Chapter 28: Stone, grass and a lighter touch Frozen 9 October 2026 at 19:36 BST (game main at 8418d4e, clean; the frozen handoff and owner-decisions.md equal HEAD). Covers sub-stage P10-I of stage P10 (polish backlog): P10p hover highlight, arrows only and double-click to move (owner request, with its second review, fix round and follow-ups), P10a violet-grey canal stone and calmer upper storeys (two review fixes), P10b lawn tone bands and turf octave, P10f irregular paving and the mauve plaza (two fix passes), P10i water sun shadow and streak HDR core (review fix), P10h art-pixel tilt-shift and photo bokeh (review fix), and the P10-I stage close. Owner answers: P10f and P10i "keep all new"; P10a, P10b, P10p and P10h "keep all" (9 October). The note on evidence paths (handoff line 318) is orchestration and left out. Original source terminology is preserved. Line numbers refer to the frozen copy. Local paths and run IDs are replaced by placeholders; moves "to the macOS Trash/" read "to the macOS Trash" and "the Desktop/" reads "the Desktop". SOURCE: docs/visual-handoff.md (frozen copy = HEAD 8418d4e, 9 October 2026, 19:36 BST), lines 205-316 205: **Beautify P10-I (parts p, a, b, f, i, h of P10) stage close (9 October 2026):** the stage-end overhead of sub-stage P10-I "Polish backlog" (step `28-P10-I` in the plan's overhead table). Opus 5.5 ran it on this Mac on the final P10-I code, `main` 862ccbe (P10h), working tree clean (`close/head.txt`). No game code changed. Evidence: `artifacts/beautify/28-P10-I/close/`. 206: - **Before side:** `close/pre-tree` from `tools/feel_snapshot.sh tree 28-P10` (= `main` a46ea21, the P9 close code). It differed from live only in the stage's 29 files (`close/diff-pretree.txt`); each of the 27 files that exist at a46ea21 is byte-equal to it (`close/git-changed-since-pre.txt`); `paving_grid.gdshaderinc` and its uid are new. 207: - **Power:** every run on AC, 100% charged, Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`); 1-min load 1.8-3.2. 208: - **Full suite** (`close/suite.sh` a, b, v; 06:56-07:04): **46 suites, 24,306 checks, 0 failures** (`close/suite/suite_all.out`). `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`. `test_creature_reactions` 5,583 checks, 0 failures in the run and the rerun; the known flaky check did not fire. 1440x900 not run (inherited stall). 209: - **Journey** (`artifacts/feel-journey/28-P10-I/`, six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): draw calls and every integer count identical to 27-P9 on all six views (`28-P10-I/counts-vs-27-P9.txt`), as every part declared (none). Frame >8 luma against 27-P9 (`28-P10-I/frame-vs-27-P9.txt`): harbour-night 3.8%, harbour-day 7.2%, styles-day 5.4%, harbour-lane 23.1%, quarter-day 32.3%, quarter-night 34.9%; within-tree re-run of harbour-lane and quarter-night 1.9% and 1.3%, counts identical (`close/rerun/`). Looked at side by side: the large shares are P10a's violet-grey stone, P10f's mauve plaza and P10b's lawn bands; harbour-day shows the P10i shadow on the water right of the island. `notes.json` has a `28-P10-I` entry (`close/journey-note.js`, backup `notes.json.bak-before-28-P10-I`); `index.html` and `journey-standalone.html` (110.86 MB) rebuilt. 210: - **Art capture and protected looks** (`close/caps.sh`, the P9 close's 101 views in groups g1-g3; 07:07-07:50; pre-1, after, pre-2, after-2, every view on try 1). Regions `close/regions.json` = the P9 close's plus the P10 parts' regions from `tools/art_regions.json`. Metrics `close/metrics/` (`metrics.sh`, `frame-table.js`, `prot.js` → `protected.md`, `elapsed.js`, `consistent.js` → `metrics/consistent.txt`, `look/consistent/*.png`). Against the previous stage close `27-P9/close/after` (the same code as pre-1, another sitting): `ab-p9close-*`, `frame-p9close-*`. Sheets `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P9-close.png`, `noise-pre.png`, `noise-after.png` (looked at: 1 and 4). 211: - **Counts** (`metrics/counts.txt`): identical on 99 views in both pairs; qn-orbit and quarter-orbit vary within each tree as before. 27-P9 close after against pre-1: identical apart from the two orbit views. 212: - **Frames:** all 101 views above both noise legs (`frame-table.txt`; smallest field-q 1.2%/1.1% against 0.2%), since canal stone is in every canal view. Systematic change (all four cross pairs, neither noise pair; `consistent.txt`): quarter-day 10.7%, harbour-lane 11.0%, harbour-day 4.6%, harbour-golden 6.1%, grove-day 51.1% (its lawn), glade-day 1.7%, glade-night 0.6% (Glade's stone house and canal roof decks, `look/consistent/glade-night.png` looked at; the parts declared both). The live capture is a median 56 ms later after start than the pre one (`elapsed-ms.txt`), so TIME-driven clouds sit at a slightly different phase between trees (harbour-night zenith 22.5% changed between trees, 2.6% between the P9 close and after). 213: - **Canal mirror columns:** p50 43.5 → 43.5 (43.4 → 43.6) quarter-night, 52.5 → 52.4 at 21:00; changed >8 26.0%/20.6% against noise 6.8%/0.9%: P10i's streak core (paler centres), declared. 214: - **HDR neon cores:** bar 70.6 → 70.8, lane sign 140.2 → 140.0, tower windows 96.0 → 95.8, all inside or at noise in value; the fish and lane-sign regions change 11.8-18.7% of pixels on the stone behind the letters (P10a), the letters themselves are unchanged (`protected-1.png`). 215: - **05:45 and 20:45 skies:** changed ≤0.9% against noise up to 0.6% (largest leg 0.87% against 0.63%), p50 equal (quarter 20:45 horizon 151.1, zenith 97.7; 05:45 horizon 113.9, zenith 56.6; Harbour 05:45 horizon 115.8). 216: - **Glade/Canal separation:** Glade night zenith `#001d45`, horizon, water, lawn, bank and crowns 0.0-0.1% changed; Harbour night zenith p50 53.9 = 53.9. 217: - **Approved tree style:** crown and conifer colours unchanged (grove-day p50 55.3 and 55.6, grove-1930 48.3 → 48.5 and 42.9 → 43.1, quarter-night undersides 27.7 = 27.7); the 5-14% of pixels that change in the crown and conifer boxes are the lawn showing between them. **The grove's lawn changed** (78% of pixels >8, golden 73%): P10b's tone bands and turf octave reach the approved grove (a canal document); P10b reported it (grove tufts 1 → 9 px). Golden Harbour lawn region p50 91.3 → 82.6, mean 86.5 → 84.0 (the dark band covers more of the region; the part held the whole lawn's linear value within +0.8%). 