Chapter 20: A mist the town can light Frozen 29 September 2026 about 06:45 BST. Implemented, reviewed and merged work; decisions by the owner (the project's creator) and review rulings are in the second source. Original source terminology is preserved; local scratch paths are removed. Line numbers refer to the frozen copies. Part codes: P2 is the beautify stage of settled choices (owner decisions D9 to D12). P2c fog lit by the city and warm chimney smoke, with two fix passes after review. P2a leaded windows, P2b fog and smoke defaults, P2d contact shade and P2e tree-foot collars are other parts of the same stage and are not quoted here. Captures use the flags --height-fog=0.3 --chimney-smoke=on ("flags"). SOURCE: docs/visual-handoff.md (frozen copy, lines as of 29 September 2026) 45: **Beautify P2c Fog lit by the city, lighter by day; warm smoke over lit houses (29 September 2026):** Opus 5.5 on this Mac. The height fog is now tinted by the town's light and drawn lighter by day. At night its colour is the hour's haze leaning toward the horizon and lifted, `haze.lerp(horizon, 0.45*windows).lightened(0.08*windows)`, so the mist glows violet (rose at 21:00) instead of only darkening. What the shader draws is the setting times the hour's share: 0.45 of it by day and at golden hour, all of it from the twilight anchors' windows (0.85, at 05:45 and 20:45) through the night. At night every chimney plume starts warm (toward `ffb070`) and cools into the dusky grey. Glade's smoke is paler. Report quick win 6 (colour, hour scaling, warm smoke base); the defaults are P2b's. The look is reached with `--height-fog` and `--chimney-smoke=on`. Branch `beautify/P2c` from `main` 791ff0c. Snapshot: `build/backup-feel/20-P2-c/` (same as `20-P2`). At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/20-P2/c/pre-tree/`. Evidence: `artifacts/beautify/20-P2/c/`. Power: AC, 100% charged, Low Power Mode 0 for every capture and the suite (`cap-power.txt`, `suite-power.txt`). 46: - **Files:** `scripts/height_fog.gd`: `HORIZON_SHARE` 0.45, `LIFT` 0.08, `DAY_SHARE` 0.45, `LIT_WINDOWS` 0.85, `share`, `drawn()`, static `night`, `hour_share` and `lit`. `configure` keeps the setting in `amount` and `tint` sets the colour and the share; the shader gets `amount*share`. Fix pass: `expose`, static `compat_out`, `compat_in`, `veil_albedo`, `SRGB_A/B/C`, and the shader's `glow_amount` (the setting) set in `configure`. `shaders/height_fog.gdshader` (fix pass): `glow_amount`, `exposure`, `compat_linear/compat_out/compat_in` and the Compatibility-only ALBEDO. `scripts/world_view.gd` (fix pass): `sync_environment` calls `fog.expose(tonemap_exposure)`. `scripts/chimney_smoke.gd`: `WARM` `ffb070`, `WARM_MIX` 0.7, `WARM_UNTIL` 0.2, and the ramp as `RAMP_OFFSETS`/`RAMP_ALPHAS` (a 0.2 point added on the old alpha line, so the opacity curve is unchanged). `tone(profile, glade)`, `warm_base`, `ramp_colors`, `paint` and `body`. The colour now lives in the ramp and the emitter's `color` is white. `scripts/look_profile.gd`: `GLADE_SMOKE_PALE` 0.4. `scripts/world_view.gd`: `sync_profile` calls `smoke.paint(profile, is_glade)`, and `smoke_color()` returns `smoke.body`. `scripts/main.gd`: `evening_report` adds `height_fog_drawn`. `tests/test_dusk.gd`. Backups: `c/backup/`. 47: - **Decisions:** (1) Warm smoke is **uniform** (the plan allowed uniform or per plume). Every plume warms alike, scaled by `HeightFog.night` (none by day or at golden hour, full from windows 0.85, i.e. 20:45). It does not wait for the houses: at 20:45 and 21:00 the warmth is already full while the dusk sequence is still lighting houses (`Dusk.progress` 0.812 and 0.832, `c/fix/probe_dusk.log.out`; it reaches 1 only at 22:00), so a plume can be warm over a house whose windows are still dark. (2) `performance.json` `evening.height_fog` still reports the **setting**, which journeys compare with the flag. The new `height_fog_drawn` reports the amount drawn at that hour. (3) The first tuning round used the full share only at Night's windows (1.0). That drew 0.78 at 21:00, and the fog-2100 mist came out no brighter than before (−0.5/−0.3 luma, `c/round1/metrics/`). The full share now starts at the twilight anchors' 0.85. 48: - **Tests:** test-first `c/P2c-testfirst-test_dusk.log(.out)`: 166 checks, 20 failures. The final tests on the pre-part tree fail 18 of 169 (`c/P2c-testfirst2-test_dusk-final-tests-on-pre-tree.log.out`). After: test_dusk 169 (was 136). New `fog_hour_checks`: all ten anchors give a defined colour and amount that follow the formula, in canal and in Glade. 13:00 and 19:30 draw 0.45. 01:00, 04:00, 05:45, 20:45, 21:00 and 22:00 draw everything. The fog thickens steadily from 19:30 to 20:45 and thins steadily after 05:45. The night mist is lighter than the bare haze. Glade's night fog comes from Glade's own palette. New `smoke_warm_checks`: the opacity curve is unchanged. At night the plume is warm at 0.06 and grey from 0.2. There is no warm base at 13:00 or 19:30. Glade's smoke is paler day and night and still dusky at night. main_checks covers the evening report. Affected suites `c/r1-*`: test_look 359, test_camera 191, test_warmup 34, test_drag_harness 124, all 0 failures. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/c/suite` (a/b/v): **40 suites, 19,187 checks, 0 failures**. `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 49: - **Changed expectations:** test_dusk "height_fog 0.6 draws the fog at 0.6" (the shader's `amount` is 0.6) → the setting stays 0.6 (`fog.amount`), and the shader draws 0.6 × the hour's share (0.6 at 21:00). "The fog is the hour's haze" (21:00) → the lit formula at 21:00. "The fog follows the hour" (13:00 haze) → the lit formula at 13:00, plus 0.45 × the setting drawn at noon. "Smoke is pale by day and dusky at night" now reads `smoke.body` (the ramp's tone), because the emitter `color` is white. main_checks: `evening_report` gains `height_fog_drawn`. 50: - **Captures** (`c/caps.sh`, one sitting 29 Sep 00:08-00:15, AC; tuning round 28 Sep 23:55-00:03; `before-1/2` from the pre-tree, `after-1/2` live, 17 custom views each, all on the first try): fog-2100 (artwork `--time=21 --height-fog=0.3`), q2100-off, qn-flags, qn-fog, q2045-flags, qd-flags, hn/hg/hd-flags, hd-fog, hg-fog, hd-off, hg-off, gn-flags, gd-flags, ln-flags, ln-fog ("flags" = `--height-fog=0.3 --chimney-smoke=on`). The tuning round's sets are in `c/round1/`. Integer counts and **draw calls are identical on all 17 views in both sets** (`metrics/counts.txt`). The only new key is `evening.height_fog_drawn`. The fog-off views are within their noise pairs (`metrics/noise.txt`). Sheets, looked at: `c/sheet/p2c-1-canal-evening.png` (fog-2100 and 20:45), `-2-harbour-day-golden.png`, `-3-night.png`, `-4-glade.png`, `-5-harbour-night-plaza.png`, `-6-glade-night-plaza.png` (full-resolution crops; the plaza lamp glow is unchanged at night: plaza region mean luma 83.2/83.1 → 84.0/83.9, Glade 100.5/100.5 → 100.3/100.3; `metrics/plaza.txt`, rect x760-1040 y440-515, `c/plaza.sh`, crop `metrics/plaza-crop.png`), and `-smoke-lane-night.png` (a dusky blue plume before; a warm base cooling to grey after). 51: - **Acceptance** (`c/acc.js` → `metrics/acceptance.txt`, `c/plume.js` → `metrics/plume.txt`): 52: - **fog-2100 mist brighter, hue toward the horizon:** measured on the mist pixels, the 19.3% of the frame where the pre-part fog moves the frame by at least 4 luma in both before sets. Mean luma 96.6/97.1 → 99.4/99.4 (+2.7/+2.3; noise pairs +0.5 and 0.0). Brighter by >1 on 71/79% of those pixels (noise 13%/2%). The mean colour's hue goes 341 → 344, toward the 21:00 horizon (hue 9, from the haze side at 336). The fog's own addition (fog − off) goes 6.4/−0.8/−1.4 (hue 5) → 10.1/2.0/−0.8 (hue 15). **Met.** 53: - **Day and golden veil at most 0.45 as strong:** the drawn amount is exactly 0.45 × the setting (tests; `height_fog_drawn` 0.135 in hd/hg). In the picture, mean |fog − off| per channel over the pixels the pre fog moves, after/before: day 0.194/0.203, golden 0.403/0.430 (off-vs-off noise 0.45 luma against a pre veil of 9.6-10.0). **Met.** Finding: the veil's picture effect is not linear in amount. The pre-part tree draws +0.0 at 0.01, about −1.4 at 0.135 and +9.2 at 0.3 (`metrics/day-veil-linearity.txt`, `c/lin.js` on this sitting's sets plus the 00:05 probes; the 0.01, 0.135 and 0.667 points are one capture each; the tuning round's copy is `round1/metrics/`), so the lighter day veil very slightly darkens the water (−1.5 luma mean) instead of milking it. The cause was not isolated (the presentation quantiser is the likely one). **At golden hour the plaza lamp glow in the fog drops with the veil: plaza region mean luma 90.1/90.2 → 60.2/60.0 (p50 83.0 → 49.9; fog-only view 90.1/90.0 → 60.2/60.3), against 38.7 with no fog, so the fog's glow over the plaza goes from +51 to +21.** By day the region goes 61.1/61.2 → 57.7/57.9 (no fog 58.7 → 58.6). `metrics/plaza.txt`. 54: - **Plumes start warm and cool to the dusk grey (uniformly):** lane night, smoke pixels at the plume bases, r−b 3.0/8.4 → 48.1/43.0 (mean `#574554` → `#7d5c4d`); above them 14.4/17.3 → 18.2/25.7. Golden and day smoke are unchanged in warmth (warm share 22.3/20.6% → 20.8/20.2% golden; 3.3/1.9% → 2.4/3.3% day). 55: - Glade: the night mist takes Glade's own bluer palette, and the frame is slightly darker (mean dluma −0.37, >8 on 1.3-1.6%). Glade by day is the lighter veil (−5.3) plus the paler smoke. 56: - **Bench:** none (the plan's P2c has none: uniform values and one ramp). Draw calls are unchanged on every view. `paint` rebuilds a five-colour array on each `sync_profile`. 57: - **Journey:** expected count deltas: none. The journey views carry no fog or smoke, and their draw calls are unchanged. Every `performance.json` gains `evening.height_fog_drawn` (0 without fog), and `art_metrics counts` prints it as `undefined -> 0`. Looks: journey views unchanged (within noise). 58: - **Review:** filled by the orchestrator. 59: - **Fix pass (orchestrator must-fix, 29 September 2026):** evidence `c/fix/`; AC, 100%, Low Power Mode 0 for every capture, probe and suite (`fix/cap-power.txt`, `fix/probe/power.txt`, `fix/suite-power.txt`). At start the live tree differed from `20-P2-c` in exactly P2c's six files; nothing unexpected. P2c as reviewed was copied to `fix/p2c-tree` before editing (backups `fix/backup/`). 60: - **Asked (1):** the lamp glow must not drop with the day/golden veil share (golden plaza 89.5 → 60.2). **Changed:** the glow has its own `glow_amount` = the setting at every hour; the veil keeps `amount` = setting × share (0.45 by day and golden). The glow now adds exactly its pre-P2c step over the pre-P2c veil at the setting (`compat_out(haze·f_glow + glow) − compat_out(haze·f_glow)`). **Result** (sitting `fix/sit2`, pre / P2c / fix, two sets each): golden plaza region 90.0/90.1 → 60.0/60.0 → **93.5/93.5**. With the lamp glow switched off (probe 4) the region is 38.1 (pre veil 0.3), 41.8 (fixed veil at 0.135), so the glow's own share is 51.9 pre and about 51.7 now: kept. The +3.5 over the target is the fixed veil itself (item 2), not the glow, so the line "within the noise pair of 89.5" is **not met** as written; the glow is. Night plaza 83.2/83.2 → 84.0/83.9 → 86.5/86.5, Glade night 100.5 → 100.4/100.3 → 104.3/104.3: the glow is unchanged at night by construction, the +2.5/+3.9 is the fixed veil. 61: - **Asked (2):** the fog must never darken. **Cause (probe, `fix/probe/summary.txt`, `fix/diffmap.js` maps):** under GL Compatibility Godot 4.6 turns ALBEDO into the stored colour per fragment before the premultiplied blend: `linear_to_srgb(srgb_to_linear(ALBEDO)·exposure)`, a cubic then a pure-power encode clamped at 0 (`drivers/gles3/shaders/scene.glsl` and `tonemap_inc.glsl`, 4.6.3-stable; tonemapping is in post because glow/SSAO are on). A null-output quad changes nothing; a constant added ALBEDO of 0.02 stores nothing, 0.03 about 1 luma, 0.05 about 10; ALPHA scales the stored frame. The premultiplied haze·f (≤0.12) sat in that crushed foot while the frame was scaled by 1 − f, so the day veil darkened (and at night the mist only darkened more than it lit). **Fix at the root:** under Compatibility the shader writes `compat_in(compat_out(haze)·f + glow step)`, the ALBEDO whose stored colour is the haze's own stored colour times f: a plain mix toward the haze. Forward+ keeps the old line. **Linearity re-measured** (`fix/metrics/fix-metrics-sit2.txt`, mask = pre fog 0.3 moves ≥4 luma in both pre sets, 43%/44% of the frame; dluma vs 6 offs, off noise ±0.02): day, drawn 0.01/0.05/0.1/0.135/0.2/0.3: pre −0.04/−1.52/−3.40/−1.51/+2.83/+9.18 → fix **+0.00/+2.42/+4.84/+6.40/+9.30/+14.01** (set 2 +0.04/+2.46/+4.84/+6.40/+9.31/+14.05); the sky (rows 0-150), which fell −4.0 at 0.1, now rises 0.0 → +3.8. Golden: pre −0.00/−1.18/−2.37/−1.13/+3.51/+10.1 → fix +0.02/+2.72/+6.52/+9.01/+12.82/+18.36. Monotone and never darker. Glade day (setting 0.3): frame −0.17 (P2c) → +3.27, darker >2 on 2.3% → 0.2%. **fog-2100** mist (18.7% mask): 98.5/98.6 → 101.0/101.0 → **106.1/106.1** (≥ 99), mean hue 342 → 345 → 345 (toward the 21:00 horizon, hue 9, from the haze, 336); the fog's own add 6.6/−0.7/−1.4 → 10.0/1.9/−0.9 → 13.7/7.4/4.2. 