Chapter 19: Finding the town in the dark Frozen 28 September 2026. Implemented, reviewed and merged work; decisions by the owner (the project's creator) are in the second source. Original source terminology is preserved; workflow run IDs and local paths are removed. Line numbers refer to the frozen copies. Part codes: P1a night sky and water, P1b twilight skylight and foliage night floor, P1c night fills and toe grade, P1c2 retune after the first review (G1), P1d quays lit by the water, P1e tide line and algae band, P1f painted metal. G1 and G2 are the two review points. The P1 stage close was still running and is not quoted. SOURCE: docs/visual-handoff.md (frozen copy, lines as of 28 September 2026) 246: **Beautify P1a Indigo-violet night sky and water (26 September 2026):** Opus 5.5 on this Mac turned the canal night's sky, horizon, haze, clouds, far city and water from blue-teal to indigo-violet (report quick win 1). This is a data-only change in the lighting profile. Key light, energy, fill, ambient, windows and lamps are unchanged. 05:45 and 20:45 change only their water. Branch `beautify/P1a` from `main` fc8650e. Files: `scripts/lighting_profile.gd` (Night preset, 22:00, 05:45 and 20:45 anchors, header note), `tests/test_time.gd` (new `palette_checks`), `docs/visual-handoff.md` (this entry). Snapshot: `build/backup-feel/19-P1/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/19-P1/a/pre-tree/`. Evidence: `artifacts/beautify/19-P1/a/`. Power: mains, Low Power Mode 0, battery 100% for captures and suite (`cap-power.txt`, `suite-power.txt`). 247: - **Values.** Night: sky `20285e`, horizon `36366a`, haze `322e5e`, cloud `4d4a7c`, city `161a38`, water `1b1f40`. 22:00: sky `1c2450`, horizon `4a4278`, haze `3a3463`, cloud `55527e`, city `1c1f3c`, water `1c2046`. Water at 05:45 is `1f2a52` and at 20:45 `2c2c58`. 248: - Round 0 used the report's colours exactly (`round0/`). The rendered zenith came out at 44.6 (target 45-55) and the horizon band at 75.5 (target 60-75). By eye, the harbour water was a saturated ultramarine (`#19205e`, 74% saturation), where the reference's water is greyer (about 49-58%). 249: - One value round changed four report colours: Night zenith `1c2452` → `20285e`, Night horizon `3a3a70` → `36366a`, Night water `1a1f4a` → `1b1f40`, 22:00 water `1b2250` → `1c2046`. The 22:00 zenith is `1c2450`, not the report's `1a2450`, because `1a2450` has hue 228.9, outside the plan's 230-245. 250: - **Tests.** 251: - Test-first log: `a/P1a-testfirst-test_time.log(.out)`, 176 checks, 9 failures (every new P1a check that has to fail before the change). 252: - Affected suites after the change (`a/affected.sh`; logs in `a/suites/r1/` for round 0 and `a/suites/r2/` for the final values): all pass. test_time 176, test_look 271, test_dusk 136, test_water 126, test_foliage 121, test_lighting 155, test_sun 72. 253: - Full suite (`VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/a/suite`, parts a, b and v, 03:52-03:58): **40 suites, 18,465 checks, 0 failures** (P0 close: 18,447, plus the 18 new checks). `--verify` at 2880x1800: exit 0, `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun: 5,549 checks, 0 failures. The known flaky check did not fire. 1440x900 was not run; its stall is inherited from before this part. 254: - New checks, at Night and at 22:00: 255: - sky hue and water hue are 230-245; 256: - horizon and haze hue are 230-250; 257: - the city is darker than the haze; 258: - the zenith is lighter than before and the horizon no brighter; 259: - key, energy and windows are unchanged. 260: - New checks at 05:45 and 20:45: sky, horizon, haze, cloud and city are unchanged, and the water hue moves from about 211 to at least 225. The `nearest_preset` pairs and the 3%-per-step cycle limit still hold. 261: - **Changed expectations:** none. No existing assertion changed. `test_time:177` (fog colour equals the haze) still holds, so the fog-colour fallback was not needed. 262: - **Captures** (`a/caps.sh`, one sitting 03:46-03:50, in this order: `before-1` from the pre-tree, `after-1`, `before-2` from the pre-tree, `after-2`; 11 views, all on the first try): quarter-night, harbour-night, lane-night, quarter-dry, glade-night, quarter-0545, quarter-2045, quarter-2200, quarter-2100, harbour-2100, quarter-day. Folder names follow the kit's `//`. Metrics are in `a/metrics/` (`measure-*`, `ab-b1-a1`, `noise-before`, `noise-after`, `frame-b1-a1`, `counts.txt`, `far-walls.txt`). Sheets, all looked at: `a/sheet/p1a-1-night.png` (colour and grey), `p1a-2-hours.png`, `p1a-3-regions.png`; the round-0 and round-1 sheets are `r0-*` and `r1-*`. 263: - Draw calls and every integer count are identical on all 11 views. 264: - quarter-day: whole frame changed 4.1%, against noise pairs of 5.3% and 3.7%. Unchanged. 265: - quarter-dry: whole frame changed 34.6% (noise 0.1% and 1.9%). quarter-night: 36.9% (noise 15.4% and 8.0%). harbour-night: 74.3% (noise 6.4% and 2.9%). 266: - **Acceptance** (before-1 / before-2 → after-1 / after-2; the noise pairs agree to within 0.2): 267: - quarter-night zenith p50: 31.1 → **49.9** (target 45-55). Hue 220.9 → **233.2** (target 230-245). 268: - quarter-night horizon band: 88.8 → **72.7** (target 60-75, and not brighter than before). 269: - Skyline against the band: skyline 60.6 → 48.5; skyline minus band −28.2 → −24.2, so the blocks stay darker than the haze. The data check also passes: city luminance 0.11 against haze 0.20. 270: - harbour-night: zenith 27.3 → 48.2, band 87.7 → 69.1. Open water `#0b2744` → `#1a2155`, hue 210 → 234, p50 33.1 → 34.3. 271: - quarter-dry: zenith 25.0 → 46.1, band 87.7 → 71.1. 272: - Far walls under the depth fog show no magenta cast. Wall-only patches (`far-walls.txt`) keep their colour exactly: 273: - harbour tower (850-880, 190-240) `#211b15` → `#211b15`; 274: - harbour house (800-840, 250-300) `#262c33` → `#262c33`; 275: - quarter back row `#1b2126` → `#1c2126` (noise-level). 276: 277: The kit regions `far_houses` (hue 248 → 263) and `back_row` (218 → 224) move only because their sky and water pixels changed; their p50 is unchanged (28.8, 18.4). 278: - 05:45 and 20:45 are unchanged except their water: zenith, band, skyline and `sky_right` are equal to within 0.3 luma (05:45 zenith 52.1 → 52.1, 20:45 band 153.0 → 153.0). The water under the boat (`mirror_seam` hue) moves 219 → 242 at 05:45 and 231 → 270 at 20:45. 279: - quarter-night below luma 0.1: 47.0 → 47.2%. That is P1c's line, not this part's. 280: - Mirror columns (a protected look): p50 35.5 → 34.5 / 35.0. The orange, cyan and pink streaks are all there at 2x. 281: - Glade night was judged too. Sky and water are now deep royal blue (water `#022948` → `#0f235a`, saturation 96 → 84%) and the zenith is 23.4 → 46.4. For the owner at G1. 282: - **Carried from P0 (orchestrator ruling):** BANK was not changed. 283: - On harbour-night, bank minus open water went **4.1 → 2.9** and bank minus water below **3.2 → 4.5** (target at least 8 each). Both still fail, and the open-water margin is smaller than before. 284: - Bank over lawn: 0.861 → 0.861 (target at most 1): holds. 285: - P0e seam fade at night (quarter-night x150-850 y880-1000): no step line at 2x, colour or grey. The mirrored violet sky shifts the region's hue from 221 to 263. Mean luma 48.9 → 46.8 (noise 48.9 / 48.5); share below 0.1 35.4 → 38.0%. quarter-dry: mean 49.3 → 47.8. 286: - **Journey:** no count deltas expected. The sky and water change on harbour-night and quarter-night; harbour-day, harbour-lane, quarter-day and styles-day are expected to be unchanged. 287: - **Bench:** none (data only, per the plan; draw calls identical). 288: - **Review:** Opus 5.5 in a fresh context [run ID removed]: acceptance met, no must-fix. Notes carried to the G1 sheet: Glade night is now a saturated royal blue (zenith hue 222 at 88% saturation, water `0f235a`), close to the canal night, which touches the protected Glade/Canal separation; 21:00 water turns violet (harbour-2100 `2f4667` -> `3a3871`); the P0c night bank line is worse on open water (4.1 -> 2.9; below 3.2 -> 4.5), so it goes to the owner at G1 per the orchestrator ruling. The orchestrator aligned the `lighting_profile.gd` comment (band 72.7-74.1 across sittings) and judged `p1a-1-night.png` by eye: the indigo-violet night reads, the Harbour and Glade water is strongly saturated, left for P1c tuning and G1. Merged to `main` on 26 September 2026. 289: - **Not done / open:** 290: - The P0c night bank line still fails (above). It goes to the owner with the G1 sheet, per the ruling. 291: - For G1: the water saturation (lowered once by eye, still more saturated than the reference); the Glade night blue; the four value-round departures from the report's colours. 292: - 22:00 horizon band is 98.0 → 90.2, above 75. The plan's band target is for Night only. 293: - **Platforms:** Mac run; Windows not run. ... 189: **Beautify P1b Twilight skylight and foliage night floor (26 September 2026):** Opus 5.5 on this Mac added a twilight skylight for the handover at 21:00 and 06:10 (report Defect 8; remaining-work item 11) and a night floor on foliage (report quick win 2, step 3). The skylight follows the plan's formula. The floor needed a different value from the report's, for a reason measured below. Branch `beautify/P1b` from `main` 19dc5b5. Files: `scripts/world_view.gd` (`TWILIGHT_SKY` 0.3 and `TWILIGHT_TINT` 0.4, `sync_profile`; after the review also the leaf's `key_light` and `key_direction`), `shaders/foliage.gdshader` (`NIGHT_FLOOR`, `NIGHT_FLOOR_FROM`, `NIGHT_FLOOR_LEAF`, `NIGHT_FLOOR_SKY`, `NIGHT_FLOOR_LOBE`, `night_floor_share`, the key allowance; `NIGHT_FLOOR_UP` removed in the fix), `tests/test_time.gd` (new `twilight_checks`), `tests/test_foliage.gd` (new `night_floor_checks`), `tools/art_regions.json` (a `glade-2100` view alias), `docs/visual-handoff.md` (this entry). Snapshot: `build/backup-feel/19-P1-b/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools, so nothing unexpected had changed. Scratch tree: `artifacts/beautify/19-P1/b/pre-tree/`. Evidence: `artifacts/beautify/19-P1/b/`. Power: mains, Low Power Mode 0, battery 100% for all captures and the suite (`cap-power.txt`, `suite-power.txt`). 190: - **Twilight skylight.** While the key light changes body, the ambient energy is `fill x (1 - GOLDEN_SHADE x golden) + 0.3 x (1 - lit)`, and the ambient colour is `ambient.lerp(horizon, 0.4 x (1 - lit))`. `lit` is 1 at all ten anchors, so every preset and anchor is exactly what it was. The term is non-zero for about 42 minutes of game time a day: two spells of about 21 minutes, centred on 21:00 and 06:10. 191: - **Glade decision (for G1): Glade gets the twilight term too**, with the same values. Its handover went just as dark (glade-2100 crowns `#180b17`), and the look's own fill scaling still applies. It is tested for both looks, and `glade-2100` is captured. 192: - **Review fix (26 September 2026, 04:40-05:20; supersedes the floor values, crown numbers and look notes below it where they differ).** The orchestrator ruled four must-fix items; this is what was asked and what changed. Evidence: `artifacts/beautify/19-P1/b/fix/`. At fix start the live tree differed from `build/backup-feel/19-P1-b/` only in the five P1b files; nothing unexpected. Backups of the reviewed files: `fix/backup/`. 193: 1. *Asked:* the floor was added on top of the lit colour, so lamp-lit crowns, bushes and ivy went grey-teal; make it a floor. *Changed:* the fragment adds only the share of the floor the leaf's own lift leaves owed (`night_floor_share`, in luma), and every light of the base pass (lamps and the rim fill) takes that share back in `light()` as its light covers it (a negative `SPECULAR_LIGHT`, telescoping over the lights), so a leaf lit past the floor ends at exactly its lit colour and a dimmer one at the floor's luma. Diagnosis on the way (`fix/diag/`): the lit bushes the review cited are lit mostly by the key light (the moon), which casts shadows and so is drawn in an additive pass of its own after the sRGB encode (confirmed in the engine's Compatibility scene shader and by `dbg-pass`: a negative there writes nothing). So `WorldView.sync_profile` now gives the leaves `key_light` (linear colour x energy) and `key_direction`, and the fragment allows for the key: a face turned to it owes the floor less by the key's own light, as `light()` gives it unshadowed, compared in display values (`dbg-scale`: the uniform equals `LIGHT_COLOR/PI`, foliage identical). A variant that also made up the floor in the key's pass where the key is shadowed (r3, r4) brightened lamp-lit bushes by about 20 luma and was dropped. *Numbers* (`fix/metrics/lit-foliage.txt`, before-1 / before-2 -> after-1 / after-2; reviewed after in brackets): Canal lamp-lit crown bottom right of qn-right, green pixels above luma 30: 37.3 / 37.3 -> 36.7 / 36.7 (52.9); the bush below it 42.9 / 42.3 -> 42.1 / 42.1 (52.6); Glade pixels above luma 55: bottom bush 83.7 -> 83.7, flower bush 91.1 -> 91.1, tower ivy 83.7 -> 83.7 (84.5), all exact (the Glade noise pair is 0 there). 