Chapter 22: A city on the horizon Frozen 2 October 2026 at 10:27 BST (game main at af4df88). Implemented, reviewed and merged work of the P4 stage (skyline, neon bay and golden hour): P4a skyline landmark row, gamma-space haze, neon x1.8 and the 3D ring finding; P4b violet city glow, two-tone cloud deck and lit far quay, with its review fix and recheck fix; P4c neon bay columns with its review fix; the P4 owner fixes (3D ring deleted, quay lamps 15% dimmer, Glade without columns); P4de golden hour with its review fix and recheck fix; and the P4 stage close (suite, journey, protected looks, whole-stage bench, owner sheet). The owner answered P4-1 to P4-7 on 30 September and signed off the P4 sheet on 1 October. "Review: filled by the orchestrator" lines are as frozen. Original source terminology is preserved; local paths are removed. Line numbers refer to the frozen copy. Part codes: P4a skyline, P4b city glow, cloud deck and far quay, P4c neon bay, P4de golden hour. SOURCE: docs/visual-handoff.md (frozen copy, 2 October 2026, 10:27 BST, game main at af4df88) 396: **Beautify P4 stage close (1 October 2026):** the stage-end overhead of P4 "Skyline, neon bay and golden hour" (step `22-P4`), run by Opus 5.5 on this Mac on the final P4 code. That code is `main` d5106f6, clean: the P4 code is c51b76c, and the commits after it touch only the water ambience and the trailer and soundtrack plans (`close/head.txt`). No game code changed. Evidence is in `artifacts/beautify/22-P4/close/`. The before side is the pre-stage scratch tree `close/pre-tree`, built from `build/backup-feel/22-P4/`. It was checked equal to `main` 9df1e18 (the P3 close plus the sign-off) on scripts, shaders, tests, data, scenes, tools and `project.godot`; only the three `.ps1` tools differ, in line endings (`close/diff-pretree-vs-9df1e18.txt`). Power was AC, 100% charged, with Low Power Mode 0 for every run (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`); the load average was about 3.5. 397: - **Full suite** (`close/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/close/suite`, 00:29-00:36): **40 suites, 20,085 checks, 0 failures**. That is P4de's recheck suite of 20,081 plus 4 from `test_scenery_audio` (221 to 225, from the audio commits). `--verify` at 2880x1800 exited 0 in 34 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures, and the known flaky check did not fire. 1440x900 was not run (the inherited stall). 398: - **Journey** (`artifacts/feel-journey/22-P4/`: the six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on the first try; counts in `22-P4/counts-vs-21-P3.txt`). Against `21-P3`, integer counts and draw calls are **identical on all six views** (harbour-day and -night 889, harbour-lane 733, quarter-day and -night 969, styles-day 469). That matches every P4 part, each of which declared no count deltas. As a re-run, the stage-end art capture took all six views again and got identical counts. By eye (`close/look/j-1.png`, `j-2.png`): 399: - The landmark tower and a violet sky with a two-tone deck behind both Harbour views and the quarter. 400: - Neon columns on the night bay. 401: - The tower as a pale silhouette by day. 402: - The town and island unchanged. 404: Frame >8 luma against 21-P3 is 2.5-3.3% on lane, 7-9% on the day views (the skyline) and 19-23% on the night views. `notes.json` has a new `22-P4` entry (backup `notes.json.bak-before-22-P4`); `index.html` and `journey-standalone.html` (106.47 MB) were rebuilt. 405: - **Art capture and protected looks** (`close/caps.sh`): one sitting, 00:38-00:56, in the order `pre-1`, `after`, `pre-2`, `after-2`. Each set has 43 views and every view worked on the first try. The views are the P3 close's 35 plus harbour-0645, the 4-yaw golden orbit (harbour-golden, orbit-g1 to g3), grove-1930, glade-1930, h0600 and h0608. Metrics are in `close/metrics/` (`close/metrics.sh`, regions `close/regions.json` = `21-P3/d/regions.json` plus the golden views, no region moved): `ab-pre1-after`, `ab-pre2-after2`, the noise pairs, `frame-*.txt`, `counts.txt`, `protected.md` (`close/prot.js`), `columns-city.txt`, and `ab-p3close-*` / `frame-p3close-*` against the previous stage close `21-P3/close/after`. 406: - Counts are identical on every view in every pairing (pre against after, the noise pairs, journey against close, and the P3 close against pre-1). Noise pairs change >8 on 0-5.6% of the frame. 407: - Protected-look sheets: `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P3-close.png`, `noise-pre.png`, `noise-after.png`, all looked at. 408: - **Canal mirror columns:** quarter-night changed >8 on 14.6/7.6% against noise 25.6/13.8% (p50 44.4 → 44.2). Quarter-21:00 changed on 15.6/8.9% against 2.4/7.0%. That is over one pair, but inside the 9.9-39% this region showed at the P3 close, and its p50 moved 52.3 → 52.7. By eye they are the same streaks. 409: - **HDR neon cores:** 0-1.4%, within noise. 410: - **05:45 and 20:45 skies:** the zeniths are identical. The horizon bands changed as the parts intended: quarter-20:45 #cd8b89 → #c38492 (P4b half tint plus the P4de belt, owner P4-5), quarter-05:45 #7a6e93 → #766791, 66% and 15% of the band. Some of each band's change is the new tower standing inside the crop. 411: - **Glade/Canal separation:** Glade's night zenith is still `#001d45` and its water unchanged; Harbour went `#293280` → `#2b337f`, and its horizon `#46458c` → `#4e3d7a` (hue 257). 412: - **Approved tree style:** grove-day crowns 0.64%, conifers 0.54% and feet 0.11% against noise 0-0.41%, with p50s identical. Quarter-night crown undersides changed 1.75/0% against 1.45/0.50%. Grove-1930 changes on 86% of the frame (the P4de lawn lift 44.6 → 78.8, the golden haze); its crowns moved +2.4 luma. 413: - Close-measured lines: 414: - Harbour-night has **12/12 neon columns** 30-120 render px long (longest 78 and 73, none over 120; pre 1/2). The part had 13-15 and 14-15 after the owner fix; the count moves a little between captures. Glade-night has 0 columns in every set. 415: - Harbour-golden: the city stands **23.9/23.6** luma under the sky (pre 13.6/12.4). The far bay went `#a49b94` → `#cfae9b` (25.0% saturation, as P4de). The lawn went 54.5 → **91.3** against water 116.3 (0.79x). 416: - **Whole-stage A/B bench** (`close/bench.sh`, `bench-t1.out`, `bench-t13.out`, `bench-t19.5.out`, `bench/summary.txt` from `close/bench-summary.js`, `bench/power.txt`). Setup: pre-stage tree against the live root, interleaved, round 0 discarded, rounds 1-3 kept, `--disable-vsync` before `--path`, 1920x1080, rain 0.45. It ran at night (`--time=1`), at 13:00 and at 19:30 (P4de's golden bench hour), on dense `artifacts/weather-pass9/dense.json` and Harbour, 01:03-01:13. vsync was disabled in every run, each run had 491 raw samples, and all ran on the first try. Raw p50 medians (pre → live, paired difference, rounds where live was slower): 418: | Scene | Night | 13:00 | 19:30 | 419: |---|---|---|---| 420: | Dense | 13.27 → 13.33 ms (+0.06, 3/3) | 13.21 → 13.24 (+0.03, 3/3) | 15.07 → 15.12 (+0.05, 3/3) | 421: | Harbour | 6.63 → 6.72 (+0.11, 3/3) | 6.61 → 6.68 (+0.06, 2/3) | 6.50 → 6.59 (+0.04, 3/3) | 423: - p95 medians: dense 15.21 → 15.43 (night), 15.23 → 15.26 (13:00), 16.99 → 17.04 (19:30); Harbour 7.88 → 8.46, 7.90 → 8.15, 7.00 → 8.01. Harbour's p95 is noisy, with round spreads up to 1.2 ms. 424: - Draw calls are unchanged: 2360/899 (night, 13:00) and 3124/814 (19:30). 425: - Reading: P4 costs a small, consistent amount, about +0.05 ms on dense (0.4%) and +0.04-0.11 ms on Harbour (1-1.6%). That is the backdrop and water shaders' extra sky work (landmark row, glow tint, deck, sun side, panel run). The per-part benches each saw it as no change beyond noise; summed over the stage it shows in 17 of 18 paired rounds. 426: - **Owner sheet** `artifacts/beautify/22-P4/close/sheet/`: before (pre-stage tree, `pre-1`) on the left, after on the right, colour then greyscale where values matter. Crops come from `close/crops.js` and pages from `close/sheet.sh`; all pages were looked at. Pages: `p4-01-overview.png`, `p4-02-P4a-skyline.png`, `p4-03-P4b-cloud-deck.png`, `p4-04-P4b-far-quay.png`, `p4-05-twilight-skies.png`, `p4-06-P4c-neon-bay.png`, `p4-07-P4c-where-not.png`, `p4-08-P4de-golden-sky.png`, `p4-09-P4de-golden-lawn.png`, `p4-10-P4de-dawn-water.png`, `p4-11-golden-orbit.png`, `p4-12-existing-scenes-1.png`, `p4-13-existing-scenes-2.png`, `p4-14-existing-scenes-3.png`. `sheet.md` has one plain paragraph per part with its numbers, then what existing scenes see, the protected looks, the bench and the suite. The plan's night sheet and golden sheet are combined here. 427: - Looked at and flagged for the owner (not changed): 428: - Glade shows the new tower row (by plan until P5d; the owner dropped P5d on 1 October 2026, so Glade keeps the tower). 429: - Golden hour's far-water rain rings dissolve into a smooth peach band near the horizon. 430: - The 06:00 water after is even but carries soft lighter swirls (no hard patches). 431: - **Not done / open:** 432: - The cool band in the water is only a faint mauve strip (P4de, open for an owner or orchestrator ruling). 433: - Puddles still mirror the untinted sky (P4b/P4c follow-up). 434: - The report's warm rim on block tops was not built. 435: - P4f was not run (owner default). 436: - The owner has not seen this combined sheet (written at the close; signed off on 1 October 2026). **The M3 exports** (owner D8: after the P3 and P4 sheets) were not built at the close and were built later the same day on the owner's request ([local path removed], from main 6fa5452). 437: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 438: - Scratch trees: `close/pre-tree` and the part trees `a/`, `b/`, `c/`, `de/` and `ownerfix/pre-tree` under `22-P4` were moved to [local path removed] with `mv` once their numbers were recorded. They are named `beautify-22-P4-close-pre-tree-011559` and `beautify-22-P4--pre-tree-011606`. Left for the orchestrator: `b/fix/reviewed-tree`, `b/fix2/fix1-tree`, `b/probe/final-pre-tree`, `b/probe/final-tree`, `c/fix/rev-tree`, `de/fix2/fix1-tree`. 439: - Platforms: Mac run; Windows not run. 441: **Recorded lake-water ambience (30 September 2026):** Sonnet 5.5 on this Mac. The synthesised lapping is replaced by a looped field recording, `assets/ambience/water.ogg` (Ogg q4, 52.0 s, 44.1 kHz stereo, 874 KB), from Freesound 573163 by TRP (CC0; credited in `THIRD_PARTY_NOTICES.txt`). Everything else in `scripts/scenery_audio.gd` (the level following visible water, the glide, the volume slider, the other beds and the score) is untouched. 442: - **Asset:** `python3 tools/make_water_loop.py` (ffmpeg, oggenc, numpy; no Godot) takes seconds 20-75 of the source, high-pass 80 Hz, low-pass 3.5 kHz, slow leveller 0.1 (400 ms RMS), splash peaks held to 16 dB over the body, -28 dBFS RMS, last 3 s crossfaded equal-power into the head. Numbers it prints and asserts: peak -6.6 dBFS, wrap step 0.0017 against a typical sample step 0.0052, 400 ms RMS range 15.3 dB (raw section 28 dB). The 43 MB source WAV lives in `artifacts/audio-src/` (git-ignored), not in the project, so it is not exported or committed (export_filter is all_resources). The first build (80 Hz high-pass only, leveller 0.3) sounded like white noise and sped up to the owner; four variants went to the owner by ear (`artifacts/audio-src/work/water_loop_{A_natural,B_gentle,C_untouched,D_dark}_x3.wav`) and **D (3.5 kHz low-pass, leveller 0.1) was chosen**; those are its defaults. 443: - **Code:** `water_path` (default `WATER_PATH`), `_load_recorded_water()` run in `_finish_render()` after the synthesis: if the file loads as an `AudioStreamOggVorbis` at least 1 s long it gets `loop = true` and becomes `streams.water`, else the synthesised bed stays (`water_recorded` says which). `_start_offset()` opens the recording at a seeded point (`water_rng`, SEED+11, separate from the bird generator so the birds' sequence is unchanged). `AMB_DB.water` -13.0 -> -1.0 (recording -28.1 dBFS RMS against the synthesised bed's -16.3, measured with `artifacts/audio-src/work/rms.gd`), then -4.7 after the owner played the build and found it about 35% too loud (0.65 x amplitude, -3.7 dB), then -7.7 because it still overpowered the score (3 dB more); the fallback keeps -13.0 (`WATER_SYNTH_DB`, `_amb_db()`). The header comment ("nothing is loaded from disk") and the README are updated. 444: - **Tests:** `tests/test_scenery_audio.gd` (run with `tools/run_tests_mac.sh $PWD/tests/test_scenery_audio.gd 170` after a headless `--import` with the pinned 4.6.3, no other Godot running): **231 checks, 0 failures**. The node-vs-synth assertion moved from water to wind; new checks that the recording loads as an ogg, loops, is at least 20 s, opens inside its length and not on the same lap twice, plays at the AMB_DB level, is what `loop_players.water` plays at its level, and that a missing `water_path` falls back to the synthesised bed at its own level. `assets/ambience/water.ogg.import` was created by the import and belongs with the file. 445: - **Only heard, not measured:** the owner heard the standalone loops with `afplay`, not in the game, so the in-game level (`AMB_DB.water` -1.0, matched by RMS) and the balance against wind, rain and the score are unjudged by ear. The seam was checked numerically only, not by listening past it. 446: - **Wind removed (same day, owner request):** the wind bed is gone: `wind` leaves `AMBIENCE`, `ALL_SOUNDS`, `LOOP_SECONDS`, `AMB_DB`, `context`, `amb_gain`, `listen()` and `render()`, with the `WIND_*` constants; `VOICES` 4 -> 3 ambience players (still within the 32-player budget). The other synthesised sounds now take their noise seeds from a list one place shorter (`SEED + index*977`), so rain, hum, the score and the birds are statistically the same but not sample-identical to before. The wind checks left `tests/test_scenery_audio.gd` and the node-vs-synth check now reads the rain bed: **225 checks, 0 failures** (231 before; the difference is the removed wind checks). The wind is not heard in any other script; the chimney-smoke lean is visual and is untouched. 