218: - **Whole-stage A/B bench** (`close/bench.sh` = the P9 close recipe; `bench-t1.out`, `-t13.out`, `-t19.5.out`, `bench/summary.txt`; 08:03-08:15, every run on try 1): pre-tree against live, interleaved, round 0 discarded, rounds 1-3, `--disable-vsync` before `--path`, 1920x1080, rain 0.45, `--motion-live`, dense `artifacts/weather-pass9/dense.json` and Harbour; 491 samples each. Raw medians pre → live (paired median; rounds live slower): 220: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 221: |---|---|---|---|---|---|---| 222: | Dense | 14.307 → 14.358 (+0.036, 3/3) | 16.168 → 16.214 (+0.022, 3/3) | 14.290 → 14.436 (+0.130, 3/3) | 16.211 → 16.296 (+0.085, 2/3) | 16.098 → 16.244 (+0.146, 3/3) | 17.950 → 18.040 (+0.109, 3/3) | 223: | Harbour | 6.712 → 6.725 (+0.021, 3/3) | 8.179 → 7.362 (−0.573, 0/3) | 6.602 → 6.811 (+0.212, 3/3) | 7.601 → 8.078 (+0.771, 2/3) | 6.464 → 6.698 (+0.210, 3/3) | 7.753 → 8.284 (−0.148, 1/3) | 225: - Draw calls unchanged (dense 2588 / 2589 / 3440, Harbour 918 / 919 / 831). The stage costs about +0.2 ms p50 at Harbour by day and at golden hour and +0.13-0.15 ms on dense (P10i's march, as the part measured), almost nothing at night. Harbour p95 spreads 7.0-8.5 ms in both trees and swings both ways; the plan's stop line (dense p95 over about +0.5 ms) is not hit. 226: - **Owner sheet** `artifacts/beautify/28-P10-I/close/sheet/` (`close/sheet.sh`, regions `close/sheet-regions.json`; before left, after right, colour plus greyscale where values matter): `p10-01-p-hover.png`, `p10-02-p-selected-placement.png` (from P10p's own captures `28-P10/p/sheetdirs`, taken before the later parts and before the P10p follow-up commit 3401cc9), `p10-03-a-stone.png`, `p10-04-b-lawn.png`, `p10-05-f-paving.png`, `p10-06-i-water-shadow.png`, `p10-07-i-streak-core.png` (close pre-1 | after), `p10-08-h-photo-bokeh.png` (P10h `h/fix/cap` pre-1 | fixed-1), `p10-09-h-tilt-grid.png` (P10h `h/cap` pre-1 | live-1; the review fix did not touch the editing lens), `p10-10-existing-scenes.png`, and `sheet.md` (one paragraph per part with its measured numbers, then the bench and the suite). The close agent looked at pages 01, 03 (first version), 05, 06 and 08. 227: - **Sheet check after the close (9 October, no Godot):** all ten pages opened at full size: labels legible, panels in the order the labels say (before left, after right; colour above greyscale where the header says so), no title cut off, no box or wireframe drawn round any object by the sheet (the only boxes are the old game's hover, cage and floating wireframes in the *before* panels of pages 01-02, which is what P10p removed). One rebuild: page 09 showed the tilt-shift rows only at 1x, where the 2x2 art-pixel blocks cannot be seen; it now adds `tilt-night:bottom_zoom` ([600,880,760,976], 3x; new region in `close/sheet-regions.json`, nothing moved; `close/sheet.sh` line for page 09 updated; same captures). Pages 01-02 were captured on 8 October 22:35 on the first P10p code (`28-P10-p-wip`); fix round 1 (eb7ea0b) and the follow-up 3401cc9 changed only input and capture bookkeeping in `scripts/main.gd` (no change to `world_view.gd`, `selection.gdshaderinc` or anything drawn while the pointer is on a part), so the hover, selection and placement look they show is the committed one, but their stone, lawn and paving predate P10a, b and f; `sheet.md` says so. Caption notes added to `sheet.md` where it quotes a region other than the one on the page (pages 04, 05), for page 06 (the shadow is clear in the boats crop, faint in `p10i_shadow` and `p10i_gright`) and for page 07 (below). 228: - **Page 07, harbour-night front water:** the cyan reflection dashes along the front lawn's edge are present in both pre captures (`close/pre-1`, `pre-2`) and in the P9 close `after`, and absent in both after captures (`close/after`, `after-2`); this is the largest cluster ([368,640-1071,959], 11,796 px) of harbour-night's 5.0% systematic change in `close/metrics/consistent.txt`, which the systematic-change list above does not name. Every part's own pair shows the dashes in both trees (`a/fix2/cap` pre-5 | live-5, `f/cap` pre-1 | live-1, `i/cap` pre-1 | live-1, the last on the 32-step P10i code), and the journey shows them in both trees but at a different place along the edge (`feel-journey/27-P9` under the lawn, `28-P10-I` further right under the boathouse). That points to the TIME-driven neon-bay dashes (re-rolled `PANEL_DASH_RATE` 1.5 times a second, `shaders/water.gdshader`) and each tree's capture clock rather than code, but it is not proven; the page-07 caption ("streak core") does not describe what that panel mostly shows. Settling it needs harbour-night captures of `28-P10-i` against `28-P10-h-pre` (or two clocks of one tree). 229: - **Not done / open:** 230: - No photo-mode, perspective, hover or Reduced/Off views in the close capture (the 101-view set has photo-golden only); P10p and P10h are shown from their parts' captures. 231: - The P10p follow-up commit (3401cc9: Select tool only, hover released over the UI, capture fade edge, wall second press) has no picture of its own on the sheet. 232: - The harbour-night cyan front-water dashes that are missing from both close after captures (sheet check above) are not explained; probably the dashes' clock, not proven. 233: - Snapshots as they are on disk (`diff -rq` on scripts shaders tests data scenes tools and `project.godot`, 9 October after the close): `28-P10` = `27-P9-post` (pre-stage); `28-P10-p-wip` = the reviewed first P10p code; `28-P10-p` = 3401cc9 = `28-P10-a-pre`; each later `28-P10--pre` is the part's pre-part tree and equals the previous part's final (`b-pre` = `a-post-fix2`, `f-pre` = `b-post`, `i-pre` = `f-post`, `h-pre` = `i-post-fix`); each `28-P10-` (a, b, f, i, h) is the post-commit snapshot of that part's final code and equals its commit (1486774, a3c2948, adb5249, 9f0f36a, 862ccbe) and its last `-post*` folder. The part entries below were reworded to match. 234: - The grove lawn change (approved tree-style view) and the darker golden-lawn median are on the sheet for the owner; nothing was changed. 235: - No export (none asked); the handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 236: - Scratch trees: `close/pre-tree` was moved to `the macOS Trash/` with `mv` once its numbers were recorded (`close/trash.txt`). The part scratch trees under `artifacts/beautify/28-P10/` (`a/pre-tree`, `a/calm0-tree`, `a/old-tree`, `b/pre-tree`, `f/pre-tree`, `f/grey-tree`, `h/pre-tree`, `h/fix/prefix-tree`, `i/pre-tree`, `i/coreoff-tree`, `i/fix/s32-tree`, `p/pre-tree`) were left in place (outside this close's folder). 