62: - **Consequence for the P2c acceptance line "day and golden veil at most 0.45 as strong":** the drawn amount is still exactly 0.45 × the setting, and the fixed veil is linear (drawn 0.135 against drawn 0.3 on the fixed shader: mean |fog − off| 0.46 day, 0.50 golden with the glow). Against the pre-part picture, whose veil was crushed, it is now 0.70 (day, 6.66 against 9.53) and 0.90 (golden, glow restored) (`fix/metrics/veil-strength-sit2.txt`). For the owner/P2b: the day veil is now a real, if light, milk (+6.4 on the water at 13:00, `fix/sheet/fix-hd-fog-water-2x.png`). 63: - **Asked (3):** the uniform-smoke reason. The `scripts/chimney_smoke.gd` comment and Decision (1) now give the real timing (above); behaviour unchanged. 64: - **Tests (written first):** `fix/testfirst-test_dusk.kept.out`: 182 checks, 8 failures (glow_amount ×5, the shader's glow line, exposure, `veil_albedo` missing). The glow-step form of the expected value was tightened after sitting 1 (`fix/sit1`, glow as `compat_out(glow)`: golden plaza 85.5, glow share 43.7) showed the plain form lost 8 luma of glow. New `fog_fix_checks`: glow_amount = setting at 13:00, 19:30, 21:00, 01:00 and after a new setting; the shader scales the glow by `glow_amount`; the fog gets the look's exposure (canal 1.0, Glade 1.5); the model of Godot's transform has the probed foot; the old ALBEDO darkens water under a thin day veil on that model (the cause); through the test's own copy of Godot's transform `veil_albedo` gives haze × f plus the glow's pre-P2c step (worst error < 0.002, both exposures, four hours); by day and golden the veil is never darker and rises steadily from 0 to 0.3; the shader carries the same constants and inverse. The GPU path itself is not testable headlessly; the probe is the record. After: test_dusk **187** (was 169), test_look 359, test_camera 191, 0 failures (`fix/test_*.log.out`). Full suite `fix/suite.sh` a/b/v (`VERIFY_RES=2880x1800`): **40 suites, 19,205 checks, 0 failures**; `--verify` 2880x1800 exit 0 in 31 s, `PIXELPUNK_INTEGRATION []`; the creature_reactions rerun 0 failures (the known flaky check did not fire). 1440x900 not run (inherited stall). 65: - **Changed expectations (fix pass):** none of P2c's old checks changed; new checks only. Captures: integer counts and draw calls identical P2c → fix on all 41 views (`fix/metrics/counts-p2c-live.txt`); fog-off views within their noise pairs (P2c → fix hd-off mean |dluma| 0.22 against 0.39 live pair). Journey count deltas: none (no fog in the journey). Sheets, looked at at 2x: `fix/sheet/fix-hg-flags-plaza-2x.png` (pre | P2c | fix: the warm lamp pools are back), `fix-hg-flags-water-2x.png`, `fix-hd-fog-water-2x.png` and `fix-hd-fog-plaza-2x.png` (the day veil is a light milk, not a dimming), `fix-off-ref-2x.png`. Bench: none (one full-screen pass gains two pow per channel and a six-step loop; draw calls unchanged). 66: - **Second fix pass (orchestrator must-fix items 1 and 2, 29 September 2026):** evidence `c/fix2/`. Power: AC, 100%, Low Power Mode 0 for every capture, test and suite (`fix2/cap-power.txt`, `power-tests.txt`, `suite-power.txt`). At start the live tree differed from `20-P2-c` in exactly P2c's six code/test files. Fix pass 1 as reviewed was copied to `fix2/fix1-tree` before editing (backups `fix2/backup/`). Final sitting `fix2/sit` (02:13-02:30): pre-1/2 (`c/pre-tree`), fix1-1/2, live-1/2, plus pre and live with the lamp glow off (`fix2/probe/tree-*-noglow`, `lamp_glow` 0), two sets each. Every view was taken on the first try. Metrics: `fix2/metrics.js` → `metrics/fix2-metrics-sit.txt`; `fix2/block.js` → `metrics/veil-strength-blocks-sit.txt`. 67: - **Asked (2):** say what the "at most 0.45 as strong" line is measured against. Once fix 1 stopped the veil darkening, it drew about 1.5 times the pre-part veil at the same setting (day 14.4 against 9.5; golden, glow off, 11.5 against 7.1). **Changed:** under Compatibility the veil's opacity is now `f × veil_match`, where `HeightFog.VEIL_MATCH` is 0.46. A setting now draws no more than its pre-P2c strength. The drawn amount does not change: it is still exactly 0.45 × the setting by day and golden hour (`height_fog_drawn` 0.135), so the line holds against both references. 0.55 came first, from fix 1's probes, and gave golden 0.450 (blocks) / 0.487 (pixels) (`fix2/sit-k055`, `metrics/fix2-metrics-sit-k055.txt`). **Result** (veil alone, against the pre-part tree in the same sitting, mean |fog − off| per channel on 8×8 block means, which removes the per-pixel TIME noise): day **0.326**, golden **0.406** (target ≤ 0.45). With the glow included: day 0.322, golden 0.636, because the glow is kept at the whole setting (item 1). Against the live shader's own full setting (drawn 0.3), the day figure is 0.47. The per-pixel metric used by fix 1: day 0.381, golden 0.466. Its noise floor is about 0.5 per channel (off/off 0.48-0.64), which is a fifth of a light veil; signed mean dluma gives 0.31 / 0.40. Fix 1 gave 0.674 / 0.887 on blocks. The veil never darkens: at every drawn amount from 0.01 to 0.45 the day and golden mean dluma is ≥ −0.04, and the share darker than off by more than 2 luma (0.7-2.4%) stays at the off-pair noise level (pre 0.8-2.2%). 68: - **Asked (1):** the golden plaza should be within the noise pair of the pre-part picture, and night should be unchanged. **Changed:** the glow's step is now taken over the pre-P2c veil in its pre-P2c colour: a new `glow_haze` uniform carries the bare haze, set in `tint`, where fix 1 used the lit colour. `veil_albedo` takes it as a sixth argument. With `VEIL_MATCH` added, the plaza region (x760-1040 y440-515), pre → fix 1 → now: golden 90.0/90.0 → 93.5/93.5 → **91.8/91.9**. Harbour night 83.2/83.2 → 86.6/86.5 → **84.9/84.9**. Glade night 100.5/100.5 → 104.2/104.2 → **103.4/103.3**. Day 61.5/61.1 → 64.9/64.8 → 63.1/63.2. The glow's own share (fog view − glow-off view, same tree) is kept: golden 51.7 → 51.8, night 40.4 → 40.5, Glade night 49.3 → 49.1. **Not met as written, and left open for an orchestrator ruling.** The pre-part numbers sit *below* what "no fog + the pre glow" gives (golden 90.3, night 84.3, Glade 102.6), because the pre veil darkened the plaza (golden −0.4, night −1.1, Glade −2.1). Under fix 1's "the fog never darkens", with the glow kept, those numbers can only be reached by trimming the glow by the veil's own lift (+1.5 / +0.7 / +1.0 now), and by a different amount per hour and look. That would be a design choice this pass should not take. 69: - **Other lines, re-measured:** fog-2100 mist (17.9% mask): 100.0/100.3 → 107.5/107.6 → **105.1/105.0**, brighter than pre by >1 luma on 76/81%, mean hue 344 → 345 (toward the horizon); the fog's own addition is 6.6/−0.7/−1.5 → 10.3/5.1/2.6. Harbour night fog − off over the frame: −1.51 (darker on 71%) → 3.59 → 2.26 (darker 1.1-1.4%). Glade day frame +5.17 → +3.27 → +1.56; Glade night +2.60 → +6.66 → +3.84. Integer counts and draw calls are identical fix 1 → now on all 23 views (`metrics/counts-fix1-live.txt`). Pre → now differs only by `height_fog_drawn`. 70: - **Tests (written first):** `fix2/testfirst-test_dusk.kept.out`: 208 checks, 13 failures (VEIL_MATCH, veil_match ×4, glow_haze ×4, the shader's veil, glow_haze and Forward+ lines, `veil_albedo` arity). New `fog_fix2_checks`: VEIL_MATCH pinned (0.55 at test-first, then 0.46); the shader gets veil_match and the bare haze at 13:00, 19:30, 21:00 and 01:00, while the drawn amount stays 0.45 / 1.0 of the setting and fog_color stays lit; the Compatibility branch scales opacity and colour by veil_match; Forward+ keeps the pre-P2c blend; through the model, `veil_albedo` with the bare haze gives the lit colour × f plus the glow's step over the bare haze (worst error < 0.002). After: test_dusk **209** (was 187), test_look 359, test_camera 191, all with 0 failures (`fix2/test_*.log.out`). Full suite (`fix2/suite.sh` a/b/v, `VERIFY_RES=2880x1800`): **40 suites, 19,227 checks, 0 failures**. `--verify` 2880x1800 exited 0 in 31 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 71: - **Changed expectations:** none of the old checks changed; this pass only adds checks. Journey count deltas: none. Sheets (2x, looked at): `fix2/sheet/hg-flags-plaza-2x.png` (pre | fix 1 | now | off: warm lamp pools as before, a touch lighter than pre), `hn-flags-plaza-2x.png`, `hd-fog-plaza-2x.png`, `hd-fog-water-2x.png` (the day veil is now a faint milk, hard to tell from off), `fog-2100-1x.png` and `full-hg-hd-1x.png`. Bench: none (one multiply; draw calls unchanged). For P2b and the owner: at the default 0.3 the day veil is now light: at 13:00 the fogged area (43% of the frame, mostly water) is about +2.9 luma, against +9.2 before P2c. 72: - **Review fixes (29 September 2026):** the review found two must-fix gaps in this entry's evidence, no code faults. (1) `metrics/day-veil-linearity.txt` was cited but existed only under `round1/metrics/`: `c/lin.js` was re-run on the final sitting into `metrics/day-veil-linearity.txt` (+0.02 at 0.01, −1.43 at 0.135, +9.18/+9.17 at 0.3, after −1.53/−1.56; the same as the tuning round within 0.1). (2) The plaza lamp-glow numbers had no file or rectangle: `c/plaza.sh` now measures a recorded rect (x760-1040 y440-515) with `art_metrics.js region` on all four sets and eight views into `metrics/plaza.txt`. The night and Glade numbers above replace the unreproducible 48.9 → 49.1-49.3 and 52.3 → 52.3 (a different, unrecorded rect); the golden drop holds (was 89.5 → 60.2, now 90.1/90.2 → 60.2/60.0). No code, test or capture changed, so no suite or Godot run was repeated. 73: - **Not done:** no per-plume warmth (the uniform route was chosen). Harbour and artwork were not benched (P2b benches the defaults). The fog's nonlinear day response was chased in the fix pass (above). The pre-tree is left for the review. Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy) 1: # Owner decisions for docs/beautify-plan.md (section 6) 2: 3: Answered by the owner in chat on 25 September 2026, while P0a ran. These override the plan's defaults where they differ. 4: 5: | # | Answer | Differs from the default? | Effect on the plan | 6: |---|---|---|---| 7: | D1 | Mains power with Low Power Mode off (the owner sets this up) | yes | Benches record the power state; never mix power states inside one A/B. The P0a baseline and pre-flight suite may have run on battery: note it, don't compare their timings with mains runs. The flaky creature check should stop failing on mains; if it still fails, record it as known. | 15: | D9 | Leaded medieval windows; no "under 0.6 m transom only" rule | no | P2a | 16: | D10 | Fog and smoke on for starters, SampleScene and the new canvas; `artwork.json` unchanged | no | P2b | 17: | D11 | Canal contact shade, cost reported | no | P2d | 18: | D12 | Collar follows the ground finish | no | P2e | 74: ## P2c ruling (orchestrator, 29 September 2026; reversible by the owner) 75: 76: - The P2c fix pass gave the lamp glow its own amount and fixed the fog's darkening under GL Compatibility (a stored-colour floor swallowed the small veil term). The golden plaza now reads 92 against 90 before P2c, and the night plazas 2-3 luma lighter, because the veil no longer darkens them; the glow's own share (fog view minus the glow-off probe) is kept at every hour (golden 51.95 vs 51.9). Ruled: item 1 is measured by the kept glow share, not the old totals. The day/golden veil is now a real light milk, at most 0.45 as strong as the full setting against both references. ADDED 29 September 2026 (later): fog and smoke by default, and the canal contact shade Frozen 29 September 2026 at 07:41 BST (game main at 20284d6). P2b (height fog and chimney smoke on for new scenes and starters) and P2d (canal contact shade, with the owner's ruling (b) applied in its third fix pass) were implemented, reviewed and merged. Line numbers refer to the frozen copy of docs/visual-handoff.md taken at 07:41 on 29 September, not the copy above. Original source terminology is preserved. The warm lamp halo in the fog at golden hour and night (P2b sheet notes) is left for the P2 stage review. P2a leaded windows and P2e tree-foot collars are not quoted here. SOURCE: docs/visual-handoff.md (frozen copy, lines as of 29 September 2026, 07:41) 33: **Beautify P2b Height fog and chimney smoke on for new scenes and starters (29 September 2026):** Opus 5.5 on this Mac. Every starter the menu opens now carries `height_fog` 0.3 and `smoke` true (report quick win 6, the defaults; owner D10). Glade starts at 0.15. A saved document without the keys still opens with no fog or smoke and saves byte for byte. Branch `beautify/P2b` from `main` 109caf0. Snapshot: `build/backup-feel/20-P2-b/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch trees: `artifacts/beautify/20-P2/b/pre-tree/` (before) and `b/bench/post-tree/` (a copy of the final live code, for the bench). Evidence: `artifacts/beautify/20-P2/b/`. Power: AC, 100% charged, Low Power Mode 0 for every capture, bench round and suite (`cap-power.txt`, `bench/power.txt`, `suite-power.txt`, `suite-final-power.txt`). 