447: - **Not run:** the full suite, `--verify`, an export (the ogg should ship under all_resources; not checked), Windows, Linux. The run printed `Resource file not found: res://assets/ui/hologram/terrain_flowers.svg` from `hologram_theme.gd` while the test built the editor; that file is not in git and this change does not touch it. 449: **Beautify P4de Golden hour: sun-side sky and golden values (30 September 2026):** Opus 5.5 on this Mac. The sky now knows where the sun is (report golden-hour pass; lens "Golden hour that knows where the sun is"). `sky.gdshaderinc` gains `sky_sun_side(height, facing, sky, sun, low_sun, shadow)` (signature after the review fix below): `low_sun` = `SunPath.golden(hour, heading)` (0 by day and at night, so the sky is exact there), `g = smoothstep(SUN_SIDE_RANGE 0.03-0.40, low_sun)` (0 at 05:45, 0.85 at 19:30, 1 at 20:45 and 06:45); away from the sun (`facing` < 0) the Belt of Venus, a pink band `BELT_BAND` 0.35 +- 0.12 of the panel (`mix(sun, BELT_PINK, 0.65)` at `BELT_GAIN` 0.46 x g x -facing), over the Earth's shadow (rebuilt by the review fix below); toward it a warm glow low down (`SUN_SIDE_GLOW` 0.35 x pow(facing, 4) x exp(-3 h)). The backdrop applies it after the city-glow tint with each column's own azimuth (`camera_yaw + pi + (x - 0.5) x SKY_PANORAMA`); the water applies it to `sky_ref` and to the haze its far water dissolves into, with the mirror ray's azimuth (no `camera_yaw`) at the height the soft read mirrors the panel from (`min(sky_h, below x PANEL_SOFT_SQUASH / SKY_HORIZON)`), so the belt lies mirrored in the far water; the dissolve's haze takes only the away side (the glitter path already lights the water toward the sun). A golden-hour weight `sky_golden_weight` (whole once low_sun passes 0.3 while the windows are under 0.66-0.85: golden hour and 06:45; 0 by day, at night, 05:45 and 20:45) times -facing lets the far water dissolve further into the low sky at golden hour (`WATER_GOLDEN_DISSOLVE` 0.95 from 0.55, reaching `WATER_GOLDEN_REACH` 0.58 of the frame from 0.50). `lighting_profile.gd`: golden anchor fog 0.78 -> 0.62, haze f0c8a6 -> ecb898; the 06:45 anchor is not retuned (owner default, fog 0.78). `ground.gdshader`: the meadow branch only lifts its albedo by `LAWN_GOLDEN_LIFT` 0.45 with the same knee and windows band (`LAWN_LIFT_*`), times (1 - ground_calm); the bank branch, earth and sand do not read `low_sun`. `WorldView.sync_profile` sets `low_sun` on the backdrop, the water and grass, earth and sand (with build variants), both looks. 450: - **Review fix (30 September 2026, two must-fix findings):** (1) *The sun side faced the moon from 21:00 to 21:30 and 05:40 to 06:10*: `sync_profile` passed `sun_direction` = `SunPath.disc_direction` (the moon once moon_weight >= 0.5, nearly opposite the sun) while `low_sun` was still over 0, so the belt and glow swapped sides as a running clock crossed 21:00 and 06:10. Now the backdrop and the water take a separate `uniform vec3 low_sun_direction` (default 0: facing 0, the old sky) = `SunPath.direction(SunPath.sun(hour, heading))`, the sun itself; `sun_direction` stays the disc. (2) *No cool band showed* (the shadow sat under 0.075-0.15 of the panel, behind the rows and mist, 60% slate x 0.33 over the peach). Rebuilt: `sky_earth_shadow(c, amount)` turns a colour to `EARTH_SHADOW` (0.64, 0.64, 0.84) at its own gamma luma (`sky_luma_match`) x `EARTH_SHADOW_DIM` 0.90 linear; `sky_shadow_side(facing, low_sun, windows)` = g x (1 - the `GOLDEN_WINDOWS` band) x `EARTH_SHADOW_GAIN` 0.85 x -facing; the sky takes it whole up to `EARTH_SHADOW_REACH`.x 0.18 of the panel (over the far rows' tops), gone by .y 0.28 (under the belt), and the backdrop's `haze` (rows, viaduct, mist, quay) takes the same amount, so the whole low band away from the sun is cool. The windows band keeps it off at 20:45 (owner P4-5: its magenta low sky stays) and from 21:00 on; whole at golden hour and 06:45. In the water the shadow lies only in the far water's dissolve haze, by the frame's own height (`WATER_SHADOW_SQUASH` 3.5: a strip about 18 px under the horizon at 1000 px) at `WATER_SHADOW` 0.4, not in `sky_ref` or by the ripples' mirror rays (trials t1-t3 put lilac blotches over the mirrored belt). Held faint because far-water saturation falls with it: 0.6 over 2.0 gave 22.5%, 0.5 over 3.0 24.3%, 0.45 over 3.5 24.9% (`de/fix/cap/t4`-`t6`). Trials `de/fix/cap/t1`-`t7` (t2 a lilac shadow, garish; t7 a darker water strip, no saturation gain), `de/fix/trial.sh`, `de/fix/m.sh`. Backups `de/backup/pre-fix/`. Snapshot diff at fix start: only this part's own files differ from `build/backup-feel/22-P4-de` (expected). 451: - **Fix tests:** test-first `de/fixfirst-test_{water,sun}.log.out` (366/2, 79/3 checks/failures); after `de/fix3-test_{water,sun}.log.out` **test_water 394 -> 406, test_sun 76 -> 79, 0 failures** (new: low_sun_direction on backdrop and water = the sun's own direction every 5 min across 20:55-21:35 and 05:35-06:15, two headings; in those windows it is over 90 degrees from the disc; under 3 degrees of turn across 20:58-21:02, 06:08-06:12, 05:38-05:42; the shadow reaches over the rows and is cool at 0.17 of the panel at facing -0.9 (blue over red, hue 190-290), a little darker (luma x0.85-1.0), the haze takes it, none toward the sun, side-on, by day, at night, 05:45, 20:45, 21:15 or 22:00, whole at 19:30 and 06:45, smooth over the evening; the water's strip line and no shadow in the ripples' sky). The windows-band checks were written with the code after trial 1 showed 20:45 turning blue, not before it. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh de/fix/suite` (a/b/v, 16:59-17:07, AC, Low Power Mode 0, `de/fix/suite-power.txt`): **40 suites, 20,077 checks, 0 failures**; `--verify` 2880x1800 exit 0 in 33 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549, 0 failures; 1440x900 not run (inherited stall). 452: - **Fix changed expectations:** committed P4a check `sky_mix_gamma(city_color.rgb, haze_color.rgb, %s)` -> `sky_mix_gamma(city_color.rgb, haze, %s)` (0.50, 0.26, 0.10) and P4b quay check `sky_mix_gamma(city_color.rgb * 0.7, haze_color.rgb, 0.35)` -> `... haze, 0.35)` (same checks; `haze` equals `haze_color` by day, at night and toward the sun). This part's own checks: `sky_sun_side` signature and calls (`horizon` dropped, `shadow` added; sky_ref passes 0.0), `EARTH_SHADOW_TOP`/`SHARE` -> `EARTH_SHADOW_REACH`/`DIM`, facing from `low_sun_direction` (backdrop, water, the heading check in test_sun). Journey count deltas: **none** (integer counts and draw calls identical on all 20 views, both sets, `de/fix/metrics/counts-1.txt`, `counts-2.txt`). 453: - **Fix captures** (`de/fix/caps.sh`, `de/fix/sit.sh`, 1600x1000, AC and Low Power Mode 0, `de/fix/power.txt`): pre-1 (`de/pre-tree`), after-1 (live), pre-2, after-2, 16:51-16:59, all first try or one retry, the 16 views above plus h2058, h2102, h0555, h0600; metrics `de/fix/summary.js` -> `de/fix/metrics/summary.txt`, `city.txt`. Looked at: `de/fix/look/final-hg-horizon-3x.png` (a lavender band behind the city under the pink belt; a faint mauve strip in the water under the quay, peach below, the mirrored pink lower), `final-handover-2x.png` (pre 20:58, live 20:58, live 21:02: the belt stays over the city across 21:00), `final-2045-0645-glade.png` (20:45 and 06:45 pre/after: as the entry above), `final-quarter-glade.png` (quarter-golden and Glade 19:30 take the same cool band). 454: - **Fix acceptance** (harbour-golden, pre-1/pre-2 -> after-1/after-2, identical in both sets): **cool band met in the sky**: crop `low` 1150,184,1500,196 #e4cbbd hue 22 -> **#cfc1e9 hue 261**, luma 202.5 -> 192.2; belt 135-160 hue 25 -> 16 over it. **In the water faint**: y 236-248 #cdc0b3 hue 30 -> #f1d3d5 hue 356 (rows 232-243 hue 311-331, sat 11-12): a mauve strip, not a clearly cool band. **Far-water saturation 9.8% -> 25.0%** (target >= 25, met at the line; 26.2% before the fix). **City 19.9 -> 23.6** under the sky (15-25; 22.9 before the fix). Lawn 54.5 -> 91.3 = 0.79x water (unchanged by the fix). Handover (review crop): 20:58 -> 21:02 live #a57990 -> #a27790, belt 153.3 -> 153.0, as pre (#a57991 -> #a27790); 21:02 low equals pre (112.0; review live had 139.7). 05:55 and 06:00 now take the sun side facing the sun (sky >8 10.5% and 32.9% against noise 1.1-1.3: the belt away from the rising sun, low_sun 0.16 and 0.26). Protected: harbour-day sky/water >8 0.34/1.94 and 0.28/0.70 against noise 0.31/2.04 and 0.34/0.80; harbour-night 1.19/2.38 and 1.33/2.54 against 1.45/4.06 and 1.10/1.10; glade-day 0.13/0.47 against 0.22/0.47; harbour-0545 1.28 / 1.06 against 1.26 / 1.14; quarter-0545 2.22 / 2.59 against 2.46 / 2.31; 20:45 sky 8.7 / 9.1 against 1.3 / 1.6 (the belt only, as before the fix; its low crop #ac7e90 -> #ac7e91, magenta kept); 06:45 sky 14.8 / 14.8 (the sun-side glow, as before the fix). 455: - **Fix bench:** not re-run. The fix adds a few ALU ops per backdrop and water pixel; draw calls unchanged. The pre-fix bench below stands for the part; the P4 stage close A/B covers the final code. 456: - **Recheck fix (30 September 2026; the recheck left three findings open):** (1) *The water's cool band is only a pale mauve strip* (harbour-golden y 236-248 hue 356, sat 12.5%): **not fixed, left open**; the reviewer's own finding says another code pass will not fix it (a stronger strip drops far-water saturation under the 25% line: fix trials t4-t6 gave 22.5, 24.3 and 24.9%) and it needs an owner or orchestrator waiver, or a ruling at the golden sheet. (2) *Hard-edged peach and blue patches in the far water at dawn, 06:00-06:12* (`de/review/re/h0600w.png`, `h0608w.png`): fixed, but the cause was not the one the finding guessed. Gating the water sky's toward-sun glow (trial A: sky_ref under `sky_golden_weight`; B: the same plus the haze's height from the frame alone; C: `sun_h` from the frame alone; and the finding's own suggestion, the away side only, `min(sun_facing, 0.0)`, run as fix-1 of this pass) left the patches as they were: crop spread luma 20-22, b-r 20-21 at 06:00 against 12/11 pre (`de/fix2/tspread.js`, `de/fix2/trial/A`-`D`). The cause was the soft read (P4c): `panel_excess` mirrors only what the panel adds over its painted sky, taken as `haze_color x 1.15`, but the panel's sky now carries the sun side, so the belt stood above that floor. The soft read then mirrored it, the painted city rows cut out of it and swayed by the swell: map-like patches. The fix: `panel_excess(..., float facing)` turns its floor the same way, `vec3 painted = sky_sun_side(sky_height(tap.y), facing, haze_color.rgb * 1.15 + sun_color.rgb * disc_halo, sun_color.rgb, low_sun, 0.0)`, and all five taps (the soft read's four and the columns') pass `panel_facing = sky_sun_facing(forward.xz, low_sun_direction)`, the fragment's flat view azimuth (the mirror ray's without the ripples). By day and at night (low_sun 0) the floor is exactly the old one. Trial D had this change as well as the away-only sky; the away-only change did nothing to the patches, so it was reverted to stay in scope (the sky_ref line is as after the review fix). (3) The stale `world_view.gd` comment ("the city silhouette on the panel") now says "the Earth's shadow on the panel". Backups `de/backup/pre-fix2/` (`world_view.gd` there already has the comment fix). Snapshot diff at the start: only this part's own files differ from `build/backup-feel/22-P4-de` (expected); nothing unexpected. 457: - **Recheck tests:** test-first `de/fix2/first2-test_water.log.out` (410 checks, 2 failures: the floor line and the taps' facing). After `de/fix2/r2-test_{water,sun,look,time}.log.out`: **test_water 406 -> 410**, sun 79, look 563, time 377, 0 failures. New: `panel_excess` takes `facing` and turns its floor; every tap passes `panel_facing` from `forward.xz`; a GDScript replica with the `at_hour` haze and sun at 06:00, 06:08, 06:12, 06:45, 19:30 and 20:45 (heights 0.05-0.60, facing -1 to 1): the old floor would mirror over 2 levels of the painted sun side at every one of those hours, the turned floor mirrors none of it. The first try, `de/fix2/first-test_water.log.out` (409/2), tested the away-only sky (reverted). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh de/fix2/suite` (17:48-17:54, AC, Low Power Mode 0, `de/fix2/suite-power.txt`): **40 suites, 20,081 checks, 0 failures**; `--verify` 2880x1800 exit 0 in 33 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549, 0 failures; 1440x900 not run (inherited stall). An earlier run of the away-only version is in `de/fix2/suite-superseded-r1/` (20,080, 0 failures). 