Chapter 23: Lights in the small places Frozen 2 October 2026 at 10:27 BST (game main at af4df88). Implemented, reviewed and merged work of the P5 stage (small items): P5a bold neon marks, P5b wet glints in the lamp pools with its review fix and recheck, P5c1 window cores and arched interiors, P5e island visible with the tray open and outlined labels, P5f pixel halo ring with two fixes and two reviews, P5c2 sun and moon disc gain with its review fix, and the P5 stage close (suite, journey, protected looks, whole-stage bench, owner sheet, and the handoff check's twelve corrections, shown in the last section as a diff of the commit). P5d was dropped by the owner on 1 October. The owner has not seen the P5 sheet; its taste items are open. Original source terminology is preserved; local paths are removed. Line numbers refer to the frozen copy. Part codes: P5a neon marks, P5b wet glints, P5c1 window cores, P5c2 sun disc, P5e dock offset and labels, P5f pixel halo. SOURCE: docs/visual-handoff.md (frozen copy, 2 October 2026, 10:27 BST, game main at af4df88) 5: **Beautify P5 stage close (1 October 2026):** the stage-end overhead of P5 "Small items" (step `23-P5`), run by Opus 5.5 on this Mac on the final P5 code, `main` 0fe2de8, clean (`close/head.txt`). The parts are P5a, P5b, P5c1, P5e, P5f and P5c2. P5d was dropped by the owner. No game code changed. Evidence is in `artifacts/beautify/23-P5/close/`. The before side is the pre-stage scratch tree `close/pre-tree`, built from `build/backup-feel/23-P5/`. It was checked equal to `main` b1e4527 (the P4 close) on scripts, shaders, tests, data, scenes, tools and `project.godot`; only the three `.ps1` tools differ, in line endings (`close/diff-pretree-vs-b1e4527.txt`). Every run was on AC power, 100% charged, with Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). The load average was 3.5-4.6. 6: - **Full suite** (`close/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/23-P5/close/suite`, 11:59-12:06): **40 suites, 21,367 checks, 0 failures**, the same as P5c2's fix suite. `--verify` at 2880x1800 exited 0 on try 1 in 32 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures, and the known flaky check did not fire. 1440x900 was not run (the inherited stall). 7: - **Journey** (`artifacts/feel-journey/23-P5/`: the six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on the first try; counts in `23-P5/counts-vs-22-P4.txt`). Against `22-P4`, draw calls and rendered objects changed as follows: 8: - quarter-day and -night 969 → 1008 9: - harbour-day and -night 889 → 893 10: - harbour-lane 733 → 737 11: - styles-day 469 → 470 13: These are exactly the declared deltas (P5a +35 on the quarter; P5b +4, and +1 on styles); every other integer count is identical. The stage-end art capture re-ran the six views and got identical counts (`close/metrics/counts.txt`). By eye (`close/look/j-1.png`, `j-2.png`): orange rings on lamps and string lights, glints on wet paving and the new sign marks at night; at thumbnail size the soft lamp haze looks smaller; the day views are nearly unchanged. Frame >8 luma against 22-P4 (`close/metrics/journey-frame-22P4-23P5.txt`) is 4.9% on lane, 7-10% on the day views and 14-20% on the night views. `notes.json` has a new `23-P5` entry (backup `notes.json.bak-before-23-P5`); `index.html` and `journey-standalone.html` (106.95 MB) were rebuilt. 14: - **Art capture and protected looks** (`close/caps.sh`): one sitting, 12:08-12:29, in the order `pre-1`, `after`, `pre-2`, `after-2`. Each set has 55 views and all worked on the first try. The views are the P4 close's 43 plus glade-golden, the P5f/P5c2 disc views (disc-q0645, disc-qg-sun145, disc-hg-sun145, disc-gg-sun145, moon-hn-t4), P5a's mark sheet and close-ups (sheet-night, sheet-day, rclose-night, tclose-night), glade-rain and quarter-night-px1. A further set, `dock`, is live with `--dock-offset` (harbour-night, harbour-day, quarter-night). Metrics are in `close/metrics/` (`close/metrics.sh`, regions `close/regions.json` = `22-P4/close/regions.json`): `ab-*`, the noise pairs, `frame-*`, `counts.txt`, `protected.md` (`close/prot.js`), `ab-p4close-*` / `frame-p4close-*` against `22-P4/close/after`, `stage-measure.txt` (`close/stage-measure.js`) and `harbour-night-haze.txt` (`close/look/haze.js`). 15: - Counts are identical in both noise pairs, and the P4 close against pre-1 is identical on every shared view. 16: - Protected-look sheets: `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P4-close.png` and `noise-pre.png` were looked at; `noise-after.png` was made but not opened. 17: - **Canal mirror columns:** the same streaks. Quarter-night p50 44.3 → 43.5. It changed >8 on 24.3% in one pair against noise of 10.7/5.5%, and on 10.5% in the other pair; the P4 close capture of the same code differed from pre-1 by 50%, so this is within the spread between sittings. Quarter-21:00: 21.3/9.8% against 14.6/3.8%. 18: - **HDR neon:** changed by P5f, as the part measured. The letter cores are within noise in the part; in the sign crops the p50 rises (FISH Harbour 56.5 → 66.0, BAR 54.2 → 70.6) because the dark 1-px outline became a coloured rim. Owner taste item. 19: - **05:45 and 20:45 skies:** the zeniths are identical. The 20:45 horizon is 1.1-1.4 luma darker (#cb8892 → #c58892 on Harbour, 32% of the band >8): the bloom lift is gone. Not visible. 20: - **Glade/Canal separation:** Glade is identical (night zenith `#001d45`, horizon, water, lawn and bank). Harbour's night zenith and horizon are within noise. 21: - **Approved tree style:** grove-day crowns 0.54%, conifers 0.64% and feet 0.11% against noise of 0-0.37%, with identical p50s. Grove-1930 is within noise. Quarter-night undersides changed 4.8/6.1% against 0.9/3.8%, with the same p50 (24.9): the ringed string lights inside the crop. 22: - Close-measured lines (`stage-measure.txt`): 23: - Quarter plaza +4.68 / +4.84 luma (45.0 → 49.7; noise -0.03/+0.13). That is P5b's +2.65 plus P5f's rings; quarter-dry is +1.54 (P5f only). 24: - Disc cores (inner half radius): Harbour golden 209.8 → 214.0, quarter golden 224.8 → 225.7, quarter 06:45 241.1 → 242.6, with saturation 44 → 40%, 42 → 38% and 29 → 20%. The moon goes 240.7 → 245.5. Glade is 193.2 in all four sets. 25: - The Harbour-night lawn pools are brighter, not dimmer (+0.9 to +2.0 mean). 26: - **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. Dense `artifacts/weather-pass9/dense.json` and Harbour ran at night (`--time=1`), 13:00 and 19:30, 12:29-12:40. vsync was disabled in every run, each run had 491 raw samples, and every run worked on the first try. Raw p50 medians (pre → live, paired difference, rounds where live was slower): 28: | Scene | Night | 13:00 | 19:30 | 29: |---|---|---|---| 30: | Dense | 13.24 → 12.87 ms (-0.37, 0/3) | 13.21 → 12.49 (-0.72, 0/3) | 15.15 → 14.78 (-0.36, 0/3) | 31: | Harbour | 6.77 → 6.46 (-0.31, 0/3) | 6.70 → 6.10 (-0.59, 0/3) | 6.51 → 6.15 (-0.35, 0/3) | 33: - p95 medians: dense 15.22 → 14.78, 15.22 → 14.43 and 17.09 → 16.66; Harbour 8.00 → 7.05, 7.89 → 7.59 (live slower in 2/3 at 13:00, round spread up to 1.3 ms) and 7.21 → 6.98. 34: - Draw calls: dense 2360 → 2362 (19:30: 3124 → 3126); Harbour 899 → 903 (19:30: 814 → 818). 35: - Reading: P5 is faster in every paired round. P5f's 12-tap ring replaces the 49-tap bloom, which more than pays for P5a's extra signs and P5b's pools. The Harbour p95 rise that P5b showed (+0.65 ms) is no longer there at stage level. 36: - **Owner sheet** `artifacts/beautify/23-P5/close/sheet/`: before (pre-stage tree, `pre-1`) on the left and after on the right, colour then greyscale where values matter; for P5e, `pre-1` editor.png against the `dock` run. Crops come from `close/crops.js` and pages from `close/sheet.sh`; all pages were looked at. Pages: `p5-01-overview.png`, `p5-02-P5a-marks.png`, `p5-03-P5b-glints.png`, `p5-04-P5c1-windows.png`, `p5-05-P5e-dock.png`, `p5-06-P5f-halos.png`, `p5-07-P5c2-disc.png`, `p5-08-existing-scenes-1.png`, `p5-09-existing-scenes-2.png`. `sheet.md` has one plain paragraph per part with its numbers, then what existing scenes see, the bench and the suite. 37: - Looked at and flagged for the owner (not changed): 38: - FISH MARKET and BAR read bolder but less crisp (P5f rim). 39: - A small box sign on the quarter's right block reads almost unlit at night after P5 (`close/look/rc-box.png`): a crown hides most of its new P5a mark, and the soft glow it had is gone. 40: - The moon's cloud dither is fainter. 41: - An orange-red rim sits beside the 06:45 disc. 42: - **Not done / open:** 43: - The owner has not seen this sheet. Open items carried from the parts: 44: - The P5b plaza line, waived at +2.65. 45: - The P5c1 core-halo line, waived. 46: - P5e with the tray open (D16 kept). 47: - P5f's `HALO_SPAN` golden ring strength and the neon outline. 48: - P5c2's silhouette rim. 49: - Glints on lawns and cornice tops. 50: - P5d was dropped by the owner. **The M3 exports** were built after the stage close started, on the owner's request, from main 6fa5452 (the signed-off P4 tree, before any P5 part): [local path removed], `PixelPunk-M3-macOS.zip` and `PixelPunk-M3-Windows.zip` (`artifacts/beautify/22-P4/m3/`; not launched, Windows not run). They do not contain P5. The M4 export after P5 is dropped (D8; the next export is M6 after P9). 51: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 52: - Scratch trees: `close/pre-tree` and the part trees `a/`, `b/`, `c1/`, `c2/`, `e/` and `f/pre-tree` under `23-P5` were moved to [local path removed] with `mv` once their numbers were recorded. They are named `beautify-23-P5--pre-tree-125153`. 53: - Platforms: Mac run; Windows not run. 55: **Beautify P5c2 Sun and moon disc gain (1 October 2026):** Opus 5.5 on this Mac (report quick win 11, the disc half; lens "Let the sun disc and the brightest window cores reach the HDR glow", revised proposal (3)). Round 1 stopped on an acceptance line: Glade's disc went grey, and canal's golden disc went cream-white. The review made both must-fix. After the review fix the line is met in canal, and Glade's disc is exactly as before. The owner is available but busy, so the plan defaults decide. The fix adds no tuned value: the plan's 1.6, 1.2 and 1.3 stand (see the owner items). 57: What changed in `backdrop.gdshader`: 58: - Before P5c2 the disc was sun_color x 1.18 for both bodies. 59: - The sun's gain is G = `DISC_GAIN_SUN` 1.6 + `DISC_GAIN_LOW` 1.2 x low_sun. low_sun is P4de's uniform (SunPath.golden, the real sun). 60: - Review fix: G now lifts only the sun colour's brightest channel, to G x that channel. That is red for every sun colour in the profiles. The other channels stay at x `DISC_OLD` 1.18. 61: - The moon takes `DISC_GAIN_MOON` 1.3 on every channel. 62: - Review fix: a new uniform `disc_hdr` turns the gain on. Its default is 0, the old x1.18 disc for both bodies. `WorldView.sync_profile` sets it from the look's engine glow (LookProfile RENDER "glow": canal 1, Glade 0), so Glade draws its disc as before. This follows the P4-6 precedent, where Glade was exempted from the city glow. 63: - Round 1's `disc_moon` (1 when `SunPath.moon_weight(hour) >= 0.5`, the test `SunPath.disc_body` makes) still decides which body the disc is. So the moon is the disc after 21:00 and before 06:10, and it flips only at the two crossings, where the disc is hidden. 64: - Reading taken: the gains *replace* 1.18. So the moon is x1.10 of its old value, not 1.18 x 1.3 = 1.53. 65: - The halo, stars, clouds and water are untouched. 67: Files: `shaders/backdrop.gdshader` (+24 -1), `scripts/world_view.gd` (+5), `tests/test_sun.gd` (+143). Snapshot: `build/backup-feel/23-P5-c2/` (= main bfb2521). At part start the live tree matched it exactly. After the fix it differs in those three files only (`fix1/diff-live-vs-23-P5-c2.txt`). Pre-part tree: `artifacts/beautify/23-P5/c2/pre-tree`. Backups: `c2/backup/` (pre-part) and `c2/backup-round1/` (before the fix). Everything below is under `artifacts/beautify/23-P5/c2/`. Round 1's evidence is in `c2/` and the fix's in `c2/fix1/`. 68: - **Review (round 1), both must-fix:** 69: - (1) In Glade the gain made the disc greyer, not brighter, at golden hour and by day. Glade uses AgX with the glow off, so extra HDR only takes colour. Core saturation fell 33% -> 19% at golden hour and 10% -> 0% by day. That changed Glade's protected day sky. 70: - (2) The canal golden disc clipped. Applying (1.6 + 1.2 x low_sun) to the whole colour clips green as well as red. Core saturation: 06:45 27% -> 3% (white); 19:30 42% -> 17% and 44% -> 19% (cream). 71: - **Fixed:** 72: - (1) `disc_hdr` is 0 in Glade, and Glade measures identical to pre (below). 73: - (2) The gain lifts the brightest channel only. Green and blue keep x1.18, so the hue holds and red alone crosses the glow threshold. 74: - No value moved. 75: - **Tests:** round 1 took test_sun 79 -> 92. The fix's test-first log, `fix1-testfirst-test_sun.log.out`: 101 checks, 6 failures. After the fix, `fix1-after-test_sun.log.out`: **101 checks, 0 failures**. New checks: 76: - `DISC_OLD` 1.18 and the `disc_hdr` uniform (default 0). 77: - The brightest-channel lines and the `disc_rgb` mix. 78: - The arithmetic for golden (ffb85c), day (fff6e0), moon (d6e0f2) and white suns. 79: - disc_hdr 0 gives x1.18 for both bodies. 80: - The view sets disc_hdr to 1 in canal and 0 in Glade at 04:00, 06:45, 13:00 and 19:30. 81: - **Full suite** (`fix1/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh c2/fix1/suite`; 11:37-11:44; AC, Low Power Mode 0; `fix1/suite-power.txt`, `fix1/suite/suite_all.out`): **40 suites, 21,367 checks, 0 failures** (round 1's 21,358 + 9). `--verify` 2880x1800: exit 0 on try 1 in 34 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun: 5,549 checks, 0 failures, and the known flaky check did not fire. 1440x900 was not run (the inherited stall). 