34: - **Where the defaults live (D10) and the blast radius.** `LookProfile.START` gains the two keys: canal night 0.3 and smoke, Glade 0.15 and smoke. New `LookProfile.ATMOSPHERE_KEYS` and `starter_atmosphere(environment)` write the look's two keys and nothing else. They are called by every menu starter: the four New canvas demonstrations (`StarterDemos._canvas`), Blank canvas (new `StarterDemos.blank()`, used by `main.gd`'s menu handler), `SampleScene.courtyard`, `harbour`, `castle` and `styles`, `ArtworkScene.build` (Canal quarter) and `AmsterdamScene.build`. They are **not** called by `DioramaState.new()` (the `scene_state.gd:230` member default is unchanged, so in-code fixtures and tests that build their own scenes are as before), `SampleScene.benchmark` (`--benchmark`) or any saved document. Blast radius: every suite, tool and `--smoke`/`--drag` fixture that builds one of those starters now gets fog and smoke. That includes the game's opening Harbour and the default `--smoke` scene, so the harbour captures and journeys, and `--look=glade` captures, now show them. Two effects follow from `choose`. Choosing a look now writes the look's fog and smoke. Switching an older document to either look adds them, as one undoable edit. Two capture fixtures that tests regenerate from `SampleScene.styles()` are kept as they were: `test_lighting` writes `artifacts/lighting/styles.json` (the styles-day capture) without the two keys, and `test_building_tools` writes `artifacts/building-tools/styles.json` the same way. Both are byte-identical to `main` after the suite. `artwork.json` is unchanged (D10). 35: - **Smoke seed:** `ChimneySmoke.SEED` (20926) with `use_fixed_seed`. **Not repeatable in 4.6.3**, so smoke is recorded as known run noise. On the plume regions (2.4-3.4% of the frame) two seeded live captures differ by mean |dluma| 3.51 (night) and 4.04 (day), >8 on 10.5/10.6%. Two unseeded pre-tree runs differ by 3.54/4.28. No smoke: 0.65/0.44 (`smoke_repeat2.js` → `metrics/smoke-repeat.txt`; map `metrics/smoke-map-harbour-day.png`, looked at: the plumes overlap with shifted fringes). 36: - **Files:** `scripts/look_profile.gd`, `scripts/starter_demos.gd`, `scripts/sample_scene.gd`, `scripts/artwork_scene.gd`, `scripts/amsterdam_scene.gd`, `scripts/main.gd` (one line), `scripts/chimney_smoke.gd`, `tests/test_starters.gd`, `tests/test_look.gd`, `tests/test_dusk.gd`, `tests/test_lighting.gd`, `tests/test_building_tools.gd`, `docs/visual-handoff.md` (this entry). Backups: `b/backup/`. 37: - **Tests:** test-first on the pre-part code, `b/P2b-testfirst-test_{starters,look,dusk}.log(.out)`: 219 checks with 28 failures, 364 with 4, and 220 with 7. All are new P2b checks; no old check failed. Later, four checks were added: the blank canvas check was split in two, "the fog is drawn" and "no chimney, so nothing smokes yet", and three checks for `lighting/styles.json`. New checks in `test_starters` `atmosphere_checks`: all 11 starters carry fog 0.3 and smoke, validate, and save and reopen byte for byte. Each keeps the rest of its environment. `DioramaState.new()` and `benchmark()` stay without the keys. `blank()` is empty plus the two keys. In the editor, Blank canvas opens with the fog drawn and no smoke (no chimney yet), and each New canvas has its demonstration's environment. `test_look` checks the START keys and that an older document switched to Glade, then to Canal, gains 0.15, then 0.3. `test_dusk` checks the seed, the Canal quarter starter smoking and fogged with no edit, the Harbour starter's keys, `artwork.json` and `lighting/styles.json` still without the keys and saving byte for byte, and in the editor that opening Harbour steps shows the Chimney smoke switch on and undoing "off" returns to the new defaults. Affected-suite logs: `b/r1-*`, `r2-*`, `r3-*`, `r4-*`. No other suite needed triage: all 40 passed on the first full run (`b/suite/`, 19,289 checks). Final full suite on the final code, `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/b/suite-final` (a/b/v): **40 suites, 19,292 checks, 0 failures** (was 19,227: test_dusk 209 → 223, test_look 359 → 364, test_starters 176 → 222). `--verify` 2880x1800 exited 0 in 31 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 38: - **Changed expectations (old → new):** test_dusk smoke_checks "without the switch there is no smoke" and fog_checks "without height_fog the fog is not drawn": `ArtworkScene.build()` → the same scene with the two keys erased (`legacy()`, what it saved before P2b). state_checks "a document without the keys restores without them … byte for byte": `SampleScene.harbour()` → `legacy(SampleScene.harbour())`. main_checks "Chimney smoke switch, off for this scene" and "undo takes both away again": restores `legacy(harbour)` instead of `harbour`, and the message now says "this older scene". test_lighting and test_building_tools write the styles fixtures without the two keys. No check was deleted. 39: - **Captures** (`b/caps.sh`, sitting 03:51-04:06, AC; before sets from `pre-tree`, after sets from the live tree, 1600x1000; every view on the first try). `before-1/2` and `after-1/2` hold 16 views each: harbour-night, -day, -golden, -2100, -lane, lane-night, glade-day, glade-night, quarter-night, quarter-day, styles-day, plus hn/hd/hg-flags, gd-flags and qn-flags. "Flags" means `--height-fog=0.3 --chimney-smoke=on`, or 0.15 with `--look=glade`, the same arguments in both trees. `starters-pre`/`starters-live` hold each starter's own exported document (`export_starters.gd` → `scenes/pre|live/`) at 13:00, 19:30, 21:00 and 01:00: 44 views each. `smokeoff-1/2`: live harbour night and day with `--chimney-smoke=off`. **Correction:** the first full suite rewrote `lighting/styles.json` with the keys before these captures, so the sitting's styles-day was fogged on both sides. After the fixture fix, styles-day was recaptured in `styles-fix/{before,after}-{1,2}` (04:20-04:21). It is unchanged: before vs after >8 on 0.80/0.14% against noise pairs of 0.85/0.21%, integer counts identical (`metrics/styles-fix.txt`). 40: - **Owner A/B sheet** (before | after, looked at): `b/sheet/p2b-01-blank.png` (new scene: a faint mist on the water, stronger at 21:00; no chimney, no smoke), `-02-demo0` to `-05-demo3` (the four New canvas demonstrations), `-06-courtyard`, `-07-harbour`, `-08-castle`, `-09-styles`, `-10-quarter`, `-11-amsterdam`, and `-12-glade-lane.png` (Glade at 0.15, harbour lane and lane-night). By day the veil is light: frame mean +1.4 to +2.1 luma, >8 on 1-3.4%. At golden hour, 21:00 and night the plumes rise warm, and the fog's lamp glow adds a warm halo round every lit door and lamp (P2c keeps the glow at the whole setting at every hour). On the small demonstrations that halo is the biggest change: Bracketed room golden >8 on 32.7%, mean +8.0. On the bigger scenes golden to night is >8 on 7-15% (courtyard, harbour, quarter, Amsterdam), 4.4-4.8% (styles) and 1.7-2.3% (castle). Four turrets (castle) and Bracketed room have no chimney, so they get fog only. 41: - **Acceptance:** 42: - **Listed starters carry height_fog 0.3 and smoke true (Glade 0.15):** met. Tests as above. The exported documents (`artifacts/beautify/20-P2/b/scenes-run1/starter-diff.txt`) differ from the pre-part ones only by the two added keys, all 11; `benchmark` is identical. 43: - **Legacy documents unchanged byte for byte:** met. `artwork.json`, `lighting/styles.json`, `dense.json`, `grove.json` and `chemistry.json`, restored and re-saved by the pre-part and the live tree: `cmp` identical (`export.sh`, second run, `scenes/`). `git diff` of both styles fixtures is empty after the final suite. 44: - **Harbour frame time under 0.5 ms (three-round median):** met, **+0.25 ms** at night and +0.23 ms at 13:00 (see Bench). 45: - **Harbour journey views change only by fog and smoke:** met. The pre-part tree with the flags against the live defaults differs by less than each tree's own noise pair (`onlyfog.js` → `metrics/only-fog-smoke-short.txt`). Harbour night >8 on 1.40/1.31% (noise 2.11/2.12%), day 0.82/0.82% (1.41/1.24%), golden 1.85/1.16% (1.90/2.42%), Glade day 0.64/0.48% (0.78/0.35%). The change itself, no flags before and defaults after: night mean +2.30 luma, >8 on 8.3%. The flag views are within noise between trees too: hn-flags 1.95/1.42% against 2.11/1.59%, and qn-flags 2.70/2.31% against 3.25/2.01% (`metrics/ab-1.txt`, `ab-2.txt`, `noise-before.txt`, `noise-after.txt`). quarter-night and quarter-day are unchanged: 3.01/2.36% and 1.53/1.05% against noise pairs of 3.51/2.78% and 1.69/0.84% (quarter-day set 2 sits inside the before pair, not the after pair). 46: - **Draw calls** (`metrics/counts-1.txt`, `-2.txt`, `counts-starters.txt`). Harbour views and glade: +2 (fog, smoke), e.g. harbour-night 909 → 911, rendered_objects +2. `evening.height_fog` 0 → 0.3 (Glade 0.15), `height_fog_drawn` 0 → 0.3 at night and 0.135 by day, `smoke_points` 0 → 7. Starters: +2 each; castle, Bracketed room and Blank +1 (fog only). Legacy-document views (quarter-*, styles-day) and every flag view: identical. 47: - **Bench** (`b/bench/p2b-bench.sh`, `analyse.js` → `bench/summary.txt`; 04:09-04:19; P1 bisect arguments `--disable-vsync --resolution 1920x1080 --quit-after 6000 -- --smoke --churn=100000 --rain=0.45 --time=1|13`, pixels 2; fields **`raw_p50_frame_ms` and `raw_p95_frame_ms`** (raw frame intervals), every run `vsync` "disabled", 491 raw samples; warm-up round 0 discarded, rounds 1-5 interleaved, every run on the first try). **Both sides ran from scratch copies** (pre-tree, and `bench/post-tree`, a copy of the final live code). The live root ran as a third side to check the location: live vs post-tree −0.00 and −0.01 ms, so the location does not bias the numbers. Harbour at night: raw p50 6.56 → 6.82 ms, paired **+0.25** (slower in 5/5; rounds 1-3 +0.25 +0.25 +0.19, median +0.25); p95 +0.76. Harbour at 13:00: 6.58 → 6.81, paired +0.23 (5/5; rounds 1-3 median +0.23); p95 +0.02. `artwork.json` with the flags against without, same tree: night 6.94 → 7.16, +0.21 (5/5); 13:00 6.96 → 7.16, +0.19 (5/5); 987 → 989 draw calls, one chimney. So the full-screen fog costs about 0.2 ms, and seven chimneys of smoke about 0.05 ms. Dense was not benched: `dense.json` is a legacy document with no fog or smoke, and its bytes are unchanged. 48: - **Journey:** expected count deltas. harbour-night, harbour-day and harbour-lane: `draw_calls` +2, `rendered_objects` +2, `evening.height_fog` 0 → 0.3, `evening.height_fog_drawn` 0 → 0.3 (night) or 0.135 (day, lane), `evening.smoke_points` 0 → 7. quarter-night, quarter-day and styles-day: none. Looks: the harbour views gain the mist, the lamp glow in it and the plumes; the rest are within noise. 49: - **Review:** filled by the orchestrator. 