237: - Platforms: Mac run; Windows not run. 239: **Beautify P10h Tilt-shift and photo depth of field on the art-pixel grid (9 October 2026):** Opus 5.5 on this Mac (plan section "P10h"; report polish backlog "Tilt-shift and photo DOF on the art-pixel grid", pixel lens F8 revised step 1; step `28-P10`, sub-batch P10-I, part h; a look change to the lens only, approved as built under the owner's standing approval if it clears review). At pixel size 2 or more both lens passes (`shaders/tilt_shift.gdshader`) now work per art pixel: each window pixel takes its art pixel's centre on the grid the frame starts at (`art_origin` = `Main.frame_origin()`, kept current by `_place_lens_grid()` from `_place_frame` and `_apply_lens`, so P9e's snap scroll moves the grid with the picture), the blur radius is read on the centre's row and the taps step in whole art pixels (`GRID_TAPS` 10 a side), so blurred regions stay whole art-pixel blocks after the nearest upscale. In photo mode's depth of field only (`Main.lens_bokeh`), a tap's weight grows with its highlight (luma over `BOKEH_KNEE` 0.75, gain 1 + `PHOTO_BOKEH` 0.5 x `BOKEH_GAIN` 8) and its fall-off flattens to a box, and the first pass writes the brightest highlight it reached as 1 - alpha (`render_mode blend_disabled`) for the second pass, so a lamp spreads into a flat square instead of a smudge or a horizontal smear; after review the bokeh (weight and box alike) fades in with the row's radius in art px, none up to `BOKEH_FADE_FROM` 3.0 and whole from `BOKEH_FADE_TO` 6.0 (smoothstep), so photo mode's gentle default depth of field (`PHOTO_GENTLE_DOF` 0.3, at most 2.3 art px) gets none and depth of field 0.6 (at most 4.7 art px) gets it only in its most blurred rows, at most 0.6 of it. Pixel size 1 (Glade's default) and `Main.LENS_ART_GRID = false` (the revert switch) run the old window-pixel blur unchanged; `PHOTO_BOKEH = 0.0` turns the bokeh alone off. The editing tilt-shift never gets bokeh. True depth of field (a photo-only CoC SubViewport) stays out of scope; the canvas back buffer is LDR. Files: `shaders/tilt_shift.gdshader`, `scripts/main.gd`, `tests/test_tilt.gd`. Snapshots: pre-part saved as `build/backup-feel/28-P10-h/` (equal to `28-P10-i`, diff -rq clean; nothing unexpected in the live tree); that pre-part tree is now `build/backup-feel/28-P10-h-pre/` (equal to `28-P10-i-post-fix`), and `28-P10-h/` now holds the final P10h code, the post-commit snapshot (862ccbe; equal to `28-P10-h-post-fix`) (checked with `diff -rq` 9 October after the stage close; scratch tree `h/pre-tree` equal to `28-P10-h-pre`); post `build/backup-feel/28-P10-h-post/` (the code as first reviewed) and after the review fix `build/backup-feel/28-P10-h-post-fix/` (equal to the live tree; it differs from `28-P10-h-post` only in `shaders/tilt_shift.gdshader` and `tests/test_tilt.gd`; scratch tree of the reviewed code `h/fix/prefix-tree`). Evidence under `artifacts/beautify/28-P10/h/` (below `h/`). 240: - **Profiles reached:** Glade (pixel size 1): none (glade-day pre -> live 5.6% of pixels changed, >8 luma 0.27%, against its noise pairs 7.4% / 0.32% pre and 5.2% / 0.32% live). Canal and Harbour (pixel size 2): wherever the lens is drawn, that is with the tilt-shift switched on and in every interactive photograph, because entering photo mode turns on the gentle depth of field 0.3 (`_photo_lens_on`; the first entry said "neither is on in their default documents", which the review found wrong). That default is now captured (`h/fix/`, below): harbour-photo (Harbour, perspective photo lens, dof 0.3, day) pre -> fixed 55.5% of pixels changed, >8 6.7%, >24 1.7% against noise 11.3% / 0.78% / 0.27% pre and 10.8% / 0.82% / 0.33% fixed, all of it the blur moving onto the art grid (no bokeh at 0.3); photo-gentle (the quarter, dof 0.3, night) pre -> fixed 74.3% / 8.1% / 1.0% against noise 6.2% / 0.27% / 0.03%. With the lens off nothing changes: harbour-day (lens off) pre -> live 8.9% / 0.93% against noise 14.6% / 2.0% pre and 9.9% / 1.1% live. 241: - **Tests:** test first `h/testfirst/h-testfirst-test_tilt.log.out` (the new checks against the unchanged main.gd: parse errors on `PHOTO_BOKEH`, `LENS_ART_GRID`, `lens_grid`, `lens_bokeh` and "Failed to load script" in the log; no exit code is recorded in `h/testfirst/`). New `grid_checks` in test_tilt: the uniforms and their old-look defaults (`art_pixels` 0, `bokeh` 0), `blend_disabled`, the floor snap, gain/knee/tap bounds, the grid per pixel size following `LENS_ART_GRID`, no bokeh in the editing lens or at pixel size 1, and a GDScript port of the grid branch on a model frame (2-px grid from (-3,-1), one lamp): both passes give whole 2x2 blocks with and without bokeh; photo bokeh is 7 x 7 art px, as tall as wide, filling its box (fill 1.00 over half the peak), corner 0.38 of peak 0.46 (editing lens: peak 0.21, a soft blob). `--verify` gains "Lens grid did not follow the pixel size and the frame origin". `h/test_tilt.log.out`: 207 checks, 0 failures (bokeh 1.0 and 0.5 tuning logs in `h/tune/`). Full suite (`h/suite.sh`, `h/suite/suite_all.out`, `h/suite.console`, `VERIFY_RES=2880x1800`, 06:14-06:23, AC, Low Power Mode 0 in `h/suite-power.txt`): **46 suites, 24,301 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (33 s), `PIXELPUNK_INTEGRATION []`, no shader error in its logs; test_creature_reactions 5,583/0 and rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Review fix: `h/fix/test_tilt-try1.log.out` (212 checks, 1 failure: the model lamp's peak 0.34 under the old 0.35 floor once the model moved to the fade's end; the floor became relative, below), `h/fix/test_tilt-fade25.log.out` (fade 2.5..4.5, 212/0, superseded), final `h/fix/test_tilt.log.out` **212 checks, 0 failures**: new checks on the fade constants (rising; `BOKEH_FADE_FROM` >= `PHOTO_GENTLE_DOF` x `TILT_REACH`; `BOKEH_FADE_TO` <= `TILT_REACH`) and on the model, at 2.34 and 3.0 art px photo bokeh blurs exactly as the editing lens; the port carries the fade. Not test first (the shader was changed first in this fix pass). Full suite after the fix (`h/fix/suite.sh`, `h/fix/suite/suite_all.out`, `h/fix/suite.console`, `VERIFY_RES=2880x1800`, 06:36-06:45, AC, Low Power Mode 0 in `h/fix/suite-power.txt`): **46 suites, 24,306 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (33 s), `PIXELPUNK_INTEGRATION []`, no shader error; test_creature_reactions rerun 5,583/0. 242: - **Changed expectations (old -> new):** test_tilt 176 -> 207 checks (new checks only), and after the review fix 207 -> 212; suite total 24,270 -> 24,301 -> 24,306. The review fix changed two model expectations: the lamp model runs at `BOKEH_FADE_TO` (6.0 art px) instead of 3.0 (half-widths 3 x 3 -> 7 x 7, peak 0.46 -> 0.28, editing lens 0.21 -> 0.16, search window +-5 -> +-7 cells), and its peak floor "> 0.35" -> "> 1.5 x the editing lens's peak" (0.28 against 0.16). Journey count deltas expected: none (no journey view draws the lens; draw calls and every integer count identical pre -> live in all five capture views, `h/metrics/counts.txt`). 243: - **Captures** (`h/cap.sh`, `h/capall.sh` -> `tools/art_capture.sh`, 1600x1000, final sitting 06:06-06:08 pre-1, live-1, pre-2, live-2, every view try 1, AC, Low Power Mode 0 in `h/cap-power.txt`; views photo-golden, photo-night (the quarter, perspective photo lens, dof 0.6, night), tilt-night (the quarter, `--tilt-shift=0.8`), glade-day, harbour-day; an earlier set at `PHOTO_BOKEH` 1.0 is in `h/cap/old-b1/`, its `void/` set was a misnamed live run). Sheets, pre-1 left, live-1 right, all looked at: `h/sheet/p10h-1-photo-night-fg.png` (x2: string-light bulbs and lanterns become flat yellow squares that keep their brightness, where pre they were dim round smudges; bright flower heads also become pale squares), `h/sheet/p10h-2-tilt-night-bottom-x4.png` (x4: pre shows sub-art-pixel streaks and gradients inside the blur, live is clean 2 x 2 blocks), `h/sheet/p10h-3-photo-golden-fg.png` (by day few taps pass the knee: the same soft blur on blocks, a few pale flower squares), `h/sheet/p10h-4-glade-day.png` (identical). Full live frames looked at: `h/cap/live-1/photo-night/courtyard.png` (the sharp band keeps the lanterns and signs; the bulbs in the blurred foreground are squares, the pink neon signs at the right, in the blurred rows, lose the letters that read faintly before, because their bright tubes now spread as squares). 