82: - **Changed expectations (round 1 -> fix, test_sun only):** 83: - "The disc draws at sun_color x disc_gain" -> "the disc draws at disc_rgb". Neither `sun_color.rgb * 1.18` nor `sun_color.rgb * disc_gain, disc)` is left. 84: - The order check now finds `sky = mix(sky, disc_rgb, disc);`. 85: - The constants check also wants `DISC_OLD = 1.18`. 86: - No expectation from before P5c2 changed. 87: - **Captures** (`fix1/caps.sh`, one sitting 11:28-11:35; AC, Low Power Mode 0, load 3.7-4.0; `fix1/cap-power.txt`): 88: - The same 15 views as round 1: `pre-1`/`pre-2` from the pre-tree and `after-1`/`after-2` live. The day probe `fix1/dayprobe.sh` ran 11:35-11:37 at 18:00 with `--sun=145`, in quarter, Harbour and Glade. 89: - Looked at: `fix1/look/q0645.png`, `qg.png`, `hg.png`, `gg.png`, `moon.png`, `day-q.png` and `day-g.png` (pre | after, x3), and `qg-wide.png`. 90: - What they show: 91: - Canal golden: the disc stays golden. It is a little brighter, with a soft warm corona. 92: - The glow also lays a warm orange-red rim on dark silhouettes right beside the disc: the facade at 06:45, the dome at 19:30, and the Harbour dormer and the smoke in front of the disc. 93: - Glade: unchanged. 94: - Canal day: the disc is as before, with a faint peach rim. 95: - Moon: slightly whiter (as in round 1). 96: - **Acceptance** (`fix1/discmeasure.js`, `fix1/metrics/disc.txt` and `disc-day.txt`; pre-1 -> after-1; the noise pairs move under 1 luma in every band): 97: - *A brighter golden disc with a soft corona: met in canal.* 98: - Core luma: quarter golden 208.3 -> 217.0 (+8.7), Harbour golden 194.3 -> 205.3 (+11.0), quarter 06:45 230.7 -> 239.3 (+8.6). 99: - Core saturation: 41% -> 38%, 44% -> 41% and 27% -> 20% (round 1: 17%, 19% and 3%). Hue: 47 -> 45, 48 -> 44 and 43 -> 43. 100: - Corona 1.05-1.5 R: +8.6, +10.2 and +5.6 luma (33.3%, 38.6% and 15.9% of px brighter by >8; noise 0-0.3%). 1.5-2 R: +4.8, +4.6 and +5.1. 2-3 R: +0.8, +0.2 and +0.2. 101: - The corona is 3-4 degrees warmer (hue 40 -> 36, 39 -> 36, 31 -> 28). It is smaller than round 1's (+20 at 1.05-1.5 R) because only red now crosses the threshold. 102: - Silhouette rim (`fix1/fringe.js`, `fix1/metrics/fringe.txt`): 103: - The facade beside the 06:45 disc: (148,135,114) -> (192,152,127), hue 36 -> 23, +22.6 luma (round 1: +41.9, hue 35). 104: - The dome edge at 19:30: +14.3 luma, hue 46 -> 41. 105: - The Harbour dormer: +10.1 luma, hue 48 -> 45. 106: - Day (18:00): quarter core 243.0 -> 244.3 and saturation 17% -> 16%; Harbour 247.3 -> 248.1 and 18% -> 16%; corona +1.1 and +0.9. Round 1 had taken day saturation down to 3%. 107: - Glade, identical: 108: - The golden core is (192,179,127) both pre and after, and every band is within 0.2 luma. 109: - The day core is (204,202,184) both pre and after, and every band is identical. 110: - *Moon x1.3:* met in the shader. On screen the core goes 243.8 -> 246.3 and saturation 5% -> 1%, as in round 1. 111: - *Nothing else moves:* met within noise (`fix1/elsewhere.js`, `fix1/metrics/elsewhere.txt`; px changed by >8 luma). For each view, the smallest of the four pre/after cross pairs is no larger than the largest noise pair. Controls, whole frame: 112: - harbour-golden 3.67 (noise 6.85) 113: - quarter-golden 1.16 (3.21) 114: - photo-golden 3.40 (3.59) 115: - glade-golden 0.29 (0.45) 116: - quarter-night 5.46 (5.87) 117: - harbour-night 2.79 (5.76) 118: - quarter-day 2.00 (4.62) 119: - harbour-day 1.23 (6.00) 121: Disc views outside 3 R: 0.29-2.64, against noise 1.34-17.48. 122: - *Counts* (`fix1/counts.js`, `fix1/metrics/counts.txt`): pre-1 -> after-1 draw calls are identical in every view. pre-2 -> after-2 differs in draw calls and rendered objects for disc-gg (757/764), disc-hg (759/766) and photo-golden (1322/1329). The same differences occur between pre-1 and pre-2, so they come from within one tree. **Journey count deltas (declared): none.** 123: - **How the gain meets P5f's halo:** 124: - The corona is the engine glow (HDR threshold 1.0, scale 2), fed by the red channel alone: red 2.4-2.8 at golden hour (feedback about 0.5-0.9) and 1.6 by day (about 0.08). 125: - P5f's halo keys on the presented colour against t (0.90 at 19:30, 0.915 at 06:45). 126: - At 19:30 the fixed core presents under t (luma 0.84), so P5f no longer rings it, as it did in round 1. 127: - At 06:45 the core sits just over t (0.93), so P5f adds a faint ring at most. 128: - By day t >= 1, so P5f adds nothing, and the 0.07 daytime bloom lift that P5f removed does not come back. 129: - Glade skips both. 130: - **Bench:** not run. The plan's P5c2 step table has no bench. The change is a step and two mixes per sky fragment, and draw calls are identical. 131: - **Not done:** no 1440x900 `--verify`, no pixels-1 capture. `pre-tree` is left for the orchestrator. The known flaky check is recorded as known (it did not fire). 132: - **Open for the owner (plan defaults taken):** 133: - (1) Glade keeps its x1.18 disc through the look's glow key (the P4-6 precedent). The alternative is the grey disc. 134: - (2) The golden corona is warmer and smaller than round 1's, and the glow paints an orange-red rim on dark silhouettes beside the disc (the 06:45 facade reaches hue 23). A golden rather than orange corona needs green over the threshold too, which brings back some whitening of the disc. That is a value judgement, so it was not taken. 135: - (3) The moon reading: x1.3 replaces 1.18, rather than multiplying it. 136: - Review (orchestrator, 1 October 2026): one fresh-context review (155 changed lines), two must-fix findings (the Glade disc went flat grey; the canal golden disc clipped to cream or white), both fixed in fix 1 without changing a plan value (disc_hdr gate through the look glow key, canal 1 and Glade 0; the gain lifts only the sun colour's brightest channel) and rechecked on disk by a fresh reviewer: no finding open, acceptance met. Rulings: (1) Glade keeps its old x1.18 disc (the P4-6 precedent; quick win 11 is tagged N and G, not Gl); (2) the moon reading x1.3 REPLACES x1.18 (the plan default). Left for the P5 sheet, undeclared by the part until the review: the engine-glow corona lifts the dark foreground silhouettes in front of the disc (round 1 about +50 luma inside 1 R; after the fix the 06:45 facade +22.6, the dome at 19:30 +14.3), clouds and smoke over the disc now read fainter than before, and the golden corona is smaller than round 1 so the plan's soft corona is modest (+8.6 to +10.2 luma at 1.05-1.5 R). Test quality note: three of the 13 round-1 checks test the test's own copy of the formula (the behaviour is covered by the source-string and disc_moon WorldView checks). No photo-golden capture with the disc in view. Platforms: Mac run; Windows not run. 138: **Beautify P5f Pixel halo ring replaces the 49-tap presentation bloom (1 October 2026):** Opus 5.5 on this Mac (report polish backlog "pixel halo ring"; lens "Pixel-art light halos", revised proposal). The owner is available but busy, so the plan defaults decide; two values moved off the lens, see Deviations. `presentation.gdshader` drops the 7x7 Gaussian (49 taps at 2 art px, threshold 0.70, /17) for a stepped halo at 1 art px. Ring 1 is the outline's 4 axis neighbours (reused, so no new taps). Ring 2 is `RING2`, the 8 taps at distance 2 (axis +-2, diagonals +-1). Each tap throws `halo_of(s, t)`: the hue of its colour above t, times `smoothstep(t, t + HALO_SPAN, luma)` with `HALO_SPAN` 0.2 (fix 1; round 2 had `smoothstep(t, 1, luma)`, see Review). Keeping that magnitude stops a pixel just over t from ringing as fully as a white-hot lamp. A ring is the brightest tap per channel. halo = (ring1 x `HALO_RING1` 0.5 + ring2 x `HALO_RING2` 0.22) x glow_strength x `HALO_GAIN` 2, times (1 - smoothstep(t, 1, here)). t = mix(`HALO_T_DAY` 1.03, `HALO_T_NIGHT` 0.80, lamps), and the taps run only while t < 1. The ink guard is `edge *= 1 - smoothstep(INK_GUARD_FROM 0.80, INK_GUARD_TO 0.92, brightest) x lamps`. Guard and halo both sit inside `if (glow_strength > 0.0)`, so Glade (glow 0) draws exactly as before. `main.gd apply_presentation` sets the new `lamps` uniform (default 0) from the hour's profile. The engine glow is unchanged; the lens's narrower candidate (level 3 0.35, intensity 0.8) was not tried. Taps per window pixel: 54 before (1 + 4 outline + 49), now 13 at night and 5 by day. Files: `shaders/presentation.gdshader`, `scripts/main.gd`, `tests/test_look.gd`, `tests/test_lighting.gd` (fixes 1 and 2 included: `git diff --numstat` 241 lines added, 20 removed in code and tests). Snapshot: `build/backup-feel/23-P5-f/` (= main 6fa5452). At part start the live tree matched it exactly (`diff -rq` on scripts, shaders, tests, data, scenes and tools: empty `f/diff-live-vs-23-P5-f.txt`). Pre-part tree: `artifacts/beautify/23-P5/f/pre-tree`. Backups: `f/backup/`. Everything below is under `artifacts/beautify/23-P5/f/`. 139: - **Deviations from the lens values (plan defaults otherwise):** 140: - (1) `HALO_WARM` (1, 0.62, 0.25). The tonemap clips lantern glass to (255, 255, 149), so the excess hue is plain yellow, and round 0's ring was olive (93, 78, 19; `look/t1-qn-lamp.png`). A hue whose red and green both stand above its blue now leans toward orange by that margin. Cyan, pink and white keep their own hue (tested). 141: - (2) `HALO_T_DAY` 0.97 -> **1.03** (round 2). At 0.97 the day's t is 0.9615, but sunlit white is clipped to luma 1.0, so round 1 ringed it by day. On quarter-day, 37.8% of the art px 1-2 from such pixels brightened by >8 (noise 2.5%), and the inked px beside bright trim went 57.7 -> 64.6 (`metrics/t1-measure-change.txt`, `t1-measure-noise-pre.txt`; round 1 is kept in `round-t1/`). With 1.03, t >= 1 at every day anchor (09:00 1.007, 13:00 1.018, 16:30 1.002) and the halo is skipped. Golden is now 0.9035 (lens 0.8765), 06:45 0.915, 05:45 0.823, 20:45 0.846, night 0.80. This is the one constant to revisit. Round 2's gate alone was a hard switch; fix 1 (below) makes the ring start from 0 at the gate. 142: - **Tests:** test-first `testfirst-test_look.log.out` (592 checks, 24 failures) and `testfirst-test_lighting.log.out` (189, 1). The `HALO_WARM` check (round 0) and the round-2 checks (t_day 1.03, the `t < 1` gate, the day anchors, clipped white by day) were written after their shader edits; no failing log exists for them. After: `after-test_look.log.out` **test_look 563 -> 597 checks, 0 failures** (fix 1: 597 -> 601, `fix1/after-test_look.log.out`); `after-test_lighting.log.out` **test_lighting 186 -> 189, 0 failures**; test_water 410 and test_time 377 unchanged (both read this shader). New checks: 143: - No 7x7 loop. One loop, over the 8 ring-2 taps. 6 texture reads. Ring 2 is exactly the 8 distance-2 offsets, in whole art px; at most 12 halo taps. 144: - The constants; t = mix(1.03, 0.80, lamps); the `t < 1` gate; no halo at the 3 day anchors and the day preset, a halo at the 7 others. 145: - The magnitude form, the warm shift, the halo formula and the ink-guard line, and their order inside the glow gate before the outline multiply. 146: - A GDScript twin of the halo on a 3x2 lantern: an orange ring at 1 px (g/r 0.4-0.75, blue 0), fainter at 2 px, nothing at 3; a hot pixel takes less of the halo, a white-hot pixel none. 147: - No pop: the ring grows monotonically from under +0.002 just over t. Cyan, pink and white keep their hue. Golden plaster, white trim and clipped white throw nothing by day. 148: - `apply_presentation` hands the hour's lamps to the pass in every editor state (canal, Glade, undo, redo). 149: - Engine glow unchanged: intensity 0.9, strength 1, bloom 0, hdr scale 2 and cap 8, levels 0/0.6/1.0/0.7/0. 150: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/23-P5/f/suite`, ended 09:58, AC, Low Power Mode 0, `suite-power.txt`, `suite.out`): **40 suites, 21,341 checks, 0 failures** (P5e's 21,304 + 37). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). The `terrain_flowers.svg` not-found error in the logs is inherited (it is in P5e's logs). 151: - **Changed expectations:** none removed or inverted; every P5f check is new. Fix 1 changed one P5f check of round 2: "a tap throws hue x smoothstep(t, 1, luma)" -> "hue x smoothstep(t, t + HALO_SPAN, luma)", and the twin `p5f_halo_of` follows the shader's `HALO_SPAN`. The old shader had no test of the 49 taps (only "glow_strength > 0.0", which still holds). 152: - **Captures** (`caps.sh`; round 2, one sitting 09:27-09:38, AC, Low Power Mode 0, load about 4.5, `cap-power.txt`): 153: - Sets: `pre-1`/`pre-2` from the pre-tree and `after-1`/`after-2` live, 13 views each: the six journey views plus lane-night, quarter-golden, glade-day, harbour-golden, photo-golden, glade-golden and glade-night. No stalls in round 2. 154: - Filmstrips: `film-pre`/`film-after` (+ `film-pre2`/`film-after2`, quarter-golden), still camera, `--churn=100000`, frames 140-540, for quarter-golden, quarter-2045 and quarter-night. 155: - Orbit: `orbit-pre`/`orbit-after`, quarter-golden with `--orbit=0.05`. 156: - Discs: `discs.sh`, `discs-pre`/`discs-after`. 157: - Framing check (`framesig.js`, `metrics/framing.txt`): `metrics/framing.txt` prints DIFF for 6 of the 13 views (glade-day, glade-night, harbour-night, lane-night, photo-golden, quarter-golden); two move 24 rows (glade-day after-1 column 5 315 -> 291, photo-golden pre-2 column 4 302 -> 326), more than the few rows of TIME noise the script allows, so the framing across the four sets is NOT clean. 158: - *Round 1 capture faults:* several runs stalled up to 4 tries on both trees, and in `round-t1/after-2` the Glade cameras were moved. The discs sitting's disc-hg pre and after are also moved, so `camcheck-pre-1`/`camcheck-after-1` replace them (4 reruns agree). 