame). Tuning rounds `c/cap/t1`-`t5` and debug `dbg1`-`dbg4` are not used for numbers. Crops looked at: `c/look/harbour-day-bank-x4.png`, `harbour-night-bank-x4.png`, `harbour-golden-bank-x4.png`, `glade-day-bank-x4.png`, `lane-bank-x3.png`, `lane-night-bank-x3.png`, `orbit-day-x3.png`, `chem-ab.png`, `chem-canalhead.png`, `chem-shore-diff.png`, `chem-diffmap.png`, `quarter-diffmap.png`; sheet `c/sheet/p3c-banks.png`. What they show: by day a warm brown bank in three or four visible strata under a ragged grass edge with a one-art-px dark shadow line (the straight ruler edge is gone); at lane zoom the lip reads as grass hanging over the edge; at night a lifted violet-grey bank with faint strata, clear of the water; at golden a sunlit orange-brown; in Glade the lime flap squares are gone and the lip is the lawn's own green. The orbit shows the same bands and lip in all three yaws (no crawl seen). 39: - **Counts** (`c/metrics/counts-set1.txt`, `-set2.txt`, identical; noise pair `counts-noise-pre.txt` identical): draw calls and rendered objects fall where a chunk's `plain` surface held only the cut: harbour-day/-night 896 -> 889, harbour-golden 821 -> 814, harbour-lane and lane-night 737 -> 733, orbit-a 920 -> 913, orbit-c 851 -> 844, styles-day 490 -> 469, grove-day 305 -> 299, chemistry-day 593 -> 584; glade-day and quarter-night identical. 40: - Whole frame >8, A/B set 1 / set 2 against noise pre / after (`c/metrics/ab1.txt`, `ab2.txt`, `noise-pre.txt`, `noise-after.txt`): harbour-day 2.65/3.36 (1.00/0.96), -night 3.19/4.77 (2.88/1.18), -golden 2.92/3.35 (1.36/0.99), harbour-lane 5.37/5.21 (1.95/1.92), lane-night 6.60/6.63 (2.09/2.15), glade-day 2.12/2.12 (0.32/0.33), styles-day 1.81/2.02 (0.22/0.17), grove-day 1.33/1.30 (0.01/0.04), chemistry-day 5.27/4.37 (3.03/1.07). The change is on the cut faces only (`chem-diffmap.png`: red only on banks). quarter-night 4.36/1.88 against 2.64/3.12: its diff map (`quarter-diffmap.png`) is rain and ring phase only, no bank structure, and its counts are identical. 41: - **Acceptance** (`c/bank.js` on the kit's `bank`, `lawn`, `water_open`, `water_below` p50; `c/metrics/bank-*.txt`; before = pre-1/pre-2, after = after-1/after-2, identical to 0.1 in both sets): 42: - **Lip, shadow line and strata instead of a slab:** met by eye (crops above). 43: - **Bank at most 0.8x lawn by day:** met, 0.884/0.891 -> **0.738** (bank 88.3 -> 73.7, lawn 99.9). 44: - **At least water + 8 at night (owner G1 Q3):** met, open water +3.6 -> **+12.4**, water below +4.5 -> **+10.2** (bank 46.2 -> 55.0; open 42.6; below 41.7 -> 44.8, it carries the bank's reflection). 45: - **Bank at most 0.8x lawn at night: not met, 0.894 -> 1.064. It cannot be met together with the line above** on this tree: 0.8 x the night lawn (51.7) is 41.4, and water + 8 is 50.6 on open water and about 52 below the bank (the water below rises about 0.3 per luma of bank). Owner decisions override the plan, so the owner's +8 was met and the plan's 0.8x at night is left failing. In tuning round `c/cap/t2` (no night lift, meadow soil about 8% lighter than the final) the night bank was 42.3 (0.82x lawn, open water -0.3), so even the unlifted soil sits near 0.8x. One constant; a question for the P3 sheet. 46: - Golden (plan P3d line, at most 1.5x lawn): 1.480 -> **1.241**. Glade (at most 1.0x the lifted lawn): 0.664 -> **0.560**. Styles `banks_front` p50 88.1 -> 88.0 (mean 90.5 -> 87.2). 47: - **Treads and shores unchanged:** by test (no mark, same UVs, vertex counts per chunk unchanged) and on `chemistry-day` (`c/metrics/chem-regions.txt`): the sand shore rect [836,430,872,470] changes on 1.39/1.11% of px only where it holds the new canal cut beside the shore (`chem-shore-diff.png`: the shore itself unchanged); the hill with the earth path [1000,150,1450,420] 1.18/1.00% against noise 1.09/0.79 (rain). 48: - **The A2 terrain reveal still matches:** by test (above). 49: - **No isolated black pixels:** `bank` and `waterline` `art.black` 0 in all four views, before and after; window px under luma 10 in the P0c lip strip unchanged (9/7/31/10, `c/metrics/lipspecks.txt`). The P0c lip-speck count (px under 25.5 and 15 darker than both horizontal neighbours) rises, 2/1/2/3 -> 4/6/10/38 (day/night/golden/Glade): these are the one-px shadow line's diagonal steps, not isolated pixels (`harbour-golden-bank-x4.png`); a look question for the P3 sheet. 50: - **Bench:** none in the part (plan: the stage bench covers the branch). The ground shader now takes the face coordinate and one `fwidth` for every ground fragment, and the bank branch only on cut faces. 