50: - **Not done:** the smoke seed does not make captures repeatable (above); smoke is recorded as run noise. No dense bench (legacy document). [side-branch note removed] `--verify` at 1440x900 was not run (the inherited stall). Platforms: Mac run; Windows not run. 11: **Beautify P2d Canal contact shade (29 September 2026):** Opus 5.5 on this Mac. Canal's "Soft contact shadows" switch now draws the ContactShade pass (report quick win 10, Defect 9; owner D11). It used to toggle only the engine SSAO. The shade is a near-black violet `(0.03, 0.02, 0.07)` in 3 levels, half strength at night; the engine SSAO is off in canal. **Owner ruling (b)** (`artifacts/beautify/owner-decisions.md` "P2d ruling"): Glade keeps the engine SSAO exactly as before P2d (on with the switch, radius 0.45, intensity 0.75), applied in fix pass 3 below; Glade is back inside noise. By day and at golden hour, sill and coping undersides, stair risers and wall feet get a crisp dark line 1-2 art pixels wide. The stop condition first hit here (section 6, "a view changes that should not have": the engine SSAO is not a no-op in this build, so removing it everywhere moved the Glade control views and lifted the canal night creases it had darkened) is closed by that ruling: Glade is restored, and the owner waived the canal night-crease line for the SSAO-darkened creases. Branch `beautify/P2d` from `main` 6adf63c. Snapshot: `build/backup-feel/20-P2-d/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/20-P2/d/pre-tree/`. Evidence: `artifacts/beautify/20-P2/d/`. Power: AC, 100% charged, Low Power Mode 0 for every capture, sweep, bench round and suite (`cap-power.txt`, `bench/power.txt`, `suite-power.txt`). 12: - **Files:** `scripts/contact_shade.gd` (`configure` gains colour, levels and night share, all with the old pass as the default; `hour(profile)` uses `HeightFog.night`; `base_strength`, `night_scale`, `night_share`, `GLADE_COLOR`). `shaders/contact_shade.gdshader` (`uniform float levels = 0.0`; `a = floor(a/max_shade*levels+0.5)/levels*max_shade`; still one `blend_mix` variant, so warm-up is unchanged). `scripts/look_profile.gd` (RENDER canal: shade 1.2, radius 0.15, min_rise 0.06, thin 2.0, `shade_color`, `shade_levels` 3, `shade_night` 0.5; Glade: its old values plus brown, 0, 1.0; `SSAO_RADIUS`/`SSAO_INTENSITY` retired; `apply_environment` sets `ssao_enabled = false`). `scripts/world_view.gd` (`sync_environment` passes the new keys and no longer toggles SSAO; `sync_profile` calls `contact_shade.hour`; `debug_shade`). `scripts/main.gd` (`--verify`: the toggle checks read `contact_shade.visible`; `evening_report` gains `contact_shade`, the strength drawn at that hour; diagnostic `--contact-shade=,,[,]`). `tests/test_look.gd`. Backups: `d/backup/`. 13: - **Decisions:** (1) **Tuned radius 0.6 → 0.15 m and min_rise 0.22 → 0.06 m** (the plan kept the report's 0.6/0.22). At 0.6 the pass drew a soft band up to 10 art px wide over whole parapet faces. Sweep `d/sweep-s1`, `sweep-s2` (`--rain=0`, live tree), measured with `thick.js` → `metrics/sweep-thickness.txt` (thickness of the darkened art-cell bands): quarter-day at 0.6: p50 2, p90 6, 51% ≤ 2 px. At 0.15/0.06: p50 1, p90 3, 88% ≤ 2 px; harbour-day 93%; golden 87%; lane-day p90 4 (`r1sheet/s1-qd-wall.png`, `s2-hd.png`). Lawn specks (art `specks4_pct`): quarter-day 0.574 → 0.668, lane-day 0.793 → 0.962 (0.6 m: 0.619/0.926). I saw no tuft speckle at 4x (`r1sheet/s1-ld-lawn4.png`). (2) Night half strength is canal only (`shade_night` 1.0 for Glade), so Glade's pass stays as it was at every hour. The strength eases with `HeightFog.night`: 1.2 up to golden hour (windows 0.66), 0.6 from 05:45/20:45 (windows 0.85) through the night. (3) Colour: the lens packet's near-black violet, not the report's `(0.11, 0.086, 0.21)`. 14: - **Tests:** test-first `d/P2d-testfirst-test_look.log(.out)`: 395 checks, 32 failures. The final tests on the pre-part tree fail 34 of 398 (`d/P2d-testfirst2-test_look-final-tests-on-pre-tree.log.out`). After: test_look **444** (was 364), 0 failures. New: shader levels, quantisation, one blend_mix variant; SSAO constants retired and off in both looks; canal 1.2/0.15/0.06/2.0, violet, 3 levels, night 0.5; Glade brown, 0, 1.0; `shade_hour_checks` (strength at all ten anchors in both call orders, eased from 19:30 to 20:45, Glade 2.2 at every hour, switch off at every hour, old 4-argument `configure` = old pass); in the editor the canal switch off/undo, 13:00 → 1.2 and night → 0.6 through undo, and the evening report. Affected suites `d/r3-*`: test_camera 191, test_warmup 34, test_dusk 223, all 0 failures. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/d/suite` (a/b/v): **40 suites, 19,372 checks, 0 failures** (was 19,292). `--verify` 2880x1800 exited 0 in 31 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures; the known flaky check did not fire. 1440x900 was not run (the inherited stall). 15: - **Changed expectations (old → new):** test_look "the engine SSAO keeps radius 0.45 and intensity 0.75 in every look" → constants retired, `ssao_enabled` false in both looks. "canal draws no occlusion pass of its own" → canal 1.2 over 0.15 m. "canal's (undrawn) pass keeps no rise threshold and no sliver skip" → 0.06 m and 2 px. Editor "canal contact shadows" (SSAO 0.45/0.75) → SSAO off; "no occlusion pass" → the pass is drawn at 1.2 × the night share, violet, 3 levels. `--verify` "Contact shadow toggle failed" / "Undo did not restore contact shadows" read `contact_shade.visible` instead of `ssao_enabled`. No check was deleted. 16: - **Captures** (`d/caps.sh`, 04:59-05:05; `before-1/2` from `pre-tree`, `after-1/2` live, interleaved, 17 views each, all first try): quarter-day/-golden/-night, harbour-day/-golden/-night, lane-day/-night, glade-day/-night, styles-day, qd/qg/qn/hn-off (`--contact-shadows=off`), q2045, q2100. Harbour, lane and glade views carry P2b's fog and smoke (starter defaults) in both trees; quarter and styles are capture documents with neither. Metrics `d/metrics/` (`ab-1/2`, `noise-before/after`, `counts-1/2`). Sheets (before | after | grey before | grey after, 2x, looked at): `d/sheet/p2d-1-quarter-day.png`, `-2-quarter-golden`, `-3-quarter-night`, `-4-harbour-day`, `-5-harbour-night-fog` (the shade sits under the fog, faint), `-6-lane-day`, `-7-glade-day-control`, `-8-quarter-day-court`, `-9-lane-night`. Frame change >8 luma, set 1/set 2 (noise before/after): quarter-day 7.0/6.9% (1.07/1.06), mean −0.58; quarter-golden 10.0/10.1% (2.02/2.06), −0.70; quarter-night 4.8/4.2% (3.04/2.05), +0.07/+0.11; harbour-day 4.2/4.1% (0.96/0.96), −0.30; harbour-night 4.8/3.3% (2.52/1.84); lane-day 15.3/15.4%; lane-night 9.1/9.2%, **+0.96/+0.84**; styles-day 2.2/2.3% (0.30/0.13); **glade-day 1.40/1.47% (0.33/0.29), +0.18; glade-night 1.72/1.84% (0.47/0.32), +0.19**. 17: - **Finding: the engine SSAO did draw.** The probe (`d/probe/ssao-tree`: the live P2d code, with SSAO left on the switch at 0.45/0.75; `metrics/ssao-probe.txt`) shows it. Against the pre tree the probe is inside noise on Glade (glade-day >8 0.33% against noise 0.30%; glade-night 0.20% against 0.28%). Probe → live, SSAO removal alone: glade-day lighter >8 on 1.06%, mean +0.16; lane-night lighter >8 on 5.6%, mean **+1.38**. In the pre tree, SSAO off → on (`metrics/ssao-effect.txt`): frame mean −0.29 to −0.48, >2 on 7.5-8.8% (noise 1.9-4.4%), darker >8 on 1.2-2.7% (noise 0.4-0.9%). It is a broad, soft darkening: enclosed courtyards −3.1 luma and lane-zoom ground (`r1sheet/pre-ssao-qd.png`, `f-lane-night.png`). This is D1's pending verdict, measured: small on the quarter and Harbour frames, visible at lane zoom, and it costs 0.1-0.3 ms (Bench). 18: - **Acceptance:** 19: - **Wall feet, sill undersides and risers: a crisp 1-2 art px shade, at most 3 levels, day and golden:** met. On the shade mask (the art cells darkened ≥ 4 in both live sets, on vs off), quarter-day is 7.3% of the frame, mean −13.8. In the sweep, thickness p50 1, p90 3, 88% ≤ 2 px (golden 87%, harbour-day 93%). 3 levels by construction (shader quantisation, tests). 20: - **Sampled night crease pixels at or below their pre-part values:** after owner ruling (b): **waived** for the creases the pre-part SSAO darkened (M3 91-92%, M3h 97-98% lighter, as reviewed), **met** for the creases the pass draws (fix pass 3: quarter-night 1.09/0.24%, harbour-night 0.60/1.03% lighter >2, inside noise). Before the ruling: **not met** (review fix: this line was booked "met on the pass"; see "Review and fix pass" for the figures on the creases the pre-part SSAO darkened, where live is lighter on 89-97% of cells by 7-10 luma). What the pass alone does: crease sampler `d/crease.js` → `metrics/crease-*.txt`. On the creases the night pass draws (mask live on vs off at night): quarter-night mean 52.9 → 44.0/44.0, lighter >2 on 0.82/0.41% (noise 1.97%); harbour-night under the fog 58.6 → 49.8/50.4, lighter >2 on 5.0/6.2% (noise 3.6/5.5%, fog and smoke). The pass itself never lifts a crease: live off → on at night, lighter >2 on 1.28/1.44% (noise 0.8-1.5%). But on the day creases (7.3% of the frame) at quarter-night, pre → live is mean −1.98/−1.91 and yet **lighter >2 on 9.1/8.9% (noise 0.8-1.5%)**; q2045 is 8.9/8.8%, q2100 9.0/9.2%. The cause is the SSAO removal: in the pre tree SSAO darkened 12% of those cells by >2. 21: - **The Scenery switch turns it off:** met. Tests (both looks, undo). Pre off → live off within noise: qd-off >8 0.83% (noise 0.85/0.95), qg-off 1.06% (1.10/1.12), qn-off 3.03% (2.82/2.41), hn-off 2.27% (1.38/2.98). `evening.contact_shade` 0. 22: - **SSAO off:** after owner ruling (b), off in canal only (code and tests); Glade keeps it as before P2d. 23: - **glade-day control pixel-identical:** after owner ruling (b), **met within noise** (fix pass 3: glade-day pre→fixed median 0.47% against pre noise median 0.57%, glade-night 0.33% against 0.26/0.21%, means ±0.01). Before the ruling: **not met** (above; the SSAO removal). Glade's pass itself is unchanged: integer counts identical, and the probe is within noise. 24: - **Draw calls** (`metrics/counts-1/2.txt`, identical): every canal view +1 draw call and +1 rendered object (e.g. quarter-night 969 → 970, harbour-night 911 → 912). Glade views identical. New `evening.contact_shade` (1.2 by day and golden, 0.6 at 20:45, 21:00 and night, 2.2 in Glade, 0 off). 25: - **Bench** (`d/bench/p2d-bench.sh`, `analyse.js` → `bench/summary.txt`; P1 bisect arguments `--disable-vsync --resolution 1920x1080 --quit-after 6000 -- --smoke --churn=100000 --rain=0.45 --time=1|13`, pixels 2; `raw_p50/p95_frame_ms`, vsync "disabled", 491 raw samples; round 0 discarded, rounds 1-5 interleaved; every run on the first try). Every side ran from a scratch copy: pre-tree, `bench/post-tree` (the live code), `off` (the post-tree with `--contact-shadows=off`) and `bench/probe-tree`. contact_shadows is on in dense.json and the Harbour starter (fog 0.3 and smoke on in every Harbour run; dense has neither). **Pass cost (post vs off), paired p50:** dense 01:00 **+0.42** (5/5; p95 +0.56), dense 13:00 **+0.29** (p95 +0.31), Harbour 01:00 +0.13, Harbour 13:00 +0.11. That is inside the report's +0.25-0.75 on dense. **P2d vs pre-part:** dense +0.00 / −0.07, Harbour +0.00 / −0.00, because the pre-part SSAO cost the same (pre vs off: dense +0.44/+0.35, Harbour +0.14/+0.13). The alternative (probe vs post) adds SSAO's +0.26/+0.30 on dense and +0.11/+0.09 on Harbour. Draw calls 2360 → 2361, 921 → 922. Windows iGPU cost unmeasured (16 depth taps). 26: - **Journey:** expected count deltas: harbour-night, harbour-day, harbour-lane, quarter-night, quarter-day and styles-day: `draw_calls` +1, `rendered_objects` +1, and a new `evening.contact_shade` (0.6 night, 1.2 day). Looks: crisp dark lines under sills, copings and risers by day (quarter-day >8 on about 7%). Also the broad SSAO darkening is gone in canal (courtyards and lane ground lighter). Glade views (glade-day, glade-night, `--look=glade`): no deltas and no visible change (owner ruling (b), fix pass 3); `evening.contact_shade` 2.2 there as before. 27: - **Review and fix pass (29 September 2026, 05:50-06:04, AC, 100%, Low Power Mode 0: `d/fix/cap-power.txt`, `suite-power.txt`).