159: - *Frame change* >8 luma, pre-1/after-1 (noise pre-1/pre-2; after-1/after-2), `metrics/frame-table.txt`: glade-day 0.52 (0.65; 1.11); glade-golden 0.35 (0.53; 0.28); glade-night 2.02 (1.08; 2.56); harbour-day 0.84 (6.72; 6.33); quarter-day 1.04 (2.79; 4.89); styles-day 4.17 (3.28; 3.11; pre-2/after-2 1.40); harbour-lane 2.59 (2.45; 2.25). 160: - The night and golden views change by design (harbour-night 8.9, quarter-night 18.1, lane-night 5.0, quarter-golden 11.7 and harbour-golden 16.4, against noise up to 13.45 in this sitting; the 18.9 on the water is round 1's figure, harbour-golden 18.94 in `metrics/t1-frame-pairs.txt`). Diff maps: `look/diff-gd-pre1-after1.png` (Glade: only chimney smoke), `look/diff-qd-pre1-after1.png` (quarter-day: only rain streaks), `look/diff-qg-pre1-after1.png` and `look/diff-hg-pre1-after1.png` (golden: lamps, the awnings' white edges, the neon rim; no plaster or wall ringed), `look/diff-qn-pre1-after1.png`. 161: - Looked at: `sheet/p5f.png` (lamp, BAR, ln_sign, fish_sign, quarter-golden, quarter-day), `look/qn-lamp.png`, `qn-bar.png`, `ln-sign.png`, `hn-fish.png`, `qn-plaza.png` (string lights and lanterns get orange halos), `qn-windows.png`, `qd-arches.png`, `gd-glade.png`, `qg-awning.png` (barely changed), `orbit-golden-after.png` and `orbit-golden-pre.png` (camera turning: no rings popping on the sunlit stone or the pale arch trim). 162: - Every integer count of `performance.json`, draw calls included, is identical in all views except harbour-lane draw calls 730 -> 737 and `audio.voices`, which differs pre-1 -> after-1 in glade-night (7 -> 5), harbour-lane (8 -> 7), photo-golden (7 -> 9) and quarter-golden (8 -> 9) (run-time dependent), and `f/pre-2/quarter-golden`, which has 1329 draw calls and 1405 rendered objects against 1361 and 1437 in pre-1, after-1 and after-2, so the pre-2 -> after-2 pair (`metrics/measure-change-2.txt`) differs by 32 draw calls. That difference is pre-1's own: pre-1/pre-2 differ the same way, pre-2/after-2 are identical (`metrics/counts-pre1-after1.txt`). **Journey count deltas (declared): none.** 163: - **Acceptance** (`measure.js`; `metrics/measure-change.txt` pre-1 -> after-1, `measure-change-2.txt` pre-2 -> after-2, `measure-noise-pre.txt`, `measure-noise-after.txt`; rec709 luma, rings by art-px distance from the hot pixels): 164: - *Lanterns get a 1-2 art px orange ring:* **met.** Quarter-night lantern, 1 px out: 30.4 -> 78.7 (+48.3), with the colour change at hue 33 (orange). 2 px out: +10.2 (hue 33). 3 px out: -1.8. The noise pair moved 0 at every distance. Over the whole frame, warm emitters 1 px out move +44.7 (change-2 +47.6) in quarter-night, +49.8 in lane-night and +53.0 in harbour-night; 2 px out +10 to +12.6; 3 px out within 1 luma. 165: - *Neon keeps its contrast:* **met for letter against cabinet, with a softer edge.** The letter cores (the protected HDR neon) stay within noise: BAR 244.7 -> 245.0, FISH lane 240.8 -> 240.7, FISH Harbour 238.6 -> 237.8. Letter minus cabinet 3 px away: BAR 195.7 -> 194.7, FISH lane 93.5 -> 97.3 (the old bloom's haze is gone), FISH Harbour 171.0 -> 171.6. The 1-px ink beside the letters is now a coloured rim, so the step from letter to its neighbour pixel shrinks: BAR 196 -> 131, FISH lane 90 -> 48, FISH Harbour 186 -> 137. The letters read bolder and glow, close to the reference's "tight saturated rim". This is a taste call for the stage sheet. 166: - *No ring pops on sunlit plaster at golden hour:* **met.** No plaster or wall is ringed in the golden diff maps or the orbit frames. Rings at golden hour sit on lamps, string lights, awning edges and neon (the 1-2 px zone round hot pixels brightened 28.7% and 28.0% against noise 18.7% and 10.8%). Shimmer (`shimmer.js`, `metrics/shimmer-buildings.txt`, `shimmer-golden-2runs.txt`): over the buildings (300,60-1600,640), the share of the ring zone that changes by >8 between filmstrip frames is golden 3.03 and 4.05 pre against 4.19 and 4.43 live (0.14-0.38 points over the pre maximum), 20:45 3.82 -> 4.31, night 7.71 -> 8.01. Whole-frame flicker is unchanged (0.93 and 0.94 -> 0.91 and 0.89). 167: - *Ink kept beside bright trim by day:* **met in round 2.** On inked art px beside a neighbour of luma >= 0.8: quarter-day 59.0 -> 59.7 (noise +8.1 and +22.3), harbour-day 74.9 -> 75.4 (noise +7.6 and +16.7), styles-day 68.1 -> 70.5 (noise +2.0 and +4.1; change-2 +2.2), harbour-lane 69.0 -> 72.0 (noise +3.1 and +2.1; change-2 +2.3). The 1-2 px zone round clipped white brightens on 1.0-3.5% of its pixels, inside noise (0.4-15%). 168: - *Cost equal or lower:* **met for dense and Harbour** (bench below). The quarter is within noise. 169: - *P5b glints and P5c1 window cores:* window panes do not ring. Only 2 art px of `windows_right` reach t, and lit pane cores stay under 0.8. Glints that reach t (the plaza's 461 hot art px include lamps and glints) take the same orange ring as lanterns (1 px out +44), and the ring re-rolls with the glint. 170: - **P5c2 baseline (sun disc and moon):** the plan's harbour-golden, photo-golden and glade-golden views do not show the disc (sun out of frame; `look/contact-pre-1.png`). Disc views are therefore added, pre and live from one sitting: `discs-*/disc-q0645` (quarter 06:45), `disc-qg-sun145` (quarter golden, `--sun=145`), `disc-gg-sun145` (Glade golden; Glade skips the halo) and `camcheck-*-1/disc-hg-sun145` (Harbour golden). The moon is in frame only in `discs-*/moon-hn-sun145-t4` (Harbour 04:00, `--sun=145`); `moon-hn-sun145` and `moon-qn-sun145` at 01:00 have it out of frame. Crops: `look/moon-pre-after.png` (the moon gains a pale 1-px rim), `look/contact-discs-after.png`. The plain golden views are in `pre-1`/`after-1` as well. 171: - **Bench** (`bench.sh`, `bench-summary.js 1,13`, `bench/summary.txt`, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, rain 0.45, 09:41-09:52, AC, Low Power Mode 0 (load not recorded in any file under `f/`; the bench was noisier than P5e's). Raw p50 medians, paired median, rounds where live was slower: 173: | Scene | Night 01:00 | Day 13:00 | 174: |---|---|---| 175: | dense | 13.49 -> 12.94 ms (-0.33, 0/3) | 13.16 -> 12.63 (-0.52, 0/3) | 176: | Harbour | 7.22 -> 7.01 (-0.17, 1/3) | 6.97 -> 6.54 (-0.43, 1/3) | 177: | quarter | 8.20 -> 8.26 (+0.03, 2/3) | 8.04 -> 8.14 (+0.17, 2/3; p95 -0.03) | 179: Draw calls are identical (2362, 903, 1026). 180: - **Taste notes for the stage sheet:** neon letters lose their 1-px dark outline for a coloured rim (bolder, softer edge); the lantern's dark mullion fills orange; golden-hour lamps and awning edges ring faintly (after fix 1, about half as strongly as round 2); the P5b glints ring orange; the moon gets a pale rim. **Open for the owner:** the golden-hour ring strength beside clipped white. Quarter-golden ring zone >8 brighter is 22.5% after fix 1 (noise 9.8% pre, 10.3% live), against 28.7% in round 2; harbour-golden is 25.8% (round 2 30.3%). `HALO_SPAN` 0.2 sets it. A smaller span gives stronger golden rings but a steeper switch-on near 16:46 and 08:40. The plan default stands until the owner decides (`fix1/look/qg-pre-round2-fix.png`). 181: - **Review round 1 (must-fix) and fix 1** (everything under `f/fix1/`): 182: - *Found:* round 2's day fix was a hard on/off gate. When lamps passes 0.13 (t passes 1, about 16:46 and 08:40), `smoothstep(t, 1, 1.0)` is 1 for clipped sunlit white however close t is to 1, and `halo_of` normalises the hue of any tiny excess. Every pixel beside clipped white jumped from no ring to a full one between two adjacent Time-of-day slider stops (16:45 and 17:00). Reviewer's numbers (`review/pop.txt`, `pop-pre.txt`): ring zone >8 brighter 16:42 -> 16:48 34.6% (quarter) and 27.8% (Harbour), against pre-tree 6.2% and 3.5%; the pink awning stripes lightened (`review/pop-qd-1670-1680.png`). The handoff had tested day only at the 13:00 preset. 183: - *Fixed:* the thrown magnitude is now an absolute excess, `smoothstep(t, t + HALO_SPAN, luma)`, `HALO_SPAN` 0.2 = 1 - `HALO_T_NIGHT`. As t reaches 1 the ring of clipped white goes to 0, so the `t < 1` gate is continuous (the 2D viewport is not HDR, so no pixel is over 1). Night (t 0.80) is unchanged exactly. Ring beside clipped white, by the formula: 16:45 0 (gate), 17:00 +0.0002, 17:15 +0.0014 (0.35/255), 08:30 +0.0002, 08:45 0 (gate); golden 19:30 +0.085 (round 2 +0.18), 06:45 +0.066 (+0.17), 20:45 +0.19 (+0.22), 05:45 +0.21 (+0.22), night +0.24 (+0.24). Golden-hour rings are about half round 2's. 184: - *Tests:* test-first `fix1/testfirst-test_look.log.out` (601 checks, 5 failures: `HALO_SPAN`, the magnitude form, the clock sweep's largest step 0.0674 at 08:36, 0.1125 beside clipped white between 16:42-17:15 and 08:30-08:48, golden +0.180). After: `fix1/after-test_look.log.out` 601, 0 failures; `after-test_lighting` 189, `after-test_water` 410, `after-test_time` 377, all 0 failures. New checks: `HALO_SPAN` = 0.2 = 1 - `HALO_T_NIGHT`; the clock sweep 16:00 -> 19:30 and 09:30 -> 06:45 (1.2-1.5 game-minute steps) climbs monotonically by under 0.004 a step; under 1/255 from 16:42 to 17:15 and 08:30 to 08:48; night full, golden under 0.6 of round 2's. 185: - *Captures* (one sitting 10:09-10:19, AC, Low Power Mode 0, load about 4.5, `fix1/power.txt`; `fix1/popcap.sh` with the reviewer's qd-t and hd-t cameras; `L-*` live, `P-*` pre-tree; 16:42 x2 each, 16:45, 16:48, 17:00, 17:15; 08:30 x2 each (+ `L-c0850`), 08:36, 08:42, 08:45, 09:00). Framing check with `f/framesig.js` (it sits in `f/`, not `f/fix1/`): all agree except `L-a0850/hd-t` (camera moved, zoomed out), replaced by `L-c0850`. `fix1/power.txt` also logs a run named `L-c0850 P-c0850` (10:18:59-10:19:25) that left no `P-c0850` folder; `L-c0850` was rerun alone (10:19:32-10:19:46), and that run is the folder on disk. No `P-c0850` set exists and none is cited. Golden and night: `fix1/golcap.sh` sets `pre-g1`, `live-g1`, `pre-g2`, `live-g2` (quarter-golden, harbour-golden, photo-golden, quarter-2045, quarter-night, lane-night, harbour-night). 186: - *Measured, the reviewer's metric* (`fix1/pop.js`, `pop.txt`; ring zone >8 brighter, quarter / Harbour): live 16:42 -> 16:48 8.65% / 3.92% (pre 2.88% / 4.26%; noise live 6.27% / 1.13%, pre 1.21% / 1.49%); 16:42 -> 17:00 12.65% / 5.66% (pre 9.89% / 3.44%; round 2 66.5% / 73.7%); pre against live at 16:48 0.97% / 3.78% and 17:00 1.74% / 2.09% (round 2 28.5% / 26.7% at 16:48). The morning frames are overcast and their clipped white is mostly rain rings on the water, which move every frame (noise pairs at 08:30 up to 29%), so pop.js cannot read the morning. 187: - *Measured, static white* (`fix1/pop2.js`, `pop2.txt`: seeds clipped in both frames, ring px beside moving white left out): 16:42 -> 16:48 live 2.53% / 3.80%, pre 2.37% / 2.85%; 16:42 -> 17:00 live 10.78% / 5.56%, pre 8.54% / 3.20%; 16:48 -> 17:00 live 5.11% / 3.70%, pre 5.12% / 3.74%; pre against live at 16:42-17:15 0.00-0.45% (quarter) and 0.66-4.73% (Harbour; mean dluma -1.40 to +1.56). Morning, across the 08:40 gate, 09:00 -> 08:30 live 0.00% / 6.90% (13 and 10 static white px), pre 1.29% / 0.85%; pre against live 08:30-09:00 0.00-2.13% / 0.00-4.72%; noise at 08:30 11.1% / 2.1% live, 14.1% / 0.0% pre. Small samples, within noise. 188: - *Looked at:* `fix1/look/awning-qd-P1670-L1670-L1680-L1700.png` (pink awning steady from 16:42 to 17:00) against round 2's `fix1/look/awning-qd-round2-1670-1680-1700.png` (stripes lighten); pink stripe luma (`awning-pink.txt`) live 174.4, 174.1, 172.5 at 16:42, 16:48, 17:00 (pre 176.9, 176.5, 174.6; round 2 174.6, 181.2, 193.3). `pop-qd-L1670-L1680.png` (most-changed window beside white is rain rings only), `pop-qd-L1670-L1700.png`, `pop-hd-L1670-L1700.png`, `qg-pre-round2-fix.png` (golden: a fainter rim than round 2), `L-a0850/hd-t` and `L-b0850/hd-t` (the moved camera). 189: - *Night and golden re-measured* (`fix1/measure-change.txt`, `-change-2`, `-noise-pre`, `-noise-live`): night is unchanged from round 2. Quarter-night lantern 1 px out 30.4 -> 78.6 (round 2 78.7), 2 px +10.0, 3 px -2.0; warm emitters 1 px +47.0 (quarter), +49.6 (lane), +52.9 (Harbour); BAR letters 245 -> 245, letter rim +65.5; FISH lane 240.8 -> 240.7; FISH Harbour 238.7 -> 238.9 (protected neon cores within noise). Golden: ring zone >8 brighter quarter-golden 22.5% and 20.5% (noise 9.8% pre, 10.3% live; round 2 28.7%), harbour-golden 25.8% and 21.3% (noise 10.9%, 4.7%; round 2 30.3%); inked px beside bright trim quarter-golden 50.0 -> 56.7 (round 2 50.3 -> 61.0; noise +3.8 and +13.1). 20:45: 36.9% and 40.4% (noise 1.6%, 6.3%), by design. 190: - *Draw calls:* identical pre and live in every fix1 capture (for example quarter-golden 1361, quarter-night 1008, qd-t 1142 and 1194). No bench: the fix adds one add per tap and changes no tap, pass or draw call; round 2's bench stands. 191: - *Full suite* (`VERIFY_RES=2880x1800 tools/suite_all.sh fix1/suite` in parts a, b, v, 10:21-10:28, AC, Low Power Mode 0, `fix1/suite-power.txt`, `fix1/suite/suite_all.out`): **40 suites, 21,345 checks, 0 failures** (21,341 + 4). The first `--verify` 2880x1800 exited 1 with the known held-raise flake ("Holding the raise brush did not build a hill: 2.05", recorded in this file before; logs kept in `fix1/verify-try1/`); the rerun passed, exit 0 in 33 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks, 0 failures. 1440x900 was not run (the inherited stall). 