194: 2. *Asked:* crown tops stayed near-black above a lifted underside and the crowns read flat; tops get at least the floor, keep the lobe split, fix the "moonlit tops" comment. *Changed:* the upturned-face gate is gone; the floor is a skylight, `NIGHT_FLOOR_SKY` = under 1.0, side 1.15, top 1.3 (steps in the banded look, smooth in Glade); it keeps the leaf clusters and the dark outline round each clump, and each lobe takes its own value (`NIGHT_FLOOR_LOBE` +-12%, keyed on the lobe seed). The outline is left out of the key estimate (the shadow map mostly shadows a silhouette), so the crown's top rim keeps its outlined floor. The comment is rewritten. *Numbers:* the review's column x=312 of quarter-night, rows 189-197: (0,4-6,2-4) before and reviewed -> (2,21,20) to (2,25,24) in `fix/after-1` and `after-2`, now above the underside at row 260 (1,16,16). The moonlit top lobe keeps its before colour (tree top pixels 41.5 / 38.4 -> 39.6 / 39.6). By eye (`fix/review/qn-tree-4x.png`, `qn-tree-4x-reviewed-vs-fixed.png`): moonlit top lobe, darker lobes of different values below, dark outlines kept. Left for the reviewer: art-grid specks on quarter-night crowns 0.048 -> 0.242% (reviewed 0.121%), crowns_low 0 -> 0.322% (from the lobe values and faces turned to the moon but in its cast shadow, which get less floor); black art pixels 65.8 -> 16.4% (reviewed 3.1%). 195: 3. *Asked:* land at about sRGB 18,27,38 (luma about 26) and nearer the leaf's hue, Glade most of all. *Changed:* `NIGHT_FLOOR` vec3(0.092,0.105,0.140), leaning `NIGHT_FLOOR_LEAF` = 0.3 (banded) and 0.6 (Glade) of the way to the leaf's own normalised hue at the floor's luma. *Numbers:* quarter-night crowns_low `#000302` -> **`#0a1c1c`**, p50 0 -> **25.5 / 25.6** (target about 26; reviewed 31.7), below luma 3 76.5 -> 0%, hue 154 -> 180 (reviewed `#162630`, about 200; the target colour is 207); quarter-night crowns `#040d0b` -> `#091e1c` p50 4.6 -> 23.6; harbour-night blossom crowns `#25152e` -> `#2a213b` (reviewed `#36324f`); glade-night crowns `#341a28` -> `#372531`, p50 21.7 -> 32.4 (reviewed `#3f3145`, 37.6); Glade tree crown (1145,405-1190,455) hue 202 reviewed -> 176. The floor is greener than 18,27,38 by the leaf share; that trade is the owner's to judge at G1. 196: 4. *Recorded, as asked:* the quay line at 21:00 (5.5 against 8) is deferred by the orchestrator to the re-measure after P1c and P1d (`artifacts/beautify/owner-decisions.md`). In the reviewed sitting the after tree's 21:00 noise pair was about four times the before tree's (20.97% against 4-5%); in the fix sitting it is 6.45% (>8: 1.52%) against the before tree's 8.70% (>8: 1.72%), so the rule stands anyway: later parts use an after-tree noise pair at 21:00. The floor is re-measured after P1c's toe and black grade. 197: - *Fix tests:* test-first logs `fix/fix-testfirst-test_foliage.log(.out)` (146 checks, 12 failures, all new), `fix/fix-testfirst2-test_foliage.log(.out)` (the key allowance: 159 checks, 14 failures), `fix/fix-testfirst3-test_foliage.log(.out)` (the Glade leaf share and the outline check, written after the calibration rounds and run against the round-2 shader: 163 checks, 5 failures, one of them a wrong line lookup in my own test, fixed). Affected suites `b/suites/fix/`: test_time 287, test_foliage 163, test_sun 72, test_look 271, test_dusk 136, test_lighting 155, test_water 126, test_tree_growth 161, all 0 failures. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/b/fix/suite`, 05:06-05:12, mains, Low Power Mode 0 (`fix/suite-power.txt`): **40 suites, 18,618 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; the test_creature_reactions rerun 0 failures (the known flaky check did not fire). 1440x900 not run (inherited stall). 198: - *Fix changed expectations:* `test_foliage.gd` `night_floor_checks` rewritten: the old "floor line with the `NIGHT_FLOOR_UP` gate" and "floor from 0.05" checks -> a floor with take-back, no gate, sky tones, lobe spread, leaf share, key allowance, floor above 0.04; `view_checks` gains two checks per hour (the leaves take the key light's linear colour x energy and its direction). test_foliage 133 -> 163 checks. 199: - *Fix captures* (`fix/caps.sh`, one sitting 05:12-05:16, mains, Low Power Mode 0, `fix/cap-power.txt`; `fix/before-1` and `before-2` from `b/pre-tree`, `fix/after-1` and `after-2` live; all 11 views, 44 runs, first try): metrics `fix/metrics/` (`acceptance.txt`, `ab-*`, `noise-*`, `counts.txt`, `lit-foliage.txt`). Draw calls and every integer count identical on all views. Reviewer crops, same regions, before | after: `fix/review/qn-tree-4x.png` (260,180-460,320), `qn-right-2x.png` (900,350-1300,600), `gn-right-3x.png` (1000,380-1300,600), and reviewed | fixed versions of each (`*-reviewed-vs-fixed.png`). Part sheet additions, all looked at: `fix/sheet/p1b-fix-1-twilight.png`, `p1b-fix-2-crowns.png`, `p1b-fix-3-night.png`. Unchanged checks: quarter-day 2.6 / 2.5% changed (noise 3.9 / 4.8%), quarter-golden 6.2 / 4.5% (noise 6.6 / 4.3%), glade-day 4.2 / 5.5%, >8 0.04 / 0.05% (noise 4.3 / 2.8%, >8 0.04 / 0.03%). Twilight lines in this sitting: 21:00 town p50 11.7 -> 20.6, quay 0.6 -> 5.5; 06:10 town 12.2 -> 21.3, quay 1.3 -> 6.9; quarter-night town 20.8 -> 23.2; mirror columns 34.4 / 34.7 -> 34.8. ... 219: - **Captures** (`b/caps.sh`, one sitting: `before-1` from the pre-tree, `after-1`, `before-2`, `after-2`; 12 views, all on the first try): quarter-2100, quarter-0610, quarter-2045, quarter-night, harbour-2100, harbour-night, glade-night, quarter-day, quarter-golden, glade-day, and glade-2100 (a `--def`). Metrics are in `b/metrics/`. Sheets, all looked at: `sheet/p1b-1-twilight.png`, `p1b-2-crowns.png`, `p1b-3-night.png`. 220: - Draw calls and every integer count are identical on all 12 views. 221: - quarter-day changed 6.0 / 5.7% (noise 5.4 / 6.3%), quarter-golden 7.6 / 3.2% (noise 4.3 / 6.9%), glade-day 10.9 / 11.1% (noise 9.1 / 9.4%, pixels changed by more than 8: 0.12% against 0.09%). All three are unchanged. 222: - By eye, 21:00 and 06:10 now show the town under a pink-violet skylight instead of black silhouettes. 223: - **Acceptance** (before-1 / before-2 → after-1 / after-2): 224: - 21:00 town p50: 11.8 / 11.7 → **20.4 / 20.4** (target at least 18). **Met.** 225: - 21:00 front_quay p50: 0.6 → **5.5** (target at least 8). **Not met.** No value was pushed, because the full night quay is 3.7 and 20:45 is 6.8. A quay at 8 would flash brighter at 21:00 than on either side of it. The quay itself is P1d's work. 226: - 06:10: town 12.2 → 21.4, quay 1.3 → 6.8. 227: - 21:00 paving 4.7 → 12.6. Share of town pixels below luma 3: 34.4 → 18.7%. 228: - Day and Golden presets: exactly unchanged in code (tests), and unchanged in the frame (within noise, above). 229: - Crown undersides at night, quarter-night crowns_low (reviewed version): `#000302` → `#162630`, p50 0 → 31.7, over the target; after the review fix `#0a1c1c`, p50 25.5 (above). 230: - Carried to P1c: quarter-night below luma 0.1 went 47.3 → 41.4% (after the fix 47.2 → 44.9%), and town p50 20.8 → 26.4 (after the fix 23.2). The floor shows on every street tree, so P1c tunes on top of it and re-measures it after its toe and black grade. 231: - Mirror columns p50 34.8 → 35.1 (kept). 232: - **Look notes for G1:** 233: - Blossom crowns keep most of their pink after the fix (harbour-night `#25152e` → `#2a213b`, glade-night `#341a28` → `#372531`); the floor leans to the leaf's hue. 234: - The floor is greener than the plan's 18,27,38 (quarter-night crowns_low `#0a1c1c`), by the leaf share; leaves turned to the moon but in its cast shadow get less floor. 235: - The 21:00 and 06:10 skylight is warm pink-violet on the walls (the horizon lerp). 236: - **Journey:** no count deltas are expected. quarter-night and harbour-night change on the crowns. harbour-day, harbour-lane, quarter-day and styles-day are expected to be unchanged. 237: - **Bench:** none, per the plan; draw calls are identical. After the review fix each foliage fragment also does three `pow`, a hash, a normalize and the key-light estimate, and each `light()` call two share evaluations and a directional test: the P1 stage bench measures it. 238: - **Review:** Opus 5.5 in a fresh context [run ID removed]: acceptance met but for the quay line, no must-fix; notes on the look. The orchestrator judged the crops by eye and ordered a fix pass [run ID removed] on four items: the additive floor greyed lamp-lit leaves; crown tops stayed black and crowns read flat; the floor landed cool and bright (luma 31.7) and turned Glade teal; record the quay deferral and the 21:00 noise rule. Recheck: all four done (lamp-lit crown 37.3 -> 36.7, crowns_low p50 25.5, Glade hues near the leaf), acceptance met with the quay line deferred. For G1: moon-facing leaves inside the crown's own shadow stay the darkest part of a crown (26% of crowns_low below the black threshold), and the floor is greener than 18,27,38. Merged to `main` on 26 September 2026. 239: - **Not done / open:** 240: - The quay line (above): deferred by the orchestrator to after P1c and P1d. 241: - The cause of the night black crush below 0.035 was not traced. 242: - The key's make-up in its own pass for shadowed faces (dropped: it cannot see the lamps and ambient, and lit bushes rose by about 20 luma). 243: - For G1: the floor's departure from the report's value, the Glade twilight decision, and the blossom and tree-style notes. 244: - **Platforms:** Mac run; Windows not run. ... 101: **Beautify P1c Night fills and violet toe grade, tuned together (26 September 2026):** Opus 5.5 on this Mac raised the canal night fills and added a violet toe grade to the presentation pass (report quick win 2, steps 1 and 2). The two were tuned together against one histogram target. Quarter-night meets it; lane-night comes out lighter than the band after two rounds, and I stopped there as the part's stop rule requires. Glade is unchanged. Branch `beautify/P1c` from `main` b3a421e. Snapshot: `build/backup-feel/19-P1-c/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/19-P1/c/pre-tree/`. Evidence: `artifacts/beautify/19-P1/c/`. Power: mains, Low Power Mode 0, battery 100% for captures, bench and suite (`cap-power.txt`, `bench/power.txt`, `suite-power.txt`). 102: - **Files:** 103: - `scripts/lighting_profile.gd`: the fills; `toe`, `black` and `shadow_tint` in all nine dictionaries; `span_at`, which is `at_hour`'s anchor search split out with unchanged results. 104: - `scripts/look_profile.gd`: Glade zeroes toe and black and keeps its pre-P1c fill and ambient (`GLADE_NIGHT_BASE`, `GLADE_TWILIGHT_BASE`, `glade_base`). 105: - `shaders/presentation.gdshader`: the grade. 106: - `scripts/main.gd`: `apply_presentation` hands the grade over. 107: - Tests: `tests/test_time.gd` (new `grade_checks`) and `tests/test_look.gd` (shader, Glade and editor checks, and `glade_pre_p1c_checks`). 108: - This entry. 109: - **Values.** 110: - Fills: Night 0.55 with ambient `7580bb` (was 0.43, `738caf`); 22:00 0.54 (0.44); 05:45 0.54 (0.48); 20:45 0.52 (0.46). 111: - Grade, added before the outline multiply: `c += shadow_tint x (toe x l x exp(-l/0.09) + black x (1-l)^6)`. 112: - Night, 22:00, 05:45 and 20:45: toe 0.55, black 0.029. Every day anchor has 0, so the pass skips the grade. `shadow_tint` is (1.0, 0.9, 1.8) at all nine dictionaries, so the blend never swings hue. 113: - The toe is capped by the test line "mid-tones above 0.3 move by less than 0.01": the report's curve at l = 0.3 gives 0.0107 x toe, plus 0.118 x black. The reachable toe is therefore about 0.55 with black 0.029. The grade then lifts pure black by 0.031. ... 