458: - **Recheck changed expectations:** none. Every earlier check stands (the P4c floor check `haze_color.rgb * 1.15 + sun_color.rgb * disc_halo` and the `panel_excess(` count >= 6 still hold). Journey count deltas: **none** (draw calls and scene counts identical on all 22 views in all six sets, `de/fix2/metrics/counts.txt`; only `audio.listens` wobbles by 1 in the pre-tree too (quarter-2045, pre-2 6 against pre-1 5); `light_select_calls` wobbles by 1 only in after-1 on quarter-2045 and quarter-golden, never in the pre-tree, and the kit treats it as a run-length key). 459: - **Recheck captures** (`de/fix2/caps.sh`, `sit.sh`, 1600x1000, AC and Low Power Mode 0 on every set, `de/fix2/power.txt`): one sitting, 17:54-18:07, all first try: pre-1 (`de/pre-tree`), fix1-1 (`de/fix2/fix1-tree`: the live tree with the water shader before this pass), after-1 (live), then pre-2, fix1-2, after-2. There are 22 views: the fix pass's 20 plus h0608 and h0612 (the plan's golden views and the handover hours, for the P4 close). The sitting before the revert is in `de/fix2/cap-superseded-r1/`. Metrics `de/fix2/summary.js` -> `de/fix2/metrics/summary.txt`. Looked at: `de/fix2/look/h0600-`, `h0608-`, `h0612-pre-fix1-after-2x.png` (pre even blue; fix-1 with the map-like patches; after even again, a touch warmer than pre), `harbour-2045-pre-fix1-after-2x.png` (the faint dark mottles fix-1 had on the right are gone; magenta water as pre), `harbour-0645-` and `harbour-golden-pre-fix1-after-2x.png` (after equals fix-1), `trialD-h0600-h0608-2x.png`. 460: - **Recheck acceptance** (the reviewer's crop 1100,245,1600,330, luma / b-r p90-p10, pre -> fix-1 -> after, the same in both halves): **06:00 12/11 -> 22/21 -> 14/11; 06:08 15/12 -> 22/20 -> 17/13; 06:12 15/14 -> 19/19 -> 18/15; 05:55 11/15 -> 16/19 -> 12/16; 20:45 14/20 -> 12/23 -> 13/20.** Water >8 fix-1 -> after is 6.3/6.1% at 06:00 and 3.7/4.0% at 06:08, against noise 1.0-1.3. The golden lines are unchanged fix-1 -> after (harbour-golden water >8 1.75/1.11 against noise 1.01-1.92; orbit-g1 to g3, harbour-0645, glade-1930 and quarter-golden also within noise): **far-water saturation 25.0%** (#cfae9b, at the line), **city 23.6**, lawn 91.3 = 0.78x water_open 116.4, sky low band #cfc1e9 hue 261, belt hue 16, water strip y 236-248 #f1d3d5 hue 356, sat 12.5% (finding 1, open). Protected: harbour-0545 and quarter-0545 fix-1 -> after: one half of each view is within noise and the other is a little above it (1.57 against a 1.27 noise ceiling on harbour-0545, 2.68 against 2.04 on quarter-0545; the other halves 1.13 and 1.52 are inside), so "within noise" holds only for one half of each; day, night and glade-day within noise. 20:45 water changes by 3.1/3.2% (the mirrored belt excess, now off the floor), moving it back toward pre (b-r spread 23 -> 20, pre 20). 461: - **Recheck bench** (`de/fix2/bench.sh`, HARBOUR_ONLY, 19:30, fix1-tree against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, AC, Low Power Mode 0, `de/fix2/bench/power.txt`): raw p50 6.447/6.492/6.473 -> 6.465/6.466/6.487 (medians 6.473 -> 6.466; paired +0.02/-0.03/+0.01, live slower in 2/3); p95 medians 7.895 -> 7.906. No change beyond noise. Draw calls 814 on both. Dense not run (the change is in the water shader only, and Harbour is the water-heavy view). 462: - **Glade:** inherits (the plan's "Glade inherits it"): the sun-side sky, the fog and haze retune (Glade's haze is its shift of ecb898) and the water dissolve. The lawn lift is gated by Glade's `ground_calm` (0.15 of it): the whole lift turned Glade's meadow lime (83.6 -> 126.8 luma, `cap/t5`, `t6`); its daylight already keeps the meadow off black. Only the city glow stays Glade-exempt (P4-6). 463: - **Deviation from the plan, measured:** no city-silhouette term. Since P4a's gamma-space haze the far city already stands 19.9 luma under the sky above it at golden hour (target 15-25); the golden haze retune takes it to 22.9. The report's term (rows x(1 - 0.4 low_sun), city toward 5e5f7e) put it 44.9 under (`cap/t1`), a 0.08/0.12 version 31.1 (`cap/t2`); both over the band, so it was removed (`cap/t3-nosil` 22.9). The report's 1-px warm rim on block tops toward the sun was not built. 464: - **Files:** `shaders/sky.gdshaderinc`, `shaders/backdrop.gdshader`, `shaders/water.gdshader`, `shaders/ground.gdshader`, `scripts/lighting_profile.gd`, `scripts/world_view.gd`, `tests/test_water.gd`, `tests/test_time.gd`, `tests/test_sun.gd`, `tests/test_look.gd`, this entry. The recheck fix touched `shaders/water.gdshader`, `tests/test_water.gd`, `scripts/world_view.gd` (a comment) and this entry. Backups `artifacts/beautify/22-P4/de/backup/`. Snapshot `build/backup-feel/22-P4-de/` (= `main` a8a27a5); at part start the live tree equalled it on scripts, shaders, tests, data, scenes, tools and `project.godot` (empty `de/diff-live-vs-22-P4-de.txt`); nothing unexpected. Evidence `artifacts/beautify/22-P4/de/`; scratch tree `de/pre-tree` (kept for the review). 465: - **Tests:** test-first `de/testfirst-test_{water,time,sun,look}.log.out` (366/15, 377/2, 76/2, 557/2 checks/failures). After (`de/r4-*.log.out` and the suite): **test_water 356 -> 394, test_time 372 -> 377, test_sun 72 -> 76, test_look 555 -> 563, 0 failures** (golden_sky_checks: the function and its lines, constants, a GDScript replica at the golden anchor: exact at low_sun 0 and side-on, unchanged at 05:45, the belt turns pink (hue toward 341, step >= 0.03), the horizon cooler away, a warm glow toward the sun that fades up, the backdrop's column azimuth and order, the water's mirror-ray facing, band height, sky and dissolve lines, no camera_yaw; golden_water_checks: the weight whole at 19:30 and 06:45, 0 at 01:00, 05:45, 13:00, 20:45, 22:00, the dissolve and reach, the city rows untouched; golden_retune_checks: fog 0.62, haze ecb898, the rest of the anchor and 06:45 unchanged, night and golden fog under the day's; low_sun_checks: the five materials take `SunPath.golden(hour, heading)` at 9 hours, 2 headings, both looks, 0 at 13:00 and 01:00, the heading turns the side, sync_profile keeps it current; lawn_lift_checks: the uniform, the line in the meadow branch only, none in the bank, no lift by day, night, 05:45 or 20:45, x1.45 at golden, smooth). `test_time.gd:176` ambient assert untouched. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/de/suite` (a/b/v, 15:50-15:57, after the t8 trial was reverted, AC, Low Power Mode 0, `de/suite-power.txt`): **40 suites, 20,062 checks, 0 failures** (P4 owner fixes 20,005; the changes named here sum to +55, the other +2 is test_weathering 505 -> 507, which loops over the `const float` lines of ground.gdshader and picked up the new LAWN_ constants); `--verify` 2880x1800 exit 0 in 33 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549, 0 failures (the known flaky check did not fire); 1440x900 not run (inherited stall). 466: - **Changed expectations:** test_water P4c checks "the columns are added after the horizon haze" and "the soft read is added ... before the haze" now anchor on the new dissolve line `color = mix(color, sun_haze, smoothstep(mix(0.50, WATER_GOLDEN_REACH, golden_w), ...) * mix(0.55, WATER_GOLDEN_DISSOLVE, golden_w));` (old: `color = mix(color, haze_color.rgb, smoothstep(0.50, SKY_HORIZON + 0.004, screen.y) * 0.55);`); same checks. No other expectation changed. Journey count deltas: **none** (integer counts and draw calls identical on all 16 views, `de/metrics/counts-1.txt`, `counts-2.txt`); the six journey views (day and night) are unchanged within noise; pixel changes only where the sun is low (golden, 06:45, 20:45, 21:00). 467: - **Captures** (`de/caps.sh`, `de/sitting.sh`, 1600x1000, AC and Low Power Mode 0 on every set, `de/cap-power.txt`): final sitting 15:41-15:49 pre-1 (`de/pre-tree`), after-1 (live), pre-2, after-2, 16 views each, all first try: harbour-golden, quarter-golden, photo-golden, the 4-yaw golden orbit (harbour-golden at yaw 0.78 plus `orbit-g1` 2.35, `orbit-g2` 3.92 toward the sun, `orbit-g3` 5.49), grove-1930, glade-1930, harbour-0645, harbour-2045, harbour-0545, quarter-0545, quarter-2045, harbour-day, harbour-night, glade-day (for the P4 golden sheet). Tuning rounds `cap/t1`-`t8` (t8 an Earth-shadow trial, reverted), `cap/ref-pre`; a first sitting before the 05:45 fix is in `cap/superseded-sit1/`. Metrics `de/summary.js` -> `metrics/summary.txt`, `de/city.js` -> `metrics/city.txt`, `de/reg.js`, `de/crop.js`. Looked at: `look/pre-1-contact.png`, `after-1-contact.png`, `final-hg-horizon-3x.png` (a pink band over the city, the far water peach with the belt mirrored, its rain rings dissolved near the horizon), `final-2045-0645-pre-after.png` (20:45 a wider pink band, same city; 06:45 a warmer glow round the sun, otherwise as before), `final-lawn-pre-after-2x.png` (the meadow a lighter green, bank unchanged), `t5-0645-g2-2x.png` and `t6-0645-g2.png` (the white-out toward the sun that made the dissolve away-side only). 468: - **Acceptance** (harbour-golden unless named; pre-1/pre-2 -> after-1/after-2): 469: - **Pink band over a cool band away from the sun, also in the water** (as first built; see the review fix above for the cool band): pink band met: sky y 135-160 hue 25.7 -> 18.6, sat 23.3 -> 27.6%, luma 214 -> 204 (both sets); mirrored in the far water: y 335-355 #aaa8a8 (sat 1%) -> #dfb4a5 (hue 15, sat 26%), y 236-256 hue 30 -> 22.5. **Cool band not met:** the sky just over the rows stays warm (y 184-196, x 1150-1500: hue 21.8 -> 20.2, re-measured by the stage-close handoff check; the first figure, 27.6 -> 26.6, was not saved in any metrics file and does not reproduce); the Earth's shadow lies under 0.15 of the panel, behind the far rows and under the mist, and is 60% slate over a warm horizon; a taller, stronger band (0.20, 0.50; `cap/t8`) did not read either and cost far-water saturation, so it was reverted. 470: - **Warm glow orbiting toward the sun:** met: orbit-g2 sky y 135-160 luma 213.1 -> 220.8 (hue 28 -> 29); harbour-0645 (camera faces the sun) 225.0 -> 232.5. 471: - **Day and night unchanged within noise:** met. frame >8 luma pre-after (two halves) against pre-pre / after-after: harbour-day 1.77 / 0.93 against 5.12 / 4.46; harbour-night 1.46 / 4.60 against 4.23 / 1.80; glade-day 0.47 / 0.68 against 0.58 / 1.16; region p50s identical (`metrics/summary.txt`). 472: - **City 15-25 luma below the sky above it:** met, **22.9** (both sets; pre 19.9) by `de/city.js` (per column of x 1000-1590: 5 rows under the block top against 5 rows over it). quarter-golden 32.6 -> 33.6 (outside the band before and after; not the plan's view for this line), photo-golden 21.0 -> 20.4. 473: - **Far-water saturation at least 25%:** met, far_bay (x 1000-1500, y 235-420) mean colour **9.7% -> 26.2%** (#a49b94 -> #cbaa96), both sets. 474: - **Golden lawn at least 0.7x water:** met, lawn_report p50 **54.5 -> 91.3** against water_open 116.3 = **0.79x** (was 0.47x), both sets. 475: - **Protected 05:45 and 20:45 skies:** 05:45 within noise (harbour-0545 sky >8 1.45 / 1.13 against 1.41 / 1.07; quarter-0545 2.79 / 3.16 against 2.91 / 3.27; the first sitting's 7.8% at 05:45 was fixed by the smoothstep range). 20:45 changes by the intended belt only: sky >8 9.4 / 9.7% (quarter-2045 6.8 / 6.7) against 1.5 / 1.2; band hue 351 -> 354, sat 31.6 -> 34.9; city 47.2 -> 46.4 under the sky, lawn 32.6 unchanged (the lift is gone by 20:45). 06:45: sky >8 14.8% (the sun-side glow), city, water_open and lawn read as before (lawn 19.1 -> 32.9, the lift). 476: - **Bench** (`de/bench.sh`, `de/bench-summary.js` -> `de/bench/summary.txt`; pre-tree against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, AC, Low Power Mode 0, `de/bench/power.txt`): raw p50 medians, 19:30: dense 15.94 -> 15.97 (paired +0.04, 2/3 slower), Harbour 7.54 -> 7.47 (-0.10, 1/3); 01:00: dense 14.25 -> 14.26 (-0.00, 1/3), Harbour 7.54 -> 7.65 (+0.05, 2/3): no change beyond noise. Draw calls 3124 / 814 (19:30) and 2360 / 899 (01:00) both sides. 477: - **For the owner sheet (golden sheet at the P4 close):** the far water's rain rings and ripples dissolve into a smooth peach band near the horizon at golden hour (the price of 26% saturation; `final-hg-horizon-3x.png`); no cool Earth-shadow band; 06:45 sky a little warmer round the sun; 20:45's pink band wider; Glade's golden water and sky follow canal's. 478: - **Review:** filled by the orchestrator. 