192: - *Journey count deltas (declared):* still none. 193: - **Recheck after fix 1, and fix 2** (everything under `f/fix2/`): 194: - *Found:* (1) `tests/test_look.gd` kept two round-1 labels saying t = mix(0.97, 0.80, lamps): the P5f header comment and the message of the check on `float t = mix(HALO_T_DAY, HALO_T_NIGHT, lamps);`. The shader and the checks' conditions already used 1.03. (2) The fix-1 Captures bullet placed `framesig.js` under `f/fix1/`; it is `f/framesig.js`. `fix1/power.txt` logs a `P-c0850` set with no folder. (3) Taste point: golden-hour rings beside clipped white are about half as strong as round 2's (quarter-golden ring zone 22.5% against noise 9.8-10.3%; round 2 28.7%). 195: - *Fixed:* (1) The header comment now says smoothstep(t, t + HALO_SPAN, luma) and t = mix(1.03, 0.80, lamps), the check message says "t = mix(HALO_T_DAY, HALO_T_NIGHT, lamps), that is mix(1.03, 0.80, lamps)", and the comment above the magnitude checks says smoothstep(t, t + HALO_SPAN, luma). No check's condition changed, and no shader or script changed. Backup: `fix2/backup/`. (2) The Captures bullet above now gives `f/framesig.js` and says what the `P-c0850` line is. (3) Not changed. This is the owner's call for the stage sheet, and the plan default (fix 1's `HALO_SPAN` 0.2) stands until the owner decides. See the taste notes. 196: - *Tests:* `fix2/after-test_look.log.out` **test_look 601 checks, 0 failures** (same count; only labels changed, so there is no test-first log). Full suite (`fix2/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh fix2/suite`, 10:36-10:42, AC, Low Power Mode 0, `fix2/suite-power.txt`, `fix2/suite/suite_all.out`): **40 suites, 21,345 checks, 0 failures**. `--verify` 2880x1800 exit 0 on try 1 in 33 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks, 0 failures. 1440x900 was not run (the inherited stall). 197: - *Captures, bench:* none. No rendered code changed, so fix 1's captures and measurements and round 2's bench stand. *Changed expectations:* none (labels only). *Journey count deltas (declared):* still none. 198: - **Review (orchestrator, 1 October 2026):** first review (fresh context): one must-fix, the hard on/off day gate at lamps 0.13 (about 08:40 and 16:46), fixed in fix 1 (HALO_SPAN 0.2, continuous) and rechecked, plus a stale-label fix in fix 2. **Second review** (independent, own scripts in `artifacts/beautify/23-P5/f/review2/`, test_look re-run: 601 checks, 0 failures): no must-fix findings; the shader is correct and continuous (swept every minute of the day: largest one-minute step beside clipped white 1.0/255), cost is equal or lower (raw performance.json match summary.txt), P5f called done. Its corrections, recorded here and not hidden: (1) the entry's golden ring-zone percentages (22.5% quarter, 25.8% Harbour) used the lens's stale t (0.877) in the seed; at the shader's real t the 1 px ring is brighter by >8 on 56.4% of quarter-golden pixels (noise 21.6% and 19.6%) and 56.8% in Harbour (noise 14.0% and 6.0%) (`review2/gold.js`); (2) golden-hour rings sit on clipped white trim, awnings and bargeboards, not on lamps (a golden lantern throws 3.6/255, clipped white 21.7/255): sunlit pale bargeboards on the Harbour house lose their ink (ink beside bright trim 63.4 -> 74.6, `review2/hg-house.png`); (3) the check at tests/test_look.gd:905 (sunlit plaster, luma 0.8575) is below t and holds whatever the shader does; (4) UNDECLARED LOOK CHANGE: removing the bloom removes its daytime lift (glow 0.07): day pixels with luma 0.85 or more are 2.9-3.4 darker (quarter-day -2.94, harbour-day -3.02, styles-day -3.37, harbour-lane -3.42; noise 0.3-1.2) and about 25% fewer clipped pixels (16:45 quarter 891 -> 682); (5) harbour-lane rendered_objects 730 -> 813 also differs, the same pre-1 artefact as its draw calls; (6) no 05:45 capture, the protected dawn sky checked by formula (under t). **Open for the owner on the P5 sheet:** golden-hour ring strength on white trim (HALO_SPAN 0.2 is the plan default; a smaller span gives stronger golden rings and a steeper switch-on near 16:46 and 08:40), the neon letters losing their crisp dark outline for a coloured rim (BAR letter-to-neighbour step -33%, FISH lane -47%), the daytime lift gone, and the golden bargeboards. 199: - **Not done:** 200: - No failing log for the warm-shift and round-2 checks (see Tests); fix 1 has one (`fix1/testfirst-test_look.log.out`). 201: - Fix 1 did not redo the round-2 filmstrips, orbit, disc captures or bench (night is unchanged; golden only grew fainter). 202: - The lens's engine-glow candidate was not tried (the plan's default keeps the glow). 203: - No pixels-1 capture. No 1440x900 `--verify`. 204: - No recording or contact strip beyond the six-frame filmstrips. 205: - `pre-tree` is left for the orchestrator. 206: - Known flaky check recorded as known (did not fire). 207: - Platforms: Mac run; Windows not run. 209: **Beautify P5e Island visible with the tray open; text outlines instead of slabs (1 October 2026):** Opus 5.5 on this Mac, overnight with the owner away, so the plan defaults decide (report quick win 14). In a person's editor the camera's `v_offset` lowers the eye by half the height the dock covers, so the town sits in the middle of the window above the dock; photo mode eases it back to 0. The three floating labels (hint, status line, part count) lose their 0.92-alpha slab and get an INK text outline. Files: `scripts/camera_rig.gd`, `scripts/world_view.gd`, `scripts/main.gd`, `scripts/editor_ui.gd`, `tests/test_camera.gd`, `tests/test_ui.gd` (249 lines added, 5 removed). Snapshot: `build/backup-feel/23-P5-e/` (= main eff775d). The live tree matched it exactly before the first edit (`diff -rq` on scripts, shaders, tests, data, scenes and tools: no differences). Pre-part tree: `artifacts/beautify/23-P5/e/pre-tree`. Everything below is under `artifacts/beautify/23-P5/e/`. 210: - **Decisions (plan defaults, owner away):** 211: - D16 (plan default): a fixed offset for the dock only. Opening or closing the part tray does not reframe. The covered height is `EditorUI.dock_cover()`: the window height minus the dock's top, 112 UI px, 0 in photo mode. The offset is `CameraRig.dock_pixels` = -(covered/2)/window_h of the frame, held in art pixels, eased with `DOCK_TAU` 0.18 s (`ease_dock`, which lands exactly on the goal) and written as `art_offset` = whole art pixels x size / viewport_h. 212: - `DOCK_MAX_SHARE` 0.25 caps the covered share. The headless probe exposed this: a window smaller than the dock asked for an offset of 1.75 frames. 213: - Automated runs (`camera_live` false) have 0 unless `--dock-offset` is passed. With the flag the offset is snapped (no ease), so `editor.png` shows the offset. `courtyard.png` is taken in photo mode, so it is 0 either way. The free photo camera always has 0. 214: - Outline: `FLOATING_OUTLINE` 3 in INK, on a `StyleBoxEmpty` with the old 6 px side margins. The value comes from three rounds on the bright Harbour sky (`look/labels-day-o4-o3-o2.png`): 4 filled the counters of the 11 px text, and 2 left grey text with a hairline ring. 215: - **Tests:** test-first logs `testfirst-test_camera.log.out` (17 failures) and `testfirst-test_ui.log.out` (10 failures). After: `after-test_camera.log.out` **test_camera 191 -> 224 checks, 0 failures**; `after-test_ui.log.out` **test_ui 333 -> 345, 0 failures**. New checks: 216: - The maths: 112 of 900 px gives -28 art px of a 450 px view; the cap; whole pixels; the ease is part-way after one frame and exact by the end. 217: - An automated run has 0 and keeps the target in the middle. 218: - `--dock-offset` snaps onto the goal, and the target sits exactly that many whole pixels above the middle. 219: - Ground picks at four points come back on their point to within 1 cm with the offset. A pick taken in the middle of the live ease does the same. 220: - `unproject_position` (which light culling uses) moves every point by the same whole pixels. 221: - Grip sizes do not depend on the offset. 222: - Photo mode gives 0, and the offset comes back on leaving it. The live ease is monotonic, whole-pixel on every frame and takes 3-120 frames. The tray does not move the view. Photo mode eases back to 0. 223: - The labels have no flat backing and an outline of at least 3 in INK. `dock_cover` matches the dock's rect, is unchanged by the tray and is 0 in photo mode. 224: - **Full suite** (`suite.sh` = `VERIFY_RES=2880x1800 tools/suite_all.sh e/suite`, 04:30-04:36, AC, Low Power Mode 0, `suite-power.txt`): **40 suites, 21,304 checks, 0 failures** (P5c1's 21,259 + 45). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 225: - **Changed expectations:** none removed or inverted; all P5e checks are new. The floating labels' slab had no test. 226: - **Captures** (`caps.sh`, one sitting 04:22-04:26, AC, Low Power Mode 0, `cap-power.txt`): `pre-1`/`pre-2` from the pre-tree and `after-1`/`after-2` live, the six journey views without the flag; `dock-1` live with `--dock-offset` (harbour-night, harbour-day, quarter-night); `res--pre`/`-dock` harbour-night at 1100x700, 1440x900 and 2880x1800. The 2880 window came up 2880x1724, so the UI scale was 1.532 under UI_GROWTH 0.8; the other two sizes are at scale 1. All at pixels 2. An earlier round with outline 4 is kept in `round-o4/`; the outline changed after it and nothing else. The cap was added after the sitting and does not act at these sizes (covered share 0.10-0.16). 227: - *Journey unchanged* (`courtyard.png`, frame change >8 luma): pre-1/after-1 harbour-day 2.95%, harbour-lane 1.97, harbour-night 2.80, quarter-day 1.09, quarter-night 2.10, styles-day 0.88. Noise pairs: pre-1/pre-2 1.10, 2.09, 4.48, 0.93, 2.42, 0.11; after-1/after-2 1.00, 1.97, 2.50, 1.07, 2.45, 1.18. All four cross pairs are in `metrics/cross-pairs.txt`. harbour-day sits above its noise pair. Its diff map `look/hd-courtyard-diff-pre1-after1.png` shows only rain rings, rain streaks and chimney smoke, with no edge moved, and the other cross pairs give 1.48-2.30. 228: - In round o4, harbour-night pre/after was 4.6-5.8% against noise 1.3-2.1, all of it the drifting cloud wisps (`round-o4/look/hn-diff-pre1-after1.png`, `elapsed_ms` 3668/3697 pre against 3749/3792 live). In this sitting it falls inside noise. 229: - Every integer count of `performance.json`, draw calls included, is identical in all six views (`metrics/counts-pre1-after1.txt`). **Journey count deltas (declared): none.** 230: - *editor.png* without the flag changes only where the slabs were (`metrics/editor-pre1-after1.txt`). Crops: `look/labels-1100-hint.png`, `labels-1440-hint.png`, `labels-2880-hint.png`, `labels-1100-bottom.png`, `labels-1100-count.png`, `labels-2880-bottom.png`, `labels-day-hint.png` (Harbour sky) and `labels-qday-bottom.png` (quarter-day water). The text reads on night water, on the pale Harbour sky and on quarter-day water. 231: - *The offset* (`shift.js`, `metrics/editor-shift.txt`): the island moves up 56 window px (28 art px) at 1600x1000 in all three views, and at 1100x700 and 1440x900. At 2880x1724 it moves 86 px (43 art px). Each matches -(112 x UI scale / 2). 232: - **Acceptance:** 233: - *Fish Market front, boathouse and boats inside the free area:* **met for the dock (the D16 free area), not met with the tray open.** The numbers are window px at 1600x1000, read off `look/hn-editor-pre-dock-crop.png`, where the dock top is the red line, the 1-row Landscape tray (y 720) is yellow and the 2-row tray (y 620) is green; they are good to about +-10 px. The boats and boathouse foot were already above the dock top (888) before; they now clear it by about 110 px instead of 55. With the Landscape tray open, the Fish Market front is free (about 510-640). The boathouse is about two-thirds above the tray (roof about 605, piles about 775). The boats (about 730-780) stay under the tray. A 2-row tray covers the boathouse as well. Reframing for the tray is the D16 alternative the plan default declined. 234: - *Journey courtyard frames unchanged:* met (above). 235: - *test_ui green:* met, 345 checks, 0 failures. 236: - **Drag harness** (`caps.sh drag`: the A0 Harbour replay, live, windowed at 1600x1000, `drag-plain` and `drag-dock` with the flag): 0 errors both ways. Each run is repeatable (the same digests as round o4), but the digests differ from the first handle step on (`metrics/drag-step-digests.txt`). They match only at `slider-scrub-height` itself, which writes an absolute height; every edit step after it (wall-node-drag, neon-drag, wall-stroke, terrain-stroke-raise, terrain-stroke-paint, undo, redo) differs again. The headless probe (`probe/dock_drag_probe.gd`, `.log.out`) sizes the difference: the dragged wall node lands at 10.9949055 m against 10.9948997 m, a 6 micron float32 difference on an art pixel of about 7.6 cm, and the height drag is digest-identical there. A shift-compare of the two final frames (`look/drag-after-shiftdiff.png`, the dock run lowered 56 px) shows the same scene, apart from the screen-space horizon band and TIME effects. No automated run can ease, so a pick during the ease is only covered by the headless test. 237: - **Bench** (`bench.sh`, `bench-summary.js`, `bench/summary.txt`, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, night, rain 0.45, 04:36-04:41, AC, Low Power Mode 0. Raw p50 medians: dense 13.34 -> 13.33 ms (paired +0.01, live slower 2/3; draw calls 2362 both), Harbour 6.79 -> 6.78 (-0.01, 0/3; 903), quarter 7.02 -> 7.02 (+0.00, 2/3; 1026). No cost beyond noise. A benchmark runs with an offset of 0. 238: - **Taste notes for the stage sheet:** the quarter's top terrace now runs into the top edge of the editor frame (`look/qn-editor-pre-dock.png`; it touched before). Under the perspective lens the offset is exact at the target's depth only; nearer and farther points parallax slightly, which matters only for a zoom pivot away from the target depth. 