127: - **Captures** (`c/caps.sh`, one sitting 05:47-05:53, in this order: `before-1` from the pre-tree, `after-1`, `before-2`, `after-2`, then `g1/stage-1` from the pre-stage tree; 11 views, all first try): quarter-night, lane-night, harbour-night, quarter-day, harbour-golden, quarter-2045, quarter-0545, quarter-2100, glade-night, glade-day, glade-2100. ... 139: - **Acceptance** (before-1 / before-2 → after-1 / after-2): 140: - Pixels below luma 0.1: 141: - quarter-night 44.9 / 45.0 → **32.2 / 32.2%**. Target 28-33: **met**. 142: - lane-night 39.3 → **25.5%**. **Missed, too light by 2.5.** Round 2 (toe 0.50, black 0.027) gave quarter 33.2 and lane 26.6. No single value puts both in the band. Stop rule applied: numbers reported, the toe not pushed. 143: - harbour-night 14.9 → 9.7; quarter-2045 50.3 → 40.6; quarter-0545 50.4 → 39.9; quarter-2100 49.9 → 35.0. 144: - Twilight against night (`c/metrics/twilight-gap.txt`): 20:45 and 05:45 minus quarter-night was +1.3 / +3.2 points on the pre-stage tree, +5.4 / +5.4 before P1c and **+8.4 / +7.7 after**. The report's "twilight is never darker than night" is not met, and P1c widened the gap. Likely cause: the plan's twilight fills (0.52, 0.54) sit below Night's 0.55 with the same toe at all four hours; not isolated. To the owner at G1 (question 10). 145: - The fills alone moved quarter only 44.9 → 43.4 and lane 39.3 → 37.3; the grade does the rest. 146: - Iron and ink at or below 0.05 (12.75): **met** at p90. 147: - iron_rail p10 0 → 6.7; quay railings p10 0 → 6.7; pipes p10 0 → 6.4; Harbour railings p10 0 → 6.7. 148: - Pixels that were black (luma below 3) before: after p50 7.2, p90 10.9, p99 12.9. 1.1% are over 12.75 (lane 1.3%, Harbour 1.5%, the noise pair 0.07-0.36%). 149: - Dark edge pixels (`acc.js` "ink": at least 0.32 darker than a neighbour and below 0.2) are a different set and rise about 7 luma: p50 17.4 → 24.2 (quarter), 16.7 → 24.2 (lane), 19.7 → 26.9 (Harbour); p90 37.2 → 40.9. 150: - Mid-tones above 0.3 move by less than 0.01: **missed for the part on all three night views; met for the grade alone** (corrected after review; the first version said "met for the grade" and quoted pooled medians). Pixels stable within both noise pairs, by before-luma band (`c/metrics/mid-bands.txt`, `mid-bands.js`, kit decoder): 151: - Whole part (before-1 → after-1), band 0.3-0.4: median |dL| 0.0162 / 0.0160 / 0.0118 (quarter / lane / Harbour); band 0.4-0.5: 0.0148 / 0.0129 / 0.0134. Share of stable mid pixels moving 0.01 or more: 48.9 / 51.8 / 46.5%. 152: - Grade alone (diag-fills → after-1), band 0.3-0.5: median 0.0042 on all three, p90 0.0081; 0.5-1.0% move 0.01 or more. Met. 153: - Fills alone (before-1 → diag-fills), band 0.3-0.5: median 0.0115 / 0.0106 / 0.0081, p90 0.015-0.017. 154: - The pooled medians first reported (0.009 / 0.011 / 0.009) include bright pixels (0.9 and up: emissive and clipped) that cannot move; they hid the miss. Noise pair before-1 → before-2 mid median 0.0000. 155: - Cause: the plan's own fill values (Night 0.55 with `7580bb`, and the twilight fills). The plan asks for both, and they do not fit together. Not tuned here (the fills are prescribed); to the owner at G1 (question 11). 156: - Windows and neon are still the brightest pixels in greyscale: 157: - windows_tower p90 167.8 → 167.9; windows_right p90 175.9 → 176.3; BAR p90 244.3 → 244.3; FISH p90 211 → 214 (Harbour) and 241 → 241 (lane). 158: - Frame p90 71.4 → 74.5. 159: - Reflection columns (quarter-night `mirror_columns`): p50 35.2 → 42.3; sd 50.8 / 50.6 → 48.1 / 48.0; p90-p10 117.6 → 109.7 (about -6%). By eye at 2x (`p1c-3-regions.png`), every streak is kept and the water between them is lifted. Judged kept; for G1. 160: - **Re-measured lines** (before → after): 161: - **Crown floor:** quarter-night crowns_low p50 25.5 / 26.3 → **35.0 / 34.8**, `#0a1d1d` → `#13242d`; crowns p50 23.7 → 33.2. P1b calibrated the floor to about 26. With the toe on top, it now sits about 9 luma higher, so **it needs re-calibrating** (the floor could come down by roughly the toe's lift). I did not change it here, as ruled. glade-night crowns stay `#372531`. 162: - **P0c night bank line (harbour-night):** bank − water_open 2.9 → 4.2; bank − water_below 4.5 / 4.7 → 5.0 (target 8). **Still not met.** Bank 37.2 → 45.1, water_open 34.3 → 40.9, water_below 32.7 → 40.1. To the owner at G1. 163: - **P1b quay at 21:00 (quarter-2100 `front_quay` p50):** 5.5 → **14.9 / 14.8** (target 8): **met**, far beyond the after-tree noise pair (quarter-2100 frame 6.7% changed, >8 1.5%, against the A/B's 94%). It does not flash: 20:45 is 16.9 and full night 12.5 (was 6.8 and 3.8). 05:45 8.0 → 18.0. Town p50 at 21:00 20.6 → 30.0. 164: - **P0e seam at night:** quarter-night `mirror_seam` share above luma 130 7.57 / 7.55 → 7.70 / 7.67 (unchanged); quarter-2100 4.83 → 4.76; p50 33.0 → 39.8 (the lift). By eye, no stepped seam. 165: - **P1a sky lines:** 166: - quarter-night zenith 49.9 → 54.1 / 54.4 (band 45-55); horizon band 74.1 / 72.6 → 74.8 (band 60-75, but "no brighter than before" no longer holds). 167: - harbour-night zenith 48.2 → 52.5; band 69.1 → 71.7. 168: - **Bench** (`c/bench.sh`, 3 interleaved rounds against the pre-tree, Night, `--disable-vsync`, 1920x1080; `bench/summary.txt`; the engine still reports vsync enabled, as in P0): 169: - dense.json median 15.03 → 14.97 ms. 170: - Harbour 8.21 → 8.32 ms (+0.11 ms, +1.3%, higher in all three rounds: the grade costs an `exp` and a `pow` per window pixel at night). 171: - One round at 2880x1800 (`bench-2880.out`): dense 18.97 → 18.97; Harbour 11.37 → 11.36. 172: - Draw calls identical (2140, 902). 173: - **G1 sheet:** `c/g1/` (`g1-1-canal-night.png`, `g1-2-dawn-dusk.png`, `g1-3-glade.png`, `g1-4-open-questions.png`, `g1.md`). It compares the pre-stage tree (`g1/stage-1` from `build/backup-feel/19-P1`) with after-1, from the same sitting. 174: - **Journey:** no count deltas are expected; draw calls are identical. harbour-night and quarter-night change (lifted darks, violet toe). harbour-day, harbour-lane, quarter-day and styles-day are expected to be unchanged. 175: - **Review:** two Opus 5.5 reviews in fresh contexts [run ID removed]; second review per amendment 14): 3 must-fix, all about what the part told the owner (the G1 sheet's outline and twilight-gap claims, and the mid-tone line booked as met); fixed in the entry and `g1/g1.md`, rechecked. The part honestly misses three lines, each traced to the plan's own prescribed values or to one setting serving two views: mid-tones above 0.3 (fills), lane-night histogram band (25.5% vs 28-33%), twilight never darker than night (gap widened to about 8 points). The crown floor now reads 35 against the planned 26. All go to the owner at G1 (questions 1-11 in `g1/g1.md`). **Not merged to `main`: held on branch `beautify/P1c` until the owner's G1 verdict (D6).** 176: - **Review fix (26 September 2026, 06:15-06:30).** The review found three must-fix items, all in what the part told the owner; no code changed. (1) `g1.md` said outlines sit near 7 and at most about 13: those figures are the pixels that were black before, and dark edge pixels rise about 7 luma (p50 17 → 24). Corrected. (2) The mid-tone line was booked "met for the grade" from pooled medians; split by band it is missed for the part on all three night views (above). Corrected here and added to G1 as question 11. (3) `g1.md` question 10 said twilight being darker than night predates P1; the same-sitting pre-stage captures show the gap was +1.3 / +3.2 before P1 and P1c widened it to +8.4 / +7.7. Corrected. Re-measured from the existing same-sitting captures with `c/metrics/mid-bands.js` and `twilight-gap.js` (the kit's decoder; the reviewer's `review2/` numbers reproduce exactly). Backups: `c/fix/backup/`. Full suite re-run on the unchanged code: `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/c/fix/suite`, 06:23-06:28, mains, Low Power Mode 0 (`c/fix/suite-power.txt`): **40 suites, 18,730 checks, 0 failures**; `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`; the test_creature_reactions rerun 0 failures. No re-capture was needed: no pixel-producing file changed. 177: - **Not done / open:** 178: - The mid-tone line, missed for the part because of the plan's fills (G1 question 11). 179: - Twilight darker than night, widened by P1c (G1 question 10). 180: - lane-night band (above). 181: - The crown floor re-calibration (P1b's foliage shader; not touched here). 182: - The P0c night bank line (4.2 / 5.0 against 8). 183: - The horizon band is slightly brighter than P1a's. 184: - These open questions from earlier parts stay open: the saturated night water at Harbour and Glade (P1a), Glade night close to canal night, moon-side crown lobes staying darkest (P1b). 185: - All are listed for G1 in `g1.md`. 186: - The known flaky check is recorded as known. 187: - **Platforms:** Mac run; Windows not run. ... 99: **Beautify G1 verdict (28 September 2026):** the owner accepted the recommendations on all 11 G1 questions (`artifacts/beautify/owner-decisions.md`, section "G1 verdict"). P1c is merged to `main`. A retune part, **P1c2**, runs before P1d: lower night fills with more of the lift from the toe (Q11), twilight never darker than night (Q10), foliage night floor back near 26 (Q6), and Glade's own bluer-green night (Q2). The night bank line (+8) moves to P3c (Q3); the 21:00 quay line is closed (Q4); Q1, Q5, Q7, Q8 and Q9 are accepted as they are. [side-branch note removed] ... 59: **Beautify P1c2 G1 retune: fills, twilight, foliage floor and Glade night (28 September 2026):** Opus 5.5 on this Mac retuned P1c to the owner's G1 verdict (Q11, Q10, Q6, Q2). The night fills come down and the toe grade does more of the lift. Twilight is now lighter than night. The crown undersides are back near 26. Glade's night has its own pre-P1 palette again. **Stop condition hit on a no-regression line:** the Harbour night bank lines fall to +3.7/+4.5 (P1c +4.2/+5.0). A 9-round sweep shows no setting that keeps them and also meets Q11 (below); it waits for an owner or orchestrator ruling. A review fix (below) removed a dark pulse at 22:00. Branch `beautify/P1c2` from `main` 4c8be01. Snapshot: `build/backup-feel/19-P1-c2/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch trees: `artifacts/beautify/19-P1/c2/pre-tree/` (19-P1-c2), `prec-tree/` (19-P1-c, pre-P1c), `stage-tree/` (19-P1, pre-stage). Evidence: `artifacts/beautify/19-P1/c2/`. Power: AC, Low Power Mode 0, charging 79-99% for rounds, captures, bench and suite (`cap-power.txt`, `rounds.txt`, `bench/power.txt`, `suite-power.txt`). The launch at 17:10 was on battery (77%); every measurement ran on AC. 60: - **Files:** `scripts/lighting_profile.gd` (fills, toe, black), `scripts/look_profile.gd` (Glade's base carries its pre-P1 night palette: `GLADE_BASE_KEYS`), `shaders/presentation.gdshader` (`GRADE_KNEE` 0.09 → 0.08), `shaders/foliage.gdshader` (`uniform float night_floor_gain = 1.0`), `scripts/world_view.gd` (`NIGHT_FLOOR_TOE` 0.40; the leaves take gain 1 − 0.40 × toe), `tests/test_time.gd` (`retune_checks`), `tests/test_foliage.gd` (`floor_gain_checks`), `tests/test_look.gd` (`glade_pre_p1_checks`), this entry. Backups: `c2/backup/`. 61: - **Values:** Night fill 0.55 → 0.51 (ambient `7580bb` kept), toe 0.55 → 0.76; 22:00 fill 0.54 → 0.66, toe 0.55 → 0.775 (review fix; 0.52 and 0.76 before it); black 0.029 → 0.0325 at Night and 22:00. 20:45 fill 0.52 → 1.00 and 05:45 0.54 → 0.90, both with toe 0.775 and black 0.0325; the view takes up to `GOLDEN_SHADE` of the 20:45 fill while the sun is low. Knee 0.08: at the old knee the grade was already at its mid-tone cap, and quarter-night stayed at 36.7% (round 1). The grade alone still passes "mid-tones above 0.3 move by less than 0.01 at every hour" (worst 0.0099). 62: - **Tests:** test-first logs `c2/P1c2-testfirst-*.log(.out)`: test_time 351 checks, 7 failures; test_foliage 190, 27; test_look 358, 16 (all new). After: test_time 351 (was 342), test_look 358 (328), test_foliage 190 (163), all 0 failures; test_dusk 136, test_lighting 155, test_sun 72, test_water 126 pass (`c2/r1-*`). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/c2/suite`: **40 suites, 18,796 checks, 0 failures** (18,797 after the review fix); the test_creature_reactions rerun had 0 failures (the known flake did not fire). `--verify` 2880x1800: the first try exited 0 after 218 s with no `PIXELPUNK_INTEGRATION` line (7 captures). The retry (`c2/verify2/`) exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. 