479: - **Not done / open:** the cool band in the water is only a faint mauve strip (a stronger one costs the 25% far-water saturation line). The recheck left this open for an owner or orchestrator waiver, or a ruling at the golden sheet; the recheck fix did not change it. Put h0600 and h0608 (`de/fix2/cap/after-1/`) on the P4 golden sheet to confirm the dawn water. The cool band in the sky is met after the review fix. Glade takes the cool band too (it inherits the sun-side sky); the report's city silhouette term and warm rim (measured unnecessary / not built); puddle.gdshader still mirrors the untinted sky (open since P4c). Stop rule not hit. 480: - **Platforms:** Mac run; Windows not run. 482: **Beautify P4 owner fixes: 3D ring deleted, quay lamps 15% dimmer, Glade without neon-bay columns (30 September 2026):** Opus 5.5 on this Mac, one fix pass after P4c for the owner's answers after P4a and P4b (`artifacts/beautify/owner-decisions.md`, P4-1 and P4-4) and the orchestrator's ruling on Glade (with owner P4-6). This one entry carries the evidence (clearer than spreading it over three); the P4a, P4b and P4c entries below each got a one-line pointer to it. 483: - **Asked and changed:** (1) **P4-1, delete the 3D skyline ring:** `world_view.gd` loses `_build_skyline`, `skyline_ring_boxes`, `RING_SEED`, the second rng stream `RING_SEED_DRESSING`, `RING_SETBACKS`, `RING_SPIRES`, `RING_NEON`, `RING_TONES`, the `skyline` member and the switch hook `if skyline != null: skyline.visible = ...` in `sync_environment`. The painted skyline, the Scenery skyline switch's effect on it (`skyline_enabled`) and `beacons_blink` are as they were; no save key ever held the ring. (2) **P4-4, far-quay lamps 15% dimmer:** `backdrop.gdshader` gains `QUAY_LAMP_DIM` 0.85: `quay_lamp = tint * lamp * QUAY_LAMP_GAIN * QUAY_LAMP_DIM * smoothstep(0.3, 1.0, night_amount)`. The band, its colour, the spacing, share, tints, `QUAY_LAMP_GAIN` 0.8, the deck, the tint and the columns' code are untouched (the columns read the lamps, so they dim with them). (3) **Glade takes no P4c columns:** `water.gdshader` gains `uniform float neon_columns = 1.0` and the columns' mask is `smoothstep(PANEL_NIGHT.x, PANEL_NIGHT.y, night_amount) * neon_columns`, so at 0 the squashed read is skipped; the soft read (P4c review fix) sits outside that mask and stays for every look. `WorldView.sync_profile` sets `neon_columns` from the look's RENDER `city_glow` (canal 1, Glade 0), beside the glow. 484: - **Files:** `scripts/world_view.gd`, `shaders/backdrop.gdshader`, `shaders/water.gdshader`, `tests/test_water.gd`, this entry and pointers in the P4a, P4b, P4c entries. Backups `artifacts/beautify/22-P4/ownerfix/backup/`. Snapshot `build/backup-feel/22-P4-ownerfix/` (= `main` 9a49b56); at the start the live tree equalled it on scripts, shaders, tests, data, scenes, tools and `project.godot` (empty `ownerfix/diff-live-vs-22-P4-ownerfix.txt`); nothing unexpected. Evidence `artifacts/beautify/22-P4/ownerfix/`; scratch trees `ownerfix/pre-tree` (the snapshot) and `ownerfix/i1-tree` (pre-tree plus the new `world_view.gd` only: the item-1 control). 485: - **Tests:** test-first `ownerfix/testfirst-test_water.log.out` (356 checks, 15 failures); after `ownerfix/r1-test_water.log.out` **test_water 347 -> 356, 0 failures**: `skyline_ring_checks` (13) replaced by `skyline_ring_gone_checks` (15: no builder, no `skyline` member, none of the seven names in `world_view.gd`, a pre-P4a scene `data/drag/cluster6.json` validates and loads with the skyline on and steady beacons, the switch hides and shows the painted panel); `far_quay_checks` +3 (`QUAY_LAMP_DIM` 0.85, the lamp line, gain still 0.8); `neon_bay_checks` +4 (the uniform, the mask line, the soft read does not read it, canal 1 and Glade 0 at 01:00, 20:45 and 22:00). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/ownerfix/suite` (a/b/v, 14:35-14:41, AC, Low Power Mode 0, `ownerfix/suite-power.txt`): **40 suites, 20,005 checks, 0 failures** (19,996 + 9); `--verify` 2880x1800 exit 0 in 33 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549 checks, 0 failures (the known flaky check did not fire); 1440x900 not run (inherited stall). 486: - **Changed expectations:** the 13 P4a ring checks in test_water ("every old box is where it was under seed 1776", "about 20% of the blocks step back", "about 5% carry a spire", "6-8 steady neon strips", "two depth tones", "the ring's neon is steady", "the list is the same every time", "the ring is not built in the running view", "built, the ring is one mesh", "the switch hides the ring", "and shows it again", the list exists, the first box is a block) -> deleted with the ring (owner P4-1), replaced by the ring-gone checks. No other expectation changed. Journey count deltas: **none** (draw calls identical); pixel changes: dimmer quay lamps and harbour columns at night and 20:45, no columns on Glade's night water. 487: - **Captures** (`ownerfix/caps.sh`, `sit1.sh`, `sit2.sh`, `sit3.sh`, 1600x1000, all first try, AC and Low Power Mode 0 on every set, `ownerfix/cap-power.txt`): one sitting 14:41-14:51: pre-1, after-1, p4c-1 (`c/pre-tree`, before P4c), pre-2, after-2 (harbour-night, quarter-night, glade-day, glade-night, harbour-day, harbour-golden, harbour-2045), p4c-2, pre-3/after-3, pre-4/after-4 (night views), then pre-5, i1-1, pre-6, i1-2 (harbour-night, quarter-night, glade-day). Metrics `ownerfix/metrics.sh`, `metrics3.sh`, `counts.sh`, `summary.js` -> `metrics/summary.txt`, `c/columns.js` -> `metrics/columns-*.txt`, `lamps.js` -> `metrics/lamps.txt`. Looked at: `ownerfix/look/hn-quay-pre-after-4x.png` (the quay lamps a little dimmer, the dark outline pixels beside them still there), `hn-bay-pre-after-2x.png` (the same harbour columns, slightly dimmer), `gn-bay-p4c-pre-after-2x.png` (Glade's bay before P4c with scattered point lights, with the P4c columns, and now plain water with the soft sky read). 488: - **Acceptance** (numbers against the orchestrator's targets): 489: - **Item 1, nothing visible changes (item-1 control against pre, same sitting):** harbour-night frame >8 luma 8.1 / 1.5% against noise pre-pre 7.9, i1-i1 1.8; quarter-night 4.4 / 2.4 against 4.3 / 2.0; glade-day 0.50 / 0.75 against 0.44 / 0.72; integer counts and draw calls identical (889, 969, 888): **met**. Full fix against pre: integer counts and draw calls identical on every view of all four pairs (`metrics/counts-*.txt`). 490: - **Item 2, lamp p99 against the town tower windows' p50** (lamp px: 40+ luma over the row median in the quay strip, free columns): harbour-night **215.1 -> 196.6** (all four sets each), quarter-night **220.4 -> 201.5**, harbour-2045 229.8 -> 215.0; quarter-night `windows_tower` p50 **122.2** both sides, so the lamps go from 1.76-1.80x to 1.61-1.65x the tower windows and stay the brightest horizon points. **Harbour-night columns 30-120 render px** (x 1240-1570, y 232-520): pre-1..4 14 / 14 / 14 / 14, after-1..4 **15 / 14 / 14 / 15**, longest 71, none over 120 (target 12-14, at least 10): met, one over the upper figure in two sets. Day unchanged: harbour-day frame >8 1.0 / 3.6% against noise 0.8 / 3.1, water_open p50 119.4 both. Golden: the lamps are half-lit at golden hour (smoothstep(0.3, 1, 0.66) = 0.52), so they dim too: the bright strip px p50 226.5 / 227.5 -> 223.1 (p99 246.4 unchanged); every region within noise (frame >8 1.4 / 2.9% against 1.1 / 2.5, skyline 0.29 / 0.19 against 0.10 / 0.00). harbour-2045 far_bay >8 4.3 / 3.1% against 1.5 / 1.5 (the dimmer half-strength columns and lamps). 491: - **Item 3, Glade without columns:** glade-night columns 30-120 px pre 4 -> **after 0** (before P4c 0). far_bay >8 pre -> after 5.7-5.9% (the columns gone; noise 0.46-0.65). Against the tree before P4c: **1.23 / 1.14%** against this sitting's pairs p4c-p4c 0.57 and after-after 0.60 / 0.65 (P4c measured 6.6-6.7% against 0.4-1.5%): inside P4c's noise range, about twice this sitting's pairs; the residual is the scattered point lights of the old point-for-point read (top strip of `gn-bay-p4c-pre-after-2x.png`), which the soft read drops in every look since the P4c review fix; far_bay p50 78.1 -> 78.1. Canal unchanged by it (the harbour-night count above). 492: - **Bench** (`ownerfix/bench.sh`, from `c/bench.sh`; pre-tree against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, 01:00, 14:51-14:55, AC, Low Power Mode 0, `ownerfix/bench/power.txt`; raw p50 per round in `ownerfix/bench/*/performance.json`): raw p50 medians dense 13.305 -> 13.281 (paired -0.005, 1/3 slower), Harbour 6.761 -> 6.744 (paired +0.006, 2/3): no change. Draw calls 2360 and 899 both sides. 13:00 not run (no change touches the day path). 493: - **For the owner sheet:** the 15% dim leaves the presentation outline's dark pixels beside the quay lamps (`hn-quay-pre-after-4x.png`); golden hour's half-lit lamps dim too. 494: - **Review:** filled by the orchestrator. 495: - **Not done / open:** nothing asked is left out. The puddle tint (P4c entry) is still open. 496: - **Platforms:** Mac run; Windows not run. 498: **Beautify P4c Neon bay: long broken columns on the water (30 September 2026):** Opus 5.5 on this Mac. At night the open bay now mirrors the far city's lights as long broken columns (report Signature B, water; lens "Neon bay", revised proposal). `water.gdshader`: where the mirror ray meets no geometry, the painted panel is read squashed toward the horizon, `SKY_HORIZON - below * PANEL_SQUASH` (0.12; was point for point, `below * 0.7`), faded over `PANEL_FADE` 0.10-0.32 of the frame below the horizon, gated at `PANEL_GATE` 0.33 (was 0.30 against a 0.30 fade). **Deviation from the lens, measured:** P4a left the far windows under the haze threshold, so the lights in the read zone (the bottom 17% of the panel) are 1-px lamps (the far quay's, the viaduct's, neon); squashed alone they mirror as 8-px dashes (`cap/t1`, `look/t1-hn-bay-2x.png`). Each read therefore takes the brightest of a run of `PANEL_RUN` 8 taps one render px apart toward the horizon (never past it, each tap depth-tested so a building in front reads nothing): a light at panel height p lights the rows from p/0.12 to (p+8)/0.12, a column 67 render px long whatever its height. Each pair of rows shivers by +-`PANEL_JITTER_PX` 1.5 render px on the mirror's own row hash, and a rain ring's push shivers the column it lands in (`+ ring.y`; P3b's rings now reach the columns); the swell's sway of the read is `PANEL_WOBBLE` 0.002 (0.008 as before made zigzags, `cap/t2`, `look/t2-hn-bay-2x.png`). Dashes drop out on `PANEL_DASH` 35% of 2-row x `PANEL_DASH_PX` 3-px cells, re-rolled `PANEL_DASH_RATE` 1.5 times a second (the lens wrote `sky_hash(...) + floor(TIME * 1.5)`, which is never below the step after the first second; the time is inside the hash). Only the city's lights: the read is night-only (`night_mask = smoothstep(PANEL_NIGHT 0.7, 1.0, night_amount)`, the windows: 0 by day, at golden hour (windows 0.66) and 06:45, 0.5 at 05:45 and 20:45, 1 at night), and the sun or moon is found where the backdrop draws it (same panel position with the yaw from the view matrix, same compressed height, same radius from the new `sun_size` uniform, same halo): its disc is cut out to `PANEL_DISC_CUT` 1.6 radii and its halo counts as painted sky (the floor is `haze * 1.15 + sun_color * halo * 0.85`; under the old `haze * 1.15` the night moon's halo passed out to several radii). The columns are added after the horizon haze (`PANEL_GAIN` 0.7 x mirror strength; was 0.8 before the haze), so they stay crisp to the waterline. `WorldView.sync_profile` sets `night_amount` = windows and `sun_size` = the panel's on the water. 499: - **Files:** `shaders/water.gdshader`, `scripts/world_view.gd`, `tests/test_water.gd`, this entry. Backups `artifacts/beautify/22-P4/c/backup/`. Snapshot `build/backup-feel/22-P4-c/`; at part start the live tree equalled it on scripts, shaders, tests, data, scenes, tools and `project.godot` (empty `c/diff-live-vs-22-P4-c.txt`); nothing unexpected. Evidence `artifacts/beautify/22-P4/c/`; scratch tree `c/pre-tree` (kept for the review). 500: - **Tests:** test-first `c/P4c-testfirst-test_water.log.out` (332 checks, 33 failures); after `c/r2-test_water.log.out` and the suite run on the final code (`c/suite/suites/test_water.log.out`) **test_water 295 -> 335, 0 failures** (`neon_bay_checks`: squash 0.10-0.14 and the old 0.7 gone; fade and gate, gate at or past the fade end; the read reaches 0.167 of the panel (quay, viaduct, front row, not the landmarks); the run: 4-12 taps giving 30-120 px, stepping to the horizon, depth-tested; jitter on the mirror's row hash; dash share, width, rate and the time inside the hash; `night_amount` uniform and the mask 0 at day, golden, 06:45 and 09:00, about half at 05:45 and 20:45, 1 at night and 22:00; `sun_size`, the backdrop's disc height, radius and halo; the disc cut; the halo floor by arithmetic (no painted halo px passes at 01:00, 05:45, 20:45, 22:00 over 1.2-16 radii; the old floor let some through); no `camera_yaw`; the add after the haze, the old add gone; gain; the depth test; the ring push; the view sets both uniforms at 7 hours). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/c/suite` (a/b/v, 13:23-13:29, AC, Low Power Mode 0, `c/suite-power.txt`): **40 suites, 19,984 checks, 0 failures** (P4b 19,944 + 40); `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 5,549 checks, 0 failures, the known flaky check did not fire. 1440x900 not run (inherited stall). 501: - **Changed expectations:** none existing. Mine, refined while tuning: "a light one render px tall mirrors as a 1/PANEL_SQUASH dash" stays, and the column length moved to the new run checks ("PANEL_RUN / PANEL_SQUASH in 30-120"). 