239: - **Review (orchestrator, 1 October 2026, owner away):** one fresh-context review (254 changed lines), no must-fix findings, no fix pass; acceptance met for the dock. Decided on plan defaults (D16: dock-only, no tray reframing). **Open for the owner on the P5 sheet:** (1) with the Landscape tray OPEN the boats (y about 730-780) still sit under the tray top at 720 and a 2-row tray also covers part of the boathouse; fixing that needs tray reframing, the D16 alternative the plan default declined, so keep D16 or reframe; there is no tray-open editor.png (the smoke run opens with the tray closed); (2) the outlined status line is harder to read than its old slab on bright quarter-day water, and outline 3 looks heavy on the Harbour day sky; (3) the drag-dock digests differ from the plain run by float32 rounding in one measured step (wall-node drag, 6 microns), inferred for the earlier steps, and automated runs without --dock-offset are unchanged (reviewer replay: all 11 digests identical); (4) a free-camera toggle during the 0.5 s ease after entering photo mode jumps by the remaining offset; (5) mid-ease light culling and grips are covered only by the 0.25 s culling refresh. 240: - **Not done:** 241: - No tray-open `editor.png`. Since the 25 September autohide, `--smoke` opens with the tray closed, and the plan names no flag to open it. The tray rects in the look crops are computed from `_fit_tray` at 1600x1000 (inspector closed, Landscape: 11 cards, 1230 x 148). 242: - The drag harness cannot run during an ease (automated runs snap), so it ran with the snapped offset instead. 243: - The tilt-shift band stays centred on the window; it is off by default. 244: - The pre-part tree `e/pre-tree` is left for the orchestrator. 245: - Known flaky check recorded as known (did not fire). 246: - Platforms: Mac run; Windows not run. 248: **Beautify P5c1 Window cores, styled glass and arched interiors (1 October 2026):** Opus 5.5 on this Mac, overnight with the owner away, so the plan defaults decide (report quick win 11, windows only; the sun and moon disc is P5c2). Three changes. (1) `window_body` lifts the middle of a lit pane after the dusk mix: `ALBEDO *= 1.0 + CORE_BOOST * lv * smoothstep(CORE_FROM, 1.0, centre)`, `CORE_BOOST` 0.3 and `CORE_FROM` 0.5 (the plan's values), `lv` the pane's own dusk level (exactly `warmth` outside the evening, so an unlit pane at dusk is not lifted). (2) A lit room's tone is hashed from the window's world position: `Architecture.room_tone(saved * center)` and `room_colour` (the old hash and lerp, moved into two functions; `ROOM_PALE` ffd6a0). It still draws nothing from the rng. Tower slots and cluster bays draw at `Vector3.ZERO` in a moved frame, so before this every one of them had one tone. (3) `ClusterArchitecture.styled_window`'s lit glass takes the same tone: `room_colour(WARM, tone).darkened(0.15)`. (4) A lit recessed room (the canal's big arched and cross-mullioned lights: interior plus 0.30 glass) leans its interior and glass toward `RECESS_AMBER` ff8a30 by `RECESS_TINT` 0.55; the light it casts (emitter and spill pool) keeps the room colour. Files: `shaders/window_body.gdshaderinc`, `scripts/architecture.gd` (constants, `room_tone`, `room_colour`, `window`), `scripts/cluster_architecture.gd` (`styled_window`), `tests/test_dusk.gd`. Snapshot: `build/backup-feel/23-P5-c1/` (= `main` d435f0b; `diff -rq` on scripts, shaders, tests, data, scenes and tools against the live tree at part start showed nothing). Scoped backups: `artifacts/beautify/23-P5/c1/backup/`. Evidence: `artifacts/beautify/23-P5/c1/` (below `c1/`). 249: - **Investigation (plan step 2):** the buff windows are not styled windows. The quarter and dense are all canal; the pale panes (lens: "large arched windows 1020-1050 x 320-380") are `Architecture.window(..., recess = true)`: an unshaded interior back wall at `glow x (0.55, 0.50, 0.42)` behind a 0.30-alpha glass in the room colour (crops `look/inv-qn-arched.png`, `look/inv-lane-full.png`). Lit pixels there read hue 30.4-33.0 at saturation 45-53%, against the `window_body` tower and timber panes at hue 27-31 and 78-88% (`metrics/acceptance-hue.txt`, pre-1). So the tint goes on every lit recessed room, arched or not, which is what makes the two families match. 250: - **Finding, BoxMesh UVs:** `MeshBatch.box` is Godot's BoxMesh, whose UVs lay the six faces out 3 x 2 (front face u 0-1/3, v 0-1/2; probe `probe/uv_probe.log.out`). The report's `centre` on raw UV would have been centred on a face corner, and the existing `reveal` vignette (`min(UV, 1-UV)`) only frames each face's left and top. The core uses the face's own coordinate, `fract(UV * vec2(3.0, 2.0))`; `reveal` is left as it was (changing it would relight every pane). 251: - **Decisions (plan defaults, owner away):** `lv` in place of the report's `warmth` (identical outside the evening). The tint covers all lit recessed rooms (above). Styled glass changes only in `styled_window`, the plan's site; `styled_bay_legacy` (architecture.gd:746), the oriel cheeks (:881), modern ribbon emitters (:4702) and the round tower facets (:4786) keep `WARM.darkened(0.15)`. Glade's windows change with canal's (not exempt in the plan). `RECESS_TINT` is this part's own constant: 0.4 (round `after-t1s`) left the lane's lower recess at hue 31.4, 0.55 (`after-t2s`) gives the numbers below; not pushed further. 252: - **Tests:** test-first `c1/testfirst-test_dusk.log.out` (231 checks, 7 failures: constants, the boost's place, the face centre, `lv`, windows at `Vector3.ZERO` in three frames all one tone, a medieval tower's lit windows 1 colour, the API gate). After: `c1/after-test_dusk.log.out`, **test_dusk 223 -> 273 checks, 0 failures**. New checks: the shader constants and the boost's order; a GDScript twin of the fragment (a full warm pane's centre crosses 1.0 in red only, its rim, an unlit pane and the +X face behave); the tone from the world position (same place through another frame, same colour); rng state after `Architecture.window` (plain and recessed, 5 seeds) and `styled_window` equal to a replica of the old draws; styled glass equal to `room_colour(WARM, tone).darkened(0.15)` and lighting in that colour; a medieval tower's windows vary and two towers or two canal houses at different places draw the same vertex counts; the recessed glass and back wall from the leaned colour, the emitter and spill pool in the room colour (the lamp and pool colour re-check after the emitter shift). 253: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/23-P5/c1/suite`, a/b/v, 03:55-04:01, AC, 100% charged, Low Power Mode 0, `c1/suite-power.txt`): **40 suites, 21,259 checks, 0 failures** (P5b's 21,209 + 50 from test_dusk). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 not run (the inherited stall). 254: - **Changed expectations:** none removed or inverted; all P5c1 checks are new. 255: - **Captures** (`c1/caps.sh`, one sitting 03:45-03:49, pre from `c1/pre-tree` = the snapshot, after live; AC, Low Power Mode 0, `c1/cap-power.txt`): sets `pre-1`, `after-1`, `pre-2`, `after-2`: the six journey views plus `lane-night` (harbour-lane at night), `quarter-night-px1` (`c1/quarter-px1.json`, artwork.json at pixels 1), `quarter-dry`, `glade-night`. Frame change >8 luma, pre-1 against after-1 (noise pre-1/pre-2; after-1/after-2), `metrics/frame-*.txt`: lane-night 7.04% (2.61; 2.00), quarter-night 4.33 (2.33; 2.18), quarter-night-px1 4.64 (2.57; 2.96), harbour-night 4.20 (3.40; 2.52), quarter-dry 2.12 (0.41; 0.43), glade-night 1.36 (0.34; 0.30), harbour-lane 3.91 (2.27; 1.91), quarter-day 1.75 (0.89; 1.13), harbour-day 1.77 (2.49; 3.75), styles-day 0.53 (0.21; 0.39). The diff maps (`look/diff-*.png`) show only window panes (the recessed ones darker and redder, tower and styled panes retoned up or down), plus rain, water and chimney smoke at noise level (`look/hn-blobs.png`: the blobs by the Harbour lamps are smoke, they differ pre-1/pre-2 too). By day the change is the lit rooms' 12% warm share and their interior tint. Looked at: `look/t1-qn-recess.png`, `look/t1-lane-recess.png` (round 1), `look/qn-tower-pre-after-px2.png`, `look/qn-tower-pre-after-px1.png`, `sheet/p5c1-night.png` (lane-night and harbour-night, pre-1 | after-1). 256: - **Acceptance** (`c1/measure.js`, `metrics/acceptance-hue.txt`; `probe/pane_probe.gd`, `metrics/pane-probe.txt`): 257: - *Styled arched windows at hue about 26-31, saturation at least 36%:* met for the recessed (buff) family, on lit pane pixels (luma >= 90; frames excluded). Quarter arches 30.6 / 49.6% -> **28.5 / 62.9%**; quarter right block 32.9 / 44.9 -> **30.5 / 58.9**; the same at pixels 1 30.4 / 48.2 -> 28.8 / 62.3 and 32.7 / 44.5 -> 30.4 / 58.4; lane upper 33.0 / 47.0 -> **30.5 / 61.0**; lane lower 32.0 / 53.0 -> **31.3 / 60.6** (the edge of "about 31"). after-2 repeats after-1 to 0.1. On the lens's whole-rect metric (frames and mullions in the mean) saturation is 26-51% -> 38-58%, all at least 36; its hue (16-29) is dragged by the dark frames. The tower and timber panes stay at hue 27-30, 77-87%. Lit luma p50 drops about 6-12 (quarter arches 154.5 -> 147.9; quarter recess_right 161.6 -> 149.9, -11.7; px1 163.7 -> 153.5, -10.2, `c1/metrics/acceptance-hue.txt`): the deeper amber is darker. 258: - *Window centres above 1.0 in red only with an orange halo at pixels 2 and 1:* **not met visibly at the plan's CORE_BOOST 0.3, not pushed.** The CPU replica (night, 17 x 17 samples over each pane's face) finds 1 of 10 `window_body` panes in the quarter above 1.0 (peak 1.013, red only), 0 of 55 in Harbour (peak 0.789: styled glass is `WARM.darkened(0.15)`, darker again by its new tone), and none in dense, which has no `window_body` panes at all (1,958 lit recessed glass panes, a family the boost does not touch and which never exceeds 1.0). The crops show a soft core (a tower pane's centre column 177 -> 221 in red at x663, edge 204) but no halo at either pixel size. For the orchestrator: at 0.6 the quarter has 5 of 10 above 1.0 (peak 1.247), Harbour still 0 of 55 (peak 0.971), red only in both. 259: - *Only a minority of panes crossing 1.0 on the dense night scene:* met trivially, 0 of 0 `window_body` panes and 0 of 1,958 recessed (above). The probe ran at RECESS_TINT 0.4; its numbers do not depend on it. 260: - **Journey count deltas (declared): none.** Every integer count of `performance.json` is identical pre-1 against after-1 in all ten views (`metrics/counts-pre1-after1.txt`), draw calls included, except `audio.voices` (quarter-dry 5 -> 6), a nested field the comparison skipped; it is run-time dependent. 261: - **Bench** (`c1/bench.sh`, `c1/bench-summary.js`, `c1/bench/summary.txt`, `c1/bench/power.txt`): pre-part tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, night, rain 0.45, 03:50-03:55, AC, Low Power Mode 0. Raw p50 medians: dense 13.37 -> 13.32 ms (paired -0.06, live slower 0/3; draw calls 2362 both), Harbour 6.79 -> 6.78 (+0.00, 2/3; 903), quarter 7.00 -> 7.00 (-0.00, 1/3; 1026). No cost beyond noise. 262: - **Review (orchestrator, 1 October 2026, owner away; reversible by one constant):** one fresh-context review (230+ changed code lines), no must-fix findings, no fix pass. Re-run by the reviewer: test_dusk 273 checks, 0 failures; the final code is covered by the 03:55 suite (40 suites, 21,259 checks, 0 failures). **Ruling on the one missed line:** the core halo (window centres above 1.0 with an orange halo at pixels 2 and 1) is met at neither size at the plan's CORE_BOOST 0.3 (quarter 1 of 10 panes above 1.0, Harbour 0 of 55). The orchestrator waives the line at the honest numbers and does not raise CORE_BOOST overnight: 0.6 would still leave Harbour at 0 of 55, so the raise changes little and risks the protected HDR neon cores. The owner may raise CORE_BOOST (and CORE_FROM) on the P5 sheet. The recessed-window warm tint (hue 28-31, saturation 59-63%) is met. Open for the P5 sheet: the position-hashed room tones darken tower and cluster panes (lit luma p50 quarter tower 150 -> 144, lane tower 109 -> 100, mean luma negative in every night view, as the plan mandates), the recessed rooms are redder and 6-12 luma darker (RECESS_TINT 0.55, owner may want 0.4), a window's colour now follows its world position so moving a house recolours its lit rooms on the next rebuild, four other styled-glass sites keep a single tone, no golden-hour or 20:45 capture. 263: - **Not done:** 264: - The core-halo line is open (above): raising CORE_BOOST is a value change beyond the plan; it waits on the orchestrator or the owner. A halo on styled panes would need about 0.7, or a lift of styled glass itself. 265: - Styled glass outside `styled_window` (four sites above) keeps one tone. 266: - No golden-hour or 20:45 capture (the boost scales with the dusk level, so it is partial there). 267: - The pre-part tree `c1/pre-tree` is left for the orchestrator. 268: - Known flaky check recorded as known (did not fire). 269: - Platforms: Mac run; Windows not run. 271: **Beautify P5b Wet glints inside the lamp pools (1 October 2026):** Opus 5.5 on this Mac, overnight with the owner away, so the plan defaults decide. Rainy night paving now glints under the lamps (report quick win 8). The light-pool shader takes a `wetness` uniform, the paving streaks' `clampf(rain*PAVING_SOAK, 0, 1)`, set on the settled and build variants alike. Everything new sits behind `if (wetness > 0.0)`, ahead of the only TIME term, so dry scenes and Glade at its rain 0 draw the old pools. Inside it, on flat faces only (`wn.y > 0.5`), glint cells light a 1-pixel-wide dash 2-4 pixel rows tall. The cells are world-anchored, view-aligned and sized in art pixels (`GLINT_COLS` 3 across by `GLINT_ROWS` 6 rows, from `fwidth` of the camera-aligned ground coordinates). `GLINT_LIT` 0.3 of the cells are lit at a time, each re-rolled every `GLINT_PERIOD` 1.2 s on its own phase. A lit dash is pushed 0.35 toward white at `GLINT_GAIN` 1.4 x sqrt(falloff), and fades in with the lamps (`GLINT_LAMPS_FROM` 0.15 to `_FULL` 0.5: none at the day's 0.05-0.12). Files: `shaders/light_pool_body.gdshaderinc` (the two wrappers are unchanged), `scripts/world_view.gd` (one `set_material_parameter`), `scripts/architecture.gd` (not in the plan's file list, see below), `tests/test_wet.gd`. Snapshot: `build/backup-feel/23-P5-b/` (equal to the live tree at part start: `diff -rq` on scripts, shaders, tests, data, scenes and tools showed nothing). 