1440x900 was not run (inherited stall). The "terrain_flowers.svg not found" error is in P1c's verify logs too. 63: - **Changed expectations:** test_time: P1c fills "Night 0.55, 22:00 0.54, 05:45 0.54, 20:45 0.52" → "0.51, 0.52, 0.90, 1.00"; `GRADE_KNEE` 0.09 → 0.08. test_look: the shader text `GRADE_KNEE = 0.09` → `0.08`. test_time's new no-dip window was first written from 20:30; it now starts at 20:45, because 20:30 is golden hour with the sun up. Review fix, test_time: "22:00 fill 0.52" → 0.66; the check "22:00's fill is lowered from P1c's 0.54" → "22:00's toe lifts at least as much as Night's"; new check "no dark pulse from 21:00 to 01:00" (352 checks). 64: - **Captures** (`c2/caps.sh`, one sitting 17:32-17:38, AC; the same 11 views as P1c; `before-1/2` pre-tree, `after-1/2` live, `prec-s1/2` pre-P1c, `stage-1` pre-stage Glade). Draw calls and all integer counts are identical on all 11 views in both pairs. glade-day, quarter-day and harbour-golden are within their noise pairs (glade-day >8: 0.16/0.31% against 0.09/0.23%). Sheets, all looked at: `c2/sheet/p1c2-1-night.png`, `-2-hours.png`, `-3-regions.png`, `-4-glade-vs-stage.png`, `-5-glade-and-day.png`, `-6-vs-pre-p1c.png`. Tuning rounds: `c2/round1`-`round9`, `c2/rounds.txt`. Handover series: `c2/series-r6`, `series-pre`. 65: - **Acceptance** (after-1/after-2; `c2/metrics/acceptance.txt`, `mid-bands.txt`, `twilight-gap.txt`, `ink.txt`): 66: - Q11 mid-tones 0.3-0.5 against pre-P1c (pixels stable in all three noise pairs): median 0.0081 / 0.0076 / 0.0073 (quarter / lane / Harbour). P1c was 0.0157 / 0.0151 / 0.0129. **Met.** Quarter-night below 0.1: 32.3 / 32.3% (band 28-33). **Met.** 67: - Q10 (share below 0.1): 20:45 31.2 and 05:45 30.9 against night 32.3 (gaps −1.1 / −1.5; P1c +8.3 / +7.6). **Met.** 21:00 is 28.2, lighter than both (see the review fix). Quay p50: 20:45 25.4, 21:00 22.7, night 13.9, so the quay runs in order. 22:00 is now 31.8 against night 32.3 (35.7 before the review fix). 06:20 is 46.5 (was 50.2). 68: - Q6: quarter-night crowns_low p50 34.1 → **25.3 / 26.2**, `#0e1a23` (P1b 25.5-26.3). **Met.** Not black. The gain scales the floor uniformly, so the lobe order (Q5) is unchanged. 69: - Q2: glade-night frame, zenith, horizon and water match the pre-stage tree: `#203647`, `#001b43`, `#405e8c`, water `#022948`, hue 207. The P1c tree had `#273355`, `#0f235a`. Glade night against stage: >8 on 0.93% of pixels (the P1b floor on the crowns). The water is saturated as before P1 (sat 96.5%) (Q1). Glade day and golden: unchanged in code (tested at 9 day hours and both presets) and within the noise pairs. 70: - No-regression lines: quay at 21:00 22.6 (target ≥ 8), met. P1a zenith 54.4 and horizon band 74.8, unchanged. Iron and ink: formerly-black pixels p90 11.1 / 11.5 / 11.8 (limit 12.75), met; iron p10 6.7 → 7.7. Windows and neon still brightest (windows_right p90 176.0, BAR 244.3, frame p90 74.1). Reflection columns sd 48.1/47.9 → 47.6/47.6, p90-p10 110.4/108.8 → 107.9/107.6 (within the pair spread of 1.6). Lane-night 25.5 → **24.4%** (1.1 lighter; Q7 accepted near 26). **Harbour bank lines 4.2/5.0 → 3.7/4.5: worse.** 71: - **Stop condition (bank lines):** the bank takes its lift from the fill; the water does not. The sweep (`rounds.txt`: fill/toe → quarter mid p50, bank open/below): 72: - .47/.515 (knee .09) → .0034, 2.7/4.3 73: - .50/.74 → .0067, 3.2/3.9 74: - .51/.72 → .0081, 3.2/3.9 75: - .51/.76 → .0081, 3.7/4.6 (shipped) 76: - .51/.775 → .0084, 3.5/4.5 77: - .52/.75 → .0104, 4.5/4.6 78: - .52/.775 → .0104, 4.2/4.7 79: - .55/.72 → .0154, 4.8/4.9 80: - .51/.76 with the old ambient `738caf` → .0199, 5.6/5.9 81: 82: Bank-below stays under 5.0 even at P1c's 0.55. Holding +4.2/+5.0 costs Q11. I did not push the fill. Q3 already asks P3c to reach +8. 83: - **Look note:** at 20:45 the 1.00 fill flattened the warm sunset side of the tower (`p1c2-2-hours.png`). Q10 needs it: a 0.90 fill left 20:45 +0.5 darker than night. Fixed by the warm 20:45 ambient (below). 84: - **Bench** (`c2/bench.sh`, 3 interleaved rounds against the pre-tree, Night, `--disable-vsync`, 1920x1080): dense.json median 14.95 → 14.92 ms; Harbour 8.26 → 8.26 ms. One round at 2880x1800: dense 19.39 → 19.60, Harbour 11.78 → 11.79. Draw calls identical (2140, 902). 85: - **Journey:** no count deltas expected (draw calls identical). harbour-night and quarter-night change (darker crowns, stronger toe). harbour-day, harbour-lane, quarter-day and styles-day are expected unchanged. 86: - **Review:** filled by the orchestrator. 87: - **Review fix (28 September, 18:15-18:35, AC 100%, Low Power Mode 0, `c2/fix/`):** the review found two must-fix items. (1) The Harbour bank lines are worse (3.7/4.5 against 4.2/5.0). The stop stands and nothing was changed for it; the 22:00 fix does not move the 01:00 captures (bank 3.7/4.5, quarter-night 32.3%, lane-night 24.4%, crowns_low 25.8/25.4, as before). The reviewer's options for the owner: accept 3.7/4.5 until P3c's +8, or take round 5 (Night fill 0.52, toe 0.75: bank 4.5/4.6, quarter mid-tone median 0.0104, over the 0.01 line; Harbour 0.0081). Neither keeps bank-below at 5.0. (2) 22:00 was a dark pulse: 35.7% below 0.1 against night's 32.3%, after a twilight near 30%. Fixed: 22:00 fill 0.52 → 0.66 and toe 0.76 → 0.775 (the cap). 22:00 has a weaker, lower moon than night (energy 0.45), so its fill makes up the difference. The value was tuned on two series, `c2/fix/series-a66` and `series-a68`. With 0.68 the path was flat but lighter at 21:15-21:30. test_time now checks it with a light proxy (`town_light`, key counted at `KEY_SHARE` 0.5, calibrated on those captures; tolerance 0.01). Test-first logs: `c2/fix/testfirst-test_time.log` (uncalibrated proxy) and `testfirst2-test_time.log` (final test on the pre-fix values: 2 failures). After: test_time 352, test_look 358, test_foliage 190, test_dusk 136, test_lighting 155, test_sun 72, test_water 126, all 0 failures (`c2/fix/*.res`). Full suite `c2/fix/suite`: **40 suites, 18,797 checks, 0 failures**. The creature rerun had 0 failures. `--verify` 2880x1800 exited 0 in 31 s with `PIXELPUNK_INTEGRATION []`. 1440x900 was not run (inherited stall). Captures: one sitting, `c2/fix/before-1`, `after-1`, `before-2`, `after-2` (pre-tree against live, 17 views: the quarter at 20:45 to 06:00 plus harbour-night and lane-night). All integer counts and draw calls are identical. Within the pairs, the share below 0.1 moves by at most 0.1. Series after, share below 0.1: 20:45 31.2, 21:00 28.2, 21:15 29.1, 21:30 28.8, 21:45 30.9, 22:00 31.8, 22:30 31.7, 23:00 31.7, 00:00 32.2, night 32.3. The town p50 falls through 35.2, 35.5, 32.9, 32.7, 31.9, 31.5, 31.6, 31.7, 31.6 and 31.8. The P1c tree ran 40.5, 35.0, 37.7, 37.3, 36.9, 36.7, 35.8, 34.8, 33.2 and 32.2 (`c2/fix/metrics/series.txt`, `noise-*.txt`, `counts.txt`). Sheets, looked at: `c2/fix/sheet/fix-1-handover-a.png` and `fix-2-handover-b.png`, with no visible step or flash from 20:45 to night. **Still open:** 21:00-21:30 (28.2-29.1) is lighter than both 20:45 and night by the frame share. It is the key's handover, when the shadows go. The P1c tree had the same dip against its neighbours (21:00 35.0 against 40.5 and 37.0). The fix makes 21:15-21:30 about 1-2 points lighter than before. The town p50 does not dip. It is left for the orchestrator: a fix needs a lighter 20:45 (it already flattens the sunset) or a new anchor. Bench not rerun: only two constants changed, and draw calls are identical. 88: - **Recheck (28 September, after 18:35):** the recheck kept both items open, and both need a ruling. (1) Harbour bank lines 3.7/4.5 against P1c's 4.2/5.0. Both before captures and both after captures give the same figures (`c2/fix/metrics/measure-*.json`), so this is not noise. The choice is the owner's or the orchestrator's: accept 3.7/4.5 until P3c's +8 (Q3), or take round 5 (bank 4.5/4.6, quarter mid-tone median 0.0104, over Q11's 0.01). (2) By frame share below 0.1, 21:00-21:30 (28.2, 29.1, 28.8) is lighter than both 20:45 (31.2) and night (32.3). The quay runs in order (25.4, 22.7, 13.9). The town p50 is flat within 0.3 from 20:45 to 21:00. The sheets show no visible step. The P1b quay flash ruling in `owner-decisions.md` is not broken. No ruling on either item was in `owner-decisions.md` at the recheck, so nothing was changed. The live tree is the review-fix state, and `git fetch` found `origin/main` already in this branch. The suites, captures and `--verify` from the review fix stand as the current evidence. 89: - **Warm 20:45 fix (28 September, 19:55-20:20, AC 100%, Low Power Mode 0, `c2/warm/`):** the orchestrator asked for the owner's option B from `owner-decisions.md` "P1c2 rulings": keep the 20:45 lift but make it warm, so shaded stone is at least as warm as before P1c at about P1c2's luma. Q10, the 20:45 sky, horizon and water, lit faces, the handover and Glade had to hold. Change: the 20:45 anchor's ambient `6f7ea0` → **`967870`**, a dusty rose near the horizon `d07a64` hue with linear luminance 0.211 against 0.209. Fill (1.00), toe (0.775), `shadow_tint` and every other anchor are unchanged. Glade keeps its own 20:45 base. Tuning: `c2/warm/rounds.txt`. `9a7482` (r1) and `94767a` (r3) left the tower's hue pinker than before P1c. Test-first logs: `c2/warm/testfirst-test_time.log` (3 failures) and `testfirst-test_look.log` (1 failure). New `warm_twilight_checks` in test_time pins `967870`, rose hue, luminance within 10%, fill 1.00, the violet tint, the other anchors' ambients and Glade's base. test_look gains one check. After: test_time 365, test_look 359, test_dusk 136, test_lighting 155, test_sun 72, test_water 126, test_foliage 190, all 0 failures (`c2/warm/after-*.res`). Full suite `c2/warm/suite`: **40 suites, 18,811 checks, 0 failures**. The creature rerun had 0 failures. `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. 1440x900 was not run (inherited stall). Captures (`c2/warm/caps.sh`, one sitting, 20:03-20:11): `before-1/2` from `c2/warm/p1c2-tree` (snapshot `build/backup-feel/19-P1-c2-warm`, = 2fba894), `after-1/2` live and `prec-1/2` pre-P1c. That is 18 views: P1c's 11 plus glade-2045, quarter-2200 and 20:50-21:30. Warmth (`c2/warm/warmth.js`, crop x350-1050 y100-600, stone pixels stable in all three pairs, `metrics/warmth.txt`), pre-P1c → P1c2 → fix: 90: - 20:45 shaded stone: R−B 15.4 → 5.7 → **30.5**; OKLab hue 23.6° → 342.4° → **26.8°**; chroma 0.035 → 0.038 → 0.047; luma 0.107 → 0.177 → **0.178**. 91: - 20:45 tower left face: R−B 9.6 → −2.0 → **21.0**; hue 15.3° → 318.5° → **23.4°**; luma 0.106 → 0.174 → 0.173. 92: - 21:00 shaded stone: R−B 4.2 → −1.0 → 15.5; hue 351.6° → 324.2° → 6.6°. 21:00 tower face: −5.9 → −13.5 → 0.5. 93: - Lit stone (luma ≥ 0.30): 0.4584 → 0.4585, not brighter. 94: 95: The fix is warmer and more saturated than pre-P1c. Sheets, looked at: `c2/warm/sheet/warm-2045-crop-2x.png` and `warm-2100-crop-2x.png` (pre-P1c | P1c2 | fix). The tower and terrace walls read warm brown-rose, not grey-violet. Q10: quarter-2045 is 31.3 / 31.3% below 0.1 against night 32.3 (P1c2 31.2 / 31.3). 20:45 sky: zenith `#586194` p50 97.5, horizon band 153.2 → 153.0, skyline 123.8 → 124.0, the same as the noise pair (124.4). Water: the shader's colour is unchanged. The mirror p50 moved 56.3 → 56.5 (noise 56.1) and turned slightly warmer (`#4d3b56` → `#523a50`) because it mirrors the warmer stone. Handover series after (share below 0.1 / town p50): 20:45 31.3/35.2, 20:50 30.8/35.6, 20:55 30.1/35.6, 21:00 28.0/35.6, 21:05 29.5/34.1, 21:15 29.2/32.9, 21:30 28.9/32.7, 22:00 31.9/31.5, night 32.3/31.8. Before: 31.2/35.2, 30.7/35.5, 30.1/35.5, 28.2/35.5, 29.5/33.9, 29.1/32.8, 28.8/32.7, 31.9/31.5, 32.3/31.8. That is within 0.3 everywhere, with no flash. Glade (glade-night, -2045, -2100, -day): code identical (tested). Its A/B spread is within the spread of glade-day, which cannot change (>8 up to 0.29% against glade-day's 0.15% across trees). No-regression lines, P1c2 → fix: quay 21:00 22.7 → 22.6 (≥ 8); bank lines 3.7/4.5 → 3.7/4.5 (ruled); crowns_low 25.3/25.8 → 25.4/25.3; lane-night 24.4 → 24.4; quarter-night 32.3 → 32.3; P1a zenith 54.4 and band 75.5 unchanged; reflection sd 47.7/47.5 → 47.4/47.7; iron p10 7.7 → 7.7; windows unchanged. Draw calls and all integer counts are identical on all 18 views. Not benched: only one colour changed, and draw calls are identical. Journey: no count deltas; nothing in the journey views is at 20:45. 