** Five must-fix findings, all on the SSAO removal. (1) The glade-day control is not pixel-identical, because the engine SSAO is not a no-op. (2)-(3) The night-crease line fails per pixel, and the summary had booked it "met". Both masks came from where the P2d pass darkens, so they missed the creases the pre-part SSAO darkened. (4) The plan's lines contradict each other: only the probe meets the Glade control and night-crease lines, and it breaks "SSAO off". (5) Four comments called the SSAO a no-op. **Fixed:** the comments (`look_profile.gd` header and `apply_environment`, `world_view.gd` `sync_environment`, `test_look.gd` render_checks, plus a correction note in `contact_shade.gd`'s D1 header) now say the SSAO draws a soft darkening that costs 0.35-0.44 ms on dense, so turning it off is a trade. The night-crease acceptance line now reads "not met". **Not changed:** `ssao_enabled = false` stays, because options (a)-(c) are a ruling for the orchestrator or owner and the plan's "SSAO off" line forbids (a). **Re-measured** in one new sitting, with pre (`pre-tree`), live and probe (`probe/ssao-tree`), two sets each (`d/fix/caps.sh`, 12 views, all on the first try), using `d/fix/ruling.js` → `d/fix/ruling.txt`. Whole frame, >8 luma: glade-day pre→live 1.20/1.40% (lighter >8 1.07/1.18%, mean +0.16/+0.17) against noise 0.51/0.30%; pre→probe 0.50/0.30% (inside noise). glade-night pre→live 1.37/1.44% against noise 0.27/0.22%; pre→probe 0.25/0.19%. lane-night pre→live mean +0.97/+0.99 (lighter >8 5.3/5.4%); pre→probe −0.36/−0.32. Night creases on 2x2 art cells, T=4: **M3**, the cells the pre tree's SSAO darkens at quarter-night (8692 cells, 2.17% of the frame): pre→live lighter >2 on **93.9/93.6%**, mean **+7.3**, >8 on 26% (noise 1.9-2.2%). q2045 89.1/89.2%, +6.8/+7.0. q2100 91.0/90.6%, +7.2/+7.3. **M3h** harbour-night under fog (6286 cells, 1.57%): **97.0/97.1%**, mean **+9.6**, >8 on 57% (noise 2.5-2.9%). pre→probe on M3: 3.7/2.5% at quarter-night, 0.7/1.7% at q2045, 1.2/1.8% at q2100, M3h 4.6/3.6%; means −0.8 to +0.3, all near noise. **M1**, the live pass's day creases (28190 cells, 7.05%), at quarter-night: pre→live lighter >2 9.9/9.0% (noise 1.1-1.6%), pre→probe 1.8/1.1%. The two masks agree with the reviewer's (8702, 29141 cells; 94.0/93.7%, 9.11/8.86%). **Options, measured:** (a) SSAO on the switch in both looks (the probe) meets the Glade control and both crease masks within about noise. It fails "SSAO off" and costs +0.24/+0.25 ms p50 on dense and +0.11/+0.09 on Harbour against pre-part (`bench/summary.txt`; total occlusion 0.59-0.68 ms on dense, inside the report's +0.25-0.75). (b) SSAO in Glade only restores the Glade control (Glade equals the probe) but **not** the night creases, which are canal views: canal equals live, so M3 stays 94% and M3h 97% lighter. (c) Accept as D1's verdict: both lines stay failed. **Day A/B for (a)**, `d/fix/ab-quarter-day.png`, `ab-quarter-golden.png` (pre | live | probe | live→probe |dluma| x4) and `-grey.png`, all looked at: the crisp 1-2 art-px lines under the fountain rim, the awning and the wall feet are the same in live and probe. The probe adds a faint broad darkening over the court walls and ground (the dark-red wash in the diff), and the lines are not blurred. On the day frame, probe→live is mean +0.39 (quarter-day) and +0.52 (golden), lighter >8 on 1.4/2.8%. Tests after the fix: test_look 444, 0 failures (`d/fix/fx-test_look.log.out`). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/d/fix/suite` (a/b/v): **40 suites, 19,372 checks, 0 failures**. `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures. 1440x900 was not run (the inherited stall). No expectation changed in the fix pass (comments only). 28: - **Recheck and escalation pass (29 September 2026, 06:06).** The recheck left three must-fix findings open, and all three need the same ruling. (1) The glade-day control is still not pixel-identical, because `look_profile.gd:188` `env.ssao_enabled = false` is unchanged: pre→live >8 on 1.20/1.40% against noise 0.51/0.30%, and glade-night 1.37/1.44% against 0.27/0.22%. (2) The night-crease line still fails: M3 is 93.9/93.6% lighter at quarter-night (+7.3), q2045 89.1/89.2%, q2100 91.0/90.6%; M3h is 97.0/97.1% (+9.6); M1 is 9.9/9.0%. (3) The plan contradicts itself: only (a) meets both lines, and it breaks "SSAO off", at a cost of +0.24/+0.25 ms p50 on dense and +0.11/+0.09 on Harbour. The reviewer's figures match `d/fix/ruling.txt`. **Not changed, left open:** each finding asks for the orchestrator's or owner's choice of (a), (b) or (c). Choosing one changes the plan's own "SSAO off" line and inverts a test, so it is not the implementer's call. The code, tests and expectations are exactly those of the fix pass: nothing in scripts, shaders, tests, data, scenes or tools is newer than `d/fix/ruling.txt`. So no suite, capture or bench was rerun, and the fix pass's full suite (40 suites, 19,372 checks, 0 failures) and 2880x1800 `--verify` still cover this tree. **After a ruling:** for (a), apply the probe's two lines, invert "apply_environment turns the engine SSAO off" and restore the retired 0.45/0.75 SSAO check in both looks; for (b), apply them to Glade only. Then rerun test_look, `tools/suite_all.sh`, the `d/fix/caps.sh` captures and `d/fix/ruling.js`. For (c), book both lines as D1's accepted verdict. 29: - **Fix pass 3: owner ruling (b) (29 September 2026, 07:00-07:35, AC, 100%, Low Power Mode 0 for every capture, suite and bench round: `d/fix3/power.txt`, `cap-power.txt`, `suite-power.txt`, `bench/power.txt`).** Asked: keep the engine SSAO in Glade only, exactly as before P2d (on with the contact-shadow switch, radius 0.45, intensity 0.75), keep it off in canal where the crisp pass draws; tests first; Glade pre-part → live inside noise; canal unchanged against P2d as reviewed; the canal night-crease line waived for the SSAO-darkened creases and still holding for the pass's own; bench pre vs fixed on Harbour and dense. At start the live tree differed from `build/backup-feel/20-P2-d/` only in the six P2d files; nothing unexpected. The reviewed P2d tree was copied to `d/fix3/p2d-tree` before any edit; backups `d/fix3/backup/`. **Changed:** `scripts/look_profile.gd`: `SSAO_RADIUS` 0.45 and `SSAO_INTENSITY` 0.75 restored; RENDER gains `"ssao"` (canal false, Glade true); `apply_environment` sets `ssao_enabled = r.ssao and contact_shadows` plus the radius and intensity (so the warm-up's Glade world gets the SSAO too, as before P2d). Comments on the SSAO updated in `look_profile.gd` (header, `apply_environment`), `world_view.gd` `sync_environment`, `contact_shade.gd`, `main.gd` `--verify` and `test_look.gd`. **Tests first** (`d/fix3/testfirst-test_look.log.out`): 451 checks, 8 failures, all new: constants 0.45/0.75; `ssao` canal false, Glade true; `apply_environment` canal off for no key, switch on and switch off; Glade on at 0.45/0.75 with no key and with the switch on (2); Glade off with the switch off; in the editor Glade's SSAO at 0.45/0.75 after choosing, redo and choosing by index, off with the switch and back on with undo. After: test_look **451, 0 failures** (`d/fix3/f3-test_look.log.out`). Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/d/fix3/suite` (a/b/v): **40 suites, 19,379 checks, 0 failures** (was 19,372). `--verify` 2880x1800 exited 0 in 32 s, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 0 failures; 1440x900 not run (inherited stall). **Changed expectations (old → new):** test_look "the engine SSAO constants are retired" → SSAO_RADIUS 0.45 and SSAO_INTENSITY 0.75; "apply_environment turns the engine SSAO off" in both looks → off in canal (3 cases), on in Glade at 0.45/0.75 with the switch, off without it; editor "the engine SSAO is off in Glade too" → "Glade keeps the engine SSAO at 0.45/0.75". Canal expectations unchanged. **Captures** (`d/fix3/fix3-caps.sh`, 07:10-07:20, sets 1-2 of pre (`d/pre-tree`), p2d (`fix3/p2d-tree`) and fixed (live), interleaved, 16 views: quarter-day/-golden/-night, q2045, q2100, harbour-day/-night, lane-day/-night, glade-day/-night, styles-day, qd/qn/hn-off and gd-off (Glade day, switch off); plus `fix3-glade.sh`, Glade-only sets 3-4 of pre and fixed; all first try). Metrics `d/fix3/fix3.js` → `fix3.txt`, `cross4.js` → `cross4.txt`. Whole frame >8 luma: **glade-day pre→fixed 16 cross pairs 0.25-1.04%, median 0.47%, mean d +0.004, against pre noise 0.34-0.92% (median 0.57) and fixed noise 0.31-0.51% (0.37)**; P2d was 1.31-1.34% (p2d→fixed), mean −0.15/−0.18. **glade-night pre→fixed 0.19-0.42%, median 0.33, mean d +0.006, against noise 0.22-0.39% and 0.18-0.23%**; P2d 1.32-1.42%. gd-off pre→fixed 0.26-0.45% (noise 0.33-0.46%). **Canal p2d→fixed within noise on every view**: quarter-day 0.81/0.97% (noise 0.92/0.86), quarter-golden 1.18/1.49 (0.96/1.36), quarter-night 1.42/1.40 (1.16/1.16), harbour-day 1.02/1.02 (0.69/0.81), harbour-night 0.92/0.85 (0.82/0.69), lane-day 1.75/1.88 (1.92/1.77), lane-night 1.28/1.23 (1.13/1.30), styles-day 0.35/0.73 (0.26/0.37), all means within ±0.09. Canal runs the same code in both trees, so the few tenths over a pair (harbour-day 1.02 against 0.81) are sitting noise; cross-tree pairs of identical Glade code reach 1.04% too. **Night creases** (2x2 art cells, T=4): on the creases the pass draws at quarter-night, less those the pre SSAO darkened (MN−M3, 13,086 cells, 3.27%), pre→fixed mean −4.3 (set 1), lighter >2 on **1.09/0.24%** against noise 1.19-1.26%: holds. Harbour-night under the fog (MNh−M3h, 4,657 cells) pre→fixed lighter >2 **0.60/1.03%** (noise 1.95%), mean −10.7/−7.3: holds. The waived SSAO creases stay as reviewed: M3 pre→fixed 91.1/92.1% lighter >2, mean +6.9/+7.3; M3h 97.7/96.5%, +9.6/+9.4; p2d→fixed on both inside noise. M1 day creases at quarter-night pre→fixed 7.7/8.4% lighter >2 (as P2d: pre→p2d 8.2%). pre-2 quarter-night is the noisy capture of its pair (water shimmer and rain, `d/fix3/sheet/fix3-qn-pre-noise.png`, looked at), so its crease rows swing by up to 13 luma; set 1 and the p2d/fixed pairs are the clean reads. **Crops at 2x, looked at** (pre | p2d | fixed | diffs, |dluma| x4): `d/fix3/sheet/fix3-glade-day.png`, `fix3-glade-night.png`: pre→fixed differs only in the chimney plume (smoke is known run noise); pre→p2d shows the broad green lift over the lawn and walls that the SSAO removal made, and it is gone. `fix3-lane-night.png`: p2d→fixed is rain streaks and the plume only; pre→fixed keeps the waived broad lift (canal lost its SSAO, as reviewed). **Draw calls:** unchanged against P2d (Harbour 922, dense 2361; Glade identical to pre). **Bench** (`d/fix3/bench/fix3-bench.sh`, `analyse.js` → `bench/summary.txt`; P1 bisect arguments, `--disable-vsync`, `raw_p50/p95_frame_ms`, 491 samples, vsync "disabled"; pre from `d/pre-tree`, fixed from `d/fix3/bench/fixed-tree`, a copy of the live code; round 0 discarded, rounds 1-5 interleaved, all first try, 07:23-07:34). fixed vs pre paired p50: Harbour 01:00 **−0.01** (0/5 slower; p95 +0.63, one slow round), Harbour 13:00 −0.01 (2/5), dense 01:00 **−0.07** (0/5), dense 13:00 −0.06 (0/5). Canal cost stays neutral. Glade was not benched (its code path is pre-part's again). 30: - **Review:** filled by the orchestrator. 31: - **Not done:** nothing from the ruling: owner ruling (b) is applied and measured in fix pass 3; the canal night-crease line stays waived for the SSAO-darkened creases by the owner. Glade was not benched (pre-part code path). No owner A/B sheet beyond the crops above (the stage sheet is the orchestrator's). The pre-tree, probe, bench, `fix3/p2d-tree` and `fix3/bench/fixed-tree` trees are left for the review. `--verify` at 1440x900 was not run (the inherited stall). Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy, 29 September 2026, 07:41) 16: | D10 | Fog and smoke on for starters, SampleScene and the new canvas; `artwork.json` unchanged | no | P2b | 17: | D11 | Canal contact shade, cost reported | no | P2d | 78: ## P2d ruling (owner, 29 September 2026: "b") 79: 80: The engine SSAO is not a no-op in GL Compatibility (P2d measured it: Glade control 1.3% of pixels over noise; canal night creases lose their soft shade; its cost about equals the new pass's). Owner choice **(b): keep the engine SSAO in Glade only** (on the Scenery switch, at its pre-P2d radius 0.45 and intensity 0.75), so Glade is unchanged; **canal uses the new crisp contact-shade pass with SSAO off**. The canal's night-crease acceptance line ("sampled night crease pixels at or below their pre-part values") is waived for the creases SSAO used to darken: the owner accepts that canal loses the soft night crease shade. It still holds for the creases the new pass draws. ADDED 30 September 2026: tree-foot collars, leaded windows, the roof-tree pit, the foliage night-floor speed fix and the P2 stage close Frozen 30 September 2026 at 07:10 BST (game main at 085517d). P2e, P2a, P2e2 and P2p were implemented and merged (P2e2 was an inline one-rule change with no separate review); the P2 stage close ran on the final P2 code. Line numbers refer to the frozen copy of docs/visual-handoff.md taken at 07:10 on 30 September, not the copies above. Original source terminology is preserved; workflow run IDs and local paths are removed. The owner (the project's creator) has not yet looked at the P2 sheet: P2-4, P2-5 and P2-6 are open. The handoff check at the close found the P2b owner-sheet figures (frame mean +1.4 to +2.1, >8 on 1-3.4%; bracketed room golden 32.7% and +8.0; golden-to-night 7-15%) not backed by a file; the chapter now says so. The resume note at line 186 (workflow, GitHub and progress-board details) is not quoted. SOURCE: docs/visual-handoff.md (frozen copy, lines as of 30 September 2026, 07:10) 195: **Beautify P2e Soil or paving collar at the tree foot (29 September 2026):** Opus 5.5 on this Mac. The pale TRIM stone ring at every tree foot is gone (report quick win 13; owner D12: the collar follows the ground finish). On meadow, earth and sand a tree now stands in a low dark soil collar (`plain`, 0.06 m at the default girth). On stone paving or a cobbled road it stands in a ring of `STONE.darkened(0.25)` sunk nearly flush with the slabs (a tree pit, `stone`). Both use the old ring's 12-sided cylinder at the old radius, so triangles are unchanged. Branch `beautify/P2e` from `main` 20284d6. Snapshot: `build/backup-feel/20-P2-e/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/20-P2/e/pre-tree/`. Evidence: `artifacts/beautify/20-P2/e/`. Power: AC, 100% charged, Low Power Mode 0 for every capture, sweep and suite (`cap-power.txt`, `sweep-power.txt`, `suite-power.txt`, `suite/suite_all.out`). 