502: - **Captures** (`c/caps.sh`, one sitting 13:03-13:09 in the order pre-1 (scratch tree), after-1, pre-2, after-2, 16 views each, all first try; `c/cap-power.txt` AC, Low Power Mode 0 (three entries mis-named by a quoting slip captured nothing, noted there)): harbour-night, quarter-night, quarter-dry, lane-night, harbour-golden, photo-golden, harbour-day, glade-day, glade-night, quarter-0545, quarter-2045, harbour-0645, harbour-0545, harbour-2045, orbit-moon (yaw 2.35, night, the moon in view), orbit-b (yaw 3.92). Phase spread (`c/qn-extra.sh`): quarter-night and harbour-night x4 more from each tree, alternating (`cap/pre-3`..`6`, `after-3`..`6`). Tuning rounds `cap/t1`-`t3`. Looked at: `look/t1-hn-bay-2x.png`, `t2-hn-bay-2x.png`, `t3-hn-bay-2x.png`, `hn-bay-pre-after-2x.png`, `t2-contact.png`, `sheet-sun.png`, `sheet-other.png`, `sun-0645-2045-moon-2x.png`, `qn-mirror-cols-4sets.png`. They show the bay under the skyline striped by thin broken warm columns (one per quay lamp) and an occasional cyan one on harbour-night; columns in the open water behind the quarter on quarter-night; half-strength columns at 05:45 and 20:45; golden, 06:45 and day unchanged; toward the moon no moon column; on the orbit views a dense cyan comb where the viaduct's lamp row is in the read. 503: - Counts: integer counts and draw calls identical pre -> after on all 16 views in both sets (`c/metrics/counts-*.txt`). 504: - Noise: within-tree pairs differ by phase (rings, wobble, the P4b deck's dither): e.g. quarter-night mirror columns >8 luma 1.3-26.5% pre-pre, 3.5-39.1% after-after over 15 pairs each (`c/pairs.js` -> `metrics/pairs-spread.txt`); single pairs (`metrics/noise-*.txt`, `ab-*`) are unreliable for "unchanged" claims, so the six-set spread is used for them. 505: - **Acceptance** (`c/columns.js` -> `metrics/columns-harbour-night.txt`, `c/sunmoon.js` -> `metrics/sun-moon-water.txt`, `c/summary.js` -> `metrics/summary.txt`, `pairs-spread.txt`): 506: - **Broken columns 30-120 render px long across the bay on harbour-night:** met. Open bay x 1240-1570, y 232-520 (a column: lamp-coloured px, chroma >= 40 and 15+ over the row median, in 4-render-px bins, gaps up to 12 render px): after-1..6 **13-15 columns 30-120 px** (35-44% of the fragments), longest 71-77, none over 120, lit share per column 0.52-0.61 (broken); pre-1..3 **0** (longest 18-21). far_bay >8 luma 23.3/24.6% against noise 2.7/9.6; the cross-tree spread 8.1-10.4% against within-tree 0.8-7.4. 507: - **No second sun or moon column at golden, 06:45 or 20:45:** met. Under the 06:45 sun (x 1230-1330, y 236-500) p50/p90/p99 165.3/197.1/210.5 pre-1 -> 165.3/197.1/210.6 after-1, column px 6.74/7.06 pre -> 6.94/6.88 after; harbour-golden near-horizon water p99 205.7/206.1 -> 205.6/205.2; photo-golden mid water p50 56.3/56.5 -> 57.0/56.4. 20:45 column px 8.33/8.36 -> 9.20/9.13 (the half-strength lamp columns; the sun is under the horizon). Night toward the moon (orbit-moon, x 985-1065): p50/p90 31.0/38.6 in all four, column px 4.38/4.42 -> 4.40/4.38. Not captured: a moon low enough to sit in the read zone (the disc cut is covered by the arithmetic checks only). 508: - **Open-water p50 by day and in glade-day within 1 luma:** met, harbour-day water_open 119.4 -> 119.4 (both sets), glade-day 114.6 -> 114.6. 509: - **Quarter-night mirror columns unchanged within noise:** met on the spread: pre-after 1.37-39.2% >8 luma (median 10.6) against pre-pre 1.32-26.5 (14.1) and after-after 3.5-39.1 (9.7); p50 44.2/44.1 -> 44.0/44.5 against noise 44.2/44.1 and 44.0/44.5; the columns are mirror hits, which the change does not touch. 510: - Protected looks: P3 rim (harbour-night waterline p50 54 -> 54, >8 0.00/1.41 against noise 0.39/1.95), HDR neon cores (fish sign >8 0.21/2.13 against 0.51/1.82; bar sign 8.25/3.30 against 0.00/11.56; tower windows 0), skies (the water discards above the horizon) within noise. 511: - **Bench** (`c/bench.sh`, `c/bench-summary.js` -> `c/bench/summary.txt`; `c/pre-tree` against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, 13:16-13:23, AC, Low Power Mode 0, `c/bench/power.txt`): raw p50 medians, night: dense 13.27 -> 13.25 (paired -0.02, 0/3 slower), **Harbour 6.68 -> 6.71 (+0.04, 3/3)**, the eight taps on open night water; 13:00 (no read by day): dense 13.15 -> 13.12 (-0.04, 1/3), Harbour 6.65 -> 6.65 (-0.00, 1/3). Draw calls 2360 and 899 both sides. 512: - **Journey:** expected count deltas **none**; pixel changes on the open night water of harbour-night and quarter-night (columns), none on the day views. 513: - **For the owner sheet:** the run of taps (a deviation from the lens's squash-only read, above); ~~the faint mauve reflection of the P4b deck's lit undersides near the horizon is gone (the read no longer reaches them; hn_horizon p50 53.0 -> 51.9, far_bay 52.2 -> 50.9, `look/hn-bay-pre-after-2x.png`)~~ (superseded: the review fix below brings it back, with the twilight cloud wash this entry did not report); the viaduct's dense lamp row makes a cyan comb of columns on some orbits (`look/sheet-sun.png` orbit-moon); ~~Glade keeps the columns (faint grey-blue, glade-night far_bay >8 6.6% against 1.5) since its panel still shows the rows until P5d~~ (**P4 owner fixes:** Glade takes no columns, gated by RENDER `city_glow` through `neon_columns`; glade-night 30-120 px columns 4 -> 0); the columns are 1 render px wide; P4f (night ripples) is at the owner's default D19: no. 514: - **Review fix (30 September 2026, Opus 5.5 on this Mac; evidence `artifacts/beautify/22-P4/c/fix/`):** the review found that the squashed read reaches only 0.167 of the panel and is night-masked (0.5 at 05:45 and 20:45), so it also removed the **twilight cloud reflection**, the mauve-pink wash across the upper Harbour water on the P3-signed-off 05:45 anchor and at 20:45, and this entry reported only "half-strength columns" there. Reviewed P4c against pre (first sitting, `c/metrics/summary.txt`): harbour-0545 far_bay p50 75.8 -> 73.1 with >8 luma on 35.1 / 35.3% (noise 2.5 / 5.4), hn_horizon 78.3 -> 76.8 on 31.6%; harbour-2045 far_bay >8 13.9 / 15.2% (noise at most 5.1), p50 103.6 -> 103.0 (`c/review/h0545-p3close-vs-pre.png`, `h2045-pre-after.png`). Fixed in `water.gdshader` and `test_water.gd` only (backups `fix/backup/`; at the start the live tree differed from `build/backup-feel/22-P4-c/` in exactly the three P4c files, `fix/diff-live-vs-22-P4-c.txt`; nothing unexpected), taking the review's alternative: a second, **soft read** keeps the painted sky and clouds: the panel read point for point as before P4c (`PANEL_SOFT_SQUASH` 0.7, faded over `PANEL_SOFT_FADE` 0.08-0.30, the swell's 0.008 sway, no run, jitter or dashes), at every hour (not night-masked), added right after the mirror mix and before the haze at `PANEL_SOFT_GAIN` 0.8 x mirror strength, exactly where and how the old read was. The lights stay the columns' alone: the soft read takes the least of four taps `PANEL_SOFT_SPAN` 1 render px left, right, above and below (one depth test at the centre), so a lamp, window or neon strip under 3 px in either direction drops out, and it stands above the same floor as the columns (a new `panel_excess()`: haze x 1.15 plus the disc's halo, the disc cut to 1.6 radii), so no sun or moon enters it. A first version with a 2-px span eroded the cloud shapes (05:45 far_bay p50 74.8 / 75.0 against pre 75.8; superseded sitting `fix/cap/superseded-span2/`, tuning `fix/cap/t1-span1`, `fix/look/t1-h0545-bay-2x.png`, `t1-hn-bay-2x.png`). The far quay's lamps, which the old read mirrored as dots just under the quay, stay out (the columns carry them). 515: - Tests: test-first `fix/testfirst-test_water.log.out` (347 checks, 12 failures); `fix/r1` found a test lookup matching the declaration (fixed, test only); after `fix/r3-test_water.log.out` and the suite **test_water 335 -> 347, 0 failures** (new: soft squash 0.7 and fade 0.08-0.30 as before, reach 0.91 of the panel, 0.008 sway, gain 0.8, added after the mirror mix and before the haze, outside the night mask, span 1-3, least of four taps, one floor function for both reads, and a tap sweep: 1x1, 1x4, 3x1, 4x2 px lights drop out, 12x8 and 40x20 clouds pass). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/c/fix/suite` (14:14-14:20, AC, Low Power Mode 0, `fix/suite-power.txt`): **40 suites, 19,996 checks, 0 failures** (19,984 + 12); `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549, 0 failures (the known flaky check did not fire); 1440x900 not run (inherited stall). 516: - Changed expectations: test_water "the read runs SKY_HORIZON - below x PANEL_SQUASH (the point-for-point 0.7 is gone)" -> "the columns' read runs ... (no point-for-point 0.7 in it)" (same check; the soft read uses the named constant). Journey count deltas: **none**; pixel change back toward pre on the open water of the twilight and night Harbour views. 517: - Captures (`fix/caps.sh`, one sitting 13:57-14:07, sets pre-1 (`c/pre-tree`), fix-1 (live), rev-1 (the reviewed P4c, scratch tree `fix/rev-tree` = pre-tree + the reviewed `water.gdshader` and `world_view.gd`), pre-2, fix-2, rev-2, the part's 16 views each, all first try (the superseded span-2 sitting had one retry); `fix/nights.sh` pre-3, fix-3, pre-4, fix-4 (harbour-night and quarter-night) at 14:10-14:12; AC, Low Power Mode 0 on every set, `fix/cap-power.txt`). **pre-1's harbour-night and quarter-night are framed off** (camera higher, `fix/look/pre1-hn-framing.png`; draw calls 913 vs 889), so those two use pre-2..4. Metrics `fix/metrics.sh`, `fix/summary.js` -> `fix/metrics/summary.txt`, `fix/tab.sh`. Looked at: `fix/look/h0545-p3close-pre-rev-fix.png` (P3 close, pre, reviewed, fix), `h0545-bay-pre-rev-fix-2x.png`, `h2045-pre-rev-fix.png`, `hn-bay-pre-rev-fix-2x.png`: the mauve cloud wash is back across the upper bay at 05:45 and the pink one at 20:45, with the half-strength columns over it; at night the faint mauve deck patches are back under the full columns. 518: - **The wash (set 1 / set 2):** harbour-0545 far_bay p50 pre 75.8 -> reviewed 73.2 -> **fix 75.3 / 75.3**, >8 pre->fix 8.8 / 10.2% (reviewed 35.4%) against noise pre-pre 4.7, fix-fix 2.6; hn_horizon 78.5 -> 76.8 -> **77.9**, >8 7.1 / 8.0% (reviewed 31.8) against 3.2 / 1.9. The pre->fix share includes the intended half-strength columns; **a residual 0.5-0.6 luma p50 under pre remains** (the 1-px erosion of the cloud edges), for the night sheet. harbour-2045 far_bay 103.8 / 104.0 -> **104.0 / 104.2** (reviewed 103.1), >8 6.0 / 6.8% against noise 2.6 / 4.3; hn_horizon 112.1 -> **112.5** (reviewed 111.9). harbour-night (pre-2..4 against fix-2..4) far_bay 52.2 -> **52.0** (reviewed 50.9), hn_horizon 53.0 -> **52.9** (51.9), >8 12.0-13.7% against noise pre-pre 7.6-8.3 and fix-fix 3.0-4.0 (the columns). 519: - The part's acceptance re-measured on the fix: **columns** (`c/columns.js` -> `fix/metrics/columns-harbour-night.txt`): fix-1..4 **12-14 columns 30-120 px** (30-34% of fragments), longest 72-92, none over 120, lit share 0.52-0.63; reviewed 12-13; pre-2..4 0-1 (longest 22-32): met. **No second sun or moon column** (`fix/sunmoon.js` -> `fix/metrics/sun-moon-water.txt`): under the 06:45 sun p50/p90/p99 165.6/197.3/207.6 pre -> 165.6/197.2-197.4/207.6-208.3 fix, column px 6.82-6.95 -> 6.93-7.22; harbour-golden near-horizon p99 205.2 -> 205.4; photo-golden mid water p50 56.3/56.4 -> 56.4/56.4; 20:45 column px 8.50 -> 9.30/9.42 (reviewed 9.29/9.33: the half-strength lamp columns); toward the moon p50 31.0-31.1 in all six, column px 4.69/4.80 -> 4.83/4.41: met. **Open-water p50 by day and glade-day:** harbour-day 119.4 -> 119.4, glade-day 114.6 -> 114.6: met. **Quarter-night mirror columns within noise:** pre->fix >8 3.5 / 15.3 / 3.9% (pre-2..4) against pre-pre 22.9 / 24.1 and fix-fix 10.6 / 12.6, p50 44.4 / 44.6 / 44.2 -> 44.1 / 44.2 / 44.0 against noise swings of 0.2-0.4: met. Protected looks: HDR neon cores (quarter-night bar_sign 0.0 / 0.0 / 5.5% against pre-pre 0.0 / 1.6, fix-fix 0.0 / 3.9; windows_tower p50 121.2 throughout), P3 rim (harbour waterline p50 54 -> 54 at night, 50.5 at 05:45, >8 at most 1.8%). Integer counts and draw calls identical pre-2 -> fix-2 and rev-1 -> fix-1 on all 16 views (`fix/metrics/counts-*.txt`; pre-1 -> fix-1 identical except the mis-framed harbour-night). 520: - Bench (`fix/bench.sh`, `fix/bench-summary.js` -> `fix/bench/summary.txt`; `c/pre-tree` against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, 14:20-14:28, AC, Low Power Mode 0, `fix/bench/power.txt`): raw p50 medians, night: dense 12.85 -> 12.88 (paired -0.00, 1/3 slower), Harbour 6.70 -> 6.75 (**+0.06, 3/3**; the reviewed P4c measured +0.04); 13:00: dense 12.82 -> 12.81 (-0.01, 1/3), Harbour 6.68 -> 6.69 (+0.01, 2/3; the soft read's four taps by day, which pre had as one). Draw calls 2360 and 899 both sides. 521: - For the night sheet (owner): the 05:45 wash is back but about 0.5 luma under pre on the far bay's median; the old read's point reflections of the quay lamps stay replaced by the columns. 522: - **Review:** filled by the orchestrator. 523: - **Not done / open:** puddles still mirror the untinted `sky_gradient` (P4b review finding): `puddle.gdshader` is not in this part's file list and the change would alter the quarter's night paving outside the columns' A/B, so it is left for a follow-up (one line: `sky_glow_tint` with the water's `glow_amount` and `city_glow`, set in `sync_profile`). No capture with a low moon in the read zone. `c/pre-tree` kept for the review. 