272: - **Why `architecture.gd`:** the plaza's lamps are standing `lantern` posts, which had no light pool at all (only an OmniLight), so a pool-shader change alone could not put glints under them. A standing lantern now carries a 3.4 m ground pool with vertex alpha 0. Dry, it adds exactly 0 under `blend_add`. Wet, the shader gives a vertex-alpha-0 pool `WET_POOL` 0.9 x wetness of strength (every other pool keeps its own vertex strength, wet or dry) plus its glints. Eave lanterns and wall-mounted lanterns are unchanged (no ground pool, no glints). 273: - **Review fix (02:30-03:35, after the orchestrator's review):** two must-fix findings. 274: - *Pools hung past edges.* Each standing lantern's 3.4 m pool is a flat quad, one mesh per lantern kind, so it cannot be clipped per lantern. Where a lantern stands within about 1.7 m of a quay or terrace edge, the pool hung past the edge. In the rain its sheen and glint dashes floated over the quay wall and parapet below (`b/review/qn-quayedge.png`, `b/review/qn-wall.png`). The review's probe (`b/review/pool_edge_probe.log.out`) found 17 of the quarter's 24 standing lanterns over a drop of more than 0.25 m or over water. 275: - *The fix:* inside the wet, flat branch only, the pool shader reads the opaque depth (`hint_depth_texture`, the same copy the water and puddles read) and rebuilds the world height of the surface behind each pool fragment, on the fragment's own view ray (Compatibility depth range; orthographic and perspective). The drop below the pool plane gives `ledge = 1 - smoothstep(LEDGE_FROM 0.14, LEDGE_TO 0.22, drop)`. It multiplies every pool's glints and the vertex-alpha-0 lantern pool's wet sheen. The other pools' own strength is untouched, so their sheen is as before, wet or dry. The dry path still reads nothing. 276: - *Why 0.14-0.22 m and not 0.08-0.16:* the first try (0.08-0.16, sets 6-8) faded the pools on flat roof terraces too. Plaza +2.08; the diff `b/look-fix/qn-plaza-after1-after6-diff.png` shows whole flat pools lost. A drop probe (`b/probe-ledge3`, the shader temporarily colouring the drop in bins, crop `b/look-fix/probe-ledge3-plaza.png`) read 0.08-0.12 m on the roof decks under terrace lanterns and under 0.04 m on the quay walk. 277: - *Result:* `b/look-fix/qn-quayedge-pre9-after1-after9.png` (pre, first pass, fix): the quay wall faces under both bench lanterns are clear, and the pools stop at the coping line. The diff maps `b/look-fix/qn-diff-quay-set1-set9.png` and `qn-diff-set9.png` show straight pool edges at every quay and terrace edge, with the overhangs gone; Harbour is in `hn-diff-set1-set9-set10.png`. 278: - *What remains:* `b/look-fix/qn-wall-pre9-after1-after9.png` keeps a few dashes on the parapet coping's top and on the terrace floor behind it, both flat. The rclose row (`b/look-fix/rclose-coping-zoom.png`) also stays. The probe on that view (`b/probe-ledge4`, crop `b/look-fix/probe-ledge4-rclose.png`) shows it is a flat cornice top 0.04-0.22 m below a pool that hangs past a terrace above. Its dashes are kept, or half kept, there: glints on a ledge a little below the lamp's own ground. They do not sit on a vertical face. A taste note for the stage sheet. 279: - **Decisions (plan defaults, owner away):** Glade glints when the user raises its rain, like the paving streaks and P0b's leaf glints; at its rain 0 it is unchanged (tested both ways). Glints fall on any flat face inside a pool, lawns included (the pool cannot tell paving from grass; Harbour's lanterns stand on lawn). The day gate was added after the first sitting showed a few glint pixels at the day's lamps. 280: - **Tests:** test-first `artifacts/beautify/23-P5/b/testfirst-test_wet.log.out` (22 failures: no uniform, no constants, no lantern pool). After: `b/after-test_wet.log.out`, **test_wet 171 -> 209 checks, 0 failures**. The new checks: 281: - The dry ALPHA expression is untouched, and the only TIME term sits behind the gate. 282: - The `wn.y` gate; the view-aligned, `fwidth`-sized cells; the constants; the lamp gate (added after the test-first run). 283: - A GDScript twin of `pool_glint` over whole cells and four times: every dash is 1 column and 2-4 contiguous rows, about 0.3 of the cells are lit, at most 10% of pixels are glints, and every cell re-rolls once per period on spread phases. 284: - wetness = the streaks' value at rain 0 to 1, on settled and build materials; Glade at 0 and at 0.45. 285: - The lantern's flat, alpha-0, centred pool; the light unchanged; the eave form without one. 286: - Affected suites: test_dusk 223, test_adapt 635, test_build_anim 422, test_lighting 186, all 0 failures (`b/after-*.log.out`). 287: - Review fix, test-first: `b/fixfirst-test_wet.log.out` (7 failures), then `b/fixfirst2-test_wet.log.out` (2 failures: LEDGE 0.08-0.16 faded a 0.12 m roof deck). After: `b/fix-test_wet.log.out`, **test_wet 209 -> 218 checks, 0 failures**. The new checks: 288: - The depth sampler, read only inside the wet, flat branch. 289: - The Compatibility depth range and both projections. 290: - The `ledge` expression on the glints and on the alpha-0 sheen; LEDGE_FROM >= 0.13 and LEDGE_TO <= 0.25. 291: - A GDScript twin of the reconstruction against known geometry, orthographic and perspective. It recovers the surface height within 1 cm. A pool over flat ground (0.03 m) or a roof deck (0.12 m) keeps everything; over a 0.3 m wall, a 1.48 m terrace drop or the canal it shows nothing. 292: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/23-P5/b/suite`, a/b/v, 02:18-02:24, AC, 100% charged, Low Power Mode 0): **40 suites, 21,200 checks, 0 failures** (P5a's 21,162 + 38 from test_wet). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). After the review fix (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/23-P5/b/suite-fix`, a/b/v, 03:11-03:17, AC, 100% charged, Low Power Mode 0, `b/suite-fix-power.txt`): **40 suites, 21,209 checks, 0 failures** (+9 from test_wet). `--verify` 2880x1800 exit 0 in 32 s, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 293: - **Changed expectations:** none removed or inverted. One P5b check changed in the fix: `(vcol.a > 0.0 ? vcol.a : WET_POOL*wetness)` -> `(vcol.a > 0.0 ? vcol.a : WET_POOL*wetness*ledge)` (the lantern pool's sheen fades past an edge). The rest are new checks. 294: - **Captures** (`b/caps.sh`, one sitting 02:02-02:08, from the pre-part tree `b/pre-tree` and live; power `b/cap-power.txt`, AC, Low Power Mode 0). Sets: 295: - `pre-1`, `after-1`, `pre-2`, `after-2`: the six journey views plus `quarter-dry`, `lane-night`, `glade-night`, `glade-rain` (Glade at rain 0.45) and P5a's `rclose-night` (wall pools in the rain). 296: - `pre-3`, `after-3`: the averaged views only. 297: - After the lamp gate, the day views of `after-1` and `after-2` were re-captured, with a fresh day set `pre-4`, `after-4`, `pre-5` (02:10-02:13). Night output is unchanged by the gate (smoothstep = 1 at lamps 1). 299: Probes from the three tuning rounds are in `b/probe*`. Crops looked at, in `b/look/`: 300: - `qn-plaza-pre1-after1.png`: warm pools and pale dashes under the plaza lanterns. 301: - `qn-pool-zoom.png`: 1x2-4 dashes and the sheen; the bin and posts occlude it. 302: - `hn-pools-pre1-after1.png`: lawn lanterns. 303: - `lane-night-pre1-after1.png`. 304: - `rclose-tl.png` and `rclose-trunk.png`: glints behind a trunk, correctly occluded; a row along a parapet coping. The review read that row as the same overhang. The fix's probe (`b/look-fix/probe-ledge4-rclose.png`) shows a flat cornice top 0.04-0.22 m below the pool, and the row stays after the fix (see Review fix). 305: - Diff maps: `qd-diff.png` and `hd-diff-set4.png` (nothing at the lantern feet by day), `hn-diff.png`, `glade-rain-diff.png`, `rclose-night-diff.png` (sign and string-light wall pools unchanged). 307: Frame change >8 luma, pre-1 against after-1 (noise pre-1 against pre-2): 308: - quarter-night 4.73% (3.31) 309: - harbour-night 9.07 (7.23) 310: - lane-night 3.53 (2.57) 311: - rclose-night 3.80 (2.01) 312: - glade-rain 4.52 (3.66) 313: - glade-night 0.40 (1.25) 314: - quarter-dry 0.000 (0.023) 315: - Day, set 4 against noise pre-4/pre-5: quarter-day 3.09 (3.35), harbour-day 1.65 (0.87; the diff map shows only chimney smoke and water), harbour-lane 1.90 (2.04), styles-day 0.40 (0.51). 317: Metrics are in `b/metrics/`. 319: Review-fix sitting (`b/run-fix-caps.sh` -> `b/caps-fix.sh`, 03:05-03:10, AC, Low Power Mode 0, `b/cap-power-fix.txt`). Sets `pre-9`/`after-9`, `pre-10`/`after-10` (journey plus the five extra views) and `pre-11`/`after-11` (averaged views). Sets 6-8 are the first try (LEDGE 0.08-0.16), kept for its evidence only. Frame change >8 luma, pre-9 against after-9 (noise pre-9 against pre-10): 320: - quarter-night 8.77% (6.90) 321: - harbour-night 3.46 (8.52) 322: - lane-night 3.84 (2.62) 323: - rclose-night 3.53 (1.50) 324: - glade-rain 5.40 (4.52) 325: - glade-night 0.59 (0.57) 326: - quarter-dry 0.61 (0.94) 327: - quarter-day 2.24 (2.43), harbour-day 3.18 (3.95), harbour-lane 2.37 (2.53), styles-day 2.73 (3.22) 329: This sitting's noise is larger than the first sitting's (quarter-dry 0.94 against 0.023), from water, rain and stars on TIME. Metrics: `b/metrics/frame-pre9-after9.txt`, `frame-noise-pre9-10.txt`, `counts-pre9-after9.txt`, `counts-noise9-10.txt`. 330: - **Acceptance** (first pass `b/measure.js`, `b/metrics/acceptance.txt`; after the review fix `b/measure-fix.js` on sets 9-11, `b/metrics/acceptance-fix.txt`): 331: - *Plaza mean up at least 3 luma:* **after the fix +2.65, missed** (3-run averages 44.97 -> 47.62). The first pass measured +3.16 (44.93 -> 48.09), but that included the sheen floating over the quay wall and the terrace overhangs inside the plaza rectangle. The flat pools are unchanged by the fix (`b/look-fix/qn-diff-quay-set1-set9.png`). Not pushed: raising `WET_POOL` (0.9) would buy the missing 0.35 luma. That is a value change beyond the review findings, so it is left to the orchestrator or the owner. quarter-dry is 45.23 -> 45.23 in all six runs of both sittings. 332: - *Glints under every lamp:* met for every standing lantern in the quarter, Harbour, lane and Glade-rain views (diff maps). Wall-mounted and eave lanterns have no ground pool and get no glints. After the fix every lantern still has its pool and glints on the ground it stands on (`b/look-fix/qn-diff-set9.png`, `hn-diff-set1-set9-set10.png`); only the parts past an edge are gone. 333: - *Lit glint pixels at most 10% of pool pixels:* **4.69%** quarter-night, 1.82% harbour-night, 5.76% lane-night, met. After the fix: **5.76%** quarter-night, 3.62% harbour-night, 5.14% lane-night, met. The pool-pixel denominators differ between sittings with the water noise. Art grid; pool = median brightening >= 2 over three pairs; glint = +25 in 2 of 3 pairs and 15 over its side neighbours. 334: - *1 px wide, 2-4 px tall:* exact by construction and in the twin test. Measured groups are 1 column in 78% (quarter), 93% (Harbour) and 90% (lane), and 2-4 rows in 45/54/73%. After the fix: 1 column in 81/90/89%, 2-4 rows in 38/59/71%. The rest are mostly 1x1: a dash's dimmer end row falls under the threshold at the pool's edge or over bright stone. By eye (`qn-pool-zoom.png`) they are 1x2-4. 335: - *Dry, Glade and wall pools unchanged:* met. quarter-dry's plaza is pixel-identical pre-1 against after-1 (0 of 109,200 px; the whole frame 20 px). glade-night is under its noise. Wall pools are gated by the `wn.y` test, and rclose shows no change on them. After the fix: quarter-dry's plaza is 45.23 in all six runs, and glade-night is 0.59% against its noise of 0.57%. 336: - **Journey count deltas (declared):** **draw_calls and rendered_objects +4** on harbour-night, harbour-day, harbour-lane, quarter-night and quarter-day (889 -> 893, 733 -> 737, 1004 -> 1008), and **+1** on styles-day (469 -> 470). The new lantern pool surface is one per lantern instance group, drawn dry too (it adds 0). Noise pair counts are identical (`b/metrics/counts-*.txt`). Unchanged by the review fix: the after-1 and after-6 counts are identical (`b/metrics/counts-after1-after6.txt`), and pre-9 -> after-9 gives the same deltas. 337: - **Bench** (`b/bench.sh`, `b/bench-summary.js`, `b/bench/summary.txt`, `b/bench/power.txt`). Setup: pre-part tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, night, rain 0.45, 02:13-02:18, AC, Low Power Mode 0. Raw p50 medians: 338: - dense 13.36 -> 13.34 ms (paired -0.08, live slower 1/3; draw calls 2360 -> 2362) 339: - Harbour 6.78 -> 6.79 (+0.01, 2/3; 899 -> 903) 340: - quarter 7.02 -> 7.00 (-0.04, 0/3; 1022 -> 1026) 341: - No cost beyond noise. 