51: - **Journey:** expected count deltas: draw_calls and rendered_objects -7 on harbour-night and harbour-day, -4 on harbour-lane, -21 on styles-day, 0 on quarter-night and quarter-day. Look change: the bank on every Harbour and styles view. 52: - **Review:** one Opus 5.5 review and a second, independent review (measuring part; [run ID removed]): no must-fix from either, every number reproduced. One stop condition, ruled by the orchestrator (`owner-decisions.md`, "P3c ruling"): the plan's night 0.8x-of-lawn line is waived in favour of the owner's G1 Q3 water + 8 line (they cannot both hold), also for P3d. Notes fixed by the orchestrator: the misplaced `BANK_NIGHT_LIFT` comment and the stale BANK and LIP_GRASS colours in comments (`ground.gdshader`, `architecture.gd`, `test_nature.gd`); test_look 542, test_nature 282, test_chemistry 394, all 0 failures after (`c/orch-*`). Recorded, not fixed: no twilight (05:45, 20:45) or Glade-night view was captured though the bank now lifts at twilight and in Glade's night palette (for the P3 close or P3d); the face-coordinate and extra `fwidth` run on every ground fragment, so the stage bench should look at ground-heavy views; the noise pairs are nearly frame-identical, so they understate cross-tree noise (the per-region conclusions were rechecked). Look questions are P3-1 to P3-3 in `owner-decisions.md`. Merged under the owner's overnight standing instruction. 53: - **Not done / open:** the night 0.8x line (above; owner or orchestrator ruling). The day soil reads warm orange-brown at lane zoom and at golden hour, and the strata are regular bands; look questions for the P3 sheet. The rest of the land side (roots, rock band, algae line, the sand and earth variants' detail, Glade calm of drips) is P3d. The experiment files `c/backup/ground.gdshader.t1` and the scratch tree `c/pre-tree` are left for the orchestrator. Platforms: Mac run; Windows not run. 5: **Beautify P3d Cake-slice banks II: roots, rock band, algae line, sand and earth, Glade calm (30 September 2026):** Opus 5.5 on this Mac. Under P3c's lip and shadow line, the cut (`ground.gdshader` `bank_albedo`) now carries, top to bottom: pale roots, one art px wide (`0.5*footprint`), hanging 0.10-0.42 m through the top strata from lanes 0.14 m apart (`BANK_ROOT_*`; 35% of lanes on meadow, half that on earth, none in sand; at Harbour zoom half of them are dropped); a rock band of stones in running rows (`BANK_STONE` 0.28 x 0.16 m cells, lit top, shaded underside, soil between them x0.85) from the water up to a wandering top `BANK_ROCK` 0.30 m over the tide top (70% of cells, sand 22% pebbles; softer at Harbour zoom); and the quay's algae waterline, the P1e tide band and pale tide line taken from `weathering.gdshaderinc` (now included) with no copies: `weather_tide(WEATHER_WATERLINE, vec3(0.0, y, 0.0), n, weather_wave(face.x), px)` and `weather_tide_albedo`, glossy by `WEATHER_TIDE_ROUGHNESS`/`_SPECULAR`; after the review and its recheck, the bank floors the band's darkening (`BANK_TIDE_FLOOR` 0.95: in the tide band the albedo keeps at least 0.95 of its soil's luminance, keeping the algae's hue, and in the rock band over it all of it, `mix(rock_band, floor_share, tide.x)`; Glade raises the band's floor toward 1 by `BANK_CALM_ALGAE*ground_calm`), see "Review fix" and "Recheck fix" below. The wave runs along the masonry face coordinate, so a bank running into a quay keeps its tide top. The line's px is `face_fw.y`, from the same `fwidth(vec2(face.x, wp.y))` P3c already took (no new derivative). Roots and stones fade with the strata's detail. The night cool and lift now apply to everything under the lip (roots, stones, algae), so the soil is exactly as before. Glade calm (`ground_calm`, 0.85): drips and roots by `quiet = 1.0-ground_calm`, stones by `BANK_CALM_STONES` 0.6 of it, the soil softened toward a less saturated brown (`BANK_CALM_SOFT` 0.8; owner question P3-1), and the algae band's darkening cut by `BANK_CALM_ALGAE` 0.6 (at full strength it was a black stripe at Glade 20:45: waterline `art.black` 0 -> 528, sitting 1); after the recheck, P3c's shadow line under the lip gives up its darkening by `BANK_CALM_SHADOW` 1.0 of the calm share (`soil*mix(BANK_SHADOW_TONE, 1.0, BANK_CALM_SHADOW*ground_calm)`: 0.93 of the soil in Glade, P3c's 0.55 on canal). `CALM_LIFT` still lifts the lip only. Report Signature A land side (roots, rock, algae waterline, variants, Glade calm), QW9 constants shared. Branch `beautify/P3d` from `main` c89ff0f. Snapshot `build/backup-feel/21-P3-d/`; at part start the live tree matched it in scripts, shaders, tests, data, scenes and tools (empty `d/diff-live-vs-21-P3-d.txt`; nothing unexpected). Scratch trees `artifacts/beautify/21-P3/d/pre-tree/` (21-P3-d) and `d/prestage-tree/` (21-P3, before P3a); evidence `artifacts/beautify/21-P3/d/`; backups `d/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, bench round and the suite (`d/cap-power.txt`, `d/bench/power.txt`, `d/suite/suite_all.out`). 6: - **Files:** `shaders/ground.gdshader`, `tests/test_weathering.gd`, `tests/test_look.gd`, `tests/test_chemistry.gd`, `docs/visual-handoff.md` (this entry). `weathering.gdshaderinc` and `weathering.gd` needed no change (the bank reads the existing contract). Scene `d/banks/banks.json` (`d/banks/make_banks.gd`): one water front in four finishes, a meadow cut, a stone quay, an earth path along the water (set directly: earth painted onto water becomes stepping stones) and a sand island's cut. 7: - **Tests:** test-first (`d/P3d-testfirst-*.log.out`): test_weathering 469 checks, 6 failures; test_look 549, 5; test_chemistry 412, 19. After the review fix (`d/fix-r2-*`): test_weathering **502, 0** (was 453; 495 before the fix; the fix's test-first run `d/fix-testfirst-test_weathering.log.out`, 499 checks, 3 failures), test_look 550, 0, test_chemistry 413, 0, test_nature 282, 0; full suite after the fix `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/d/suite-fix` (a/b/v, 06:27:57-06:33:47, AC, Low