96: - **Not done:** Lane-night is 1.1 lighter (Q7 accepted). The bank lines and the 21:00-21:30 share are ruled accepted ("P1c2 rulings"). The known flaky check is recorded as known. 97: - **Platforms:** Mac run; Windows not run. ... 41: **Beautify P1d Quays lit by the water (28 September 2026):** Opus 5.5 on this Mac. Quay and pier faces are no longer near-black. They are lighter stone, and a water-bounce term now lights upright masonry that stands in water (report Defect 4, steps 1 and 2). The lift comes almost entirely from the lighter albedo; the bounce at the report's strengths adds under 0.5 luma (see "Finding"). Branch `beautify/P1d` from `main` 4c7791a. Snapshot: `build/backup-feel/19-P1-d/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/19-P1/d/pre-tree/`. Evidence: `artifacts/beautify/19-P1/d/`. Power: AC, 100%, Low Power Mode 0 for the tuning rounds, captures, bench and suite (`cap-power.txt`, `rounds-power.txt`, `bench/power.txt`, `suite/suite_all.out`). 42: - **Files:** `shaders/weathering.gdshaderinc` and `scripts/weathering.gd`. These hold the shared "stands in water" gate `weather_in_water(ground)` / `Weathering.in_water` and the bounce mask `weather_bounce` / `Weathering.bounce`. Five new constants are on the parity list: `WATERLINE` 0.15, `WET_GROUND` 0.75 (WATER_Y + 0.6), `DRY_GROUND` 1.6 (LAND_Y), `BOUNCE_REACH` 2.5 and `UPRIGHT` 0.6. The other files: `shaders/masonry_body.gdshaderinc` (`uniform vec3 bounce_color = vec3(0.0)`, `uniform float bounce_strength = 0.0`, `EMISSION += ALBEDO*bounce_color*bounce_strength*weather_bounce(weather.x,wp,wn)`), `scripts/world_view.gd` (`BOUNCE_DAY` 0.12, `BOUNCE_NIGHT` 0.18, `BOUNCE_HORIZON` 0.5; `sync_profile` sets stone, stone_uv and their build variants; the strength follows the hour's windows between the Day and Night presets and is 0 with the water off), `scripts/architecture.gd` (`QUAY_FACE` = STONE.darkened(0.05); the terrain quay face is stamped with ground WATER_Y), `tests/test_weathering.gd` (`water_checks`, plus the parity pairs), `tests/test_lighting.gd` (`bounce_checks`) and this entry. Backups: `d/backup/`. Scene: `d/gate.json` (`d/make_gate.gd`). 43: - **Decisions:** (1) Plaster walls in water: the gate is shared through the include, and only masonry takes the bounce (the owner default; plaster_body is unchanged and a test holds it). (2) The gate counts ground up to WATER_Y + 0.6 as fully in the water, fades out to LAND_Y, and treats the no-ground fallback (0) as in the water. That fallback already measures damp from the water, and it covers bridge piers and plinths. The terrain quay face is stamped at **WATER_Y**, not the plan's WATER_Y + 0.6, which I read as the gate's full-strength ground. Its damp band moves only from 0.25-1.25 m to 0.40-1.40 m. That avoids the plan's named risk: the damp band moving to 1.0 m and darkening the face this part lifts. (3) The quay albedo is **STONE.darkened(0.05)**, not the report's STONE. With STONE (`round1`, `round-k0`), the quarter-night quay reached p50 26.2 and the frame fell to 27.3-27.4% below 0.1, under the owner's G1 Q11 band of 28-33. darkened(0.1) gave a p50 ratio of 0.545. Rounds: `round-a-d10`, `-b-d10k6`, `-c-d20k12` and `-d-d05`, logged in `rounds-power.txt`. (4) The bounce colour is passed as sRGB components, a plain vec3 without `source_color`, so the report's k keeps its calibration. 44: - **Tests:** test-first logs `d/P1d-testfirst-test_weathering.log(.out)` and `d/P1d-testfirst-test_lighting.log(.out)`, both parse errors on the missing members (exit 3). After: test_weathering 308 (was 189), test_lighting 186 (was 155), plus test_chemistry 350, test_masonry 501, test_wet 145, test_ivy 148 and test_look 359, all 0 failures (`d/r2-*`). `test_chemistry.gd:518-520` holds. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/d/suite`: **40 suites, 18,961 checks, 0 failures**. The test_creature_reactions rerun had 0 failures (the known flake did not fire). `--verify` 2880x1800 exited 0 in 31 s with `PIXELPUNK_INTEGRATION []`. 1440x900 was not run (the inherited stall). 45: - **Changed expectations:** none in existing tests. My own new check changed after tuning: "the quay face is STONE" → "the quay face is QUAY_FACE (STONE.darkened(0.05)), lighter than STONE.darkened(0.2)". 46: - **Captures** (`d/caps.sh`, one sitting 20:40-20:46; `before-1/2` from the pre-tree, `after-1/2` live; 14 views: quarter night/day/golden/20:45/21:00/22:00, harbour-night/day, lane-night, chemistry-day, glade-night/day, gate-night/day; plus `before-j1/2` and `after-j1/2` for styles-day and harbour-lane). Integer counts and draw calls are identical in all 16 views. Harbour, lane, chemistry, Glade and styles are within their noise pairs; they have no terrain quays. Harbour bank lines are unchanged at 3.7/4.5. The gate scene: the land house has 0 px changed. The sunk house (ground 0.9, gate 0.92) is also unchanged, because inside 2.5 m of the water it has arched shop windows, not masonry. Its quays lift (night p50 37 → 45.7 on land, 35.2 → 40.8 lowered; `metrics/gate.txt`). Sheets, looked at: `d/sheet/p1d-1-pier-and-land.png`, `p1d-2-hours.png` and `tune-d05.png`. Frames looked at: `after-1/quarter-night`, `after-1/gate-night`, `after-1/gate-day` and `round-k3/quarter-night`. 47: - **Acceptance** (`d/metrics/acceptance.txt`, `quay.txt`, `hue.txt`): 48: - Quarter-night quay against paving: the kit's `quay_over_paving` goes 0.567 → **0.725/0.723** (target ≥ 0.55; P1c2 had already reached 0.567 on this measure). The p50 form goes 0.379 → **0.641/0.643**, and front_quay p50 goes 13.9 → 23.6. Met. 49: - Left pier lifted: night 10.5 → 20.0, day 11.8 → 24.7. Met. 50: - Land walls unchanged: land_wall p50 38.9/38.7 → 38.7/38.7 at night and 66.5/66.3 → 66.8/67.3 by day. Changed >8 is 0.07-3.3% against the noise pair's 0.2-2.0%. Met. 51: - No violet cast: quay hue at night goes 232.5 → 229.6 (land walls 221.5). At 21:00 it goes 284.5 → 298.0, with saturation 34.9 → 26.4, which moves it toward the walls' 340. Met. 52: - No-regression: the quay at 21:00 is 22.5 → 33.1 (closed line ≥ 8). The quay runs 38.9 (20:45), 33.1 (21:00), 25.7 (22:00), 23.6 (night) against 25.4, 22.5, 16.1, 13.9 before, so the order holds and there is no flash. Frame below 0.1: night 32.3 → **28.6** (Q11 band 28-33, holds with 0.6 to spare); 20:45 31.3 → 28.5 (Q10 holds by 0.1); 21:00 28.1 → 25.0; 22:00 31.8 → 27.8; lane-night 24.4 → 24.4. 53: - **Finding (for G2):** at k 0.12/0.18 the bounce is almost invisible. With the same albedo and k 0 (`round-e-k0-final`), the quay p50 is 23.5 against 23.6, the pier 19.7 against 20.0, and 20:45 38.7 against 38.9. Emission is added in linear light, and the night bounce colour `28 2a 55` is dark. At 17× the report's k (`round-k3`, with STONE) the waterline gradient shows (quay 26.2 → 33.5), but the frame drops to 24.5% below 0.1. Under the P1c toe grade, the albedo does what the report expected the bounce to do. A visible "lit from below" gradient would need k at 10-15× the report's value with the albedo left at darkened(0.2); that is a G2 call. 54: - **Bench** (`d/bench.sh`, 3 interleaved rounds against the pre-tree, Night, rain 0.45, `--disable-vsync`, 1920x1080; `d/bench.out`): dense.json median 14.99 → 15.02 ms; Harbour 8.19 → 8.21 ms. Draw calls are identical (2140, 902). 55: - **Journey:** no count deltas expected (draw calls identical). quarter-night and quarter-day change: quay and pier faces are lighter, and their mirror in the canal lifts slightly. harbour-night, harbour-day, harbour-lane and styles-day are expected unchanged (within noise here). 56: - **Review:** filled by the orchestrator. 57: - **Not done:** the tide line and algae band (P1e). A plaster wall standing in water was not captured, because plaster takes no bounce here. The plan offers a second review as optional; the orchestrator decides. Platforms: Mac run; Windows not run. ... 24: **Beautify P1e Tide line and algae band (28 September 2026):** Opus 5.5 on this Mac. Every upright masonry or plaster face standing in the water now carries a dark, glossy algae band from the waterline to a wandering tide top (0.41-0.49 m) and a pale tide line one art pixel wide along that top (report quick win 9, Defect 4 step 3). First, as part of this part, the P1d review's gate defect is fixed: geometry carrying no ground (0) now reads its own height as its ground, so island kerbs and edging no longer pass the in-water gate. Branch `beautify/P1e` from `main` 5900b68. Snapshot: `build/backup-feel/19-P1-e/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/19-P1/e/pre-tree/`. Evidence: `artifacts/beautify/19-P1/e/`. Power: AC, Low Power Mode 0 for the captures, bench and suite (`cap-power.txt`, `bench/power.txt`, `suite-power.txt`). 25: - **Files:** `shaders/weathering.gdshaderinc` and `scripts/weathering.gd`: `weather_in_water(ground, y)` / `Weathering.in_water(ground, point_y)` (the no-ground fallback reads the point's height, `NO_GROUND` 0.001), and `weather_wave`, `weather_tide_top`, `weather_tide` (returns band, line), `weather_tide_albedo` with their mirrors `tide_top`, `tide`, `tide_albedo`. Constants on the parity list: `TIDE` 0.30, `TIDE_WAVE` 0.08, `TIDE_LINE` 0.03, `TIDE_PITCH` 1.3, `ALGAE` (0.42, 0.58, 0.50), `ALGAE_MIX` 0.85, `TIDE_LIFT` 0.35 (the review's night value, not the report's 0.20), `TIDE_ROUGHNESS` 0.30, `TIDE_SPECULAR` 0.5; `WATERLINE` 0.15 was already there. `shaders/masonry_body.gdshaderinc` and `shaders/plaster_body.gdshaderinc` take the tide from the include; they keep no copy of its numbers, and the MOSS_SHELTER literal and the wet roughness line are untouched. `scripts/architecture.gd` `supports()`: a building's pillar standing in the canal is stamped with ground WATER_Y, like a terrain quay. Before, it carried the island's 1.6 and missed both band and bounce. `tests/test_weathering.gd`. Backups: `e/backup/`. Scene: `e/tide.json` (`e/make_tide.gd`). It has a quay, ground sculpted down in three steps (1.3, 0.9, 0.5) with a freehand stone wall across them, a canal house on the lowest step, a medieval house hanging over the water on pillars, stepping stones, a plinth bench and a causeway. 26: - **Gate fix (P1d review):** the fallback rule was the choice between stamping the generators and limiting the fallback. Stamping LAND_Y on kerbs, edging and props would also have given them a damp foot and moss, which is a look change on every kerb. The fallback rule changes only the gate. The review probe (`d/review/probe_ground.gd`, unchanged; `e/probe-review-post1.log.out`) counts land-lit vertices: courtyard 60 → **0**, harbour 11,532 → **0**, styles 576 → **0**, benchmark 0, castle 1,968 → 24. All 24 castle vertices lie below the terrain surface (y 1.48), so they are never seen. `e/probe_e.gd` adds decorations (rebuilt from the state, because MultiMesh transforms read back as the origin in headless runs) and a buried test. It gives 0 visible land-lit and 0 land-tide vertices in all seven sample scenes (`e/probe-post3.log.out`; before: `probe-pre.log.out`). 27: - **Tests:** test-first log `e/P1e-testfirst-test_weathering.log(.out)` (parse error on the missing members, exit 3). After: test_weathering 452 (was 308), plus test_masonry 501, test_ivy 148, test_chemistry 350, test_wet 145, test_lighting 186 and test_look 359, all 0 failures (`e/r1-*`, `e/r2-*`). My first body edit broke test_wet's source check "ROUGHNESS = clamp(mix(0.80, 0.45, wet)". The line is now kept and the tide mix follows it. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/e/suite`: **40 suites, 19,105 checks, 1 failure**: test_ivy "dense: a cold ivy pass is under 250 ms (365.6)", a timing check on code this part does not touch. It passed in `r1` and twice on rerun (`e/r3-test_ivy.log.out`). The test_creature_reactions rerun had 0 failures (the known flake did not fire). `--verify` 2880x1800 exited 0 in 33 s with `PIXELPUNK_INTEGRATION []`; 1440x900 was not run (the inherited stall). 