342: - After the review fix (`b/bench-fix.sh`, `b/bench-fix-summary.js`, `b/bench-fix/summary.txt`, `b/bench-fix/power.txt`; same setup, 03:17-03:22, AC, Low Power Mode 0): 343: - dense 13.25 -> 13.32 ms (paired +0.06, live slower 3/3; pre rounds spread 0.03; draw calls 2360 -> 2362) 344: - Harbour 6.78 -> 6.78 (+0.01, 2/3; p95 7.30 -> 7.96, slower 3/3; 899 -> 903) 345: - quarter 7.00 -> 7.00 (+0.00, 1/3; 1022 -> 1026) 346: - Reading: dense costs about 0.06 ms (0.5%) at rain 0.45, a little over its round-to-round spread. The cause is the depth fetch and reconstruction per wet pool fragment; the depth copy itself is already made for the water and puddles. Harbour's p95 rose on every round while its p50 did not move. 347: - **Recheck (03:28, escalation pass):** the recheck left three items open. None is a code defect this part can fix within the plan, so no code, test or value changed in this pass, and no suite, capture or bench was re-run. All three stay **open** for the orchestrator or the owner: 348: - *Plaza mean +2.65 against the plan's +3 (open, needs a decision).* The reviewer re-ran `b/measure-fix.js` and got `b/metrics/acceptance-fix.txt` exactly (`b/review/measure-fix-rerun.txt`). Closing the 0.35 luma gap needs a value change: a `WET_POOL` raise above 0.9, re-measured, with the quay walls and parapets re-checked. Under the run's stop rule it is reported, not pushed. To close P5b, either the owner accepts +2.65 or the orchestrator or owner approves a `WET_POOL` raise. 349: - *Glint dashes on flat cornice and coping tops a little below a pool that hangs past a terrace (open, taste note for the stage sheet).* rclose: `b/look-fix/rclose-coping-zoom.png`, reviewer's crop `b/review/rc2-tree.png`. The probe `b/look-fix/probe-ledge4-rclose.png` shows flat tops 0.04-0.22 m below the pool, not vertical faces. Removing them would mean a tighter LEDGE band, which faded flat roof decks in the first try (sets 6-8). That is a design choice for the owner, not a fix. 350: - *Bench watch at the stage close (open).* After the fix, Harbour p95 7.30 -> 7.96 ms (slower 3 of 3) and dense p50 +0.06 ms (3 of 3), both at rain 0.45 (`b/bench-fix/summary.txt`). Recorded above, not changed. 351: - Evidence check: `shaders/light_pool_body.gdshaderinc` was written at 03:23, after the fix's captures (03:05-03:10), suite (03:11-03:17) and bench (03:17-03:22). Its content equals `b/backup/light_pool_body.fix2.gdshaderinc` (03:22:56), and against `b/backup/light_pool_body.fix1.gdshaderinc` it differs only in `LEDGE_FROM`/`LEDGE_TO` (0.08/0.16 -> 0.14/0.22) and comment lines. So sets 9-11, `b/suite-fix` and `b/bench-fix` match the live code. The reviewer's `b/review/recheck-test_wet.log.out` (03:26) reads 218 checks, 0 failures. 352: - **Review (orchestrator, 1 October 2026, owner away; ruling reversible by one constant):** one fresh-context review, two fix rounds (the pools floating over quay walls fixed with a depth-based ledge fade; the handoff no longer says none were seen). Final recheck: both must-fix findings fixed, no code defect open. **Ruling on the one missed line:** the plaza mean is +2.65 luma against the plan's +3 (the first pass reached +3.16 only because sheen floated over the quay walls). The orchestrator accepts +2.65 as the honest number and does NOT raise WET_POOL (0.9) overnight: the glints, not a brighter sheen, carry the look, and a raise would need quay walls re-checked. The owner may raise WET_POOL on the P5 sheet. Open for the owner on the P5 sheet: glint dashes on flat cornice and coping tops just below overhanging pools, glints on lawns inside a pool, no golden-hour or 20:45 capture yet (full-strength glints from lamps 0.5; add quarter-golden and quarter-2045 pairs at the stage close), perspective (photo) cells warp with depth, only 26-65% of measured glint groups (final code, `b/metrics/acceptance-fix.txt` 26.1/51.8/64.5%; the first pass was 33-67%) are exactly 1 x 2-4 px (the rest are dim 1x1 ends). Watch at the stage close: Harbour p95 7.30 -> 7.96 ms and dense p50 +0.06 ms at rain 0.45, slower in 3 of 3 rounds. 353: - **Not done:** 354: - Open acceptance line: the plaza mean is +2.65, short of the +3 target (see Recheck). It waits on an owner or orchestrator decision. 355: - The crawl check is deferred to P8a's recipe: `--filmstrip` needs a scripted edit (`--drag`), so it does not run without one. P8a also puts the 1.2 s re-roll under the motion preference (D17). 356: - Taste notes for the stage sheet: glints on lawns as well as paving; a lantern pool that hangs past a terrace over a flat ledge 0.04-0.22 m lower still glints on that ledge (the rclose cornice, above); the first pass's pools did float over quay walls and parapets in quarter-night, and the review fix removes that; the GPU `fract(sin)` hash at large cell indices showed no pattern in these views. 357: - The pre-part tree `b/pre-tree` is left for the orchestrator. 358: - Known flaky check recorded as known (did not fire). 359: - Platforms: Mac run; Windows not run. 361: **Beautify P5a Bold neon marks instead of the ladder glyph (1 October 2026):** Opus 5.5 on this Mac, overnight with the owner away, so the plan defaults decide (owner D15 as answered: drop the frame tubes on marked blades, no `icon` param). A sign without text no longer draws the four-rung ladder (report quick win 4: 13 of the quarter's 14 blades showed it). It draws one of 12 mark sets picked by `posmod(seed, 12)` (`Architecture.neon_mark_set`): 8 bold 3x5 marks (`NEON_MARKS`: cup, fish, star, moon, arrow, heart, cross, note) drawn by the FONT tube code (the lettering loop is now `glyph_tubes`, shared), and 4 plain tube patterns (`NEON_TUBE_SETS`: double, bar_dots, bar_dash, double_dots; 4 of 12 = 33%, the report's "about 35%"). A blade's mark fills the face width (cell 0.15 m before the size scale: **0.1125 m at size 1**, 2% under the 0.115 m lettering floor, the plan default; 0.15 m at size 2) with a short tube above and below; a bar's mark fills the face height (0.124 m before scale: **0.093 m at size 1**, 0.124 at size 2; the plan did not cover bars, and only growing the cabinet would give more) with a tube run either side. Cabinets, bounds, emitters and the save format are unchanged. Files: `scripts/architecture.gd` (constants and `neon_mark_set`, `neon_mark_cell`, `neon_marks`, `glyph_tubes`; the textless branch of `decoration`), `scripts/world_view.gd` (not in the plan's file list, see below), `tests/test_scene.gd`. `scene_state.gd` and `editor_ui.gd` are untouched (no `icon` param). Snapshot: `build/backup-feel/23-P5/` (= `main` b1e4527; the live tree matched it on scripts, shaders, tests, data, scenes and tools at part start). Scoped backups: `artifacts/beautify/23-P5/a/backup/`. 362: - **Why `world_view.gd`:** the view instances decorations by kind, variant and `seed%4` and draws each group's mesh from an example at `seed%4 + 128`, so in game every textless sign showed one of only four sets (heart, cross, double_dots, note): the first after-capture showed exactly that (`a/probe-after-4sets/sheet-night/courtyard.png`). A textless sign's mark set now joins its instance key, and its example keeps its own seed (two lines). A set never spans two `seed%4` groups (4 divides 12), so no group splits beyond the set. 363: - **Tests:** test-first `artifacts/beautify/23-P5/a/testfirst-test_scene.log.out` (parse error: the API did not exist). After: `a/after-test_scene.log.out`, **test_scene 633 -> 1,710 checks, 0 failures**. New `neon_mark_checks`: 6-8 marks, each a bold 3x5 glyph (at least 6 lit cells, all 3 columns, 4+ rows) and all distinct; the tube share is 30-40%. For both kinds, sizes 1-4 and every set: the set is `posmod(seed, n)`; the drawn tubes are never the pre-P5a ladder (rebuilt in the test from the snapshot code); they stay within the face and inside the rim; the cabinet AABB, the single emitter and `dimensions` equal the pre-P5a values; the same seed draws the same thing; the 12 sets all draw differently; drawing writes nothing into the part. Mark cells are checked at each size, and marked blades reach past the old frame tubes. Lettered signs draw exactly as the old loop did. A negative seed picks `posmod`. The quarter's 13 textless blades show 6 distinct marks and 5 blades with a tube pattern (only 4 tube patterns exist; the 6 and 5 are from the capture and not recorded on disk, the test asserts only textless == 13, at least 6 marks and 2-6 tube-pattern blades). The quarter saves byte-identically (Persistence's format) with no `icon`. The `world_view` keying is checked as source. Those two source checks and the `seed%4` check were added after the four-set probe, so they have no test-first log. 364: - **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/23-P5/a/suite`, a/b/v, 01:40-01:46, AC power, 100% charged, Low Power Mode 0): **40 suites, 21,162 checks, 0 failures**, which is the P4 close's 20,085 + 1,077 from test_scene. `--verify` 2880x1800 exit 0 in 32 s, `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). 365: - **Changed expectations:** none removed or inverted. New checks only. 366: - **Captures** (`a/caps.sh`, one sitting 01:30-01:34, `pre-1` and `pre-2` from the pre-part tree `a/pre-tree`, `after` and `after-2` live; power in `a/cap-power.txt`, AC, Low Power Mode 0). Views: the six journey views plus `sheet-night`/`-day` (the mark sheet `a/sheet/marks.json` from `a/sheet/make_marks.gd`: every set on blades and bars at sizes 1 and 2), `rclose-*` (the right block's signs and BAR) and `tclose-*` (the round tower's signs), close-ups at camera size 8 (the validator's minimum). Crops looked at: `a/look/sheet-night-pre-1.png` (every sign a ladder), `sheet-night-after.png`, `sheet-day-after.png`, `blades-after-night.png`, `bars-after-night.png` (12 different sets, readable at size 1), `quarter-night-pre-1.png`, `quarter-night-after.png` and `quarter-day-after.png` (no ladder left; marks and tube patterns), `rclose-night-after.png` and `tclose-night-after.png`. Frame change >8 luma, pre-1 against after (noise pre-1 against pre-2): 367: - quarter-night 4.59% (2.04) 368: - quarter-day 2.81% (2.04) 369: - rclose-night 5.72% (1.61) 370: - tclose-night 6.13% (1.93) 371: - sheet-night 13.39% (1.84) 372: - harbour-night 2.06% (3.95), harbour-day 1.45% (1.11), harbour-lane 1.99% (2.16), styles-day 0.11% (0.12): within noise at frame level. Harbour's textless signs do change, but they are a few pixels at that zoom. 374: Metrics are in `a/metrics/`. 375: - **Acceptance:** 376: - *No ladder glyph remains:* met. No seed, kind or size draws it (test). No ladder is visible in any after view. 377: - *6-8 bold marks or tube patterns by seed:* 8 marks plus 4 tube patterns. The quarter shows 6 marks and 5 tube patterns on its 13 textless blades. 378: - The fallback (tube-only) was not needed: the sheet reads as distinct signs, not noise. 379: - Taste notes for the stage sheet: 380: - The cup (a martini glass, `111/101/010/010/111`) reads as an "I" at quarter zoom. 381: - The tube patterns put less light on a blade than the ladder did (rclose-night mean luma -1.2). 382: - **Journey count deltas (declared):** quarter-night and quarter-day **draw_calls 969 -> 1004 and rendered_objects 1045 -> 1080** (+35 each: textless signs that shared a `seed%4` mesh now draw their own set). harbour-night, harbour-day, harbour-lane and styles-day are identical (`a/metrics/counts-pre1-after.txt`; noise pair counts identical). 383: - **Bench** (`a/bench.sh`, `a/bench-summary.js`, `a/bench/summary.txt`, `a/bench/power.txt`). Setup: the pre-part tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, night, rain 0.45, 01:35-01:40, AC, Low Power Mode 0. Raw p50 medians: 384: - dense 13.33 -> 13.35 ms (+0.01, 3/3) 385: - Harbour 6.78 -> 6.79 (+0.03, 3/3) 386: - quarter 6.93 -> 6.99 (+0.06, 3/3; draw calls 987 -> 1022) 387: - Dense and Harbour draw calls are unchanged (2360, 899). Harbour p95 is noisy (round spread 1.5 ms). 388: - Reading: the quarter pays about 0.9% for its 35 extra draw calls; dense and Harbour show no change beyond noise. 389: - **Review (orchestrator, 1 October 2026, owner away):** one fresh-context review (329 changed lines), no must-fix findings, acceptance met; reviewer logs test_scene 1710 checks and test_building_tools 687, 0 failures. Ruled: plan defaults D15; the placement-ghost signature that omits the seed (cosmetic) and the source-string guard on the world_view fix are left as notes; tube-pattern blades carry less light, the cup reads as an I at quarter zoom, size-1 bars take 0.093 m cells: all on the P5 sheet. 390: - **Not done:** 391: - The mark sheet was captured in one sitting with two rounds per side (`after`, `after-2`), not two tuning rounds. One retune happened before the final captures: cup and arrow were redrawn after the first sheet, where the arrow read as a second cross. 392: - The pre-part scratch tree `a/pre-tree` and `a/probe/` are left for the orchestrator. 393: - Known flaky check recorded as known (did not fire). 394: - Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 2 October 2026, 10:27 BST, game main at af4df88) 151: ## Orchestrator rulings during the overnight window (1 October 2026, owner away, decisions delegated) 153: | Part | Ruling | 154: |---|---| 155: | P5a | Plan defaults D15 (frame tubes dropped on marked blades, no icon param); size-1 bars take 0.093 m cells. Taste notes (tube-pattern blades carry less light, cup reads as an I) go on the P5 sheet. | 156: | P5b | The plaza line (+3 luma) is waived at the honest +2.65: WET_POOL stays 0.9 (reversible, one constant). Glints on lawns inside a pool and on flat cornice tops are left for the P5 sheet. | 157: | P5c1 | The core halo line is waived at the honest numbers (quarter 1 of 10 panes above 1.0, Harbour 0 of 55 at CORE_BOOST 0.3); no raise overnight (0.6 would still give Harbour 0 of 55 and risks the HDR neon cores). Recessed-window warm tint kept at RECESS_TINT 0.55 (the owner may prefer 0.4). | 158: | P5d | Not started overnight: it changes a Glade look the owner approved on 24 September, so it waits for the owner. | 159: | P5e | Plan default D16 (dock-only, fixed offset, no tray reframing) applied. Open for the owner: with the Landscape tray open the boats still sit under the tray, so keep D16 or reframe for the tray. | 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. 