196: - **Files:** `scripts/tree_growth.gd` (`FOOT_KEY`/`FOOT_PAVED`, `PAVED_FINISHES` [1, 3], `COLLAR_SOIL`, `COLLAR_H`, `PIT_*`; `ring_triangles()`, `pit_color()`, `foot_of(state, o)`, `foot_collar()`, called from `_build_body` and `build_conifer` in place of the TRIM ring). `scripts/world_view.gd`: `_grown_key` adds `params.foot` and a `_paved` stand suffix; `_foliage_mark` and `_deco_signature` gain `foot_of`. `scripts/tree_house.gd`: `draw_host` sets the flag and `host_key` includes it. `scripts/architecture.gd` (two comments only). `tests/test_tree_growth.gd`, `tests/test_foliage_merge.gd`. Backups: `e/backup/`. 197: - **Decisions:** (1) **Soil colour tuned 4a3d30 → 2f271f.** At the report's 4a3d30 the sunlit collar top rendered `#897457` (luma 118) against a sunlit lawn of about 100, so it read as a tan disc lighter than the grass. At 2f271f it is 74-76 (`e/sheet/r1-grove-4a3d30-vs-2f271f.png`, `sweep-s2f271f/`; round 1 = `after-1/2`). One constant. (2) **Flag and fallback:** `params.foot = "paved"` is set only on the example a mesh is drawn from (the grown group's example, the host example). It is never stored. The soil collar is both the fallback and what an example with no flag draws. The fallback covers a tree with no terrain cell under its foot: parented (roof, planter), afloat, on a plinth, on stepping stones, over water. (3) Both collars are drawn with the growth log off, so they stand at the foot from the start and never morph (zero morph, `b.w` 0). (4) Pit top `max(0.03·√girth, 0.2·trunk r)`, depth 0.1·√girth below, so the trunk's lifted foot stays covered. 198: - **Tests:** test-first `e/P2e-testfirst-test_tree_growth.log(.out)` 220 checks, 36 failures; `e/P2e-testfirst-test_foliage_merge.log(.out)` 81, 1 failure. After: test_tree_growth **254, 0** (`e/r2-*`), test_foliage_merge **81, 0**. New `collar_checks`: colours against their ground; soil and pit for tree 4.2/14, blossom 8, conifer 4.2/12 (vertices, height, radius, sunk, trunk foot covered, no TRIM anywhere); triangles identical soil/paved; paved ring carries no morph; `foot_of` for finishes 1-5 and the fallbacks; a WorldView paint dab under a standing tree regroups it (`_paved` key, flag on the example, +1 mesh surface), the lawn tree keeps its group, a derived (synth) tree on paving is flagged, and the release keeps it; nothing is saved; undo restores the soil group and the scene byte for byte; a host tree's collar and `host_key` follow the paint. Affected `e/r1-*`: test_tree_house 160, test_tree_grips 72, test_foliage 190, test_release_spread 107, test_build_anim 422, all 0. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/e/suite` (a/b/v): **40 suites, 19,473 checks, 0 failures**. `--verify` 2880x1800 exited 0 in 33 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures, and the known flaky check did not fire. 1440x900 was not run (the inherited stall). 199: - **Changed expectations (old → new):** test_tree_growth `_morph_ok` "every plain vertex shares the tree's growth total" → all but the soil collar's 216 vertices, which carry none. The growth-anchor loop skips the collar. "the stone foot ring does not grow" → "a tree on grass draws no stone foot ring", plus "the ring sunk in paving does not grow". "the foot ring carries neither [CUSTOM2/3]" → "a tree on grass merges with no stone surface", plus the paved stone surface carrying neither and one surface more. test_foliage_merge "its stone foot is common to every variant" → plain common and no stone on grass, with a paved tree's stone common. The plinth check reads a missing stone group as 0. No check was deleted. 200: - **Captures** (`e/caps.sh`; final sitting sets 3 and 4, `before-3/4` from `pre-tree`, `after-3/4` live, interleaved, 7 views: grove-day, glade-day, harbour-day, quarter-day, styles-day, lane-day, quarter-night; all first try). Frame >8 luma, pre→live set 3/4 (noise before/after): grove 0.59/0.57% (0.02/0.00), glade 0.66/0.58 (0.36/0.71), harbour 0.94/1.30 (0.97/1.03), lane 1.96/2.11 (1.93/2.02), quarter-day 1.43/1.67 (1.01/1.06), quarter-night 2.77/3.37 (2.45/2.74), styles 0.70/0.39 (0.19/0.40). The styles diff (`e/sheet/p2e-4-styles-diff.png`) shows change only at the tree feet, plus rain streaks. Sheets, all looked at: `e/sheet/p2e-1-grove-feet.png` (round 1), `r1-grove-4a3d30-vs-2f271f.png`, `p2e-2-quarter-feet.png`, `p2e-3-lane-foot.png`, `p2e-5-quarter-day-night.png`, `p2e-6-glade-harbour.png`, `p2e-7-roof-tree.png`. 201: - **Acceptance:** 202: - **No pale ring on grass, earth or sand:** met. Foot mask `e/collar.js` → `e/metrics/collar-34.txt` (changed >8 in both pairs, ≤3 in noise; non-green pixels). Collar luma p50/p90 against green ground within 5 px: grove 28/76 vs 59/100 (old ring 55/144), glade 48/72 vs 80/124 (old 130/163), harbour 59/77 vs 77/98 (old 89/161), lane 73/136 vs 87/149 (old 93/177). No earth or sand tree in these views; test-covered. 203: - **A darker sunk ring on paving:** met. quarter-day ring p50 48 (old 91) against ground p50 54, and the old pale disc is gone. quarter-night 26 (old 55) against 33. styles 73 (old 111) against 113. It reads as a dark pit flush with the slabs (`p2e-5`). 204: - **Triangles identical:** met. Tree 4.2 m 1,892, 14 m 7,704, blossom 8 m 4,144, conifer 4.2 m 2,390, 12 m 4,378, the same as the pre-part log (`P2E_TRIANGLES` against test-first), and the same on grass and paving. Budget worst is still 2,404 / 11,800. 205: - **Draw-call change reported** (`e/metrics/counts-3/4.txt`, identical): a tree group on grass loses its stone surface, and a paved group gains a grown group. draw_calls glade 903 → 888, grove 329 → 305, harbour-day 912 → 897, lane 752 → 738, quarter-day/-night 970 → 980, styles 488 → 491; rendered_objects move by the same amounts. Count probe (`e/probe/`, pre `pre.log` / live `live.log`), tree groups and their mesh surfaces: harbour 7 groups, 21 → 14 surfaces; quarter 7 → 10 groups (5 paved), 21 → 25; grove 12, 36 → 24; styles 5 → 7 groups (2 paved), 15 → 16; 4 host meshes. 206: - **Bench:** none as an A/B (the plan's count probe instead; same cylinder, fewer or +3-10 draw calls). 207: - **Journey:** expected count deltas: harbour-night, harbour-day and harbour-lane `draw_calls` and `rendered_objects` about −14 to −15; quarter-night and quarter-day +10; styles-day +3. Looks: dark soil discs replace the pale stone rings on the lawns, and a dark sunk ring replaces them on quarter paving. 208: - **Edge shade (orchestrator note):** P2d's crisp shade needs a rise of 0.06 m. The soil collar's front edge shows a 1-art-px dark line at lane zoom, which reads as the collar's shaded lip. The paved ring (0.03 m proud) draws none, and its own dark side reads as the pit edge (`p2e-3`, `p2e-5`). 209: - **Owner questions:** (a) the soil 2f271f instead of the report's 4a3d30 (reason above); (b) roof trees (the fallback) stand in a mid-brown soil bed on the stone terrace (`p2e-7`, quarter-day). Should they take the paved pit instead? Changing it is a one-line rule in `foot_of`. 210: - **Review:** one Opus 5.5 review in a fresh context (run [run ID removed]): no must-fix. Noted, not fixed: the quarter pit numbers include the 7 roof trees' soil beds; a paved ring near a quay edge may show its buried part on the bank face (plausible, unconfirmed). Orchestrator look check of the grove and roof-tree crops. Merged to `main` 6c4cb7a on 29 September 2026. Owner questions for the P2 sheet: soil colour 2f271f; roof trees stand in a mid-brown soil bed on stone terraces. 211: - **Not done:** no earth or sand tree was captured (covered by tests). No A/B bench (the plan asks for a count probe). The grove view has no night capture. Scratch trees are left for review. `--verify` at 1440x900 was not run (the inherited stall). Platforms: Mac run; Windows not run. 167: **Beautify P2a Leaded medieval windows instead of the X (29 September 2026):** Opus 5.5 on this Mac. Every medieval light is now leaded: a 0.05 m square mullion on its centre line and a transom across it at light×0.18 over its middle, 0.04 m over the glass (report quick win 5; owner D9: no "under 0.6 m transom only" rule). The X of two 0.035 m trim beams corner to corner is gone from the small houses, merged small houses, tower slot lights, placed windows and dormers. WP4 had kept it there for parity only (the composed fronts were already leaded). The X survives only as wall timber, in `Architecture.medieval_panel`'s 0.10 m braces. Branch `beautify/P2a` from `main` 50ac3a7, restarted cleanly. The WIP on `beautify/P2a-wip-0929` (35b42bd) was read for ideas only, and its evidence was moved to `artifacts/beautify/20-P2/a/old-wip-0929/`. Snapshot: `build/backup-feel/20-P2-a/`. At part start the live tree matched it in scripts, shaders, tests, data, scenes and tools; nothing unexpected. Scratch tree: `artifacts/beautify/20-P2/a/pre-tree/`, rebuilt this session. Evidence: `artifacts/beautify/20-P2/a/`. Power: battery, Low Power Mode 0 for every capture (`cap-power.txt`, 72→71%). The suite started on battery at 70%. The power state at its `--verify` is not recorded in a file. No bench was run. 168: - **Files:** `scripts/cluster_architecture.gd`: `styled_window` loses its `leaded` parameter and the beam branch, and the medieval branch always draws the mullion and transom. Header and branch comments rewritten. `scripts/architecture.gd`: `styled_bay_legacy`, the lone small house's bay (a second X the plan's file list did not name), draws the same leading at z 0.155, 0.04 m over its glass at 0.115, in place of the beams at 0.16. Its header comment is updated. `styled_lights` no longer passes `leaded`. `tests/test_building_tools.gd`, `tests/test_cluster_walls.gd`. `artifacts/building-tools/tests.json` (rewritten by the suite: 669 → 687 checks). Backups: `a/backup/`. 169: - **Tests:** test-first `a/P2a-testfirst-test_building_tools.log(.out)`: 687 checks, 17 failures. `a/P2a-testfirst-test_cluster_walls.log(.out)`: 249 checks, 2 failures. After: test_building_tools **687, 0** (`a/r1-*`) and test_cluster_walls **249, 0** (`a/r2-*`). New `leaded_window_checks` uses a `LeadProbe` (a MeshBatch that counts the 0.035 m diagonal trim beams, the leading boxes and the 0.10 m panel braces). It covers `styled_window` at 0.45 and 1.25 m, square and arched. It covers `styled_bay_legacy` square and arched. It builds whole a lone 4×4 house, a solid-walled 7 m house with a placed window and a dormer, a lone tower with its slot lights and a placed window, a merged 4×4 pair, a house clustered with a tower, Harbour and the styles showcase. In each, zero X beams, one mullion per transom, and the placed window, dormer and tower lights add leaded lights. Pre-part counts: X beams 32/4/26/48/56/54/40. After: mullions 16/2/13/24/28/47/20. A composed 10×8 block keeps 60 leaded lights and 19 timber braces, and Harbour keeps 13 braces. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/a/suite` (a/b/v): **40 suites, 19,492 checks, 0 failures** (P2e 19,473, +19 new). `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures, and the known flaky check did not fire. 1440x900 was not run (the inherited stall). 170: - **Changed expectations (old → new):** 171: - test_building_tools "A 4 x 4 medieval house keeps the X of beams across its lights" (trim vertices at the beam's end > 0) → "draws no X of beams" (= 0), plus "its lights are leaded: a mullion and a transom". 