28: - **Changed expectations:** test_weathering: "`Weathering.in_water(0.0)` is 1: geometry carrying no ground is in the water" → "geometry carrying no ground reads its own height: in the water up to 0.75 m, dry from LAND_Y (kerbs at 1.64-1.78)". The include text `float weather_in_water(float ground)` → `float weather_in_water(float ground, float y)`. `Weathering.in_water(ground)` → `in_water(ground, point_y)`; the other P1d gate checks pass a height and are otherwise unchanged. **Correction to the report's prior-work note:** the old waterline band was not lost. It is the no-ground fallback's damp band (`weather_height` measures from 0), and it still applies. The tide line is new. 29: - **Captures** (`e/caps.sh`, one sitting 21:25-21:34, AC). There are three sets, `before-1..3` from the pre-tree and `after-1..3` live. Views: quarter night/day/golden/05:45/20:45, harbour-night/day, lane-night, chemistry-day, styles-day, tide-night/day. Every capture was compared with every other (`e/metrics/pairs.txt`), which exposed framing flakes where the whole frame shifts: chemistry-day a1, harbour-night a1 and b3, lane-night a2 and a3, styles-day a1, tide-night b3. Those frames are excluded. The flakes are in both trees, and the other captures of the same trees match. Integer counts and draw calls are identical on all 12 views (`counts-2.txt`). Lane-night, chemistry-day and styles-day are within their noise pairs. Harbour-day and harbour-night are within their noise pairs on the whole-frame count only; the boathouse pillars change for real (see "Review fix"). Sheets, looked at: `e/sheet/p1e-1-quarter-day-waterline.png`, `-2-quarter-night-waterline.png`, `-3-tide-day-left.png`, `-4-tide-night-left.png`, `-5-tide-day-lowered.png`, `-6-tide-night-lowered.png`, and `try1-ab-day.png` / `try1-ab-night.png`. 30: - **Acceptance** (`e/metrics/acceptance.txt`, `band.txt`, `quick.txt`, `pier-column-x150.txt`): 31: - **Band on everything in the water:** upright weathered triangles over water that cross the band are all gated in 8 scenes (quarter 451, tide 401, castle 160, harbour 72 …), 0 missed. Before, 249 were missed (building pillars stamped 1.6). **Met**, with one limit on the tide line: stepping stones stand 0.13-0.17 m out of the water, below the tide top, so their whole side is band and they have no line. 32: - **The line reads at night as a thin highlight:** in the tide scene the line pixels go 35.8 → 45.2 (day 79.0 → 93.2) and the band pixels go 36.6 → 25.1 (day 68.3 → 50.4); the noise pair moves 0.05-0.2% of pixels. On the quarter canal quay (column x 700) the band goes from 85-89 to 42-43 by day and from 45 to 23 at night; the line row goes 12 → 23 at night. **Met.** The part's named risk shows as expected: at night the band on a dark face goes near black (stepping stones and the pillar feet read as dark silhouettes), and the line and gloss carry it. 33: - **Nothing on land walls:** 0 land-tide vertices (probe). Quarter land_wall p50 is 38.7 → 38.7 at night and 66.5 → 67.3/66.3 by day. In the lowered-ground sheet the freehand wall on the sculpted steps is unchanged, because its band lies below its own ground. **Met.** 34: - **P1d lines still hold:** quay_over_paving 0.747/0.729 → **0.735/0.730** (≥ 0.55; p50 form 0.627/0.622). Left pier: night 20.0 → 18.9, day 25.0 → 24.3 (pre-P1d 10.5 / 11.8). Quarter-night below 0.1: 28.6 → **29.0** (band 28-33). Q10: 20:45 28.8 and 05:45 28.3, both at or under night's 29.0. **Met.** 35: - **Bench** (`e/bench.sh`, 3 interleaved rounds against the pre-tree, dense.json, Night, rain 0.45, 1920x1080; `e/bench.out`): median 15.72 → 15.58 ms, draw calls 2140 = 2140. performance.json reports vsync "enabled", as in P1d's bench, and p95 is pinned at 16.67 ms. 36: - **Journey:** no count deltas expected (draw calls identical). quarter-night and quarter-day change: a dark band and a pale line at every quay and pier foot, and the pillars in the canal are lighter above the band (their damp now rises from the waterline). **harbour-day and harbour-night change at the boathouse pillars** (the stone pillars under the boathouse standing in the water, around x 850-980, y 770-835): lighter over their whole visible height and a darker foot, the same change as the quarter canal pillars. Treat this as expected, not a regression. harbour-lane and styles-day are expected unchanged. 37: - **Review fix (28 September 2026):** the review found that this entry booked harbour-day and harbour-night as unchanged, but the boathouse pillars (cluster masonry, now stamped ground WATER_Y by `supports()` and gated at g 0.15, `e/probe-post3.log.out`) change for real. The whole-frame >8 count hid it because the water dominates. Fix: documentation only, in "Captures" and "Journey" above; no code changed. Re-measured on the capture sets from the 21:25 sitting (`e/review/col.js`, `stable-diff.js`, `pillar-box.js`). Harbour-day, pillar column x 922, identical in sets 1-3: rows 804-827 go 30.8/15.1/20.5/15.6 → 44.5/25.1/32.0/25.8 (+8 to +14 luma), and the foot rows 830-831 go 26.1 → 7.5-7.7. The before frames are identical at every row, so this is not noise. Harbour-night (sets 2 and 3; after-1 is a framing flake): rows 804-826 +2 to +8, row 830 16.6 → 8.5. In the pillar box (800-1100, 700-880), px changed by more than 8 in both same-tree pairs: day 440 lighter (21.2 → 33.9) and 128 darker (24.4 → 7.6); night 132 lighter, 52 darker. The rest of the frame's stable changes are thin rain streaks (`e/review/cell-992-672.png`, `cell-768-384-night.png`). New sitting 21:56-21:58 on AC, Low Power Mode 0 (`e/caps-fix.sh`, `cap-power.txt`): `before-j1/2` and `after-j1/2` for harbour-lane and harbour-day. Harbour-day repeats the pillar change (the same cells). Harbour-lane: before-j2 is a framing flake. Before-j1 against after-j1 differs on 1.60% against the after pair's 0.57%, but the extra is water shimmer and flower and leaf sparkle of both signs, with none on stone (`e/review/diff-harbour-lane.png` against `noise-harbour-lane.png`), so it is TIME phase. Counts and draw calls are identical (`e/metrics/counts-j.txt`). Styles-day has 0 stable px changed. Sheet, looked at: `e/sheet/p1e-7-harbour-pillars-day-night.png`. Suite rerun `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/e/suite-fix`: **40 suites, 19,105 checks, 0 failures** (the earlier test_ivy timing failure did not recur). `--verify` 2880x1800 exited 0 in 34 s with `PIXELPUNK_INTEGRATION []`, and the creature_reactions rerun had 0 failures. 38: - **Review:** filled by the orchestrator. 39: - **Not done:** no plaster wall stands in the water in any scene. Plaster walls start at their ground plus the lift, and the lowest sculpted ground is 0.5 m, above the tide top. Plaster therefore carries the code for parity, but `tide.json`'s medieval house shows the band only on its stone pillars. The tide does not switch off with the water off (bounce does); a drained canal keeps its tide marks. Harbour was not benched (the plan asks for dense only). Platforms: Mac run; Windows not run. ... 5: **Beautify P1f Painted metal (28 September 2026):** Opus 5.5 on this Mac. Railings, sign cabinets, posts, street furniture and pipe bores now draw in a new `painted_metal` material (owner D7: a material, not a flag): the surface shader at kind 4, metallic 0.18 and roughness 0.55 instead of bare metal's 0.65 and 0.42 (report quick win 3). Pipe bores lean 0.2-0.4 of the way from PIPE to COPPER by a hash of their route. Collars, brackets, elbows, copper vents, flashing, lamp cages, hooks, grates, window bars, antennas and roof trims stay bare `metal`. **Stop rule hit (owner call at G2):** the painted railings and the cabinets' painted rim and sides lighten but stay below 0.05 by day, because their colour METAL `#232a3e` is itself near-black; I did not repaint them. The black mass of each blade sign is its face (SIGN_FACE `#0e0d1c`, dark on purpose behind the neon), not painted metal. Pipes lift clearly. The part also finds why the P1d/P1e benches looked vsync-bound, and fixes the report (below). Branch `beautify/P1f` from `main` 52a7d20. Snapshot: `build/backup-feel/19-P1-f/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/19-P1/f/pre-tree/` (its `scripts/main.gd` was replaced by this part's report-only main.gd after the captures and before the bench, so both sides of the bench report the same fields; original in `f/backup/pre-tree-main.gd.orig`). Evidence: `artifacts/beautify/19-P1/f/`. Power: AC, 100%, Low Power Mode 0 for captures, bench and suite (`cap-power.txt`, `bench/power.txt`, `suite-power.txt`, `vsync/power.txt`). 6: - **Files:** `shaders/surface_body.gdshaderinc` (kind 4: `PAINTED_METALLIC` 0.18, `PAINTED_ROUGHNESS` 0.55, `METALLIC=surface_kind==3 ? 0.65 : (surface_kind==4 ? PAINTED_METALLIC : 0.0);`, SPECULAR 0.5 for kinds 3 and 4). `scripts/architecture.gd`: `PAINTED` ("painted_metal"), `PAINTED_KIND` 4, `PIPE_COPPER` (0.2, 0.4), `pipe_tint(points)`, the material in `materials()`, and 73 call sites retagged (`f/retag.js` lists them by line): `sign_cabinet`, the fence runs in `fittings`, `railing`, `stair_look` banister material for iron and modern (spines stay bare), `flight_rails`, `stair_flight`, the modern helix guard, `walkway`, the modern bridge rail, the modern roof-edge rails (`styled_building`, `styled_tower_roof_edge`), and the decorations balcony brackets, fence preview, bench, umbrella, flag, tank (whole tank, so it stays one surface), bistro chair, pergola table, tree seat, cafe table, lantern post (post and arm; the cage stays bare), ladder, utility box, quay fence, bollard and litter bin. `scripts/cluster_architecture.gd` (the modern glass parapet's top rail). `scripts/build_anim.gd` (`painted_metal` in `MATERIALS`, so painted pieces are revealed with their wall). `scripts/world_view.gd` (painted metal takes the rain). `scripts/main.gd` (report only, below). Tests below. Backups: `f/backup/`. 7: - **Vsync finding (the P1e review's bench warning):** vsync was already off in the P1d and P1e `--smoke --disable-vsync` benches. Two things made them look vsync-bound. (1) performance.json's `"vsync"` was a label, "disabled" for `--benchmark`/`--drag` and "enabled" for everything else, whatever the window did. (2) The frame samples are `_process` delta, and Godot's delta smoothing (`application/run/delta_smoothing`, on by default) snaps delta to the display refresh (120 Hz here) with or without vsync, which pins p95 at 16.667 (2 x 8.333). Probe (`f/vsync/probe.sh`, `probe_vsync.gd`, windowed, 640x360, a spinning box): without the flag the window is vsync mode 1 with raw frame p50 6.9 ms; with `--disable-vsync` it is mode 0 with raw frames of 0.5 ms, and the smoothed delta still snaps (p95 1.389 ms = 1/720 s). No script or project setting turns vsync back on (the only `window_set_vsync_mode` is main.gd's `--benchmark`/`--drag` line). Fix, report only; the game's vsync and delta are unchanged: `"vsync"` is now the window's real mode (`vsync_name(DisplayServer.window_get_vsync_mode())`), and the report adds `raw_mean_frame_ms`, `raw_p50_frame_ms`, `raw_p95_frame_ms` and `raw_sample_frames` (`Time.get_ticks_usec` intervals on the same frames as the old samples, `frame_stats`) and `delta_smoothing`. Old fields are kept as they were, so older benches stay comparable. Benches should read the raw fields; the raw mean carries the capture frames' outliers, so raw p50 is the steadiest. 8: - **Tests:** test-first logs `f/P1f-testfirst-test_scene.log(.out)` (parse errors on `PAINTED_KIND`, `pipe_tint`, `PIPE_COPPER`), `-test_build_anim` (1 failure: no painted build variant), `-test_warmup` (2 failures), `-test_drag_harness` (no `vsync_name`; the run hung on the script error and timed out, exit 124). After: test_scene 619 (was 588; new `painted_metal_checks`: the material and kind, metallic at most 0.2 and roughness 0.55 read from the shader, railings, posts and furniture all painted, tank legs, the cabinet, the pipe bore painted with its tint between 0.2 and 0.4 and the collars bare iron, copper vent, lantern cage and tree-house flashing bare, banisters painted with bare spines, BuildAnim and the rain list), test_build_anim 422 (was 413: one new check plus the MATERIALS loops over the new entry), test_warmup 34 (was 32: painted metal warmed on a mesh, an instanced MultiMesh and as its build variant), test_drag_harness 124 (was 118: vsync names, `frame_stats`, the report's fields, and the game's own vsync untouched), test_weathering 453 (was 452: a building now has one more group in its per-group loop). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/f/suite` (split a/b/v): **40 suites, 19,154 checks, 0 failures**. `--verify` 2880x1800 exited 0 in 33 s with `PIXELPUNK_INTEGRATION []`. The test_creature_reactions rerun had 0 failures (the known flake did not fire). 