524: - **Platforms:** Mac run; Windows not run. 526: **Beautify P4b Violet cloud deck lit from below; lit far quay (30 September 2026):** Opus 5.5 on this Mac. The city now lights the night sky by hue, not by an added glow (report Signature B, lens "City glow", revised proposal). `sky.gdshaderinc` gains `sky_glow_tint(height, base, glow, amount)`: the sky leans toward `city_glow` scaled by `sky_luma_match` to the sky's own luma on gamma-encoded values, by `GLOW_TINT` 1.0 at the horizon, fading out between `GLOW_REACH` 0.20 and 0.80 of the panel (the zenith keeps its colour). The backdrop applies it right after `sky_gradient`; the water applies it to `sky_ref` (same function). `WorldView.sync_profile` sets the derived uniforms on both: `glow_amount = smoothstep(0.7, 1.0, windows)` (0 by day and golden hour, 0.5 at 05:45 and 20:45 = the half tint, owner default D20, 1 at night) times the look's new RENDER key `city_glow`, and `city_glow = horizon.lerp(8a5a9c, 0.5)` (`CITY_GLOW`, `CITY_GLOW_SHARE`, `GLOW_WINDOWS`, `glow_amount_of`, `city_glow_of`); no profile key, every anchor exact. Clouds (`backdrop.gdshader`): undersides take the glow's hue at the cloud's luma (`CLOUD_UNDERLIGHT` 0.55 up to `UNDERLIGHT_REACH` 0.60); at night only (`night_deck = smoothstep(0.5, 1.0, glow_amount)`, 0 at twilight) the near deck's threshold is `0.50 - rain * mix(0.16, DECK_RAIN 0.46, night_deck)` and the far deck's `0.46 - rain * mix(0.14, DECK_FAR_RAIN 0.30, night_deck)`, the deck is drawn as a dark mass with lit undersides (review fix, below; it replaced a sink toward the zenith colour, `DECK_TOP_DARK`, that made the deck vanish), and the three band steps are dithered on a centred 2x2 ordered pattern (`sky_bayer`) laid on each layer's own panel pixels, so it scrolls and drifts with its layer. The far quay: after the mist, behind the skyline switch, a band `QUAY_PX` 2-4 render px tall stepping per `QUAY_BLOCK_PX` 40, colour `sky_mix_gamma(city_color * 0.7, haze, 0.35)`, scrolled with the front row; a 1-px lamp every `QUAY_LAMP_PX` 13 on its top row, `QUAY_LAMP_SHARE` 60% lit, warm (ffe6a0/ffb860), 7% cyan and 7% pink, `QUAY_LAMP_GAIN` 0.8 x smoothstep(0.3, 1.0, windows), added last (**P4 owner fixes, owner P4-4:** times `QUAY_LAMP_DIM` 0.85; harbour-night lamp p99 215.1 -> 196.6). By day and at golden hour the sky and clouds draw exactly as before; the quay band shows at every hour. 527: - **Glade (decision here, reversible):** RENDER `city_glow` is canal 1.0, Glade 0.0, so Glade's night keeps its bluer-green sky (owner G1 verdict Q2: not canal's violet); a first version tinted glade-night's band 216 -> 236 hue (`cap/sitting1-superseded/`). The report says "Glade behaves the same": owner question. The quay shows in Glade with the rows (until P5d, which the owner dropped on 1 October 2026). 528: - **Review fix (30 September 2026, Opus 5.5 on this Mac; evidence `artifacts/beautify/22-P4/b/fix/`):** the review found the heavier night deck heavier only in the shader's mask: sunk toward the zenith colour and luma-matched to the tinted sky, it merged into the sky (`review/visible.txt`: harbour-night cloud px >= 5 luma off their row's clear sky 18.4 -> 7.7% of the panel, quarter-night 18.7 -> 1.5%). Fixed in `backdrop.gdshader` (and its tests in `test_water.gd`; backups `fix/backup/`) only, all of it inside `if (night_deck > 0.0)` (day and twilight take the old code path exactly): each near-deck pixel is body or underside by the deck's slope (`near_n - below_n` against the noise `DECK_UNDER_PX` 4 render px lower, one extra noise sample, at night only); the bar drops by up to `DECK_UNDER_SLOPE.x` -0.60 toward the horizon (fading out by `DECK_LIT_REACH` 0.05-0.60), so the low deck is lit from below nearly all over and higher up the masses are dark with lit bottoms. Body = the tinted clear sky (`sky_clear`, taken right after the tint) x `DECK_SHADE_LOW` 0.70 low to `DECK_SHADE` 0.50 high (an indigo mass under the sky's luma), moonlit near the moon as before; underside = `sky_luma_match(city_glow, sky_clear)` x `DECK_UNDER` 1.30 low to `DECK_UNDER_HIGH` 1.35 high (violet, lighter than the sky; superseded by the recheck fix below: 1.15 up to 0.35 of the panel). Night step opacity `DECK_EDGE` 0.45 / midway / `DECK_OPACITY` 0.90 (by day 0.62 x step), so the thinnest step stands off the sky. Far layer: heavier (`DECK_FAR_RAIN` 0.30), `DECK_FAR` 0.80 of that opacity, body `DECK_FAR_HAZE` 0.35 toward the sky, lit as an underside below `DECK_FAR_LIT` 0.20-0.40 of the panel through a 2x2 ordered dither on its own pixels; the night band reaches `DECK_BAND_TOP` 0.92 (0.82 by day) before the top fade. Tuning rounds `fix/cap/u1`-`u13` (`fix/tune.sh`, a fresh probe each round); `u9` tried step-edge undersides (band 54-61, dropped); a superseded final sitting of an intermediate version is in `fix/cap/superseded-u8/`. Tests: test-first `fix/testfirst-test_water.log.out` (first draft, 288 checks, 9 failures) and `fix/testfirst-final-test_water.log.out` (the final checks against the reviewed backdrop in the scratch tree `fix/reviewed-tree`: 290 checks, 13 failures); after `fix/r3-test_water.log.out` **277 -> 292, 0 failures**. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/b/fix/suite` (12:03-12:09, AC, Low Power Mode 0, `fix/suite-power.txt`): **40 suites, 19,941 checks, 0 failures**; `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549, 0 failures (the known flaky check did not fire); 1440x900 not run (inherited stall). 529: - Changed expectation: test_water "at night the deck's tops sink toward the zenith" -> "the deck no longer sinks toward the zenith colour"; new: coloured from `sky_clear`; body and underside constants; underside judged `DECK_UNDER_PX` lower with a slope bar lowered toward the horizon; night block skipped by day and twilight; day and twilight mixed exactly as before; far layer heavier at night only; night band top; far layer lit low down by a dither; step opacities; at the night preset, every height 0.12-0.9, the thinnest body step >= 3 luma under the clear sky, the thinnest underside step >= 3 over it, the far step >= 2 under it; undersides violet (hue 250-290). 530: - Captures (`fix/caps.sh`, one sitting 11:53-12:02, AC, Low Power Mode 0, `fix/cap-power.txt`): `fix/cap/pre-1`, `fix-1`, `rev-1` (the reviewed P4b, `fix/reviewed-tree`), `pre-2`, `fix-2`, `rev-2`, 13 views each (quarter-night, harbour-night, quarter-night-r35, lane-night, glade-night, orbit-a/-b2/-c, beached-night, quarter-2045, harbour-0545, harbour-day, quarter-day-r35), all first try; live probe `probe/fix-final`. Looked at: `fix/look/night-panel-2x.png` (harbour pre / reviewed / fix), `quarter-r35-panel-2x.png`, `lane-orbitc-beached.png`, `u13-panel-2x.png`, `u8-panel-2x.png`, `u3-panel-2x.png`. They show dark indigo masses up high with lit violet bottoms, and a violet lit layer low over the skyline, where the reviewed version showed a smooth gradient; the dither edges are plainer now that the deck has contrast. 531: - **Cloud coverage on visible pixels** (`fix/visible.js` = `review/visible.js` on any set -> `fix/metrics/visible.txt`; set 1 / set 2): harbour-night pre 26.4 / 26.4 (>= 3 luma), reviewed 15.2 / 15.3, **fix 42.8 / 42.8** (>= 5: 18.4 -> 7.8 -> **41.1 / 41.1**; >= 8: 13.4 -> 3.3 -> 32.3), mask 44.3%; target over 40%: **met on harbour-night**. quarter-night-r35: pre 26.9, reviewed 10.0, fix **35.3 / 35.3** (>= 5: 34.6), mask 67.3%: under 40 on this metric, because rows with fewer than 20 clear px are skipped (31% of the panel is cloud in such rows). With the clear-sky level taken from the nearest row that has one (`fix/visible-fb.js` -> `visible-fallback.txt`): quarter-night-r35 44.7 -> 25.5 -> **66.0** (>= 5: 32.1 -> 13.4 -> 64.5), quarter-night 46.0 -> 30.8 -> **73.3**, harbour-night unchanged (42.9). Split (`visible.txt`): harbour-night cloud px darker than the sky by >= 3: 23.1%, lighter: 19.7%. The reviewer's per-band contrast (`fix/contrast.js` -> `contrast.txt`): harbour-night cloud minus clear y60-120 +7.9 -> +2.8 -> **-7.0** (dark masses), y180-215 +9 -> +6.5 -> **about +11** (lit), band sd 5.4 -> 4.2 -> 10.3 (two tones). 532: - **Horizon band within 2 luma of post-P1** (superseded by the recheck fix below, which meets it on the kit's p50) (`fix/metrics/sky.txt`): on the **mean**, met on every night view: harbour-night 70.6 -> 71.2 (set 2 70.5 -> 71.1), quarter-night 73.6 -> 74.2 (72.8 -> 74.1), r35 72.3 -> 72.8, orbit-a 69.2 -> 70.0, orbit-b2 71.0 -> 71.6, orbit-c 72.2 -> 72.1, beached-night 63.9 -> 63.6. On the **p50**, met on harbour-night 71.5 -> 70.9 and the orbits (-0.4, +0.8, +1.7), **not met** on quarter-night 74.0 -> 80.8 (+6.8), quarter-night-r35 73.6 -> 77.2 (+3.6) and beached-night 70.3 -> 59.8 (40% of its band region is the grass slope). A two-tone deck makes the band's p50 land in one tone (`fix/bandhist.js`: quarter-night's band is 90% cloud), so for this line I report the mean; owner or orchestrator call. Hue above the skyline 257.7-258.5 (target 250-280), met. 533: - Day and twilight unchanged: panel rows (y < 230) changed > 8 luma, reviewed -> fix against the reviewed noise pair: harbour-0545 1.00 / 1.15 against 1.10, harbour-day 0.19 / 0.15 against 0.17, quarter-day-r35 0.43 / 0.29 against 0.35; the rest of the frame is rain and ring phase (`fix/metrics/day-twilight.txt`). Integer counts and draw calls identical on all 13 views, reviewed -> fix and pre -> fix (`fix/metrics/counts-*.txt`). 534: - Dither and deck move with the clouds (`crawl.js` on `fix/cap/fix-1`, `fix-2` -> `fix/metrics/crawl.txt`): yaw +0.02 best shift -14 px (r 0.86), yaw +0.08 -60 px (r 0.70-0.78), orbit-c -56/-60 px (r 0.28-0.31); r at 0 is now 0.39-0.42 on the +0.08 pair (the static height ramps; the peak stays at the layer's shift). Noise pairs (`fix/driftpair.js` -> `fix/metrics/pairs.txt`, panel x 1320-1540, px changed >= 3 luma): final sitting quarter-night pre 8.3, reviewed 2.0, **fix 4.4** (>= 12 luma: 1.8%), harbour-night 0.1 / 0.8 / **2.0**. In the superseded sitting the quarter-night pair of the fix differed on 22-26% (reviewed 7%, pre 3%), mostly by 3-12 luma with no sideways slide that aligns them: when that view's capture time varies, the dither and the lit edges move with the drifting deck, and at the new contrast that shows; the owner should judge the motion on a film. 535: - Bench (`fix/bench.sh`, `fix/bench-summary.js` -> `fix/bench/summary.txt`; reviewed tree ("pre") against the fix, interleaved, round 0 discarded, night only (the fix is night-gated), `--disable-vsync`, 1920x1080, rain 0.45, 12:09-12:13, AC, Low Power Mode 0, `fix/bench/power.txt`): raw p50 dense 13.34 -> 13.38 (paired +0.05, 2/3 slower), Harbour 6.74 -> 6.69 (-0.02, 0/3); draw calls 2360 and 899 both sides; no measurable cost. The bench wrote its runs into `b/bench/` at first and overwrote the original P4b `*-t1` run folders there before they were moved to `fix/bench/`; `b/bench/summary.txt` keeps the original P4b numbers (noted in `b/bench/power.txt`). 536: - Journey: expected count deltas **none**. For the owner sheet: the deck is now a two-tone mass (dark indigo tops, violet lit bottoms, the low deck lit nearly all over); its dithered edges show more; the quarter's band p50 question above (answered by the recheck fix below). 537: - **Recheck fix (30 September 2026, Opus 5.5 on this Mac; evidence `artifacts/beautify/22-P4/b/fix2/`):** the recheck confirmed the deck is visible now, but found that the review fix lit the horizon band: on the kit's statistic (`tools/art_regions.json` horizon_band `luma.p50`, tol 2), quarter-night 74.0 -> 80.8 and quarter-night-r35 73.6 -> 77.2, a lighter lilac band behind the skyline (report: "hue change, not an additive glow"); the entry had switched to the mean after p50 failed. Also open: the dither checker and the deck's motion need a film for the owner; beached-night's band p50 for the owner sheet; the bench overwrite (disclosed); the reviewed acceptance block's horizon line not marked superseded. Fixed in `backdrop.gdshader` and `test_water.gd` only (backups `fix2/backup/`; at the start the live tree differed from `build/backup-feel/22-P4-b/` in exactly the eight P4b files, `fix2/diff-live-vs-22-P4-b.txt`; nothing unexpected): the undersides' lift is `DECK_UNDER` **1.30 -> 1.15** up to 0.35 of the panel and rises to `DECK_UNDER_HIGH` 1.35 only between the new `DECK_UNDER_REACH` 0.35-0.45 (before: over `UNDERLIGHT_REACH` 0.05-0.60, so the band had about 1.32). The underside is lighter than the body everywhere, and in the band a few luma over the clear sky (the test's arithmetic: at most 3.9 at full opacity, 7.5 at 1.30); only up high is it clearly lighter. The finding asked for "about 1.0x"; tuning rounds (`fix2/tune.sh`, `fix2/cap/v1`-`v4`, probes `probe/fix2-v1`-`v4`) show why not lower: at 1.00 the band p50 fell to 70.3 (-3.7, a miss the other way) and harbour-night visible coverage to 25.9% (the deck vanishing again, the first finding); at 1.10 quarter-night-r35 missed by 2.1; 1.15 is the lowest value that meets both. 538: - Tests: test-first `fix2/testfirst-test_water.log.out` (295 checks, 5 failures, on the review fix's backdrop); the band check fails with `DECK_UNDER` put back to 1.30 in a scratch copy (`fix2/check130-test_water.log.out`: "most 7.5"); after `fix2/r3-test_water.log.out` **292 -> 295, 0 failures**. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/b/fix2/suite` (12:40-12:46, AC, Low Power Mode 0, `fix2/suite-power.txt`): **40 suites, 19,944 checks, 0 failures**; `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,549, 0 failures (the known flaky check did not fire); 1440x900 not run (inherited stall). 539: - Changed expectations (test_water): "undersides the glow lifted over the sky (DECK_UNDER > 1)" -> "at about the sky's luma low down (DECK_UNDER 1.0-1.2), lifted over it high up (DECK_UNDER_HIGH > DECK_UNDER)"; "a lit underside's thinnest step >= 3 luma lighter than the clear sky at every height" -> ">= 3 lighter than the body's thinnest step at every height", plus "up to 0.35 (the band) 0-5 luma over the clear sky at full opacity" and "from 0.6 up >= 3 lighter than the clear sky"; new: the lift ramps over `DECK_UNDER_REACH`, which starts at or above 0.35; heights 0.23 and 0.35 added to the sweep. 540: - Captures (`fix2/caps.sh 1`, `2`, one sitting 12:28-12:36, AC, Low Power Mode 0 for all six sets, `fix2/cap-power.txt`): `fix2/cap/pre-1`, `fix2-1` (live), `fix1-1` (the first review fix, scratch tree `fix2/fix1-tree`), `pre-2`, `fix2-2`, `fix1-2`, the same 13 views as the fix, all first try; live probe `probe/fix2-final`. Looked at: `fix2/look/quarter-night-panel-2x.png`, `harbour-night-panel-2x.png` (pre / fix1 / fix2) and `r35-pair-2x.png`: the lilac layer over the skyline is gone; the low deck is a muted violet at about the sky's value with darker masses in it, the upper deck as in the fix. 541: - **Horizon band within 2 luma of post-P1 (p50, `fix2/metrics/sky.txt`, set 1 / set 2):** met on every Canal night view. quarter-night 73.8 -> **74.4 / 74.3** (fix1 80.8; P1's "60-75, no increase": within the pre noise pair's 0.2 plus 0.4), quarter-night-r35 73.6 -> **73.2 / 73.2** (fix1 77.2), harbour-night 71.6 -> **70.9** (fix1 70.9), orbit-a 71.3 -> 70.9, orbit-b2 71.9 -> 70.9, orbit-c 73.6 -> 72.7 / 73.9. The **mean** now sits under pre by 0.9-4.1 (quarter-night 72.5 -> 69.8, r35 72.7 -> 68.8, harbour 70.5 -> 68.4, orbit-c 72.4 -> 68.3): the dark bodies pull it down; on the mean the band is darker, not lighter. **beached-night** (not a must-fix; for the owner sheet): p50 70.3 -> 59.8 (fix1 59.7), mean 63.4 -> 61.3; about 40% of its region is the grass slope, and the dark bodies move the median. Hue above the skyline 257.3-258.0 (target 250-280), met. 542: - Deck still visible (`fix/visible.js` -> `fix2/metrics/visible.txt`): harbour-night cloud px >= 3 luma off their row's clear sky **41.6 / 41.6%** (fix1 42.8, pre 26.4; target over 40, met), >= 5: 35.9 (fix1 41.1); split dark 23.1 / light 18.6. quarter-night 40.7 (fix1 41.4); quarter-night-r35 30.4 / 35.5 (fix1 35.2; with the fallback clear level `visible-fallback.txt` 54.8 / 64.8, fix1 65.9). Per band (`fix2/contrast.js`): harbour-night y120-180 cloud minus clear +3.3 -> **-0.2** (the band at the sky's value), y180-215 +11.5 -> +4.7, y60-120 -7.0 -> -7.1 (unchanged dark masses). 543: - Day and twilight unchanged: panel px (y < 230) changed > 8 luma (`fix2/panel8.js` -> `fix2/metrics/day-twilight-panel8.txt`), fix1 -> fix2 against the fix2 noise pair: glade-night 0.26 / 0.18 against 0.17, harbour-0545 0.81 / 1.02 against 0.89, harbour-day 0.27 / 0.27 against 0.28, quarter-2045 1.23 / 1.48 against 1.35, quarter-day-r35 0.38 / 0.38 against 0.40 (the change is night-gated). Integer counts and draw calls identical on all 13 views, fix1 -> fix2 and pre -> fix2, both sets (`fix2/metrics/counts-*.txt`). 544: - Noise pairs (`fix/driftpair.js` -> `fix2/metrics/pairs.txt`, panel px changed >= 3 luma): quarter-night pre 3.5, fix1 3.9, **fix2 5.6**; harbour-night 0.03 / 1.2 / **0.4**; quarter-night-r35 3.4 / 4.3 / **12.6**. **The dither checker and the motion (not a must-fix; owner look):** a film for the owner, `fix2/film/film.sh` (the drag harness with a no-op undo as the filmed step, `fix2/film/deck-film.json`, `--filmstrip` every 200 ms for 8 s, 1600x1000, night; digest unchanged by the undo): `fix2/film/quarter-night-fix2-panel.mp4` (the panel strip at 2x) and `-frame.mp4`, the same for `quarter-night-pre` (the pre-part tree) and `harbour-night-fix2`; frames in the folders beside them; looked at `fix2/film/look-qn-fix2-0-400-800ms-2x.png`. Frame-to-frame change (`fix2/film/filmstep.js` -> `filmstep.txt`, panel px >= 3 luma per 200 ms): quarter-night pre median 6.1% (max 9.1), **fix2 16.6% (max 40.7)**, harbour-night fix2 7.0% (max 35.3); >= 12 luma: 1.9 / 2.9 / 0.8%. The dither sits on each layer's own pixels, so as a layer drifts by a render px its 2x2 checker steps with it, and at the deck's contrast that reads as a shimmer on the step edges; whether it stays is the owner's call on the film (a design change beyond this fix otherwise). 545: - Bench (`fix2/bench.sh`, `fix2/bench-summary.js` -> `fix2/bench/summary.txt`; `fix2/fix1-tree` ("pre") against live, interleaved, round 0 discarded, night, `--disable-vsync`, 1920x1080, rain 0.45, 12:46-12:50, AC, Low Power Mode 0, `fix2/bench/power.txt`; it writes only into `fix2/bench/` and refuses a folder that already holds runs): raw p50 dense 12.83 -> 12.83 (paired +0.00, 2/3 slower), Harbour 6.68 -> 6.69 (+0.00, 2/3); draw calls 2360 and 899 both sides; no measurable cost. The overwrite of the original P4b t1 bench folders by the first fix (above) stays as disclosed; `b/bench/NOTE-t1-runs.txt` now says so beside the summary. 546: - Journey: expected count deltas **none**. For the owner sheet: the band behind the skyline is back to the pre value (p50) and a violet hue; the deck's dither shimmer (film); beached-night's band p50 70.3 -> 59.8 (grass in the region). Scratch tree `fix2/fix1-tree` kept for the review. 547: - **Files:** `shaders/sky.gdshaderinc`, `shaders/backdrop.gdshader`, `shaders/water.gdshader`, `scripts/world_view.gd`, `scripts/look_profile.gd`, `tests/test_water.gd`, `tests/test_time.gd`, `tests/test_look.gd`, this entry. Backups `artifacts/beautify/22-P4/b/backup/`. Snapshot `build/backup-feel/22-P4-b/`; at part start the live tree equalled it on scripts, shaders, tests, data, scenes, tools and `project.godot` (empty `b/diff-live-vs-22-P4-b.txt`); nothing unexpected. Evidence `artifacts/beautify/22-P4/b/`; scratch tree `b/pre-tree` (kept for the review). 548: - **Tests:** test-first `b/P4b-testfirst-test_water.log.out` (277 checks, 31 failures), `-test_time` (372, 6), `-test_look` (553, 1). After: test_water **241 -> 277**, test_time **365 -> 372**, test_look **552 -> 555**, 0 failures (`b/r4-*`, `b/r5-test_water*`). New: `cloud_deck_checks` and `far_quay_checks` (test_water: the tint function, both shaders take it at the named places, luma kept within 1/255 at night at every height and hue 250-280 at h 0.29 by the shader's arithmetic, zenith and glow 0 unchanged, underlight, night-gated deck exactly the old threshold by day and at twilight, dither on the layer's pixels and the old floor at 0, quay size, lamps, order after the mist and rain, behind the switch, dark at night, no lamps by day), `glow_checks` (test_time: both uniforms at 58 hours, presets 0/0/1, 05:45 and 20:45 = 0.5, no glow key), `glow_checks` (test_look: RENDER key, Glade 0 at every anchor, canal 1 at night). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/b/suite` (a/b/v, 11:00-11:07, AC, Low Power Mode 0, `b/suite-power.txt`): **40 suites, 19,926 checks, 0 failures** (P4a 19,880 + 46); `--verify` 2880x1800 exit 0 in 34 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 5,549 checks, 0 failures, the known flaky check did not fire. 1440x900 not run (inherited stall). 549: - **Changed expectations:** none existing. My own new checks refined during tuning: dither `min(floor(x*3 + bayer), 3)` -> centred `clamp(floor(x*3 + (bayer - 0.5)), 0, 3)` (uncentred it brightened the deck); `DECK_RAIN` range <= 0.4 -> <= 0.5 (0.30 gave 39% cover on harbour-night); `GLOW_TINT` <= 0.9 -> <= 1.0; `GLOW_REACH.y` <= 0.8 -> <= 0.82 (float32); luma match linear -> gamma-encoded (`sky_luma_match`), underlight `city_glow.rgb` -> `sky_luma_match(city_glow.rgb, lit_cloud)`; test_look Glade "takes the glow from its own horizon" -> "takes no glow". 550: - **Captures** (`b/caps.sh`, one sitting 10:41-10:49 in the order pre-1 (scratch tree), after-1, pre-2, after-2; 20 views each, all first try; `b/cap-power.txt` AC, Low Power Mode 0): quarter-night, harbour-night, quarter-dry, quarter-0545, quarter-2045, harbour-0545, harbour-2045, harbour-day, quarter-day, quarter-day-r35, quarter-night-r35, glade-day, glade-night, lane-night, harbour-golden, photo-golden, orbit-a/-b2/-c (yaw 0.70/0.80/0.86, night), beached-night (the P3a rim). Superseded sitting (before the Glade exemption and the last tuning) in `b/cap/sitting1-superseded/`; tuning rounds `b/cap/t0-pre`-`t10`. Cloud probe (`b/probe.sh`: a copy of each tree whose backdrop draws cloud/city masks) in `b/probe/final-pre`, `b/probe/final`. Looked at: `b/look/night-panel-2x.png`, `quay-6x.png`, `twilight-panel.png`, `day-golden-glade-panel.png`, `orbit-night-panel-2x.png`, `t10-qn-sky-2x.png`, `probe-qn.png`, `diff-qd35-set1.png`. They show a violet band behind the skyline, a softer dithered deck in place of the grey wisps, the quay as a thin dark line with warm dots (and their faint reflections), by day a grey line and nothing else, golden and Glade unchanged but for the quay, 05:45 a little more violet and 20:45 a little more magenta at the horizon. 551: - Counts: integer counts and draw calls identical pre -> after on all 20 views in both sets and both noise pairs (`b/metrics/counts-*.txt`). 552: - Change by band (`b/metrics/bands.txt`, >8 luma, set 1/set 2 against noise pre/after): panel above the quay rows within noise or up to 0.7 point over it (quarter-night 1.94/1.82 against 1.28/1.64, lane-night 1.54/1.30 against 0.70/0.95, quarter-dry 0.65/0.68 against 0.02 = the dither); the quay rows 28-55% (new content); below the horizon within noise except rain-ring and streak phase (`look/diff-qd35-set1.png`). 553: - **Acceptance of the reviewed version** (superseded for the deck and the horizon band by the review fix and the recheck fix above; the quay, hue, dither and day lines stand) (`b/sky.js` -> `metrics/sky.txt`, `bandsplit.js` -> `band-split.txt`, `lower-panel-split.txt`, `coverage.js` -> `coverage.txt`, `quay.js` -> `quay.txt`, `crawl.js` -> `crawl.txt`; set 1, set 2 within 0.4): 554: - **Horizon band within 2 luma of post-P1 (superseded, twice: the review fix raised it to 80.8, the recheck fix brought it back to 74.4; see those blocks):** met on the reviewed version. quarter-night p50 73.8 -> **75.5** (mean 72.9 -> 72.5); harbour-night 71.6 -> **71.7** (70.9 -> 70.1); quarter-night at rain 0.35 73.6 -> 74.8. The rise is more cloud, not lighter pixels: clear sky 70.6 -> 70.7 and cloud px 77.3 -> 76.2 (p50). 555: - **Hue 250-280 above the skyline:** met, horizon band 241.3 -> **251.9** (quarter-night), 240.6 -> **253.6** (harbour-night), 241.2 -> 252.5 (r35); orbit 253-254. Margin is small on quarter-night. 556: - **Cloud coverage over 40% at rain 0.35 at night, violet undersides:** on the probe mask only (the review found the deck not visible; see the review fix), harbour-night 35.3 -> **44.0%** of the panel's free columns (41.7% of the px the city leaves open), quarter-night at rain 0.35 54.2 -> 63.3%, at its own rain 0.45 56.8 -> 69.1%; cloud px in the lower panel hue 241.4 -> **252.6** at luma p50 72.6 -> 71.2. 05:45, 20:45 and day coverage unchanged (35.3-35.4, 56.7). 557: - **Day unchanged:** met outside the quay rows. harbour-day and quarter-day(-r35) zenith and horizon band identical to 0.1 luma and 0.5 degree hue; panel >8 change 0.12/0.14 against noise 0.13/0.14. The quay adds a grey line by day (skyline region p50 219.4 -> 217.5). 558: - **Dither does not crawl:** met. Its high-frequency pattern moves with the cloud layer: yaw +0.02 -> best shift -14 px (the far layer's 14.1), yaw +0.08 -> -60 px (56 expected on whole render px), r 0.47-0.79; r at 0 shift -0.14 to 0.09 (a screen-fixed dither would peak at 0). 559: - **Far quay a 2-4 px dark band with 1-px lamps:** met. Band 4-8 screen px = 2-4 render px on harbour-night (4 px on 458 of 630 free columns; the 3-4 px steps partly hide against the rows' facades), luma at the horizon 51.7 -> 39.3; lamps 2x2 screen px (1 render px), 26 px apart (13 render px), brightest 215-220. Out by day (0 lamps), dim at golden hour. 560: - **05:45 and 20:45 within noise except the half tint:** zenith identical, horizon band luma +0.3 / -0.1 (Harbour), hue 256.9 -> 263.8 (05:45) and 1.7 -> 349.2 (20:45); panel >8 change within noise (harbour-0545 0.96/1.21 against 1.02/1.15). 561: - Panel lights (`metrics/panel-lights.txt`): panel p99 86.3 -> 86.1 (harbour-night); quarter-night 88.4 -> 102.6 in set 1 and 87.0 in set 2 (rain streaks); the lamps set the new maximum (162 -> 220 on quarter-night), above the town tower windows' p50 121 and under `windows_right`'s max 247. 