172: - test_cluster_walls "A merged pair of 4 x 4 medieval houses keeps the X of beams" → "draws no X", plus "leads its lights". Its WP4 comment is rewritten. 173: - test_cluster_walls P0d pinned plank events (pieces unchanged): "medieval tower" 2 → **4** events (126 pieces), "medieval building with a dormer" 18 → **19** (307). The message now reads "sounds as pinned". Probe `a/probe/audio_probe.gd` → `audio-pre.out` / `audio-live.out`. The 10 tower transoms start their reveal higher up the light than a beam did, so they form their own click buckets (t 0.254, 0.670). The dormer's transom moves one piece into a new bucket at t 1.261. 174: - No check was deleted. 175: - **Captures** (`a/caps.sh`, interleaved `before-1`, `after-1`, `before-2`, `after-2` under `a/cap/`: harbour-night, lane-night, harbour-day, styles-day, lane-day). All were captured first try except after-1 harbour-night, which was re-captured minutes later (see "tool defect" below). Whole frame >8 luma, pre→live set 1/2 (noise before/after): harbour-night 9.90/2.78% (9.65/1.89), harbour-day 2.26/2.19 (13.41/13.37), lane-night 2.54/3.05 (2.95/2.96), lane-day 3.89/2.75 (3.11/2.01), styles-day 1.53/1.40 (2.00/0.66). The frames are within or just above noise (lane-day 3.89% against noise 3.11/2.01, harbour-night 9.90% against 9.65/1.89); the change is confined to the lights. Region changes >8 (`a/metrics/regions-summary.txt`, local regions `a/p2a_regions.json`), set 1/2 against noise before/after: lane-night windows_medieval 7.67/8.08 (1.25/1.64), lane-day 8.48/8.38 (0.85/0.51), styles medieval house 5.68/5.56 (0.54/0.15), harbour-night tower 6.32/4.10 (3.36/1.89), harbour-day tower 6.04/5.82 (4.00/4.13). Region p50 moves ≤0.4 luma. Sheets, all looked at: `a/sheet/p2a-1-lane-windows.png` (X → cross-leaded lights on the lane tower, night and day), `p2a-2-harbour-night-s1.png`/`-s2.png`, `p2a-3-day.png` (Harbour tower and house, styles house). 176: - **Acceptance:** 177: - **No medieval light anywhere has diagonal beams:** met. `LeadProbe` counts 0 X beams in every caller and scene above, and the crops show none. 178: - **The X survives only as wall timber in `Architecture.medieval_panel`:** met (the 0.10 m braces are counted above and still drawn). 179: - **Journey counts change only in vertices on medieval scenes:** better than the line. Vertex counts are unchanged (a beam and a box are both 36 vertices; Harbour 189,396 and styles 36,318 before and after). `draw_calls` and `rendered_objects` are identical on harbour-night, harbour-day and styles-day. Lane moved between runs of one tree too (draw calls 2-7; rendered_objects up to 83 on lane-night, 731 against 814; look note for the P2 sheet: at night the middle Harbour tower light shows its transom but its mullion barely reads in the glow, `a/review/hn-tower-after-1.png`, geometry correct) (`metrics/counts-noise-*.txt`). 180: - **Bench:** none (plan). Vertex and draw-call deltas are zero. 181: - **Journey:** expected count deltas are zero on all six views. Looks: the medieval lights on harbour-night, harbour-day, harbour-lane and styles-day show a mullion and transom instead of the X. 182: - **Tool defect (fixed afterwards by the orchestrator in 76f9c36: `return 0` after the loop; merged with P2a):** `tools/art_capture.sh` `wait_free` returns 1 whenever it had to wait for another Godot. Its `while` loop ends on the failed `[ $n -gt 900 ]` test. So a view that waits is skipped silently with `fail=1`, as after-1 harbour-night was on its second try. 183: - **Review:** one Opus 5.5 review in a fresh context ([run ID removed], notes in `a/review/`): no must-fix; six notes, all dealt with by the orchestrator: the stale "kept verbatim" comment in `styled_bay_legacy` reworded; "frames inside noise" reworded to within or just above noise; the lane rendered_objects swing (up to 83 between runs of one tree) recorded; the night look note kept for the owner sheet (the middle Harbour tower's mullion barely reads in the glow); the `art_capture.sh` defect fixed; the owner-answers edit committed separately. Merged to `main` 1ca7b9f. 184: - **Not done:** no bench (the plan asks for none). No 1440x900 `--verify` (the inherited stall). The owner sheet's optional "transom only under 0.6 m" is not offered: D9 declined it. The pre-tree is left for review. Platforms: Mac run; Windows not run. 159: **Beautify P2e2 Roof trees take the paved pit (29 September 2026):** orchestrator, inline on this Mac (a one-rule change, no workflow run). Owner answer P2-3 (`artifacts/beautify/owner-decisions.md`, "P2 owner answers"): the roof trees that stood in a mid-brown soil bed on stone terraces (the P2e fallback) take the sunk paved pit. Owner P2-1 (fog lamp glow) and P2-2 (soil 2f271f) are keep, no change. Branch `beautify/P2e2` from `main` 1ca7b9f. Snapshot: `build/backup-feel/20-P2-e2/`; scratch tree `artifacts/beautify/20-P2/e2/pre-tree/`; evidence `artifacts/beautify/20-P2/e2/`. Power: AC, 100% charged, Low Power Mode 0 (`cap-power.txt`, `suite-power.txt`). 160: - **Change:** `TreeGrowth.foot_of` (`scripts/tree_growth.gd`): a tree whose parent is a building (`state.is_building`, a roof or terrace) returns `FOOT_PAVED`, whatever the ground under the building. Afloat and plinth trees keep the soil collar even with a parent, and so does a tree whose parent is no building. Comments updated. No other code changed; triangles are unchanged (same cylinder). The example flag and group key paths (`WorldView._grown_key`, `TreeHouse.draw_host`) already read `foot_of`, so they follow. 161: - **Tests:** test-first `e2/testfirst-test_tree_growth.log(.out)`: 256 checks, 1 failure (the changed roof-tree expectation). After: test_tree_growth **256, 0**; test_tree_house 160, test_foliage_merge 81, test_tree_grips 72, test_foliage 190, test_release_spread 107, all 0 (`e2/r1-*`). Changed expectation: "a tree on a roof over paving takes the soil fallback" becomes "takes the paved pit"; new checks: an afloat tree with a parent keeps soil, a tree with a non-building parent keeps soil. 162: - **Look:** quarter-day, before from the pre-part tree and after live, same sitting (`e2/before-1`, `e2/after-1`; quarter-night not compared by eye). The brown bed under the terrace trees at lower left and lower right is now a dark sunk ring flush with the slabs and reads as a tree pit (`e2/sheet/qd-left-*.png`, `qd-right-*.png`). The roof tree at top right has its foot hidden by the parapet, so its change is not visible (`qd-top-*`). Draw calls and counts were not benched: same cylinder, one material group swap per roof tree. 163: - **Review:** none run: an inline one-rule change by the orchestrator, checked by tests written first (256 checks, one failure before), the six affected suites, the full suite below and a look at the quarter roof trees. Merged to `main` d3a063e. 164: - **Not done:** no bench, no quarter-night look, no `--verify` at 1440x900 (the inherited stall). Platforms: Mac run; Windows not run. 165: - **Full suite:** `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/e2/suite`: **40 suites, 19,494 checks, 0 failures** (`e2/suite/suite_all.out`; 19,492 at P2a plus the 2 new checks). `--verify` 2880x1800 `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures. 145: **Beautify P2p Foliage night floor speed fix (29-30 September 2026):** Opus 5.5 on this Mac. The P1b night floor's work in `shaders/foliage.gdshader` now sits behind one uniform branch, `night_floor_on()` (`night_amount > NIGHT_FLOOR_FROM && night_floor_gain > 0.0`, true exactly where the floor's weight is above 0), in the fragment and in `light()`. By day, at golden hour and wherever the weight is 0, none of it runs. With the floor on, it is the P1b/P1c2 code line for line, plus four exact trims: `key_view` is taken in the fragment's floor block, no longer a vertex-to-fragment varying; the floor's luma is taken once and handed to `light()` (`night_floor_share(owed, floor_luma)`); `to_display(key_luma)` runs only where `key_luma > 0` (`to_display(0)` is 0); a light skips the take-back when nothing is owed before it (both shares would be 0). Owner "yes speed fix" after the P1 bisect (`artifacts/beautify/19-P1/bisect/bisect.md`); not a plan part, scoped by the `p` note in `beautify-stage.js`. Branch `beautify/P2p` from `main` d3a063e. Snapshot: `build/backup-feel/20-P2-p/`; at part start the live tree matched it in scripts, shaders, tests, data, scenes and tools (nothing unexpected). Scratch tree `artifacts/beautify/20-P2/p/pre-tree/`; evidence `artifacts/beautify/20-P2/p/`; backups `p/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture, bench round and the suite (`p/power.txt`, `cap-power.txt`, `bench/power.txt`, `suite/suite_all.out`). 146: - **Files:** `shaders/foliage.gdshader`, `tests/test_foliage.gd`. 147: - **Tests:** test-first `p/P2p-testfirst-test_foliage.log(.out)`: 208 checks, 14 failures. After: test_foliage **208, 0**; test_foliage_merge 81, 0; test_look 451, 0 (`p/r1-*`). Orchestrator fix after the review (the per-anchor "off" check could not fail): it now asserts `windows <= NIGHT_FLOOR_FROM` at every anchor between 08:00 and 17:00, so test_foliage is **206, 0** (`p/r2-test_foliage.log`), and the suite line below was run before it. New `floor_skip_checks`: the switch reads only the two uniforms; a mirror shows it is on exactly where `smoothstep(from,1,night) x gain > 0` (10 night amounts x 6 gains) and off at every anchor where the floor is 0; the fragment's and `light()`'s floor work each sit in an `if (night_floor_on())` block; the parity lists (13 fragment statements, 3 `light()` statements from P1b/P1c2) are in those blocks verbatim; nothing of the floor runs outside them or in `vertex()`; the glint's `leaf_hue` is taken before the branch. Full suite `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/p/suite` (a/b/v): **40 suites, 19,512 checks, 0 failures** (19,494 + 18 new). `--verify` 2880x1800 exit 0, `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 148: - **Changed expectations (old -> new):** "night_floor_share(float owed, vec3 floor_colour)" -> "(float owed, float floor_luma)"; the fragment adds `night_floor_share(night_floor.a,night_floor.rgb)` -> `(night_floor.a,floor_luma)`; the owed line `to_display(floor_luma)-to_display(key_luma)` -> `to_display(floor_luma)-key_display` with `if (key_luma > 0.0) key_display=to_display(key_luma);`. No check deleted. 149: - **Captures** (`p/caps.sh`, interleaved `before-1`, `after-1`, `before-2`, `after-2`: quarter-day, quarter-night, quarter-2045, harbour-night, glade-night, lane-night; `before-3`/`after-3` for glade-night and harbour-night). Metrics `p/metrics/summary.txt`, `three.txt`, `counts.txt`. Integer counts and draw calls identical on every view. Changed% / >8% A/B against the noise pairs (before, after), whole frame: quarter-day 5.3/1.18 and 16.1/2.77 (14.1/2.27, 6.6/1.36); quarter-night 8.8/3.16, 4.9/2.54 (9.7/3.62, 10.5/3.26); quarter-2045 7.8/1.98, 5.1/1.47 (5.0/1.50, 7.9/1.97); harbour-night 14.7/2.54, 30.6/4.29, 11.9/2.02 (18.3/2.80, 28.4/4.06); lane-night 7.1/1.93, 5.7/2.05 (6.2/1.96, 6.7/2.25). Crowns and crowns_low are within their pairs in most sets, not all (see Acceptance; e.g. quarter-night crowns_low >8 1.40/0.91 against 0.67/1.97). glade-night sits a little above its pairs in two of three sets (frame >8 0.57/1.15/1.39 against 0.83/0.87/0.80/0.58; crowns mean dluma −0.045/−0.026). Its diff map shows only chimney smoke, water and rain moving, nothing on the crowns (`p/sheet/diff-glade-b3-a3.png`). The quarter maps show rain streaks and glint specks, as their noise maps do (`diff-quarter-night-b2-a2.png`, `diff-quarter-day-b2-a2.png` against `-a1-a2`). Signed crown changes flip sign between sets. Crop sheet, looked at: `p/sheet/p2p-crowns-set1.png` (identical by eye). 150: - **Acceptance:** 151: - **Pixels unchanged within the noise pairs:** met in most sets, not all (corrected by the P2 close handoff check; the first wording said met on all six views). In `p/metrics/summary.txt` and `three.txt`, quarter-day set 2 reads above both noise pairs (frame >8 2.77% against 2.27/1.36; crowns 2.77 against 2.29/1.61), harbour-night set 2 too (frame 4.29 against 2.80/4.06), quarter-day crowns_low set 1 (2.25 against 1.88/1.35), and glade-night in 2 of 3 sets. The diff maps show only rain, water, smoke and glint specks moving, nothing on the crowns' shape, and crown signs flip between sets. The day path is exact by construction (weight 0 added exactly 0). 