169: | P5c2 | Plan values kept (sun gain 1.6 + 1.2 x low_sun on the brightest channel only, moon x1.3 replacing x1.18). Glade keeps its old disc (P4-6 precedent, via the look glow key). Open for the owner on the P5 sheet: the orange-red rim the engine glow lays on dark silhouettes beside the disc, fainter clouds and smoke over the disc, a modest corona. | SOURCE: handoff-check corrections in the P5 close commit (git diff 0fe2de8 af4df88 -- docs/visual-handoff.md, removed lines marked "-", replacement lines "+"; long lines cut at 600 characters) - - Framing check (`framesig.js`, `metrics/framing.txt`): the 13 views agree across the four sets. + - Framing check (`framesig.js`, `metrics/framing.txt`): `metrics/framing.txt` prints DIFF for 6 of the 13 views (glade-day, glade-night, harbour-night, lane-night, photo-golden, quarter-golden); two move 24 rows (glade-day after-1 column 5 315 -> 291, photo-golden pre-2 column 4 302 -> 326), more than the few rows of TIME noise the script allows, so the framing across the four sets is NOT clean. - - The night and golden views change by design (harbour-night 8.9, quarter-night 18.1, lane-night 5.0, quarter-golden 11.7 and harbour-golden 16.4, against noise up to 13.5 and 18.9 on the water). Diff maps: `look/diff-gd-pre1-after1.png` (Glade: only chimney smoke), `look/diff-qd-pre1-after1.png` (quarter-day: only rain streaks), `look/diff-qg-pre1-after1.png` and `look/diff-hg-pre1-after1.png` (golden: lamps, the awnings' white edges, the neon rim; no plaster or wall ringed), `look/diff-qn-pre1-after1.png`. + - The night and golden views change by design (harbour-night 8.9, quarter-night 18.1, lane-night 5.0, quarter-golden 11.7 and harbour-golden 16.4, against noise up to 13.45 in this sitting; the 18.9 on the water is round 1's figure, harbour-golden 18.94 in `metrics/t1-frame-pairs.txt`). Diff maps: `look/diff-gd-pre1-after1.png` (Glade: only chimney smoke), `look/diff-qd-pre1-after1.png` (quarter-day: only rain streaks), `look/diff-qg-pre1-after1.png` and `look/diff-hg-pre1-after1.png` (golden: lamps, the awnings' white edges, the neon rim; no plaster or wall ringed), `look/diff-qn-pre1-afte - - Every integer count of `performance.json`, draw calls included, is identical in all views except harbour-lane draw calls 730 -> 737. That difference is pre-1's own: pre-1/pre-2 differ the same way, pre-2/after-2 are identical (`metrics/counts-pre1-after1.txt`). **Journey count deltas (declared): none.** + - Every integer count of `performance.json`, draw calls included, is identical in all views except harbour-lane draw calls 730 -> 737 and `audio.voices`, which differs pre-1 -> after-1 in glade-night (7 -> 5), harbour-lane (8 -> 7), photo-golden (7 -> 9) and quarter-golden (8 -> 9) (run-time dependent), and `f/pre-2/quarter-golden`, which has 1329 draw calls and 1405 rendered objects against 1361 and 1437 in pre-1, after-1 and after-2, so the pre-2 -> after-2 pair (`metrics/measure-change-2.txt`) differs by 32 draw calls. That difference is pre-1's own: pre-1/pre-2 differ the same way, pre- -- **Bench** (`bench.sh`, `bench-summary.js 1,13`, `bench/summary.txt`, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, rain 0.45, 09:41-09:52, AC, Low Power Mode 0, load 4.2-5.3 (noisier than P5e). Raw p50 medians, paired median, rounds where live was slower: +- **Bench** (`bench.sh`, `bench-summary.js 1,13`, `bench/summary.txt`, `bench/power.txt`): pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync`, 1920x1080, rain 0.45, 09:41-09:52, AC, Low Power Mode 0 (load not recorded in any file under `f/`; the bench was noisier than P5e's). Raw p50 medians, paired median, rounds where live was slower: - - *Looked at:* `fix1/look/awning-qd-P1670-L1670-L1680-L1700.png` (pink awning steady from 16:42 to 17:00) against round 2's `look/awning-qd-round2-1670-1680-1700.png` (stripes lighten); pink stripe luma (`awning-pink.txt`) live 174.4, 174.1, 172.5 at 16:42, 16:48, 17:00 (pre 176.9, 176.5, 174.6; round 2 174.6, 181.2, 193.3). `pop-qd-L1670-L1680.png` (most-changed window beside white is rain rings only), `pop-qd-L1670-L1700.png`, `pop-hd-L1670-L1700.png`, `qg-pre-round2-fix.png` (golden: a fainter rim than round 2), `L-a0850/hd-t` and `L-b0850/hd-t` (the moved camera). + - *Looked at:* `fix1/look/awning-qd-P1670-L1670-L1680-L1700.png` (pink awning steady from 16:42 to 17:00) against round 2's `fix1/look/awning-qd-round2-1670-1680-1700.png` (stripes lighten); pink stripe luma (`awning-pink.txt`) live 174.4, 174.1, 172.5 at 16:42, 16:48, 17:00 (pre 176.9, 176.5, 174.6; round 2 174.6, 181.2, 193.3). `pop-qd-L1670-L1680.png` (most-changed window beside white is rain rings only), `pop-qd-L1670-L1700.png`, `pop-hd-L1670-L1700.png`, `qg-pre-round2-fix.png` (golden: a fainter rim than round 2), `L-a0850/hd-t` and `L-b0850/hd-t` (the moved camera). -- **Drag harness** (`caps.sh drag`: the A0 Harbour replay, live, windowed at 1600x1000, `drag-plain` and `drag-dock` with the flag): 0 errors both ways. Each run is repeatable (the same digests as round o4), but the digests differ from the first handle step on (`metrics/drag-step-digests.txt`). They match again after `slider-scrub-height`, which writes an absolute height. The headless probe (`probe/dock_drag_probe.gd`, `.log.out`) sizes the difference: the dragged wall node lands at 10.9949055 m against 10.9948997 m, a 6 micron float32 difference on an art pixel of about 7.6 cm, and the heigh +- **Drag harness** (`caps.sh drag`: the A0 Harbour replay, live, windowed at 1600x1000, `drag-plain` and `drag-dock` with the flag): 0 errors both ways. Each run is repeatable (the same digests as round o4), but the digests differ from the first handle step on (`metrics/drag-step-digests.txt`). They match only at `slider-scrub-height` itself, which writes an absolute height; every edit step after it (wall-node-drag, neon-drag, wall-stroke, terrain-stroke-raise, terrain-stroke-paint, undo, redo) differs again. The headless probe (`probe/dock_drag_probe.gd`, `.log.out`) sizes the difference: th - - *Styled arched windows at hue about 26-31, saturation at least 36%:* met for the recessed (buff) family, on lit pane pixels (luma >= 90; frames excluded). Quarter arches 30.6 / 49.6% -> **28.5 / 62.9%**; quarter right block 32.9 / 44.9 -> **30.5 / 58.9**; the same at pixels 1 30.4 / 48.2 -> 28.8 / 62.3 and 32.7 / 44.5 -> 30.4 / 58.4; lane upper 33.0 / 47.0 -> **30.5 / 61.0**; lane lower 32.0 / 53.0 -> **31.3 / 60.6** (the edge of "about 31"). after-2 repeats after-1 to 0.1. On the lens's whole-rect metric (frames and mullions in the mean) saturation is 26-51% -> 38-58%, all at least 36; i + - *Styled arched windows at hue about 26-31, saturation at least 36%:* met for the recessed (buff) family, on lit pane pixels (luma >= 90; frames excluded). Quarter arches 30.6 / 49.6% -> **28.5 / 62.9%**; quarter right block 32.9 / 44.9 -> **30.5 / 58.9**; the same at pixels 1 30.4 / 48.2 -> 28.8 / 62.3 and 32.7 / 44.5 -> 30.4 / 58.4; lane upper 33.0 / 47.0 -> **30.5 / 61.0**; lane lower 32.0 / 53.0 -> **31.3 / 60.6** (the edge of "about 31"). after-2 repeats after-1 to 0.1. On the lens's whole-rect metric (frames and mullions in the mean) saturation is 26-51% -> 38-58%, all at least 36; i -- **Journey count deltas (declared): none.** Every integer count of `performance.json` is identical pre-1 against after-1 in all ten views (`metrics/counts-pre1-after1.txt`), draw calls included. +- **Journey count deltas (declared): none.** Every integer count of `performance.json` is identical pre-1 against after-1 in all ten views (`metrics/counts-pre1-after1.txt`), draw calls included, except `audio.voices` (quarter-dry 5 -> 6), a nested field the comparison skipped; it is run-time dependent. -- **Review (orchestrator, 1 October 2026, owner away; reversible by one constant):** one fresh-context review (230+ changed code lines), no must-fix findings, no fix pass. Re-run by the reviewer: test_dusk 273 checks, 0 failures; the final code is covered by the 03:55 suite (40 suites, 21,259 checks, 0 failures). **Ruling on the one missed line:** the core halo (window centres above 1.0 with an orange halo at pixels 2 and 1) is met at neither size at the plan's CORE_BOOST 0.3 (quarter 1 of 10 panes above 1.0, Harbour 0 of 55). The orchestrator waives the line at the honest numbers and does no +- **Review (orchestrator, 1 October 2026, owner away; reversible by one constant):** one fresh-context review (230+ changed code lines), no must-fix findings, no fix pass. Re-run by the reviewer: test_dusk 273 checks, 0 failures; the final code is covered by the 03:55 suite (40 suites, 21,259 checks, 0 failures). **Ruling on the one missed line:** the core halo (window centres above 1.0 with an orange halo at pixels 2 and 1) is met at neither size at the plan's CORE_BOOST 0.3 (quarter 1 of 10 panes above 1.0, Harbour 0 of 55). The orchestrator waives the line at the honest numbers and does no -- **Review (orchestrator, 1 October 2026, owner away; ruling reversible by one constant):** one fresh-context review, two fix rounds (the pools floating over quay walls fixed with a depth-based ledge fade; the handoff no longer says none were seen). Final recheck: both must-fix findings fixed, no code defect open. **Ruling on the one missed line:** the plaza mean is +2.65 luma against the plan's +3 (the first pass reached +3.16 only because sheen floated over the quay walls). The orchestrator accepts +2.65 as the honest number and does NOT raise WET_POOL (0.9) overnight: the glints, not a bri +- **Review (orchestrator, 1 October 2026, owner away; ruling reversible by one constant):** one fresh-context review, two fix rounds (the pools floating over quay walls fixed with a depth-based ledge fade; the handoff no longer says none were seen). Final recheck: both must-fix findings fixed, no code defect open. **Ruling on the one missed line:** the plaza mean is +2.65 luma against the plan's +3 (the first pass reached +3.16 only because sheen floated over the quay walls). The orchestrator accepts +2.65 as the honest number and does NOT raise WET_POOL (0.9) overnight: the glints, not a bri -- **Tests:** test-first `artifacts/beautify/23-P5/a/testfirst-test_scene.log.out` (parse error: the API did not exist). After: `a/after-test_scene.log.out`, **test_scene 633 -> 1,710 checks, 0 failures**. New `neon_mark_checks`: 6-8 marks, each a bold 3x5 glyph (at least 6 lit cells, all 3 columns, 4+ rows) and all distinct; the tube share is 30-40%. For both kinds, sizes 1-4 and every set: the set is `posmod(seed, n)`; the drawn tubes are never the pre-P5a ladder (rebuilt in the test from the snapshot code); they stay within the face and inside the rim; the cabinet AABB, the single emitter a +- **Tests:** test-first `artifacts/beautify/23-P5/a/testfirst-test_scene.log.out` (parse error: the API did not exist). After: `a/after-test_scene.log.out`, **test_scene 633 -> 1,710 checks, 0 failures**. New `neon_mark_checks`: 6-8 marks, each a bold 3x5 glyph (at least 6 lit cells, all 3 columns, 4+ rows) and all distinct; the tube share is 30-40%. For both kinds, sizes 1-4 and every set: the set is `posmod(seed, n)`; the drawn tubes are never the pre-P5a ladder (rebuilt in the test from the snapshot code); they stay within the face and inside the rim; the cabinet AABB, the single emitter a -- **Review:** filled by the orchestrator. +- **Review (orchestrator, 1 October 2026, owner away):** one fresh-context review (329 changed lines), no must-fix findings, acceptance met; reviewer logs test_scene 1710 checks and test_building_tools 687, 0 failures. Ruled: plan defaults D15; the placement-ghost signature that omits the seed (cosmetic) and the source-string guard on the world_view fix are left as notes; tube-pattern blades carry less light, the cup reads as an I at quarter zoom, size-1 bars take 0.093 m cells: all on the P5 sheet. - - Glade shows the new tower row (by plan until P5d). + - Glade shows the new tower row (by plan until P5d; the owner dropped P5d on 1 October 2026, so Glade keeps the tower). - - The owner has not seen this combined sheet. **The M3 exports** (owner D8: after the P3 and P4 sheets) are not built. + - 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). -- **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). +- **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). -- **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 r +- **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 r ADDED 5 October 2026: the owner's answer on the P5 sheet, from a later frozen copy of artifacts/beautify/owner-decisions.md (Frozen 5 October 2026 at 17:45 BST (game main at 648feab).) 175: ## P5 sheet answered (owner, 2 October 2026, in chat) 177: "All P5 changes approved", after the seven open taste decisions were listed with their sheet pages (`artifacts/beautify/23-P5/close/sheet/`, `close/look/rc-box.png`). Kept as built: P5a neon marks (cup reading as an I, tube-pattern blades slightly dimmer) and the dim small box sign on the quarter's right block; P5b plaza glints at +2.65 (WET_POOL 0.9; lawn and cornice glints left); P5c1 window cores (CORE_BOOST 0.3, RECESS_TINT 0.55; plan line waived); P5e dock-only offset (D16; boats still sit under the open Landscape tray); P5f halo rings (HALO_SPAN 0.2, the neon outline now a coloured rim); P5c2 disc gains (1.6 / 1.2 / 1.3, orange-red rim beside the 06:45 disc, fainter moon cloud dither). The P5 stage is signed off.