562: - **Bench** (`b/bench.sh`, `bench-summary.js`, `bench/summary.txt`, `bench/power.txt`; scratch tree against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, AC, Low Power Mode 0, 10:53-11:00): raw p50 medians, night: dense 13.27 -> 13.23 (paired +0.01, 2/3 slower), Harbour 6.62 -> 6.64 (+0.04, 2/3); 13:00: dense 13.21 -> 13.21 (+0.02, 2/3), Harbour 6.68 -> 6.67 (-0.03, 1/3). Draw calls 2360 and 899 both sides. No measurable cost. 563: - **Journey:** expected count deltas **none**; pixel changes in the panel strip and the quay rows of all six views, hue only in the night sky. 564: - **For the owner sheet:** (the deck: see the review fix); the quay lamps take the presentation outline's dark cross (P4a kept far windows under the knee to avoid rings; the lamps are meant to be the brightest horizon points for P4c); the quay line by day; Glade exempted; the water's tint is small (open water hue 237 -> 242 at the same luma, `metrics/water-hue.txt`) because the mirror ray looks high up the panel. 565: - **Review:** filled by the orchestrator. 566: - **Not done / open:** the stretched columns are P4c. Puddles mirror the plain sky gradient (not in the plan's file list). `b/pre-tree` and the probe trees `b/probe/*-tree` kept for the review. 567: - **Platforms:** Mac run; Windows not run. 569: **Beautify P4a Skyline: landmark row, gamma-space haze, neon x1.8, stepped 3D ring (30 September 2026):** Opus 5.5 on this Mac. The painted panel (`backdrop.gdshader`) gains a landmark row between the far row and the middle row: `LANDMARK_FREQ` 6 slots a frame, `LANDMARK_SHARE` 40% taken, tops at `LANDMARK_HEIGHT` 0.45-0.85 of the panel, on whole render px: a shaft that steps back twice toward the crown, a 1-px mast with a red aviation beacon on 70%, floor-run windows (a lit run every third px row) and on about half a 1-px vertical cyan or magenta neon line; it scrolls at `turn * 0.58`. The rows sit in their haze by a gamma-space mix (`sky_mix_gamma` in `sky.gdshaderinc`, `SKY_GAMMA` 2.2): far 0.50, landmarks 0.38, viaduct 0.28, middle 0.26, near 0.10 (before, linear: far 0.62, viaduct 0.40, middle 0.38, near 0.18). Neon is `x mix(1, NEON_NIGHT_GAIN 1.8, night_amount)`. Far windows no longer add as glow: `sky_window` lifts the facade by `WINDOW_LIFT` 0.12 luma in gamma space (times its row's light x2, at most 1), under the presentation outline's 0.14 knee, so the outline no longer rings them (report: "keep window luma within 0.14 of the facade"); they are dimmer than before. Owner D14: the beacons and half the landmark neon blink only when the new optional scene key `beacons_blink` is on (Scenery switch "Blinking skyline beacons", `--beacons=blink`); absent it is steady and a document saves byte-identically (version 2, validated, undoable). The 3D ring (`_build_skyline`) now takes its boxes from `WorldView.skyline_ring_boxes()`: seed 1776 keeps every old box, a second stream (1777) adds 20 set-back stages, 5 spires and 7 steady neon strips (`plain`, no flickering lamp material) on blocks with no mast, and blocks take two depth tones; the skyline switch hides the ring whenever it exists. **Superseded by the P4 owner fixes (owner P4-1):** the ring, its box list, both rng streams, its tests and the switch hook are deleted; the painted skyline replaces it. 570: - **Finding: the 3D ring was never drawn.** `_build_skyline` had no caller in any snapshot under `build/` or since the first commit (2064bfc); the report's "100 near-identical boxes" at `world_view.gd:624-650` were the panel's rows. A probe with the call added (`a/probe/`, scratch tree since moved to the Trash) shows why: the near arc (49-64 m out) stands between the camera and the island and covers the front of Harbour and most of the quarter (`a/probe/harbour-night/courtyard.png`, `quarter-night/courtyard.png`). An orthographic view has no "behind the island" at that radius. So the ring stays unbuilt (tested); the owner decides whether to delete it or draw only its far arc another way. 571: - **Files:** `shaders/backdrop.gdshader`, `shaders/sky.gdshaderinc`, `scripts/world_view.gd`, `scripts/scene_state.gd`, `scripts/editor_ui.gd`, `scripts/main.gd`, `tests/test_water.gd`. Backups `build/backup-beautify-P4a/`. Snapshot `build/backup-feel/22-P4/`; the live tree matched it on scripts, shaders, tests, data, scenes, tools and `project.godot` at part start. 572: - **Tests:** test-first `a/P4a-testfirst-test_water.log.out` (227 checks, 30 failures). test_water **191 -> 241, 0 failures** (`skyline_panel_checks`, `skyline_switch_checks`, `skyline_ring_checks`: row constants, words in the landmark body, behind the switch, the gamma helper and the three shares, the old linear shares gone, night row tones, neon gain, the window lift under the outline knee read from `presentation.gdshader`, the beacon switch through validate/JSON/undo/view/UI/args, every old ring box unchanged under 1776, counts, two tones, one mesh, not built, hidden by the switch). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/22-P4/a/suite` (a/b/v, 09:53-09:59, AC, Low Power Mode 0): **40 suites, 19,880 checks, 0 failures** (P3 close 19,830; only test_water differs); `--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). 573: - **Changed expectations:** my own test-first line "the rows' total spread is wider than before" was wrong (gamma mixing darkens the far row most: night far/middle/near 38.8/33.3/29.6 against 43.1/37.8/32.7 in data) -> "the far row stands >3 luma further from the haze and adjacent rows keep >=2.5 apart". No existing assertion changed. 574: - **Captures** (`a/caps.sh`, one sitting 09:36-09:43 in the order pre-1, after-1, pre-2, after-2, pre-3, after-3; pre from the scratch tree `a/pre-tree`; views harbour-night, quarter-night, lane-night, harbour-day, harbour-golden, photo-golden, glade-day, glade-night, quarter-day, plus `after-1/harbour-night-blink`). Framing flakes (the known kit issue) were frequent, so each view uses the set whose pre and after share a framing (`a/framing.js`, pairs in `a/metrics/pairs.txt`, symlinked under `a/cap/pair/`). An earlier sitting before the window fix is in `a/cap/sitting1-superseded/`; tuning rounds `a/cap/t1`-`t3`. Looked at: `a/look/panel-.png` (pre | after strip, all nine), `a/look/t3-hn-panel-4x.png`, `a/look/hn-panel-right-4x.png` (ringed windows, superseded), `a/look/t1-day-golden.png`, `a/look/qn-full.png`, `a/look/diff-.png`. They show one tower on harbour-night (right), two on quarter-night, a pale silhouette tower by day and at golden hour, the rows darker and stepped, no rings round far windows. Glade shows the tower (until P5d, which the owner dropped on 1 October 2026, so it stays). Photo-golden barely changes (panel blurred and fogged). 575: - Counts: integer counts and draw calls identical pre -> after in all three sets on every view (`a/metrics/counts-set1..3.txt`); within-tree count flakes seen in sitting 1 only. 576: - Change by band (`a/metrics/bands.txt`, >8 luma, pair against a same-framing noise pair): panel strip harbour-night 2.95% (noise 0.63), quarter-night 4.04 (1.94), harbour-golden 7.77 (0.78), glade-night 9.34 (0.28), harbour-day 1.25 (0.19); below the panel every view within its noise (e.g. harbour-night 1.75 against 1.67) except photo-golden 2.24 against 1.53 (mirrored panel lights, not isolated). 577: - **Acceptance** (`a/accept.js`, `a/towers.js`; `a/metrics/acceptance.txt`, `towers.txt`, `day-golden-columns.txt`): 578: - **Skyline 10-20% of the frame on harbour-night:** met, beacon at y 96 = **13.4%**, crown 12.2% (same on quarter-night); before, the tallest were the rows' masts at about 10.4%. Area: the city covers a smaller share of the free columns' strip (14.3 -> 2.5% by the scan, which counts ringed dark window pixels as city before), not a target. 579: - **At least three separable layers and vertical neon:** met with small steps. Night tones (`a/metrics/*`, column runs at x 1310-1560): sky 71-85, far row 49, landmarks 47, middle 43, near 39-41 (before far 53, middle 47, near 35-42). The landmarks are only 2 luma from the far row; they separate by standing against the sky. Neon px (saturated, luma >90) 116 -> 124 on harbour-night; the right strip reads brighter. 580: - **Panel lights below the town windows:** met, panel p99 97.9 -> **86.3** (harbour-night), 119.6 -> 88.4 (quarter-night) against quarter-night `windows_tower` p50 121.2; the brightest panel px are neon/beacon (max 217). 581: - **Far windows not ringed:** met, ringed window px 41.9% -> **0.8%** (harbour-night), 38.5 -> 0.8 (quarter-night), 5.3 -> 1.9 (harbour-golden). 582: - **Day and golden rows read as silhouettes:** met, harbour-day middle row 17 -> 24 luma under the sky, far row 2 -> 7, tower 13; golden middle 22 -> 33, far 5 -> 14, tower 21. 583: - **Existing box placement unchanged:** tested (every old box, 1776). 584: - **Blink switch:** `after-1/harbour-night-blink` caught the beacon off (luma 45.8 at (1448,96) against 113.2 steady). 585: - **Bench** (`a/bench.sh`, `a/bench/summary.txt`, `a/bench/power.txt`; scratch tree against live, interleaved, round 0 discarded, `--disable-vsync`, 1920x1080, rain 0.45, AC, Low Power Mode 0): raw p50 medians, night: dense 13.41 -> 13.41 (paired -0.03, 1/3 slower), Harbour 7.51 -> 7.20 (-0.14, 0/3); 13:00: dense 13.40 -> 13.45 (+0.06, 3/3), Harbour 7.63 -> 7.54 (-0.09, 1/3). Draw calls 2360 and 899 both sides. No measurable cost. 586: - **Journey:** expected count deltas **none**; pixel changes in the panel strip of every journey view (the skyline shows in all six). 587: - **Review (orchestrator, 30 September 2026):** one Opus 5.5 review in a fresh context (378 changed lines; no second review, the plan does not mark P4a as a measuring part and the review returned findings); no must-fix findings, acceptance met (skyline 13.4% of the frame, outline rings on far windows 41.9% -> 0.8%, panel lights below the town windows, old box placement unchanged). Five non-blocking findings, all for the owner sheet or this entry: (1) at night the gamma haze packs the rows 2-6 luma apart, so the three layers separate by small steps; (2) the tower windows share the far-window lift and read as faint strips; (3) far windows no longer add as glow, so night panel windows nearly vanish (glade-night panel strip 9.34% against 0.28% noise); (4) the 3D ring is not drawn, no caller of `_build_skyline` exists, the owner decides to delete it or draw only a far arc; (5) the photo-golden over-noise change below the panel is rain streaks and rings, not mirrored panel lights, and the reviewer captured the missing 05:45 and 20:45 views (`a/review/`), skies unchanged, rows darker silhouettes. Reviewer re-ran test_water on the live tree: 241 checks, 0 failures. 588: - **Not done / open:** the 3D ring is not drawn (finding above; owner question). The waterfront quay and cloud deck are P4b, the stretched columns P4c. The landmark count per frame is the report's (40% of 6 slots): Harbour shows one tower at the default yaw. `a/pre-tree` kept for the review. 589: - **Platforms:** Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 2 October 2026, 10:27 BST, game main at af4df88) 137: ## Owner answers after P4a and P4b (owner, 30 September 2026, in chat, on the before/after pages in `artifacts/beautify/22-P4/owner-preview/`) 139: | # | Question | Verdict | 140: |---|---|---| 141: | P4-1 | The 3D skyline ring is never drawn (`_build_skyline` has no caller) | **Delete it** (code, tests, the switch hook; the painted skyline replaces it) | 142: | P4-2 | Night panel windows dimmer so far windows carry no outline rings | **Yes, accept** | 143: | P4-3 | Night cloud dither shimmers slightly as the deck drifts | **Keep** | 144: | P4-4 | Quay lamps are the brightest panel pixels and get a dark outline cross | **Dim by 15%** | 145: | P4-5 | 20:45 sunset hue about 12 degrees toward magenta (half tint, D20) | **Yes, keep** | 146: | P4-6 | Glade exempt from the city glow (also no heavier deck or dither) | **Yes, keep exempt** | 147: | P4-7 | Commit each P4 part locally as it finishes | **Yes** (commit only; push still needs a request) | 149: Orchestrator plan: P4-1 and P4-4 are applied as small fix passes after P4c finishes (P4c is editing the same shaders); P4a and P4b were committed from their part snapshots. 161: ## P4 sheet signed off (owner, 1 October 2026, in chat) 163: "Sign off, pause there." The combined P4 sheet (`artifacts/beautify/22-P4/close/sheet/`, 14 pages) is signed off as it stands, including the golden water as built (pink band, only a faint mauve strip where a cool band was wanted, smooth peach far water without rain rings), the P4b and P4c looks listed on the sheet, and Glade's inherited sky, fog and water. The owner paused the run there: the M3 export (which this sign-off unblocks) is NOT built, and P5d, P5f, P5c2 and the P5 close are not started. Nothing is running. 164: | P5f | Plan defaults: HALO_SPAN 0.2, HALO_T_DAY 1.03 (lens 0.97 moved because clipped sunlit white rang by day); engine glow candidate not taken. Second review run: no must-fix. Open for the owner on the P5 sheet: golden-hour ring strength on white trim, neon letters losing the dark outline, the daytime bloom lift (about 3 luma on the brightest pixels) gone. | 166: ## P5d dropped (owner, 1 October 2026, in chat) 168: "No need to do the glade hills." P5d (Glade hills instead of the grey city) is skipped; Glade keeps its current horizon (it shows the P4a landmark tower row). P5c2 and the P5 close follow without it.