Power Mode 0): **40 suites, 19,825 checks, 0 failures**, `--verify` 2880x1800 exit 0 in 31 s, `PIXELPUNK_INTEGRATION []`, the creature_reactions rerun 0 failures, the known flaky check did not fire; a comment-only edit to `ground.gdshader` after it, rechecked by `d/fix-r3-*` (502, 550, 413, 0 failures). Before the fix (`d/r2-*`): test_weathering **495, 0** (was 453), test_look **550, 0** (was 542), test_chemistry **413, 0** (was 394), test_nature 282, 0. New: `bank_tide_checks` (test_weathering: the include, the tide call and albedo, gloss, the rock band on `weather_tide_top`, the px, no copy of the tide's numbers, a bank's band and line equal a quay's at every height and wave), `bank_calm_checks` (test_look: quiet share, drips, roots, stones, soil, algae calmed; only the lip takes CALM_LIFT's share), P3d constants and per-finish shares, one-art-px roots, detail fade, stone size in test_chemistry. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/d/suite` (a/b/v, 05:43-05:49): **40 suites, 19,818 checks, 0 failures** (19,749 + 69). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 8: - **Changed expectations (old -> new):** test_chemistry P3c "the bank's footprint comes from the face coordinate": `length(fwidth(vec2(face.x, wp.y)))` -> `vec2 face_fw = fwidth(vec2(face.x, wp.y));` and `float face_footprint = length(face_fw);` (same derivative, now also the tide line's px). My own new test_weathering check "no copy of `= 0.30;`" was refined after tuning: every `const float` line of ground.gdshader is checked for the tide's 0.15/0.30 except `BANK_ROCK` (the rock band's own height, 0.30 by tuning). No check deleted. Review fix: seven new checks in `bank_tide_checks` (the floor constant in [0.75, 1), applied after the algae tint against the soil's luminance, Glade raising it, the rock band floored too, the quay's algae on the packed soil falls under it (0.512 of the soil) and the floored band keeps its luminance and green); my own new "floor applied" check was refined after the 0.80 round from `max(1.0, tide.x*BANK_TIDE_FLOOR*...)` (band only) to `max(1.0, max(tide.x, rock_band)*floor_share*...)` (band and rock band; Glade's glade-2045 foot stayed black on the band-only floor). Recheck fix: that check again, `max(1.0, max(tide.x, rock_band)*floor_share*...)` -> `max(1.0, mix(rock_band, floor_share, tide.x)*...)` (the rock band over the tide band at the whole soil); new: `BANK_TIDE_FLOOR` >= 0.95, the rock band's packed soil keeps its soil's luminance (test_weathering), Glade calms the shadow line and keeps it at >= 0.9 of the soil at `LookProfile.RENDER.glade.ground_calm` (test_look). No check deleted. 9: - **Captures** (`d/caps.sh`, `tools/art_capture.sh`; final sitting 05:22-05:31 in the order `pre-1` (scratch tree), `after-1`, `pre-2`, `after-2`, 24 views each, then `prestage-1`/`-2` (pre-P3 tree, quarter-night, -0545, -2045); all first try): harbour-day, -night, -golden, **harbour-0545, harbour-2045, glade-night, glade-2045** (the twilight and Glade-night views P3c did not take), harbour-lane, lane-night, glade-day, styles-day, grove-day, chemistry-day, quarter-night, quarter-0545, quarter-2045, banks-day/-night/-golden/-glade (the four presets), and a 3-frame orbit at night and by day (`orbit-a`/`-c`, yaw 0.70/0.86, with harbour-night/-day in the middle). Sitting 1 (05:07-05:17, before `BANK_CALM_ALGAE`) is kept in `d/cap/sitting1-superseded/`; tuning rounds `d/cap/t1`-`t5` not used for numbers. Crops looked at (pre | after): `d/look/p3d-harbour-day.png`, `-night`, `-golden`, `-0545`, `-2045`, `p3d-glade-day.png`, `-night`, `-2045`, `p3d-harbour-lane.png`, `p3d-lane-night.png`, `p3d-banks-day.png`, `-night`, `-golden`, `-glade`, `banks-sand-earth.png`, `banks-sand.png`, `banks-junction-day.png`, `orbit-day.png`, `t5-glade-2045.png`; diff maps `diffmap-styles-day.png`, `-grove-day`, `-harbour-day`, `-quarter-night.png`. What they show: at lane zoom pale roots under a ragged lip, a row or two of grey stones, the lower stones and soil green-dark with algae over the rim; at Harbour zoom a fine stipple of roots and stones and a dark foot line (as submitted this line was black; after the review fix it is a green-brown foot, see "Review fix"); at night the stones read as pale blue-grey chips (strong at lane-night); Glade: no drips or roots, fewer smooth stones, a calmer brown (#6B3D15 -> #5C4021, sat 80% -> 64%). Earth: its own red-brown soil, two roots, stones; sand: tan with a few pebbles and the algae foot. The orbit shows the same bands in all three yaws (no crawl seen). styles-day and grove-day change only on the bank feet; quarter-night is rain and ring phase only. 10: - **Counts** (`d/metrics/counts-set1.txt`, `-set2.txt`, `counts-noise-pre.txt`): integer counts and draw calls identical on all 24 views. 11: - Whole frame >8, set 1 / set 2 against noise pre / after (`d/metrics/summary-frame.txt`): harbour-day 3.14/2.77 (2.60/1.81), -night 2.51/2.42 (2.77/1.75), -golden 4.34/2.53 (5.13/1.13), harbour-lane 3.65/3.64 (2.31/2.05), lane-night 3.72/4.09 (2.24/1.98), glade-day 1.98/1.72 (0.70/0.36), glade-night 1.66/1.27 (0.58/0.36), styles-day 2.32/1.07 (1.12/0.10), grove-day 0.50/0.44 (0.06/0.01), quarter-night 2.54/2.11 (3.25/1.97; unchanged), banks-glade 5.71/5.36 (0.25/0.15). 