1440x900 was not run (the inherited stall). 9: - **Changed expectations:** test_scene "Buildings generate wrapped fittings" `groups.has("metal")` → `has("painted_metal")`; the tower fence count reads `painted_metal` (was `metal`); the neon cabinet is `painted_metal` (was `metal`); the pipe bore reads `groups.painted_metal` and the bent pipe's bounds merge `painted_metal` and `metal` (was `metal`). test_meet: the modern helix guard is drawn in `["glass","painted_metal"]` (was `["glass","metal"]`); Q1c "with no building at either end the banister is iron" counts METAL colours in `painted_metal` (was `metal`); "every stock connection draws byte for byte as before Q1c" now draws through `PaintedAsMetal` (painted pieces go back into the metal group), so the pre-Q1c fingerprints still match: geometry, colours and order are unchanged, only the material moved. test_building_tools: the parapet railing at a stair landing is counted in `painted_metal` (was `metal`). test_chemistry: `metal_in` counts `metal` and `painted_metal` (was `metal`). My own new check changed once: the cafe table has bare metal (its lamp cages), so it moved out of the "no bare metal" list into its own check. 10: - **Captures** (`f/caps.sh`, one sitting 22:22-22:28 on AC; `before-1..3` from the pre-tree, `after-1..3` live; views quarter day/night/05:45/20:45, harbour day/night, styles-day, lane-night; plus `before-j1`/`after-j1` harbour-lane at 22:33-22:34). Whole-frame counts are noisy in this sitting (`metrics/allpairs.txt`): framing flakes in before-3 (quarter views), after-2 lane-night, after-3 styles-day, and harbour-night sets 2 against 1 and 3. Region stats agree across sets to 0.1-0.5 luma (`metrics/metal.txt`), so the acceptance is read on regions. Sheets, looked at: `f/sheet/p1f-1-quarter-day.png`, `p1f-2-pipes-harbour-day.png`, `p1f-3-night.png`; frame looked at: `f/after-1/quarter-day/courtyard.png` (copper-brown bores read on every facade; the railing comb still reads black, and the blade signs read black by their SIGN_FACE fronts). 11: - **Acceptance** (`f/metrics/acceptance.txt`; luma 0-255, 0.05 = 12.75): 12: - **Brighter by day:** quarter-day pipes p50 0.4 → 10.6-10.9, mean 4.6-5.9 → 17.1-17.8, below 0.05 93-95% → 58-59%. **Met.** Top-terrace railing (`iron_rail`) p10 0 → 4.3-5.2, p25 0 → 8.1; harbour-day terrace railing p25 0.1-0.8 → 11.1; harbour quay fences p10 0 → 2.2; the FISH MARKET cabinet p25 0 → 8.6-9.0; right block p25 10.9 → 15.8. Measurably brighter, but still below 0.05. Quarter-day quay fences p25 12.1-12.5 → 12.9-13.0 (at the noise edge). Blade-sign cabinets, measured on the painted body's own pixels (the rim and side sliver that change, stable in both noise pairs; `f/fix/cabinet-parts.txt`, mask `f/fix/cab-qd-body-mask.png`): quarter-day 772 px, p50 0 → 5.4, mean 0.1 → 5.7; harbour-day 64 px, p50 0 → 12.1. **Met (brighter), still below 0.05.** The `cabinets` regions' own p10/p25 (0 → 0) cannot judge this: 71% of the quarter-day region is the sign face and ink, which this part does not touch. **Railings and cabinet bodies near-black: the part's stop rule, for the owner (below).** 13: - **Iron at night at or below 0.05 with the P1c grade:** quarter-night iron_rail p10 and p25 7.7 → 7.7, quay fences p10 7.7 → 7.7, harbour-night terrace railing p10 6.7 → 6.7. **Met**; the railings still read as dark lines against the windows (`p1f-3-night.png`). The bores are lighter at night (p50 8.1 → 13.6) because their albedo is. 14: - **Draw calls unchanged:** **not met, as the material route implies.** A mesh that holds both painted and bare metal draws one more surface: quarter 914 → 969 (+55), 05:45 884 → 939, 20:45 915 → 970, harbour-day/night 892 → 909 (+17), lane-night and harbour-lane 735 → 749 (+14), styles-day 484 → 487 (+3); dense bench 2140 → 2360 (+220), Harbour bench 902 → 919. In artwork.json the mixed decorations are the 22 pipes, 3 cafe tables and a pergola (`f/probe_mixed.log.out`); dense has none, so its +220 are building cluster meshes whose railings now sit beside bare hooks, bars, handles, lamp cages and trims. 15: - **Collars and copper bare:** test_scene. **Met.** 16: - **No-regression:** quarter-night below 0.1 29.0 → 28.7-28.8 (28-33); 20:45 28.8 → 28.2, 05:45 28.3 → 27.7, both at or under night (Q10). **Met.** 17: - **Bench** (`f/bench.sh`, 3 interleaved rounds against the pre-tree, Night, rain 0.45, `--disable-vsync`, 1920x1080; `f/bench.out`). Every run reports vsync "disabled" and delta smoothing on. Medians, raw frame intervals: dense p50 13.73 → 14.24 ms (+0.51; live above pre in each round, by 0.12-0.61), raw p95 15.58 → 16.22; smoothed mean 15.71 → 16.15 and smoothed p95 16.667 on both (the pin). Harbour raw p50 7.74 → 7.26, raw p95 8.40 → 8.36: no measurable change. The dense cost is the +220 draw calls. 18: - **Journey:** expected draw-call deltas: harbour-night and harbour-day +17, harbour-lane +14, quarter-night and quarter-day +55, styles-day +3. Looks: quarter-day copper-brown pipe bores on every facade; small lifts on the terrace railings and in the right block; harbour-day the terrace railing and the FISH MARKET cabinet slightly lighter. Night views are within their noise pairs on the metal regions. performance.json in every capture gains the `raw_*` and `delta_smoothing` keys, and `vsync` now says "disabled" or "enabled" truthfully. 19: - **For the owner (G2):** painted metal lights its paint, but the paint is METAL `#232a3e`, so railings and the cabinets' painted rim and sides stay near-black by day (cabinet body p50 5.4 quarter-day, 12.1 harbour-day). A lighter paint colour for the painted pieces (for example a mid blue-grey or green-black, one constant) would lift those and finish quick win 3; it would not lift the black face of each blade sign, which is SIGN_FACE, dark on purpose behind the neon (a separate question, P5a's); the plan's stop rule says ask rather than repaint. Separately, the +220 draw calls on dense (+0.5 ms raw p50) are the material route's cost; painting the few bare fixings inside building clusters (hooks, bars, handles) would win part of it back, at the cost of the report's "keep the fixings bare". 20: - **Review fix (28 September 2026):** the review found two must-fix items. (1) `tests/test_drag_harness.gd` had lost the space in its count line (`DRAG_TESTS124 checks`), so the cloud tally (`tools/suite_linux.sh:38`, the field before `checks,`) read it as 0 and dropped 124 checks. Fixed: `print("DRAG_TESTS ",...)` restored. That tally applied to the old summary gives 19,030 checks, and 19,154 on the new one (`f/fix/cloud-tally.txt`). (2) This entry judged the cabinets on the `cabinets` regions and blamed METAL for their black. Those regions are mostly the sign face (SIGN_FACE, `interior`) and ink, so the painted body shows only as a 1-2 px top edge and side sliver. Fixed in the Acceptance and G2 text above, and the body is now measured on its own pixels (`f/fix/cabinet-parts.js`; same capture sets as above, no new sitting because no code that renders changed; mask sheet looked at: `f/fix/cab-qd-body-mask.png`). test_drag_harness rerun: 124 checks, 0 failures (`f/fix/fix-test_drag_harness.log.out`). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/f/fix/suite` (a/b/v, AC, Low Power Mode 0, `f/fix/suite-power.txt`): **40 suites, 19,154 checks, 0 failures**, `DRAG_TESTS 124 checks` on line 20. `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun had 0 failures. 1440x900 not run (the inherited stall). No expectation or journey delta changes. 21: - **Review:** filled by the orchestrator. 22: - **Not done:** no lighter paint colour (stop rule above). The antenna, lamp cages, chimney flues and roof vents were left bare as fixtures, not posts. Glade was not captured (the plan does not ask for it). Harbour was benched as well as dense. Platforms: Mac run; Windows not run. ... SOURCE: artifacts/beautify/owner-decisions.md (frozen copy) 31: ## G1 verdict (owner, 28 September 2026: "accept recommendations") 32: 33: Given on the G1 sheet `artifacts/beautify/19-P1/c/g1/g1.md` (11 questions). P1c is accepted and merged; the changes below go into a new retune part, **P1c2**, which runs before P1d because P1d's quay ratio sits on the fills and toe. 34: 35: | Q | Verdict | Where it goes | 36: |---|---|---| 37: | 1 | Keep the saturated night water at Harbour (`#212660`) and Glade | no change | 38: | 2 | **Glade's night gets its own bluer-green palette back** (not derived from canal's violet haze); Glade day and golden stay exactly unchanged | P1c2 | 39: | 3 | Accept the night bank at +4.2 / +5.0 for now; **P3c (cake-slice banks) must meet bank >= water + 8 at night** | P3c | 40: | 4 | Quay at 21:00 line closed (14.9, no flash) | closed; re-measured only as a no-regression check | 41: | 5 | Moon-side crown lobes stay the darkest part of a tree | no change | 42: | 6 | **Retune the foliage night floor down** so quarter-night crown undersides are back near P1b's ~26 with the toe on | P1c2 | 43: | 7 | Accept lane-night at 26% below 0.1 (quarter in band) | accepted | 44: | 8 | Accept the slightly brighter horizon (inside P1a bands) | accepted | 45: | 9 | Accept the ~5% loss of reflection-column spread | accepted | 46: | 10 | **Raise the twilight fills (or twilight toe) so 20:45 and 05:45 are never darker than night** (share below 0.1 at or under quarter-night's) | P1c2 | 47: | 11 | **Lower the night fills and let the toe do more of the lift**, so mid-tones 0.3-0.5 move by less than 0.01 against the pre-P1c tree while quarter-night stays in 28-33% below 0.1 | P1c2 | 48: 49: ## P1c2 rulings (28 September 2026) 50: 51: - **Owner: 20:45 look, option B.** P1c2's 20:45 fill of 1.00 (needed for Q10) turned the warm sunset stone on shaded faces grey-violet (`artifacts/beautify/19-P1/c2/sheet/p1c2-7-2045-tower-crop.png`). Keep the lift but make it warm: the 20:45 fill/ambient leans toward the sunset (pink-mauve), so shaded sides read warm, not grey. Q10 must still hold. Done as an orchestrator fix pass on P1c2 before the merge. 52: - Orchestrator (reversible): the Harbour night bank lines at +3.7 / +4.5 (P1c +4.2 / +5.0) are accepted; holding them breaks Q11, and P3c already owes +8 at night. 53: - Orchestrator (reversible): 21:00-21:30 sitting 2-4 points lighter than 20:45 and night (share below 0.1) is accepted; no visible step, P1c had the same shape, and Q10 only forbids twilight being darker. 54: 55: ## G2 verdict and open P0 sheet questions (owner, 28 September 2026: "accept recommendations and move to P2 next") 56: 57: Answered before the G2 sheet was finished, on the orchestrator's list of questions. Any of these can be reversed when the owner sees the sheet. 58: 59: | # | Question | Verdict | 60: |---|---|---| 61: | G2-1 | P1f painted_metal raises dense draw calls 2140 -> 2360 (+0.5 ms raw p50) | Accept for now; revisit in a later speed pass | 62: | G2-2 | Repaint railings and cabinet bodies lighter? | No: keep the near-black paint (the 1-px iron railings are protected) | 63: | G2-3 | P1f also painted street furniture | Keep | 64: | G2-4 | P1d water bounce under 0.5 luma at plan strength | Keep it subtle | 65: | G2-5 | Quay about 1.5x the paving at 05:45, 20:45 and golden hour | Keep as is (the orchestrator had asked to judge it on the picture; the owner accepted without it) | 66: | G2-6 | Stepping stones: tide band, no line, dark at night | Fine | 67: | G2-7 | In-water pillars lighter over the shaft, darker foot | Fine | 68: | P0-1 | Far roofs with hidden tiles show a dark flat deck | Accept | 69: | P0-2 | Glade lip flaps read as bright lime squares | P3c redoes the lip | 70: | P0-3 | Harbour daytime water: soft darker band under the island's reflection | Keep | 71: | P0-4 | Stars in rain (D5: half at 0.35, gone by 0.6) | Keep | 72: | Gate | Blocking stop before P2 | Passed: go to P2 | ... ADDED 29 September 2026: the night stage close Frozen 29 September 2026 about 06:45 BST. The whole-stage measurements taken at the end of the night stage on 28 September. Line numbers refer to the frozen copy of docs/visual-handoff.md taken on 29 September, not the 28 September copy above. Local scratch-folder paths are removed. The frame-time cost it reports was still being investigated at this freeze and is quoted as the stage close recorded it. SOURCE: docs/visual-handoff.md (frozen copy, lines as of 29 September 2026) 75: **Beautify P1 stage close (28 September 2026):** the stage-end overhead of P1 "Night palette and values" (step `19-P1`), run by Opus 5.5 on this Mac on the final P1 code (`main` b8df6d0, clean tree; `close/head.txt`). No game code changed. Evidence: `artifacts/beautify/19-P1/close/`. Before side: the pre-stage scratch tree `close/pre-tree` from `build/backup-feel/19-P1/` (checked equal to tag `beautify-19-P1` = fc8650e on scripts, shaders, data, scenes, tests and `project.godot`). Power: AC, 100% charged, Low Power Mode 0 for every run (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). 