244: - **Review fix captures** (`h/fix/cap.sh`, `h/fix/capall.sh`, 1600x1000, one sitting 06:31-06:36, pre-1, prefix-1 (the reviewed code), fade25-1 (fade 2.5..4.5, superseded), pre-2, prefix-2, fade25-2, fixed-1, fixed-2, every view try 1, AC, Low Power Mode 0 in `h/fix/cap-power.txt`; views photo-night (dof 0.6), photo-gentle (dof 0.3, the default), harbour-photo, glade-day). Sheets pre-1 | prefix-1 | fixed-1, all looked at: `h/fix/sheet/fix-1-photo-night-neon-x2.png` (the review's region x 1120-1440, y 560-820: prefix blows the BAR and cross signs into pink-white slabs; fixed keeps them pink with the B, the A and the cross's arms reading about as pre, the R as soft as pre; at fade 2.5..4.5 the lower letters were still slabs, so the fade moved to 3..6), `h/fix/sheet/fix-2-photo-gentle-neon-x2.png` (dof 0.3: prefix fills the letters, fixed reads BAR and the cross crisply, sharper than pre), `h/fix/sheet/fix-3-photo-night-fg-x2.png` (the foreground still spreads bulbs and flower heads into squares, smaller and dimmer than prefix), `h/fix/sheet/fix-4-harbour-photo.png` (dof 0.3 by day: no visible change at full frame). Measured (`h/fix/metrics.sh`, `h/fix/metrics/fix-metrics.txt`): neon region, pixels changed by > 24 levels against pre-1: photo-night prefix 21.5% -> fixed **12.0%**, photo-gentle 12.8% -> **2.7%** (noise 0.0%; the rest is the grid). Bright pixels (luma >= 200) in rows 560-820: photo-night pre 1.43%, prefix 3.76%, fixed 2.02%; photo-gentle 2.34%, 4.31%, **2.65%**. Lamp bokeh, photo-night rows 760-1000, luma >= 160: pre 1.12%, prefix 6.59%, fixed 3.83% of pixels, 19 / 67 / 53 blobs, median aspect 1.20 on all three; photo-gentle 2.06% / 5.90% / 2.15% (no bokeh at 0.3). Blocks, rows 860-1000: photo-night 35.0% -> 100.0%, photo-gentle 18.7% -> 100.0%. fixed-1 and fixed-2 give identical bokeh numbers. Glade-day pre -> fixed 23.3% / >8 0.59% against fixed's own noise pair 28.9% / 0.73% (pixel size 1 does not reach the changed branch; this sitting's glade noise ran high). 245: - **Acceptance, measured:** (1) blurred regions stay pixel-size blocks (`h/blocks.js`, `h/blocks.sh`, `h/metrics/blocks.txt`; share of non-flat 2 x 2 blocks whose pixels agree within 2 levels): tilt-night bottom rows 860-1000 38.6% -> **100.0%**, top rows 0-120 46.6% -> **100.0%**, photo-night rows 860-1000 35.1% -> **100.0%**, photo-golden 36.5% -> **100.0%**; the sharp band (rows 470-530) 100% on both; pre-2/live-2 the same to 0.2. (2) lamp bokeh square, not a smear (`h/bokeh.js`, `h/metrics/bokeh.txt`; photo-night rows 760-1000, luma >= 160): bright pixels 1.11% -> 6.57%, blobs over 16 px 19 -> 67, median bounding-box aspect 1.20 -> 1.20 (no horizontal stretch), median fill 0.75 -> 0.78 (neighbouring bulbs merge into one blob, so fill under-reads the squares the sheet shows); photo-golden (luma >= 200) aspect 1.90 -> 1.17. (3) Glade at pixel size 1 unchanged: inside the noise pair (above). The alpha hand-over was not probed on its own: the squares in sheet 1 show the vertical pass weights the spread lamps, so the back buffer carries the alpha on this Mac. 246: - **Bench** (`h/bench.sh`, the P10i recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, 1920x1080, `--disable-vsync`; pre-part tree against live interleaved, warm-up round 0 discarded; `h/bench.out` rounds 1-3, `h/bench2.out` rounds 4-6, `h/bench/summary.txt`; AC, 100%, Low Power Mode 0 in `h/bench/power.txt`, 06:08-06:14): pooled 6 rounds **photo-night (perspective, dof 0.6, 01:00) p50 8.231 -> 8.113 ms (-1.4%), p95 9.487 -> 9.533 (+0.5%); tilt-day (Harbour, tilt 0.8, 13:00) p50 7.645 -> 7.431 (-2.8%), p95 8.399 -> 8.030 (-4.4%)**; by set photo p95 -0.2% and +1.2%, tilt p95 -4.5% and -4.3%. Draw calls 1154 / 921 on both trees. The grid branch takes one tap per art pixel instead of one per 1.5 window px, so it is a little cheaper. Dense fixture not benched (the lens is off there). 247: - **Review:** filled by the orchestrator. The review's two must-fix findings and what was done: (1) photo bokeh filled the neon lettering in lightly blurred rows, and photo mode's default 0.3 depth of field got it on every photograph: fixed by the radius fade above (`BOKEH_FADE_FROM` 3.0, `BOKEH_FADE_TO` 6.0; `FROM -1.0, TO 0.0` restores the unfaded bokeh); (2) the profiles line wrongly said Canal and Harbour change only with a lens switched on, and the 0.3 default was not captured: line corrected and the default captured (photo-gentle, harbour-photo). 248: - **Not done:** step 2 of the report item (true depth of field from a photo-only CoC SubViewport, L) is out of scope by the plan. The bokeh strength is a taste call: `PHOTO_BOKEH` 1.0 (first try, `h/cap/old-b1/live-1/photo-night/`) made the squares too dominant and was lowered to 0.5; bright flower heads and window glints also become squares (luma alone cannot tell them from lamps). No second look pass (time limit 07:15). After the review fix: not re-benched (two constants and a smoothstep per fragment, fewer weighted taps than before; the first bench stands as the upper bound), draw calls not re-counted (shader-only change), depth of field 0.6 now gets 0.6 of the bokeh at most (whole only from about 0.77). Orbit shimmer of the blurred band not judged (stills only). 249: - **Platforms:** Mac run; Windows not run. 251: **Beautify P10i Water receives the sun's shadow; streak HDR core (9 October 2026):** Opus 5.5 on this Mac (plan section "P10i"; report polish backlog "Water receives the sun's shadow (a screen-space march toward the sun)", water lens M0 (`missed_by_reviewer`), and "Reflection streak HDR core", water lens F8; step `28-P10`, sub-batch P10-I, part i; a look change, approved as built under the owner's standing approval if it clears review). (1) The water is unshaded, so every shadow stopped at the cut edge. Each water fragment now marches the depth buffer from its own point toward the sun (`sun_shadow_march`, new in `shaders/mirror.gdshaderinc`) and, where the ray passes behind the town (at most `SUN_SHADOW_THICK` 5 view metres behind a surface), the body and the mirrored sky darken by `WATER_SUN_SHADOW` 0.30 and the sun's glitter and path go out; the mirrored town (a reflection) keeps its colour. The ray is the true sun (`low_sun_direction`, which `WorldView.sync_profile` already sets on the water at every hour), floored at sin 3 degrees like the key light (`SUN_SHADOW_FLOOR`), and the shadow comes in over sun height y 0..0.10 (`SUN_SHADOW_GATE`), so with the sun down (every night look, 05:45, from 21:00) nothing is marched. 