12: - **Acceptance** (`d/accept.js` with `d/regions.json`, the kit's regions plus the new views; `d/metrics/acceptance-raw.txt`; pre-1 -> after-1; set 2 identical to 0.1 except left pier 19.3 in after-2): 13: - **Bank at most 0.8x lawn by day:** met, 0.738 -> **0.759** (bank 73.7 -> 75.8). 14: - **At least water + 8 at night:** met, open water +12.4 -> **+15.1**, water below +10.2 -> **+12.9** as first submitted (bank 55.0 -> 57.7); with the review fix +12.2, with the recheck fix +12.0 (see those bullets). The plan's 0.8x at night is waived (P3c ruling): 1.064 -> 1.116. 15: - **At most 1.5x lawn at golden:** met, 1.239 -> **1.255**. 16: - **At most 1.0x the lifted lawn in Glade:** met, day 0.560 -> **0.572**; Glade night 0.771 -> 0.817 (bank 62.8, water +29.0). 17: - **No isolated black pixels: not met as first submitted; met on the bank foot after the review fix** (see "Review fix" below for the whole-foot numbers). The first submission claimed `bank` and `waterline` `art.black` 0 on every view; that held only because the kit's fixed 16-px rects miss the foot, and it was wrong: the frames had a 1-2 px black line along every bank foot. P0c lip specks (`d/metrics/lipspecks.txt`): day 4 -> 7, night 6 -> 7, golden 10 -> 11, Glade 38 -> 35 (P3-2). 18: - **No crawl:** by eye across the orbit (above); roots and stones sit on the face coordinate and fade with the footprint. 19: - **Bank tide line matches the quay's** (`d/tidematch.js`, `d/metrics/tidematch.txt`, banks scene, columns either side of the junction): rim-to-line-bottom 19 px on the bank (x 604-616) and 19 px on the quay (x 630-640), 21 px further along the quay (x 650-660, the wave), by day, golden and at night (night read by hand at x 612 and 640, 19 and 19); Glade 19-20 against 21. Columns where a stone top is the largest step read 26-29 (x 608). 20: - **P1d's quay ratio after P3a's night shallows:** quarter-night `quay_over_paving` pre-P3 0.738 -> after P3a-c 0.729 -> after P3d **0.729** (target >= 0.55; P1e left 0.735/0.730); p50 form 0.627 -> 0.639; front_quay p50 22.9 -> 23.4; left pier 19.0 -> 19.5; frame below 0.1 28.8 -> 28.6% (Q11 band 28-33). The shallows did not undo the lift. Q10 on the quarter: 05:45 27.9%, 20:45 28.4%, at or under night's 28.6%. 21: - **Twilight and Glade night (owner Q10; `d/metrics/q10-harbour-glade.txt`):** on Harbour the frame share below 0.1 is 7.8% (05:45) and 7.4% (20:45) against 7.1% at night, Glade 20:45 11.7% against 10.3%; the same order before P3d (7.6/7.2 against 6.9; 11.5 against 10.2), so inherited, not from this part. The bank itself at twilight is 48.9/48.7 against 57.7 at night (46.8/47.0 against 55.0 before), while the lawn is 31.2/32.4 against 51.7: the whole Harbour frame is darker at twilight than at night. Not tuned here; a question for the P3 sheet. 22: - **Bench** (`d/bench/bench.sh`, `uptime.txt`, `summary.txt`): trees preP3c (21-P3-c), postP3c (21-P3-d) and now (live copy), lane, grove and styles at 13:00, rain 0.45, `--disable-vsync`, 1920x1080, 5 interleaved rounds after a warm-up, raw_* fields, AC, load average 6.2 at start. Raw p50 medians: lane 6.36 / 6.37 / 6.39 ms (P3c +0.01, P3d +0.03, 4 of 5 rounds slower), grove 3.94 / 3.93 / 3.92 (P3c -0.01, P3d +0.00), styles 4.20 / 4.17 / 4.17 (P3c -0.03, it also drops 21 draw calls; P3d +0.00). Draw calls 747/741/741, 322/315/315, 497/476/476. 23: - **Journey:** expected count deltas zero on all six views (draw calls identical). Look change: the bank feet on harbour-night, harbour-day, harbour-lane and styles-day; quarter-night and quarter-day unchanged. 24: - **Review fix (30 September, 05:55-06:35):** the review found two must-fix items. (1) The bank's algae foot, `weather_tide_albedo` on the already dark soil plus the x0.85 packed soil between the stones, drove the foot below the kit's black threshold (luma < 12.75) on the headline views: a continuous 1-2 px black line along every bank foot, lit and shaded, and a black honeycomb between the stones on earth at night (review: `d/review/newdarkstats.js`, masks `d/review/dm-*.png`, crops `d/review/crop-*.png`). (2) This entry's "`art.black` 0" line was contradicted by its own captures. Cause, by one-view experiments (`d/exp.sh`, `d/cap/fx-*`): not the gloss (`fx-nogloss` unchanged, 672 new black), the algae mix (`fx-noalgae` 100; a foot px pre 50 -> algae 12 -> no algae 35: the grade's toe crushes the 0.61x luminance to about 0.3x in sRGB). Fix: `BANK_TIDE_FLOOR` (above) and `BANK_LUMA`, `ground.gdshader` only; tried 0.80 band-only (`fx-f80`: glade-2045 still 883), 0.80 with the rock band and the Glade raise (`fx-f80b`, then a full sitting `d/cap/fpre80-*`/`fix80-*`: styles-day 67 and chemistry-day 35 left on the bank), 0.90 (`fx-f90`, final). Final sitting 06:17-06:26 in the order `fpre-1` (scratch tree), `fix-1`, `fpre-2`, `fix-2`, 24 views each, AC 100% Low Power Mode 0 (`d/cap-power.txt`), all first try. 25: - **Whole-foot measure** (`d/fix/footblack.js`, `d/metrics/fix-footblack.txt`): the bank mask is every pixel P3d touches (changed > 8 luma pre -> after in both review sittings while the two pre sittings agree within 4, grown 1 px: the roots, rock band, algae band and foot, 9,000-59,000 px a view); new black = black in both fix sittings and in neither fpre sitting (the review's measure), noise = black in fpre-2 not in fpre-1. New black on the bank, old P3d -> fix (noise): harbour-day 661 -> **0** (2), chemistry-day 2613 -> **0** (4), grove-day 1152 -> **0** (0), styles-day 455 -> **21** (0), harbour-golden 371 -> **22** (38), harbour-night 419 -> **3** (6), orbit-c-day 797 -> **7** (25), banks-night 1995 -> **2** (0), banks-golden 1468 -> **11** (0), banks-day 434 -> **0**, glade-2045 793 -> **176** (17), glade-day 91 -> 85 (12), glade-night 86 -> 86 (9); every other view 0-4 (lane-night 4 against noise 2, the rest at or under their noise). Whole frame, same measure: glade-2045 1216 -> 195 (noise 75), harbour-day 668 -> 36 (136), chemistry-day 2620 -> 12 (196). 