76: - **Full suite** (`close/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/19-P1/close/suite`, 23:10-23:16): **40 suites, 19,154 checks, 0 failures**. `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun had 5,549 checks and 0 failures. The known flaky check ("The re-plan is applied after the edits") did not fire. The count is the same as P1f's `f/fix/suite`. 1440x900 was not run (the inherited stall). 77: - **Journey** (`artifacts/feel-journey/19-P1/`: the six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on the first try). Against `18-P0` (`19-P1/counts-vs-18-P0.txt`), only `draw_calls` and `rendered_objects` change, and exactly as P1f declared: harbour-day and -night 892 → 909, harbour-lane 735 → 749, quarter-day and -night 914 → 969, styles-day 484 → 487. P1a-P1e declared no count change. Re-run: the art capture below took every journey view again and matched on every count but one. The `after` set's quarter-day came out with a different, zoomed-in camera and 966 draw calls (`close/look/flake-qd.png`). Two single-view re-runs (`close/rerun-qd`, `rerun-qd2`) gave 969, the same as `after-2`, so this is a capture flake, not a count change. It is the "framing flake" P1e and P1f saw, in a stronger form. By eye (`close/look/j-1.png`, `j-2.png`): violet night sky and water, lifted darks, violet-grey ink, lighter quays with a tide band, and copper pipe bores by day; harbour-day and styles-day barely change. `notes.json` has a new `19-P1` entry (backup `notes.json.bak-before-19-P1`). `index.html` and `journey-standalone.html` were rebuilt (the standalone is now 105 MB). 78: - **Art capture and protected looks** (`close/caps.sh`; one sitting 23:20-23:27, in the order `pre-1`, `after`, `pre-2`, `after-2`; 24 views each, all on the first try: the kit's 21 baseline views plus harbour-lane and P1e's tide scene by day and night). Metrics are in `close/metrics/`: `ab-pre1-after`, `ab-pre2-after2`, noise pairs `ab-pre1-pre2` and `ab-after-after2`, `frame-table.txt`, `acceptance.txt` (`close/acc.js`), `protected.md` (`close/prot.js`), `tide.txt`, `counts.txt`, and `ab-p0close-*` against the previous stage close `18-P0/close/after`. The pre-stage tree's counts equal `18-P0/close/after` on all 22 shared views. The pre-stage noise pair changes >8 on 0.03-9.0% of the frame. The after pair changes 0.03-10.2%: the quarter night and twilight views run 3.7-10.2% from rain and TIME phase. Region medians in the after pair agree within 1.1 luma, with two exceptions: quarter-night `iron_rail` (2.3) and the small `trees_q` region at 21:00, which is unstable in both pairs (5.3 after, 38.1 pre). `after/quarter-day` is the camera flake, so `after-2` and `rerun-qd` stand in for it (`close/after-ok/` links them). Protected-look sheets: `close/protected/protected-1.png` to `-3.png`, plus `noise-pre.png`. Before = `pre-1`. 79: - Canal mirror columns (quarter-night `mirror_columns`): every orange, cyan and pink streak is there by eye at 2x. The water between them is lifted and violet (p50 35.8 → 42.5), and the spread is lower: sd 51.0/50.7 → 47.0/46.8, p90−p10 117.6/116.1 → 105.2/103.3. That is about 8%, against the ~5% (sd 48.0) the owner accepted at G1 Q9. Quarter-2100: p50 38.6 → 51.4. 80: - HDR neon cores: bar sign p90 244.3 → 244.3, fish sign p90 211.0 → 212.3, windows 175.9 → 176.0. The cores are unchanged. The regions change on 38-73% (>8) because the dark surround lifts under the toe, not the cores. 81: - 05:45 and 20:45 skies: the horizon bands are unchanged (114.7 → 114.9, 153.0 → 153.3), and the 20:45 zenith and sky_right are within noise. **The 05:45 zenith is 4.5 lighter (52.1 → 56.6, `#273662` → `#2b3a6a`, 26% >8 against 0.3-7.8% noise).** P1a had left it at 52.1. The lift is P1c2's 05:45 toe (0.775) and fill (0.90), the Q10 fix. It is still inside P1a's night band (45-55 applies to Night, not dawn), and it is recorded here for the owner. 82: - Glade/Canal separation: Glade night's zenith `#001b43`, horizon `#405e8c` and water `#022948` are identical to before P1 (0.00-0.34% >8). Canal night is `#273180` / `#222762`, so the two nights are clearly apart. Glade day lawn `#637f19` and bank `#6b4a2a` are unchanged (0.00%). By day the Harbour lawn changes on 1.52% (the pre pair also 1.52%), and the bank on 1.65% (pre pair 1.04%, after pair 0.61%). Both are 0.00% in pre-2 → after-2 and have equal medians, so they count as noise. Glade-night's frame changes on 0.99% against 0.09-0.31% noise: that is the foliage night floor on its crowns, as P1c2 recorded (0.93%). 83: - Approved tree style (grove-day): crowns 0.24% (noise 0.26%, 0.18%), conifers 0.14% (0.25%, 0.30%), feet and lawn 0.00%, colours equal. At night the crowns lift on purpose: quarter-night crowns `#030e0b` → `#101f27`, undersides `#000302` → `#0e1a23`. 84: - **Stage acceptance re-read on the final code** (`close/metrics/acceptance.txt`; pre-1/pre-2 → after/after-2; quarter-day from after-2 and rerun-qd): 85: - P1a: quarter-night zenith 31.1 → 54.4 (hue 220.9 → 233.8), horizon band 88.8 → 74.8/75.5, skyline − horizon −28.2 → −21.8/−22.5. Harbour zenith 27.3 → 52.6, band 87.7 → 71.7. Far walls: back_row hue 217.7 → 241.3, far_houses 248.3 → 261.7. There is no magenta cast. 86: - P1b: 21:00 town p50 11.8 → 37.7 (≥ 18), 21:00 front quay 0.6 → 32.9. The quay runs 38.4 (20:45), 32.9 (21:00), 25.3 (22:00), 23.0 (night), in order. crowns_low 0 → 25.8 (`#0e1a23`). 87: - P1c/P1c2: below 0.1, quarter-night 47.0 → 28.7/28.8 (28-33), lane-night 39.8 → 24.4, harbour-night 15.7 → 8.4. Q10: 20:45 28.2 and 05:45 27.7/27.6, both at or under night; 21:00 25.0 and 22:00 27.9. Iron p10 0 → 7.7 (≤ 12.75). Ink (outline pixels) p50 16.8 → 25.1 (lane 16.4 → 24.5, Harbour 19.9 → 27.2). Mid-tones 0.3-0.5 against the pre-stage tree: the median shift is 0.040 on quarter-night, 0.009 on lane and 0.050 on Harbour. That is not Q11's measure: Q11 is judged against the pre-P1c tree (P1c2: 0.0081/0.0076/0.0073), and this includes P1a's recolour and P1d's quays. 88: - P1d: quarter-night quay_over_paving 0.43 → 0.73/0.72 (≥ 0.55), p50 form 0.14 → 0.63, left pier 1.4 → 19.0, quay hue 230. Quarter-day 0.45 → 0.59/0.60 (after-2 and rerun-qd), left pier 11.9 → 24.3/24.5. 89: - P1e: `metrics/tide.txt`. On the tide scene by day, band pixels darken 37.1 → 27.2 (3.3% of the rect; noise 0.3%), while 15% lighten from P1d's stone lift. At stage level the band split is not the tide alone; the part's own numbers stand. 90: - P1f: quarter-day pipes p50 0.4 → 10.6, and below 0.1 94.7 → 74%. Right block p25 10.8 → 15.8, Harbour terrace railing p25 0.1/0.8 → 11.1. Harbour cabinets p25 0 → 0: the region is mostly the sign face, as P1f's review found. Quarter-night pipes 0.2 → 13.6, iron p25 0 → 7.7. 91: - **P0c night bank line (plan: re-measured after P1a):** harbour-night bank 37.2 → 45.6/45.0, open water 33.1 → 41.3, water below 34.0 → 40.4/40.5. **Bank − open 4.1 → 4.3/3.7, bank − below 3.2 → 5.2/4.5. The target of 8 is still not met.** The two after captures differ by 0.6 on the bank itself, so this line's noise is about ±0.6. It sits where P1c2 left it (3.7/4.5), and the owner moved +8 to P3c (G1 Q3). The other P0c lines hold: bank/lawn 0.89 at night, 0.86 by day, 1.48 at golden. Glade is unchanged (+14.9/+5.7). 92: - **P0e seam fade at night (plan: re-checked after P1a):** in the P0e region (x150-850, y880-1000) at quarter-night, the share above luma 130 is 7.79/7.67 → 7.74/7.60. The corner (x95-195, y945-1000) is 7.02/7.16 → 6.98/6.95. That is unchanged: the seam does not come back under the new night. The region's mean lifts 48.9 → 55.4 (the toe), and the mirror_seam hue goes 221 → 259 (the violet sky it mirrors). quarter-dry and quarter-2100 read the same (7.8 → 7.7; 5.3/4.8 → 5.1/5.1). By day and golden the region is 3.4/3.8 → 3.5 (after-2) and 11.3/11.1 → 11.3/11.1. 93: - **Whole-stage A/B bench** (`close/bench.sh`, `bench.out`, `bench-t13.out`, `bench-harbour-confirm.out`, `bench/summary.txt`, `bench/power.txt`). It ran 3 interleaved rounds, pre-stage tree against live, with `--disable-vsync` before `--path`, at 1920x1080, rain 0.45, at Night (`--time=1`, as the P1 part benches) and at 13:00 (the P0 stage bench's hour). Power: AC, 100%, Low Power Mode 0 at every start and end. The pre-tree's `scripts/main.gd` got only P1f's report-only diff (`close/backup/p1f-report-only-main.patch`; original `backup/pre-tree-main.gd.orig`), applied after the captures. With it, both sides report the raw frame intervals. The only remaining main.gd difference is P1c's grade wiring, which stays out of the pre side. Every run reports vsync "disabled" and delta smoothing on, with 491 samples. Raw p50 medians: 94: - Night, dense.json: 13.15 → 13.70 ms (+0.55, +4.2%). Round differences +0.58, +0.73, −0.63. Raw p95 15.49 → 15.48. P1f alone measured +0.51, which matches its +220 draw calls (2140 → 2360). 95: - **Night, Harbour: 6.55 → 7.49 ms in rounds 1-3 (+1.20, +0.89, +0.28).** Three confirm rounds (`HARBOUR_ONLY=1 ROUNDS="4 5 6"`) gave +0.29, +0.22, +0.33. Over 6 rounds, **6.54 → 7.36 ms (+0.8, live slower in every round)**, raw p95 8.22 → 9.15. Draw calls 902 → 919. The part benches did not show this: P1b had none, and P1c, P1c2 and P1d read the smoothed delta, which snaps to 8.33 ms on Harbour. P1f's raw Harbour showed no change of its own. The most likely sources are the per-pixel night work P1 added: P1b's foliage floor (three `pow`, a hash and the key estimate per fragment), P1c's grade in the full-screen pass, and P1d's bounce. This was not isolated. 96: - 13:00, dense: 13.14 → 14.21 ms (+1.07; rounds +1.45, +0.05, +0.54), raw p95 15.59 → 17.10. 13:00, Harbour: 8.05 → 8.18 (+0.13; rounds +0.13, −0.02, +0.71). At 13:00 one round per tree drew more (dense pre 2184 against 2140, live 2463 against 2360, Harbour live 936 against 919). That happened in both trees, so it is not P1. 97: - **Owner sheets** `artifacts/beautify/19-P1/close/sheet/`. Before (pre-stage tree) is on the left and after on the right, with colour and greyscale where values matter. Crops are from `close/crops.js` and the pages from `close/sheet.sh`; all pages were looked at. Stage pages: `p1-1-overview.png`, `p1-2-P1a-night-sky.png`, `p1-3-P1b-twilight.png`, `p1-4-P1c-night-values.png`, `p1-5-P1c2-retune.png`, `p1-6-P1d-quays.png`, `p1-7-P1de-pier-and-tide.png`, `p1-8-P1f-painted-metal.png`, and `p1-9` to `p1-11-existing-scenes-*.png`. G1 (whole night look): `g1-1-canal-night.png`, `g1-2-dawn-dusk.png` (05:45, 20:45, 21:00), `g1-3-glade-night.png`. G2: `g2-1-views.png` (quarter-day, harbour-day, quarter-night), `g2-2-quays-and-tide.png`, `g2-3-regions.png`. `sheet.md` has one plain paragraph per part with the numbers, then the bench and the suite. 98: - **Not done / open:** 99: - The Harbour night cost (+0.8 ms raw p50, above) is not isolated to a part. P2 or the orchestrator should bench P1b's floor and P1c's grade on their own if it matters. 100: - P1f's G2 question is still open: the railings and cabinet bodies stay near-black by day because METAL `#232a3e` is. 101: - The P0c night bank line is at +4.3/+5.2 (after) and +3.7/+4.5 (after-2) against 8. It is owed by P3c (G1 Q3). 102: - The reflection-column spread is down about 8% against before P1, more than the ~5% accepted at G1. The 05:45 zenith is 4.5 lighter. Both are for the owner. 103: - The camera flake in `after/quarter-day` (a different zoom, 3 fewer draw calls) was seen once in 5 live captures of that view. Its cause was not chased. 104: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. P1g (M1 builds) is not run (D8). 105: - Scratch trees: the `pre-tree` folders under `19-P1` (`a`, `b`, `c`, `c2`, `d`, `e`, `f` and `close`) were moved to the system trash with `mv` once their numbers were recorded, named `beautify-19-P1--pre-tree-