24 steps from 3 render px, the stride growing x1.14 a step (about 475 render px, about 36 m of ray at Harbour; 32 steps at x1.10 before the review fix, see below), a hit fading over the last half of the reach (`SUN_SHADOW_SOFT`) so an over-long golden shadow thins out, nothing above 30 m (`SUN_SHADOW_TOP`); under the orthographic camera each step is affine (no matrix), photo mode projects each step. Deviation from the report: a growing stride instead of 24-32 fixed 3-px steps (fixed steps reach only about 8 m of ray, too short for golden-hour shadows). (2) Each reflection streak (water and wet paving, same shader) gets the report's core: near its mirror point (|t| < 0.5) its middle (edge < 0.04..0.22) pales toward `tint*0.5 + (0.95, 0.88, 0.72)` by up to 0.65 and is lifted by up to x1.6, so centres read pale yellow-white and bloom in their own colour; edges and ends keep the squared tint. **Reverting:** `WATER_SUN_SHADOW = 0.0` in `shaders/water.gdshader` (no march: the old water exactly); `STREAK_HDR_CORE = false` in `shaders/reflections.gdshader` (the old streak exactly); the tests read both. Files: `shaders/mirror.gdshaderinc` (`sun_shadow_march`), `shaders/water.gdshader` (`WATER_SUN_SHADOW`, `SUN_SHADOW_*`, the march after the sky's sun alignment, glitter and path x (1 - in_shadow)), `shaders/reflections.gdshader` (`STREAK_HDR_CORE`, `STREAK_CORE_*`), `tests/test_water.gd` (`p10i_checks`). No GDScript, no new fixture or region (nothing for `tools/repo_evidence.sh`). Snapshots: pre-part saved by me as `build/backup-feel/28-P10-i/` before the first edit (`diff -rq` against `28-P10-f` and `28-P10-f-post` on scripts shaders tests data scenes tools and `project.godot` empty, and the live tree equal to it: nothing unexpected, `i/diff-live-vs-28-P10-i.txt`); that pre-part tree is now `build/backup-feel/28-P10-i-pre/` (equal to `28-P10-f-post`), and `28-P10-i/` now holds the final P10i code, the post-commit snapshot (9f0f36a; equal to `28-P10-i-post-fix`) (checked with `diff -rq` 9 October after the stage close); post-part `build/backup-feel/28-P10-i-post/` (the code as first reviewed, 32 steps) and after the review fix `build/backup-feel/28-P10-i-post-fix/` (equal to the live tree; it differs from `28-P10-i-post` only in `shaders/water.gdshader` and `tests/test_water.gd`). Scratch trees `artifacts/beautify/28-P10/i/pre-tree` (equal to `28-P10-i-pre`), `i/coreoff-tree` (P10i code with `STREAK_HDR_CORE` false) and `i/fix/s32-tree` (28-P10-i-post, the 32-step code). Evidence `artifacts/beautify/28-P10/i/`; file backups `i/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, every bench run and the suite (`i/cap-power.txt`, whose set labels are wrong: the final sitting's rows 05:04-05:10 are pre-1, live-1, pre-2, live-2, coreoff-1, coreoff-2; `i/bench/power.txt`; `i/suite-power.txt`). 252: - **Tests:** test first on the unchanged shaders `i/testfirst/i-testfirst-test_water.log.out` (489 checks, 23 failures, all P10i). New checks: the march's signature and affine orthographic steps; strength 0.25-0.35, 24-32 steps from 3 px, growth, thickness, top, soft fade; the gate on `low_sun_direction.y` and the floor = sin(`SunPath.LIGHT_FLOOR`); no march with the switch at 0 or the sun down; body, sky, glitter and path darkened; for every quarter hour the shadow is 0 whenever `SunPath.sun` is at or below the horizon (and 0 at 01:00, 05:45, 22:00; whole at 13:00 and 19:30); a GDScript port of the march over an analytic depth buffer (a 4 x 4 x 5 m block in water, orthographic camera pitched 32 degrees, sun to the camera's right): at 16 degrees the water 3 m and 12 m from the block on the far side is shaded, 24 m away and on the sun's side lit; at 52 degrees 2 m shaded, 6 m lit; the streak core text, an amber streak's centre past 1.3 in two channels, saturation down by more than 0.2, hue kept, edges and ends exactly the old colour. Full suite on the final code (`i/suite/suite_all.out`, `i/suite.console`, `i/suite.sh`, `VERIFY_RES=2880x1800`, 05:10-05:19): **46 suites, 24,270 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (33 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and rerun 5,583/0 (known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). 253: - **Changed expectations (old -> new):** none changed; test_water 462 -> 491 checks (new checks only); suite total 24,241 -> 24,270. Review fix: the P10i bound on `SUN_SHADOW_GROWTH` `<= 1.12` -> `<= 1.14` (a new P10i check, not an old expectation); counts unchanged. 254: - **Review fix (9 October, 05:28-06:00):** the review found one must-fix: the acceptance line "p95 within +5%" was not met at Harbour by day (32 steps at x1.10: p95 +6.2% over 6 rounds, +7.0% and +6.0% in the earlier sets, p50 a steady +3.5%, +0.23 ms). Fixed by a cheaper march: `SUN_SHADOW_STEPS` 32 -> 24 and `SUN_SHADOW_GROWTH` 1.10 -> 1.14 in `shaders/water.gdshader` (reach about 600 -> 475 render px, about 46 -> 36 m of ray at Harbour; `SUN_SHADOW_TOP` 30 m, thickness, soft fade and strength unchanged), the test bound widened to growth `<= 1.14`; backup of the reviewed shader `i/fix/water-before-fix.gdshader`. test_water `i/fix/test_water.log.out`: 491 checks, 0 failures (the analytic block's 16-degree shadow still shades 3 m and 12 m out and not 24 m; the 52-degree one 2 m, not 6 m). Full suite on the fixed code (`i/fix/suite/suite_all.out`, `i/fix/suite.sh`, `VERIFY_RES=2880x1800`, 05:40-05:48): **46 suites, 24,270 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and rerun 5,583/0. Bench (`i/fix/bench6.sh`, the same recipe as `bench5.sh`, Harbour day 13:00 and golden 19:30, pre-tree against the live tree interleaved, warm-up round 0 discarded; `i/fix/bench6.out` rounds 16-21, `i/fix/bench7.out` rounds 22-27; `i/bench/summary.txt`; AC, 100%, Low Power Mode 0 in `i/bench/power.txt`): pooled 12 rounds **Harbour day p50 6.621 -> 6.822 ms (+3.0%, +0.20 ms), p95 8.110 -> 7.878 (-2.8%); Harbour golden p50 6.556 -> 6.710 (+2.4%, +0.15 ms), p95 7.652 -> 7.749 (+1.3%)**; by set, day p95 +1.1% and -3.2%, golden p95 +8.5% and -5.1% (golden's p95 swings both ways between sets: single live runs 7.2-9.3 ms, pre runs 7.1-8.2 ms). Draw calls 919 / 831 on both trees. Dense fixture and Harbour night not re-benched (the change only removes march steps; the march never runs at night). Look (same sitting 05:49-05:52, `i/fix/cap.sh`, `i/fix/capall.sh`: pre-1, s32-1, live-1, pre-2, live-2; harbour-golden, harbour-day, glade-day, quarter-golden, quarter-day, photo-golden; `i/fix/cap-power.txt` AC, Low Power Mode 0): shadow boxes (`i/fix/measure.js` -> `i/fix/measure.txt`), linear Y against the pre pair, 32-step -> 24-step: harbour-day right of the island -23.8% -> -23.9%, by the boats -23.3% -> -26.3%; harbour-golden behind the courtyard wall -19.3% -> -19.0%, behind the round tower -14.3% -> -14.3%; glade-day -17.5% -> -17.5%; open-water controls within 0.3%. Pixels moved over 8 luma (`i/fix/diff.sh`, `i/fix/diff/changed.txt`), s32 -> live against the two noise pairs: harbour-golden 3.82% (noise 1.85%/0.92%), harbour-day 1.06% (0.83%/0.58%), glade-day 0.58% (0.26%/0.26%), quarter-golden 1.95% (0.74%/0.31%), quarter-day 1.22% (0.67%/1.17%), photo-golden 0.09% (0.02%/0.03%). Looked at: `i/fix/diff/harbour-golden-s32-live.png` (scattered rain-ripple rings over all the water and smoke, no shadow band moved), `i/fix/diff/harbour-day-s32-live.png` (one-pixel lines along the island's cut edge and round the boats, smoke), `i/fix/diff/harbour-golden-pre-live.png` (the shadow bands left and right of the island still there), crops `i/fix/crop/hd-boats-{s32-1,live-1}.png` and `i/fix/crop/hg-right-{s32-1,live-1}.png` (the same shadows, no new banding or hatching). Profiles reached by the fix: Canal/Harbour and Glade as above (Glade unchanged in its box); night untouched. 255: - **Captures** (`i/cap.sh`, `i/capall.sh` -> `tools/art_capture.sh`, 1600x1000; final sitting 05:04-05:10, pre-1, live-1, pre-2, live-2, coreoff-1, coreoff-2; harbour-golden, -day, -night, quarter-golden, -day, -night, glade-day, lane-night, photo-golden, every view try 1; earlier sets from code tuned since are in `i/cap/old/`). Sheets, all looked at (pre-1 left, live-1 right): `i/sheet/p10i-1-shadow-pre-vs-live.png` (harbour-day right of the island: the tower, tree and island edge cast a shadow over the water to the right, away from the sun; harbour-golden left: a band on the water behind the courtyard wall; glade-day: the same as harbour-day), `i/sheet/p10i-2-shadow-quarter-photo.png` (quarter-golden: the bay right of the town and the canal by the left quay darken; photo-golden (perspective): the boat's shadow on the canal), `i/sheet/p10i-3-streak-core.png` (quarter-night canal: streak centres pale yellow-white and blooming, edges amber; harbour-night front water: the boathouse columns slightly paler). Diff maps, pre-1 -> live-1, over 8 luma (`i/diffmap.js`; magenta brighter, green darker; looked at): `i/look/d-harbour-day.png`, `d-harbour-golden.png`, `d-quarter-golden.png`, `d-glade-day.png` (glade-day: the only structured change is the shadow band right of the island and round the boathouse, the rest is smoke); maps from the earlier sittings in `i/look/old/`. Most of the golden island's shadow falls on water the island itself hides from the Harbour camera; what shows is the bands at the left and right ends. Draw calls and every integer count identical pre -> live in all 9 views (`i/metrics/counts.txt`). 256: - **Acceptance, measured** (`i/measure.js`, `i/metrics/measure.txt`; `i/streaks.js`, `i/metrics/streaks.txt`; frame changes `i/metrics/*.txt`): (1) **the island and town cast a shadow on the water away from the sun in harbour-golden and harbour-day: met.** Boxes in the shadow, pre pair -> live pair: harbour-day right of the island luma 90.0/90.0 -> 80.6/80.6 (linear Y -23.8%), by the boats 78.5/79.0 -> 69.9/69.7 (-23.2%); harbour-golden behind the courtyard wall 118.8/118.9 -> 108.2/108.2 (-19.4%), behind the round tower 108.2/108.2 -> 101.3/101.3 (-14.4%); open-water controls in every view within 0.1 luma (noise pairs 0.0-0.5 luma). Whole frame, pixels changed by more than 24 luma, noise pair -> pre/live pairs: harbour-day 0.52% -> 1.59%/1.64%, harbour-golden 0.44% -> 1.07%/1.42%. **Profiles reached:** Canal (Harbour as above; quarter-golden 0.42% -> 1.31%/1.43%, quarter-day 0.22% -> 1.53%/1.47%) and **Glade** (the shadow is not look-gated: glade-day right of the island 101.7 -> 92.6, linear Y -17.5%; frame 0.16% -> 1.10%/1.14%); photo-golden 0.01% -> 0.47%/0.47%. (2) **night unchanged within noise: met** for the shadow (it cannot run with the sun down, by the gate and its test). With the core off (`i/coreoff-tree`, the same sitting), pre-1 -> coreoff-2 over 8 luma: harbour-night 0.57%, quarter-night 0.31%, lane-night 1.80% against the noise pair 1.01%, 0.91%, 1.57%; mean dluma 0.0. The first core-off run (coreoff-1, the first run after the tree's shader cache changed) showed harbour-night at 6.2%, all in the TIME-driven clouds, neon-bay dashes, ripples and rain (`i/look/d-hn-pre1-coreoff1.png`; coreoff-2 shows only rain and smoke, `i/look/d-hn-pre1-coreoff2.png`; both looked at); coreoff-1 -> coreoff-2 is 5.97%, so it was that run's clock, not the code. With the core on, night changes only in the streaks (harbour-night >8 luma 1.26%/1.38% against noise 1.01%). (3) **streak centres pale and bloom: met.** Quarter-night canal (0,780)-(600,1000), pixels over luma 140: 6,492/6,456 -> 8,568/8,588, their mean saturation 0.529/0.531 -> 0.448/0.447, luma p99 212 -> 227/228; core off: 6,492, 0.530, 212 (the old streak). Harbour-night front water: 2,480/2,488 -> 2,612/2,628, saturation 0.442 -> 0.434. (4) **p95 within +5%: met after the review fix** (24 steps, x1.14; pooled 12 rounds: Harbour day -2.8%, golden +1.3%; dense fixture +0.9% and Harbour night -2.1% from the first code). With the reviewed 32-step code it was not met for Harbour by day (+6.2%): see Bench and Review fix. The p95 is two-moded (single runs 7.0-7.4 or 8.2-9.3 ms on both trees), so one 6-round set can still land over +5% (golden +8.5% in rounds 16-21). 257: - **Bench** (`i/bench.sh`, P10f's recipe with the hour per run: 13, 19.5, 1; interleaved pre/live, round 0 discarded; `i/bench*.out`, `i/benchsum.js`, `i/bench/summary.txt`): **final code** (affine steps; rounds 7-9 and 13-15, `bench3.sh`, `bench5.sh`): Harbour day p50 6.627 -> 6.858 ms (+3.5%, +0.23 ms), p95 7.569 -> 8.042 (+6.2%); Harbour golden p50 6.565 -> 6.736 (+2.6%, +0.17 ms), p95 7.758 -> 7.885 (+1.6%). **First code** (a matrix per step; renders the same; rounds 1-3 all scenes, 4-6 Harbour day and golden): dense by day p50 14.309 -> 14.462 (+1.1%), p95 16.243 -> 16.383 (+0.9%); Harbour night p50 6.718 -> 6.739 (+0.3%), p95 8.457 -> 8.277 (-2.1%); Harbour day over 6 rounds p95 +6.0%, golden +11.7%. The affine steps did not lower the p50 (+0.23 ms either way), so the cost is the 32 depth reads; a trial with 24 steps (growth 1.14, rounds 10-12) saved only about 0.02 ms p50 and is the code kept by the review fix. **Review fix code** (24 steps, x1.14; rounds 16-27): Harbour day p50 +3.0% (+0.20 ms), p95 -2.8%; golden p50 +2.4% (+0.15 ms), p95 +1.3% (details under Review fix). Draw calls 2,589 / 919 / 831 / 918 in every run. Where the cost lies: only water fragments by day and golden hour, about +0.2 ms p50 at Harbour; nothing at night. 258: - **Journey count deltas expected:** none (no node, mesh or draw call). Pixels change on the water next to the island by day (harbour-day, quarter-day, harbour-lane) and on the streak centres at night (harbour-night, quarter-night); styles-day only where it shows water beside a building. 