26: - **What is left** (`d/fix/joined.js`, `d/metrics/fix-joined.txt`, maps `d/fix/j-*.png`): on Glade 173 of 176 (20:45), 84 of 85 (day) and 86 of 86 (night) touch black that was already there: they are the P3c shadow line under the lip, which is black in Glade before P3d too (460-470 black px in the mask before and after), running on where the calm took the grass drips away; not the foot and not isolated (`d/fix/look-g2045.png`: pre | old P3d | fix | map). styles-day 21 (17 isolated, about 5 art px) and harbour-golden 22 (4 isolated) sit on faces in deep shadow that were 13-21 luma before (threshold 12.75), or on a sign and a bush the mask's 1 px growth reaches (`d/fix/z-styles-f90.png`, `d/fix/z-bm-hg.png`). The foot line and the earth honeycomb are gone: `d/fix/look-harbour-day.png`, `d/fix/look-earth-night.png` (pre | old P3d | fix), `d/fix/z-lane-3way.png` (lane: pre | 0.80 | 0.90), `d/fix/z-banks-3way.png` (banks-day: old P3d | 0.80 | 0.90). The algae band now reads by its green-teal hue, gloss and pale tide line rather than by darkness; at lane zoom it is still a visible band at the foot. 27: - **Re-measured with the fix** (`d/fix/accept-fix.js`, `d/metrics/fix-acceptance-raw.txt`, fpre-1 -> fix-1, set 2 the same to 0.1): bank at most 0.8x lawn by day 0.738 -> **0.759**; night bank over water open +12.4 -> **+15.1**, below +10.2 -> **+12.2** (owner's +8, met); golden 1.241 -> **1.255** (<= 1.5); Glade day 0.560 -> **0.572** (<= 1.0), Glade night 0.771 -> 0.817; quarter-night `quay_over_paving` 0.729 -> **0.734/0.729** (>= 0.55); frame below 0.1 on the quarter 28.6% at night, 27.9% at 05:45, 28.5/28.4% at 20:45. Kit rects `bank`/`waterline` `art.black` still 0 on the canal views (glade-2045 bank 66 -> 58); the banks scene's `bank_earth` rect has 50 (golden) and 99 (night) black px before and after alike (not from this part). Counts and draw calls identical on all 24 views in both sets (`d/metrics/fix-counts-set1.txt`, `-set2.txt`). Journey deltas unchanged (zero). No new bench: the fix adds two dot products and a max in the bank branch only. 28: - **Recheck fix (30 September, 06:37-07:10):** the recheck (`d/review/recheck/nb.txt`: new black = black in both fix sittings, in neither fpre sitting, whole frame) left three items. (1) styles-day: 44 new black px (11 art px, 9 free-standing) on the shaded bank faces, pre 14-17 -> 12-13 luma: the 0.90 floor on faces that sit just over the threshold. (2) Glade: with the drips calmed away, P3c's shadow line (black under Glade's grade) ran on unbroken: 195 new black px on glade-2045 (noise 75/60), 93 glade-day, 87 glade-night. (3) This entry's suite window and the pre-fix night figure (both corrected above). Fix, `ground.gdshader` only: `BANK_TIDE_FLOOR` 0.90 -> **0.95**, the rock band over the tide band floored at the whole soil (`mix(rock_band, floor_share, tide.x)`: packed soil and stone undersides never darker than the soil; the stones read by their lit tops), and `BANK_CALM_SHADOW` 1.0 (Glade's shadow line at 0.93 of the soil; canal keeps P3c's 0.55). Found by two-capture experiments (`d/exp2.sh`, `d/fix/exp2nb.js`, `d/cap/fx-r1-*` rock floor 1.0 + calm shadow 0.6: glade-2045 still 80 on the lip, lip line 9-14 luma against soil 20; `fx-r2-*` the final values). Tests first: `d/recheck-testfirst-test_weathering.log.out` 504 checks, 2 failures; `-test_look` 552, 2. After: `d/recheck-*`: test_weathering **505, 0**, test_look **552, 0**, test_chemistry 413, 0, test_nature 282, 0. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/d/suite-recheck` (06:47:12-06:52:59, AC 100%, Low Power Mode 0): **40 suites, 19,830 checks, 0 failures** (19,825 + 5), `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`, creature_reactions and its rerun 0 failures, the known flaky check did not fire; 1440x900 not run (inherited stall). Sitting 06:53-07:02 in the order `rpre-1` (scratch tree), `rfix-1`, `rpre-2`, `rfix-2`, 24 views each, all first try, AC 100% Low Power Mode 0 (`d/cap-power.txt`; `caps.sh` now prints the set name). 