152: - **Crown lines hold:** quarter-night crowns_low p50 24.9 -> 24.9 in both sets (`#0e1923`/`#0e1a23`; target about 26; P1c2 measured 25.3-25.8 in its sitting), p10 11.9 -> 11.9, black art pixels 18.2/17.9 -> 18.0/18.2%; crowns p50 27.8 -> 27.8. Moon-side lobes: unchanged code and values, so still darkest (`metrics/crowns-quarter-night.txt`). 153: - **Bench** (`p/bench/bench.sh`, `analyse.js` -> `summary.txt`; copies of the bisect bench's arguments; scratch trees `pre`, `after` = live copied, `nofloor` = live with `night_floor_on()` returning false (today's floor-free cost), `19-P1-b`/`19-P1-c` with the P1f report-only `main.gd` patch; `--disable-vsync`, raw_* fields, 491 samples; round 0 discarded; dense 5 rounds, Harbour 9). **Dense 13:00: the fix saves 0.19 ms (0/5 rounds slower; 14.13 -> 13.95); floor left +0.04 over floor-free (4/5; −0.05 to +0.06), was +0.22 (5/5).** **Dense night: saves 0.17 ms (0/5; 14.17 -> 13.97); left +0.04 (4/5), was +0.20.** P1b re-measured this sitting: +0.27 (13:00) / +0.26 (night), matching the bisect's +0.24. So about 85% of P1b's cost is back at 13:00 and at night. The +0.04 left is the same day and night, so it is the branch's presence (register load), not floor arithmetic. Harbour does not resolve this sitting (1-minute load 3.71-15.39 and 5-minute 4.29-7.33 across the Harbour rounds, `p/bench/runs/*/meta.txt`; per-tree spreads 0.35-0.74 ms): 13:00 pre -> after +0.03 (5/9), night −0.02 (4/9), and even the P1b step reads −0.08 at night (4/9 rounds; −0.14 was round 1 alone). The bisect needed 8 quiet rounds for Harbour's +0.15. Draw calls identical (2361 / 907). 154: - **Bench verdict (orchestrator, after the review):** dense is **mostly met**: the 13:00 cost is within 0.04 ms of a floor-free copy of today's tree, not "within noise" (the tree spread is 0.06-0.07 ms, and the +0.04 is slower in 4 of 5 rounds), and the named reference `19-P1-b` cannot be used directly because it lacks the P1c-P2 content (draw calls 2140 -> 2361). **Harbour is not resolved** in this sitting (a high and varying load, spreads 0.35-0.74 ms); its saving is unmeasured, and a quiet re-measure is left to the P2 close bench. 155: - **Journey:** expected count deltas zero on all six views; no look change. 156: - **Review:** one Opus 5.5 review in a fresh context ([run ID removed]): no must-fix. Noted: the vacuous anchor test (fixed above); the early-out is exact only while every base-pass light adds non-negative luma (no negative lights exist in scripts or scenes); the moon-side lobes claim rests on unchanged code, not a per-lobe measure. Committed and merged by the orchestrator under the owner's overnight standing instruction (30 September 2026, "automate all choices until 7am"), now recorded in `artifacts/beautify/owner-decisions.md`. 157: - **Not done:** the last +0.04 ms on dense would need a floor-free shader variant swapped by `WorldView` at dusk (a second compile and warm-up); not tried. Harbour's saving is not measured to its size; it needs a quieter sitting. The bench and pre trees are left for review. Platforms: Mac run; Windows not run. 100: **Beautify P2 stage close (30 September 2026):** the stage-end overhead of P2 "Owner-decision batch" (step `20-P2`), run by Opus 5.5 on this Mac on the final P2 code (`main` a50ef91; untracked `.claude/worktrees/` and two plan docs only; `close/head.txt`). No game code changed. Evidence: `artifacts/beautify/20-P2/close/`. Before side: the pre-stage scratch tree `close/pre-tree` from `build/backup-feel/20-P2/`. It was checked equal to `main` 791ff0c on scripts, shaders, tests, data, scenes and `project.godot`; only the `.ps1` tools differ, in line endings. 791ff0c has the same code as the P1 close's b8df6d0. Power: AC, 100% charged, Low Power Mode 0 for every run (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). 101: - **Full suite** (`close/suite.sh` a, b, v = `VERIFY_RES=2880x1800 tools/suite_all.sh artifacts/beautify/20-P2/close/suite`, 00:49-00:55): **40 suites, 19,510 checks, 0 failures**. That is P2p's 19,512 less the 2 checks the orchestrator's review fix removed from test_foliage (206). `--verify` 2880x1800 exited 0 in 32 s with `PIXELPUNK_INTEGRATION []`. The creature_reactions rerun had 5,549 checks and 0 failures. The known flaky check did not fire. 1440x900 was not run (the inherited stall). 102: - **Journey** (`artifacts/feel-journey/20-P2/`, the six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all first try; counts `20-P2/counts-vs-19-P1.txt`). Against `19-P1`, only these change: 103: - harbour-day and -night: draw_calls 909 → 897. That is P2b +2 (fog, smoke), P2d +1 (contact shade) and P2e −15 (tree groups lose their stone surface). `height_fog` 0 → 0.3, `height_fog_drawn` 0.3 at night and 0.135 by day, `smoke_points` 0 → 7. 104: - harbour-lane: 749 → 738 (+2 +1 −14). 105: - quarter-day and -night: 969 → 970. P2d +1; P2e's +10 was taken back by P2e2, which turned the roof trees' soil groups into paved ones. The e2 entry did not declare that −10; its own captures show 980 → 970. 106: - styles-day: 487 → 491 (+1 +3). 107: - `rendered_objects` moves the same way. Every change is one the parts declared or explain. 108: - Re-run: the stage-end art capture below took all six views again, with identical integer counts. 109: - By eye (`close/look/j-1.png`, `j-2.png`): mist, plumes and a strong warm lamp glow in the mist on Harbour at night; plumes and crisp shade lines by day; dark tree feet; leaded windows. The quarter and styles barely change. 110: - `notes.json` has a new `20-P2` entry (backup `notes.json.bak-before-20-P2`). `index.html` and `journey-standalone.html` (105.5 MB) were rebuilt. 111: - **Art capture and protected looks** (`close/caps.sh`; one sitting 00:57-01:06 in the order `pre-1`, `after`, `pre-2`, `after-2`). There are 25 views each, all first try: the kit's 21 baseline views, harbour-lane, and fog-2100, qn-flags and qg-flags (the quarter with `--height-fog=0.3 --chimney-smoke=on`). 112: - Metrics are in `close/metrics/` (`close/metrics.sh`): `ab-pre1-after`, `ab-pre2-after2`, noise pairs `ab-pre1-pre2` / `ab-after-after2`, `frame-*.txt`, `counts.txt`, `protected.md` (`close/prot.js`), and `ab-p1close-*` against the previous stage close `19-P1/close/after-ok`. 113: - The pre-stage tree's counts equal `19-P1/close/after-ok` on all 22 shared views. Noise pairs change >8 on 0.00-3.9% of the frame (pre) and 0.01-3.4% (after). 114: - Count changes the journey does not cover: glade-day/-night 901 → 888, grove-day 328 → 305 (P2d +1, P2e −24, as P2e measured) and harbour-golden 833 → 822. harbour-2100 is 878 → 865, one more than the day view's −12. chemistry-day is 595 → 594; no part captured chemistry, and it is not chased. Every other quarter view is +1. 115: - Protected-look sheets: `close/protected/protected-1.png` to `-3.png`, plus `noise-pre.png` and `noise-after.png`, all looked at. 116: - **Canal mirror columns:** all streaks are there by eye at 2x (`close/look/mirror.png`). p50 43.0 → 42.7, but changed >8 on 8.0/9.1% against noise of 1.3/3.0%: the sparkle pattern, not the columns. quarter-2100 is 8.4/7.0% against 5.7/4.9%. 117: - **HDR neon cores:** the fish sign is unchanged (p50 56.5, 1.7% >8). The bar sign is 53.7 → 54.2 (1.0/1.7%). The quarter tower windows are 121.2 → 121.2 (3.6%, P2a's leading). lane-night's `ln_sign` is 132.7 → 139.6 (12.7%): the fog glow lifts the surround. 118: - **05:45 and 20:45 skies:** unchanged. Zenith 56.6 → 56.6 and 97.5 → 97.5, horizon bands 114.9 and 153.0; changed >8 is within or near the noise pairs. 119: - **Glade/Canal separation:** Glade night zenith `#001b43` → `#001d45` and horizon 88.2 → 89.1 (both 0.00% >8), water 32.2 → 33.8. Harbour night zenith `#273180` → `#293280`. They are still clearly apart. Glade day lawn and bank change 0.00%. The Glade frames change 1.3% (day) and 6.0% (night) from P2b's fog 0.15 and smoke; P2d left Glade within noise. 120: - **Approved tree style: a finding.** The grove (a canal document with no fog or smoke) changes by more than any part reported. Crowns change >8 on 10.8%, conifers 13.0% and the frame 7.4%, against noise of 0.00-0.14%. Means move −0.4/+0.4, and the lawn's hatched pattern is gone (`close/look/grove-crowns.png`, sheet page 11). Apart from the intended tree feet (p50 70.8 → 56.3, 71%), it comes from P2d's canal SSAO removal. P2e's pre-tree (after P2d) already differs from `pre-1` on 7.1% of the grove frame, and P2e's after set differs from `after` on 0.06%. P2d did not capture the grove. The owner waived the SSAO-darkened night creases but has not seen the grove. 121: - quarter-night crowns and undersides are within noise (P2p). 122: - **Whole-stage A/B bench** (`close/bench.sh`, `bench-t1.out`, `bench-t13.out`, `bench/summary.txt` from `close/bench-summary.js`, `bench/power.txt`). The pre-stage tree runs against the live root, interleaved, round 0 discarded and rounds 1-3 kept, with `--disable-vsync` before `--path`, 1920x1080, rain 0.45, at Night (`--time=1`) and 13:00. Both trees report raw frame fields (since P1f). vsync was "disabled" in every run, there were 491 raw samples, and every run was first try. Raw p50 medians: 123: - **Dense, night: 13.47 → 13.19 ms (−0.31; 0/3 rounds slower); p95 15.64 → 15.19.** 124: - **Dense, 13:00: 13.55 → 13.16 ms (−0.40; 0/3); p95 15.64 → 15.10.** 125: - **Harbour, night: 6.59 → 6.65 ms (+0.06; 3/3); p95 7.31 → 8.39.** One round per side jumps by about 1 ms, so the p95 is noisy. 126: - **Harbour, 13:00: 6.58 → 6.61 ms (+0.03; 3/3); p95 7.64 → 7.96.** 127: - Draw calls: dense 2360 → 2361, Harbour 919 → 907. 128: - Reading: the whole of P2 costs Harbour less than P2b's fog and smoke alone (+0.25). That fits P2p saving on Harbour at night (unresolved in P2p) and 15 fewer tree draw calls, but it was not isolated. The dense scene gains 0.3-0.4 ms, mostly from P2p, with P2d's shade about paid for by the SSAO it replaced. 129: - The live side ran from the project root; P2b measured root and scratch copy within 0.01 ms. 130: - **Owner sheet** `artifacts/beautify/20-P2/close/sheet/`: before (pre-stage tree, `pre-1`) is on the left and after on the right, colour then greyscale where values matter. Crops are from `close/crops.js` and the pages from `close/sheet.sh`; all pages were looked at. 131: - Pages: `p2-01-overview.png`, `p2-02-P2c-fog-lit-by-the-city.png`, `p2-03-P2c-warm-smoke-and-lamp-glow.png`, `p2-04-P2b-fog-and-smoke-on.png`, `p2-05-P2d-contact-shade.png`, `p2-06-P2e-soil-collars.png`, `p2-07-P2e-tree-pits-and-lawn.png`, `p2-08-P2e2-roof-trees.png`, `p2-09-P2a-leaded-windows.png`, `p2-10-P2p-no-look-change.png`, `p2-11-finding-grove-ssao.png`, `p2-12-existing-scenes-1.png`, `p2-13-existing-scenes-2.png`. 132: - `sheet.md` has one plain paragraph per part with its numbers, then the protected looks, the bench and the suite. 133: - The strongest visible change is the lamp glow in the mist at golden hour and night. On lane-night the lawn and path turn olive-amber: frame mean +7.4 luma, 26% >8, lamp crops +41 to +47. The owner kept it (P2-1), but only from the P2b/P2c crops; this sheet shows it at lane zoom. 134: - **Handoff-check corrections (30 September, orchestrator; the check listed 15 claims a file did not back):** fixed in place: P2p's pixel acceptance and its Harbour load and P1b-step figures; P2a's power line, tool-defect status and review; P2c's fog-off noise and Glade night figures; P2d's gd-off range; P2e2's review line; this resume note. **Not backed by a file, left as written and flagged:** P2b's owner-sheet figures (frame mean +1.4 to +2.1, >8 on 1-3.4%; bracketed room golden 32.7% and +8.0; golden-to-night 7-15%, 4.4-4.8%, 1.7-2.3%), which are not in `20-P2/b/metrics/`; the P2b styles-fix sitting's Low Power Mode state and the P2e sweep's power source (`b/cap-power.txt`, `e/sweep-power.txt` hold no such line). Treat those as unverified. 135: - **Not done / open:** 136: - The grove finding above: the owner may want to see page 11. Changing it would reopen owner ruling (b) for canal. 137: - chemistry-day's −1 draw call and harbour-2100's extra −1 were not traced to a part. 138: - The Harbour share of P2p's saving is still not measured on its own. 139: - The P0c night bank line is still owed by P3c (G1 Q3). 140: - No milestone export at the P2 close (D8). 141: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 142: - Scratch trees: the `pre-tree` folders under `20-P2` (`a`, `b`, `c`, `d`, `e`, `e2`, `p` and `close`) were moved to [local trash folder] with `mv` once their numbers were recorded, named `beautify-20-P2--pre-tree-