259: - **Not done / open:** the p50 cost stays about +0.2 ms at Harbour by day (+3.0%) and +0.15 ms at golden hour; the p95 line is met on 12 pooled rounds but one 6-round golden set read +8.5% (two-moded p95). Dense fixture and Harbour night not re-benched after the fix. Glade night, lane-day, harbour-lane and styles-day not captured. The golden-hour island shadow is mostly hidden behind the island at the Harbour camera (no change asked). The wet-paving streaks take the core too (same shader, x wetness), not separately captured. Review: filled by the orchestrator. Platforms: Mac run; Windows not run. 261: **Beautify P10f Irregular paving slabs and a lighter mauve plaza (9 October 2026):** Opus 5.5 on this Mac (plan section "P10f"; report polish backlog "Irregular paving slabs ... with a lighter mauve `TERRACE`" and the surfaces lens finding F7; step `28-P10`, sub-batch P10-I, part f; a look change, approved as built under the owner's standing approval if it clears review; the owner A/B the plan asks for is on the sheet). The flagstones (every masonry paving top: terrain flagstone cells, canal roof decks and the quarter's plaza, stair treads) are laid as random ashlar by a new shared include `shaders/paving_grid.gdshaderinc` (`flag_cell`): courses in pairs 0.375 + 0.625 m deep, slabs 0.5-0.875 m long with a random start per course, 15% split near the middle; every joint on the eighth-metre grid of the puddle mask, and `puddle.gdshader`'s mask takes its flagstone id from the same `flag_cell`, so puddles still fill whole stones (setts keep their own grid). Joint and pixel footprint are metric per slab (`block_shade_px`, `bevel_joint_px`; `block_shade` and `bevel_joint` now call them with their own `fwidth`, so walls and setts draw as before). Per-stone spread 0.16 -> 0.10 (`PAVING_SPREAD`), warm/cool cast halved (`PAVING_CAST` 0.5), worn-edge ring kept. Palette: `TERRACE` 777b80 -> 83808a (+12.5% linear Y) and `GROUND_COLORS[0]` 858b8d -> 908b94 (+4.6%). Deviations from the report: courses 0.375/0.625 and lengths in eighths instead of 0.42/0.58 and 0.45-0.90 m (so the puddle grid holds); the quay walk takes 908b94, not 8b878f, which is 2.3% darker than 858b8d in linear light (the part asks for lighter). **Reverting:** pattern `PAVING_SLABS = false` in `shaders/paving_grid.gdshaderinc` (the old code is kept verbatim on that branch in both shaders); palette `Architecture.PAVING_MAUVE = false` (the tests follow both). Files: `shaders/paving_grid.gdshaderinc` (new), `shaders/masonry_body.gdshaderinc`, `shaders/puddle.gdshader`, `shaders/surface_detail.gdshaderinc`, `scripts/architecture.gd` (`PAVING_MAUVE`, `TERRACE`, `GROUND_COLORS[0]`), `tests/test_look.gd`, `tests/test_wet.gd`, `tests/test_building_tools.gd`, `tests/test_dusk.gd`, `tools/art_regions.json` (new `p10f_plaza`, `p10f_plaza_wide` in `_quarter`; nothing moved), `tools/repo_evidence.sh` (new fixture `artifacts/beautify/28-P10/f/geo-pre-mauve.json`). Snapshots: pre-part saved by me as `build/backup-feel/28-P10-f/` before the first edit (it was missing; `diff -rq` against `28-P10-b` and of `28-P10-b-post` against the live tree on scripts shaders tests data scenes tools and `project.godot` empty: nothing unexpected, `f/diff-live-vs-28-P10-f.txt`); that pre-part tree is now `build/backup-feel/28-P10-f-pre/` (equal to `28-P10-b-post`), and `28-P10-f/` now holds the final P10f code, the post-commit snapshot (adb5249; equal to `28-P10-f-post`) (checked with `diff -rq` 9 October after the stage close); post-part `build/backup-feel/28-P10-f-post/` (equal to the live tree at the time). Scratch trees `artifacts/beautify/28-P10/f/pre-tree` (equal to `28-P10-f-pre`) and `f/grey-tree` (P10f code with `PAVING_MAUVE` false). Evidence `artifacts/beautify/28-P10/f/`; file backups `f/backup/`. Power: AC, Low Power Mode 0 for every capture set, the bench and both suites (`f/cap-power.txt`, `f/bench/power.txt`, `f/suite-power.txt`, `f/suite2-power.txt`). 262: - **Tests:** test first on the unchanged code `f/testfirst/f-testfirst-test_look.log.out` (736 checks, 27 failures, all P10f) and `f/testfirst/f-testfirst-test_wet.log.out` (480, 5). New checks: the include, its constants and switch, the old grid as the switch-off branch, masonry and puddle both lay flagstones by `flag_cell`, `block_shade`/`bevel_joint` as wrappers; a GDScript port of `flag_cell` over a 20 x 20 m area: every eighth-metre puddle cell inside one stone, courses 0.375/0.625, at least 5 slab lengths, 12+ length/offset pairs in 40 courses, split share 10-20%; the palette by the switch (mauve, b > r > g, lifts in linear Y). Full suite on the final code (`f/suite2/suite_all.out`, `VERIFY_RES=2880x1800`, 03:50-04:00): **46 suites, 24,241 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and rerun 5,583/0 (known flaky check did not fire). The first full run (`f/suite/`) failed 12 pins in test_building_tools and 1 in test_dusk: geometry fingerprints that hash vertex colours, moved by `TERRACE`. With `PAVING_MAUVE` false (`f/grey-tree`, `f/greypins.sh`) both pass (930/0, 461/0), so the palette is the only difference; re-pinned as P10a did. 1440x900 `--verify` not run (inherited stall). 263: - **Changed expectations (old -> new):** test_building_tools P7f, P7f trim and P7g pins for canal12, canal4, canal7, cl_canal -> new `*_MAUVE` dicts (from the live code; `p10a_pins` picks them when `STONE_VIOLET` and `PAVING_MAUVE` are on); test_dusk P7b fingerprints `GEO_PRE_VIOLET` -> `GEO_PRE_MAUVE` (`f/geo_mauve.gd`, from the violet file). test_look 703 -> 736 checks, test_wet 475 -> 480. P10a's pinned text `block_shade(stone,vec2(0.028),0.16,...)` still holds (switch-off branch). 264: - **Captures** (`f/cap.sh`, `f/capclose.sh` -> `tools/art_capture.sh`, 1600x1000; one sitting 03:28-03:35, pre-1, live-1, grey-1, pre-2, live-2; quarter-night, -day, -golden, harbour-day, -night, -lane, styles-day, glade-day, every view try 1; close plaza views by day and night at rain 0.6, `c-pre-1`, `c-grey-1`, `c-live-1`). Sheets, all looked at: `f/sheet/p10f-1-quarter-pre-vs-live.png`, `p10f-2-close-pre-vs-live.png`, **owner A/B** `p10f-3-owner-ab-palette-close.png` and `p10f-4-owner-ab-palette-views.png` (slabs with the old greys against slabs with the mauve, quarter roof decks, Harbour, styles, Glade). The plaza and quay walk read as long rectangular slabs in courses, not a diamond checker; decks and the plaza are lighter and faintly mauve; a puddle's edges follow slab joints (`f/look/c-live-day-puddle.png`). Draw calls and every integer count identical pre -> live in all 8 views (`f/metrics/counts.txt`). Changed beyond noise: quarter views (paving), Harbour paving, steps and decks (`f/look/diff-harbour-day.png`); styles-day 1.07% of pixels over 8 luma against noise 1.33%. **Glade's canal roof decks change visibly** (the terrace and fish-market decks take the mauve `TERRACE`): glade-day whole frame 0.82% against noise 0.75%, but the frame hides a deck-only change; on two deck boxes (505,320)-(625,375) and (650,545)-(790,600) pre -> live changes 21.3% and 19.7% of pixels by more than 8 luma, mean +3.18 and +3.07 luma (pair 2: 21.3%/19.7%, +3.18/+2.81), against the noise pairs 0.0%/0.4%, mean about 0; the slabs alone (pre -> grey-1) change 4-5% with mean +0.2/-0.05, so the lift is the palette (`f/review/fix-decks.sh`, `f/review/fix-decks.txt`; map `f/review/d-glade-day-p1l1.png`; at full scale `f/sheet/p10f-5-owner-ab-glade-decks.png`, panels labelled on the sheet: A before P10f | B slabs with the old greys (`PAVING_MAUVE` false) | C live, slabs and mauve, 2x, with the open owner question and the deck numbers in its footer). 