29: - **New black, the recheck's measure** (`d/fix/rnb.js`, `d/metrics/recheck-nb.txt`; before -> after the recheck fix, noise rpre/rfix pair): styles-day 44 -> **8** (24/0; 2 art px: one on a tree canopy that is 12 in two older pre sittings, one at the bridge post's foot 15 -> 13, the only bank px left), glade-2045 195 -> **5** (100/96), glade-day 93 -> **8** (95/124), glade-night 87 -> **18** (58/121), all with 0 free-standing blobs; every other view at or under its noise (quarter-night 204 against 1380/1136: rain and rings; it has no bank px). On the bank mask (`d/fix/rfootblack.js`, `d/metrics/recheck-footblack.txt`): styles-day 21 -> 4, glade-2045 176 -> 0, glade-day 85 -> 7 (noise 13), glade-night 86 -> 2, harbour-golden 22 -> 0, the rest 0-12 at or under noise except banks-golden 11 (noise 0; whole frame 20 against 16/16): 5 art px of the earth bank's P3c shadow line where two roots start under it, 13.4 -> 12.2 luma, the same in every P3d sitting and not a recheck item; left (`d/look-recheck/banks-golden-mask.png`: rpre-1 | rfix-1 | mask). Looked at: `d/look-recheck/styles-banks.png`, `glade-2045-lip.png`, `glade-2045-lip-z.png`, `glade-day-lip.png` (rpre-1 | review fix | recheck fix): styles' shaded faces keep their pre tone; in Glade the lip now meets the soil with no dark line (the lip still reads, lighter over the brown). 30: - **Acceptance re-measured** (`d/fix/accept-fix.js` with `SETS=rpre-1,rpre-2,rfix-1,rfix-2`, `d/metrics/recheck-acceptance-raw.txt`; rpre-1 -> rfix-1 / rfix-2): day bank 0.738 -> **0.766/0.760** (<= 0.8x lawn); night over water open +12.4 -> **+15.1**, below +10.2 -> **+12.0/+11.9** (>= +8); golden 1.241 -> **1.255** (<= 1.5); Glade day 0.560 -> **0.580** (<= 1.0), Glade night 0.771 -> 0.840 (water +30.8); glade-2045 kit `bank` `art.black` 66 -> **0**; twilight Harbour bank 46.8/47.0 -> 48.9/48.7; quarter-night `quay_over_paving` 0.734/0.729 -> 0.724/0.732 (>= 0.55), frame below 0.1 28.5-28.6% at night, 27.9% at 05:45, 28.4-28.5% at 20:45 (Q10 order kept). Lip specks (`d/metrics/recheck-lipspecks.txt`): harbour day 4 -> 7, night 6 -> 7, golden 10 -> 11, **Glade 38 -> 3** (P3-2: Glade back to its pre-P3c 3). Counts and draw calls identical on all 24 views in both sets (`d/metrics/recheck-counts-set1.txt`, `-set2.txt`); journey deltas zero. No new bench: the fix swaps a max for a mix and adds one mix in the bank branch. 31: - **Review:** one Opus 5.5 review, then two fix passes and two rechecks ([run ID removed]; P3d is not a measuring part, so no second reviewer). Round 1 found two must-fix items: a black 1-2 px line along every bank foot (algae mix on the dark soil; the fixed 16-px regions had missed it) and the false "black 0" claim. Fixed with `BANK_TIDE_FLOOR` (0.95) and a floor on the rock band; then Glade's lip shadow line, which ran on black where the calm removed the drips, was calmed (`BANK_CALM_SHADOW`); rechecked: nothing open. Left, recorded: 8 new black pixels on styles-day (2 art pixels, one on a tree canopy) and 5 art pixels on the earth bank's shadow line where two roots start; twilight is slightly darker than night on Harbour and in Glade (inherited from before P3d; owner Q10, for the P3 sheet); the night stones read as pale chips at lane-night; the canal soil stays warm at lane zoom (P3-1). The bench was not re-run after the fixes (a `max` swapped for a `mix`). Committed and merged on 30 September 2026 at 07:10 on the owner's instruction ("merge and update journal"), after the overnight standing instruction had ended. 32: - **Not done / open:** P3-1: canal's day soil still reads warm orange-brown at lane zoom and golden (lane `bank` sat 51%, not in this part's scope; Glade softened) and the strata stay regular bands under the new roots and stones. The night stones read strongly at lane-night (pale chips); a look question. Harbour and Glade twilight are darker than night by the frame share (inherited, above). After the review and recheck fixes the algae band on the bank is a hue band rather than a dark band, and the rock band's packed soil is no darker than the soil (the stones read by their lit tops); in Glade P3c's shadow line under the lip is gone (0.93 of the soil; Glade lip specks 38 -> 3): look questions for the P3 sheet (P3-2). Left: banks-golden's 5 art px on the earth bank's shadow line (Recheck fix). `d/cap-power.txt` headers for the fix sittings have an empty set name (caps.sh printed `$1` after `shift`, the P3c slip, fixed for the recheck sitting); the order is fpre-1, fix-1, fpre-2, fix-2 by time. Scratch trees `d/pre-tree`, `d/prestage-tree`, the bench trees `d/bench/trees` and the fixes' experiment captures `d/cap/fx-*` (with the recheck's `fx-r1*`, `fx-r2-*`) and `fpre80-*`/`fix80-*` left for the orchestrator; backups of the recheck's starting files `d/backup/*.recheck-start`. Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 30 September 2026, 07:10) 18: | D12 | Collar follows the ground finish | no | P2e | 19: | D13 | Beached boats: no lift | no | P3a | 92: ## Overnight standing instruction (owner, 30 September 2026) 94: "Automate all choices until 7am": until 07:00 on 30 September the orchestrator takes the plan's defaults and its own judgement on choices, commits and merges to local `main`, and pushes nothing. Under it: P2a, P2e2 and P2p were committed and merged (P2a and P2e2 merges were run by the owner in the terminal after the permission classifier blocked them), the P2 close was run, and the handoff check's corrections were applied. The owner is to look at the sheets on return; every choice below is reversible. 104: ## P3a beached-boat rule (orchestrator, 30 September 2026, under the overnight standing instruction; reversible) 106: The plan's gate before P3a is the beached-boat rule. Not answered by the owner, so the plan's default (the lead's rule) applies: no lift; the hull skirt sits inside the terrain; only the raised side of a tipped hull is checked; emitter and streak positions stay where they are. 108: ## P3c ruling: the night bank lines (orchestrator, 30 September 2026, under the overnight standing instruction; reversible) 110: The plan's P3c and P3d acceptance line "bank at most 0.8x lawn at night" cannot be met together with the owner's G1 Q3 line "night bank at least water + 8" (night lawn 51.7, so 0.8x is 41.4, below open water 42.6). Owner decisions override the plan: the owner's line stands, the plan's 0.8x night line is **waived** for P3c and P3d. P3c delivers night bank +12.4 over open water and +10.2 over the water below (target +8), which is 1.064x the lawn. `BANK_NIGHT_LIFT` (`shaders/ground.gdshader`, 0.36) is the one constant if the owner wants the bank nearer the lawn: about 0.25-0.28 would land near 1.0x lawn with little margin over +8. Kept at 0.36 for the margin between sittings (spread up to 1.3 luma). 112: ## Open for the owner after P3c (look questions for the P3 sheet) 114: | # | Question | State | 115: |---|---|---| 116: | P3-1 | The bank soil reads warm orange at lane zoom, golden hour and in Glade (Glade day #6B3D15), and the strata are regular bands. Keep, or cooler and less saturated (P3d also handles Glade calm)? | Open | 117: | P3-2 | The one-pixel shadow line raises the P0c lip-speck count (Glade 3 to 38, golden 2 to 10); they are line steps, luma 24-25, not black. Accept? | Open | 118: | P3-3 | P3a's night rim is a strong lavender line; the P3b day rings over the mirrored town are brighter with white dashes. Keep? | Open | ADDED 30 September 2026: the owner's answers to P3-1, P3-2 and P3-3 Appended after the fact-check. Game main at bcd4a7e (07:17 BST, 30 September), which only appended these answers to artifacts/beautify/owner-decisions.md. The header above says "P3-1, P3-2 and P3-3 are open" as frozen at 07:10; these answers supersede it. The P3 stage close itself had still not been merged when this was appended and is not reported. Only the P3 look row is quoted here; the P2 rows are in the chapter 20 notes, and the P3 close, P4a gate and GitHub rows are not journal content. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 30 September 2026, main at bcd4a7e) 120: ## Owner answers after P3d (owner, 30 September 2026, in chat) 122: | # | Question | Verdict | 123: |---|---|---| 127: | P3-1 to P3-3 | Bank soil warm orange, lip specks, lavender night rim and brighter day rings | **Keep all as they are now** (to be looked at again on the P3 sheet if the owner wishes) | ADDED 2 October 2026: the P3 stage close and the sign-off of the P3 sheet Appended by the 2 October edition. The header above says the P3 stage close had not run; it ran on 30 September (07:25-07:47) and the owner signed off the P3 sheet the same day. Original source terminology is preserved; local paths are removed. Line numbers refer to the frozen copy of docs/visual-handoff.md taken at 10:27 BST on 2 October 2026 (research/update-2026-10-02-P3close/ in the journal folder), not to the 30 September copy above. SOURCE: docs/visual-handoff.md (frozen copy, 2 October 2026, 10:27 BST, game main at af4df88) 591: **Beautify P3 stage close (30 September 2026):** the stage-end overhead of P3 "Islands that sit in water" (step `21-P3`), run by Opus 5.5 on this Mac on the final P3 code (`main` bcd4a7e; untracked `.claude/worktrees/` and two plan docs only; `close/head.txt`). No game code changed. Evidence: `artifacts/beautify/21-P3/close/`. Before side: the pre-stage scratch tree `close/pre-tree` from `build/backup-feel/21-P3/`, checked equal to `main` 5b163ad (the P2 close) on scripts, shaders, tests, data, scenes, tools and `project.godot`; only the three `.ps1` tools differ, in line endings (`close/diff-pretree-vs-5b163ad.txt`). Power: AC, 100% charged, Low Power Mode 0 for every run (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`); load average about 4 during the bench (other apps busy). 592: - **Full suite** (`close/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/21-P3/close/suite`, 07:25-07:31): **40 suites, 19,830 checks, 0 failures**, the same count as P3d's recheck suite. `--verify` 2880x1800 exited 0 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 not run (the inherited stall). 593: - **Journey** (`artifacts/feel-journey/21-P3/`, the six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all first try; counts `21-P3/counts-vs-20-P2.txt`). Against `20-P2`: harbour-day and -night draw_calls 897 → 889, harbour-lane 738 → 733, quarter-day and -night 970 → 969, styles-day 491 → 469; `rendered_objects` moves the same way. Exactly the declared deltas: P3b −1 on every rainy view with water, P3c −7 (Harbour), −4 (lane), −21 (styles), 0 (quarter); P3a and P3d 0. Re-run: the stage-end art capture took all six views again with identical integer counts. By eye (`close/look/j-1.png`, `j-2.png`): a thin lit rim and lighter, textured bank sides on the Harbour views (lavender at night), rings on the open water, the quarter's quays and boat rimmed; its town and styles' houses unchanged. `notes.json` has a new `21-P3` entry (backup `notes.json.bak-before-21-P3`); `index.html` and `journey-standalone.html` (105.98 MB) rebuilt. 594: - **Art capture and protected looks** (`close/caps.sh`; one sitting 07:34-07:47 in the order `pre-1`, `after`, `pre-2`, `after-2`; 35 views each, all first try): the kit's 21 baseline views, harbour-lane, harbour-0545, harbour-2045, glade-2045, the orbit (orbit-a/-c at night and by day), P3a's beached-boat scene (day, night) and P3d's banks scene (day, night, golden, Glade). Metrics in `close/metrics/` (`close/metrics.sh`, regions `21-P3/d/regions.json`): `ab-pre1-after`, `ab-pre2-after2`, noise pairs `ab-pre1-pre2` / `ab-after-after2`, `frame-*.txt`, `counts.txt`, `protected.md` (`close/prot.js`), `bank-lines.txt`, and `ab-p2close-*` against the previous stage close `20-P2/close/after`. 595: - The pre-stage tree's counts equal `20-P2/close/after` on all 22 shared views. Count changes beyond the journey: harbour-golden and -0545 −8, harbour-2100 −7, lane −5, orbit −8, grove −6, chemistry −10, photo and the rainy quarter hours −1, banks −13 (Glade −6), beached −5; glade views and quarter-dry identical. after vs after-2 and pre-1 vs pre-2 identical everywhere. 596: - Noise pairs change >8 on 0.06-3.8% (pre) and 0.05-6.0% (after) of the frame. quarter-dry 1.25% against 1.23% (within noise). 597: - Protected-look sheets: `close/protected/protected-1.png` to `-3.png`, `protected-4-vs-P2-close.png`, `noise-pre.png`, `noise-after.png`, all looked at. **Canal mirror columns:** held; p50 42.8 → 44.0, >8 on 16.2/44.7% against noise 14.0/39.0% (the rings on the columns). **HDR neon cores:** fish, bar and lane signs and the tower windows within noise (fish sign 3.5/0.2% against 2.0/1.8%). **05:45 and 20:45 skies:** zenith and horizon p50 identical (56.6, 114.9, 97.5, 153.0-153.1). **Glade/Canal separation:** Glade night zenith `#001d45` and horizon 89.2, Harbour `#293280`, unchanged; Harbour night open water 42.6 → 42.6 (>8 on 6.2% against 2.3-3.4%: the rings). **Approved tree style:** grove crowns 0.01%, conifers 0.02%, feet 0%, lawn 0% (the grove frame's 1.4% against 0.06% noise is its bank feet); quarter-night crown undersides 3.0% against 3.6%. 598: - Bank lines at the close (`bank-lines.txt`, set 2 identical): Harbour day 0.883 → 0.766 of lawn; night +3.6 → +15.1 over open water, +4.5 → +12.0 over the water below, 1.116x lawn (the plan's 0.8x at night waived, P3c ruling); golden 1.477 → 1.255; Glade day 0.664 → 0.580 of the lifted lawn. Quarter-night front quay p50 22.9 → 23.4 over paving 36.5. 599: - **Whole-stage A/B bench** (`close/bench.sh`, `bench-t1.out`, `bench-t13.out`, `bench/summary.txt` from `close/bench-summary.js`, `bench/power.txt`): pre-stage tree against the live root, interleaved, round 0 discarded, rounds 1-3 kept, `--disable-vsync` before `--path`, 1920x1080, rain 0.45, Night (`--time=1`) and 13:00, dense `artifacts/weather-pass9/dense.json` and Harbour. vsync disabled in every run, 491 raw samples, all first try. Raw p50 medians: 600: - **Dense, night: 13.46 → 13.44 ms (−0.01 paired; 1/3 rounds slower); p95 15.24 → 15.15.** 601: - **Dense, 13:00: 13.45 → 13.41 ms (−0.00; 1/3); p95 15.14 → 15.18.** 602: - **Harbour, night: 7.85 → 7.63 ms (−0.13; 0/3); p95 9.07 → 8.98.** 603: - **Harbour, 13:00: 7.76 → 7.63 ms (−0.13; 1/3); p95 9.14 → 8.97.** 604: - Draw calls: dense 2361 → 2360, Harbour 907 → 899. 605: - Reading: P3 costs nothing measurable; Harbour is slightly faster (8 fewer draw calls, no CPU splashes), inside its round spread (up to 0.45 ms). Harbour's absolute times are about 1.2 ms above the P2 close's for the same pre-stage code (6.59-6.65 there): machine load in this sitting, not a change; compare only inside one bench. 606: - **Owner sheet** `artifacts/beautify/21-P3/close/sheet/`: before (pre-stage tree, `pre-1`) left, after right, colour then greyscale where values matter; crops from `close/crops.js`, pages from `close/sheet.sh`; all pages looked at. Pages: `p3-01-overview.png`, `p3-02-P3a-rim-swash-shallows.png`, `p3-03-P3a-hulls-and-quays.png`, `p3-04-P3a-P3d-glade.png`, `p3-05-P3b-rain-rings.png`, `p3-06-P3c-cake-slice-cut.png`, `p3-07-P3d-roots-rock-algae.png`, `p3-08-banks-night-golden.png`, `p3-08b-banks-twilight.png`, `p3-09-P3d-glade-calm.png`, `p3-10-orbit.png`, `p3-11-existing-scenes-1.png`, `p3-12-existing-scenes-2.png`, `p3-13-existing-scenes-3.png`. `sheet.md`: one plain paragraph per part with its numbers, what existing scenes see, things to look at, protected looks, bench and suite. This is the plan's combined P3 sheet (harbour-day/-night/-golden, styles-day, glade-day, grove-day, the quarter-night boat and canal in rain, the orbit). 607: - Looked at and flagged for the owner (not changed): at 05:45 and 20:45 the Harbour bank reads 1.57x and 1.50x the dark lawn (before 1.45x, 1.39x; Glade 20:45 1.45x), page 8b; at lane-night the bank is a pale lavender-blue block with light stone chips, lighter than the lawn, page 7. 608: - **Not done / open:** 609: - The owner has kept P3-1 to P3-3; the plan's gate "the P3 sheet before P4b" still wants the owner to see this sheet before P4b starts (P4a may run meanwhile). 610: - No milestone export at the P3 close (D8: M3 comes after P4). 611: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 612: - Scratch trees: `a/pre-tree`, `b/pre-tree`, `c/pre-tree`, `d/pre-tree`, `d/prestage-tree` and `close/pre-tree` under `21-P3` were moved to [local path removed] with `mv` once their numbers were recorded, named `beautify-21-P3--pre-tree-