265: - **Acceptance, measured** (`f/measure.js`, `f/metrics/measure.txt`): (1) **no visible checker: met.** Plaza patch autocorrelation at the old grid's four lattice lags, quarter-night 0.539/0.518 (pre pair) -> 0.177/0.181 (live pair); day 0.73/0.66 -> 0.25/0.25; golden 0.63/0.63 -> 0.21/0.16. The top remaining peak is the 2-art-px lag (-4,0), 0.57 in pre too. (2) **puddles fill whole stones: met** by construction (test: every mask cell inside one stone) and by eye. (3) **paving/wall luma toward 1.6: moved toward it, not reached.** Quarter-night, against the P10a wall regions: plaza 0.86 -> 1.04 (luma 30.5 -> 36.8; linear Y 0.73 -> 0.97), quay walk 1.41 -> 1.50; day plaza 0.88 -> 1.10. Of the plaza's night lift, the pattern alone (fewer joints) gives 30.5 -> 32.1 (grey-1), the palette the rest. The report's "about 1.2" used other wall samples. 266: - **Bench** (`f/bench.sh`, P10b's recipe, interleaved, round 0 discarded, `f/bench.out`): medians of rounds 1-3, dense p50 14.376 -> 14.320 ms, p95 16.186 -> 16.183; Harbour p50 6.713 -> 6.715, p95 7.486 -> 8.333 (single runs 7.25-8.51 on both trees). Draw calls 2,588 and 918 in every run. Within noise. 267: - **Journey count deltas expected:** none (no node, mesh or draw call). Pixels change on every masonry paving top, canal roof deck and stair tread, Glade's roof decks included (about +3 luma on deck pixels); the rest of the Glade frame is within noise. 268: - **Not done / open:** the owner A/B on the palette (it recolours every canal roof deck; one switch); **Glade/Canal separation (a protected look) for the owner A/B:** Glade's canal roof decks (terrace, fish market) take the mauve `TERRACE`, about +3 luma and faintly mauve on deck pixels (`f/sheet/p10f-5-owner-ab-glade-decks.png`; sheet 4 shows Glade only at 1/3.1 scale, where the decks cannot be judged), and Glade paving takes the mauve `GROUND_COLORS[0]` too (its calm ground hides the slabs); both revert with `Architecture.PAVING_MAUVE = false`; this is an owner decision, still open (see the second fix pass); terrain paving keeps its per-cell lightness step by mark; no perspective or photo-mode capture. 269: - **Review fix pass (9 October, 04:08-):** the review found that "Glade within noise" rested on the whole-frame 0.82%/0.75% and hid a deck-only change (every Glade canal roof deck lightens with `TERRACE`). Wording fixed in the captures, journey and open lines above, the deck boxes re-measured on the same sitting's captures (`f/review/fix-decks.txt`) and a full-scale deck strip added for the owner A/B (`f/sheet/p10f-5-owner-ab-glade-decks.png`, looked at). No code changed (live tree equal to `28-P10-f-post`). Full suite re-run (`f/suite3/suite_all.out`, `f/suite3.sh`, `VERIFY_RES=2880x1800`, 04:09-04:18, AC, Low Power Mode 0, `f/suite3-power.txt`): 46 suites, 24,241 checks, 0 failures; `--verify` 2880x1800 exit 0 try 1 (33 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions rerun 5,583/0. 270: - **Second fix pass (9 October, 04:19-04:30):** the recheck left two items open. (1) Sheet 5 had no panel labels (the order was only in this entry): rebuilt by `f/review/deck-strip.js` with a title band, a label over each panel (A before P10f, old grid and `TERRACE` 777b80; B P10f slabs, 777b80, `PAVING_MAUVE` false; C live, slabs and 83808a) and a footer with the owner question and the deck numbers from `f/review/fix-decks.txt`; same captures and crop, 2056x796; the unlabelled sheet and script kept in `f/review/fix2/`; looked at. (2) Whether Glade's canal roof decks take the mauve `TERRACE` is the Glade/Canal separation question: an owner decision, **left open, not taken here** and not booked as a defect. The live code keeps the mauve on every canal roof deck, Glade's included; `Architecture.PAVING_MAUVE = false` reverts the palette everywhere (sheet 5 panel B); keeping the mauve in the Quarter but not in Glade would need a per-profile split, which is beyond this part's plan. No code changed (live tree equal to `28-P10-f-post` on scripts shaders tests data scenes tools and `project.godot`). Full suite re-run (`f/suite4/suite_all.out`, `f/suite4.sh`, `f/suite4.console`, `VERIFY_RES=2880x1800`, 04:20-04:29, AC, 100% charged, Low Power Mode 0, `f/suite4-power.txt`): 46 suites, 24,241 checks, 0 failures; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions rerun 5,583/0. Review: filled by the orchestrator. Platforms: Mac run; Windows not run. 272: **Beautify P10b Lawn tone bands and a turf octave that survives zoom (9 October 2026):** Opus 5.5 on this Mac (plan section "P10b"; report polish backlog "Lawn pixel tone bands" and the surfaces lens finding F4; step `28-P10`, sub-batch P10-I, part b; a look change, approved as built under the owner's standing approval if it clears review). On canal lawns the meadow's smooth drifts become three hard-edged tones (lush, middle, sunny: `lawn_tone`, at the bins' centres so the mean holds) and its dry straw three steps (none, half, whole: `lawn_step`), with the zone borders dithered by a world-anchored 0.25 m cell hash (`LAWN_STIPPLE`, amplitude 0.6 of a step, faded out between footprints 0.25 and 0.5); and a turf octave of 0.30 m clumps (`LAWN_TURF`, crisp crowns `LAWN_TURF_EDGE` 0.3-0.7, gap 0.86 / crown 1.06 divided by their mean 0.960) holds to footprint 0.14-0.40, where the 0.17/0.09 m clumps had already faded (they keep their 0.06-0.22 window). Albedo only, meadow branch only, behind a uniform branch on the new ground uniform `lawn_bands` (default 0 = the old lawn). Glade (`lawn_bands` 0), the bank branch (`bank_albedo`), earth and sand are untouched. One tuning change from the report: the crown edge `LAWN_TURF_EDGE` (round t3: tufts median 2 -> 3 px, one-pixel runs 34% -> 32%), and the drift rounds to bin centres rather than `floor(t*3+d)/3` (which left two tones on Harbour). **Reverting:** canal's `"lawn_bands": 1.0` -> 0.0 in `LookProfile.RENDER` (the old lawn exactly; the tests follow the value). Files: `shaders/ground.gdshader` (`lawn_bands`, `LAWN_*`, `lawn_tone`, `lawn_step`, meadow branch), `scripts/look_profile.gd` (`lawn_bands` key: canal 1, Glade 0), `scripts/world_view.gd` (`sync_profile` sets `lawn_bands` on grass, earth, sand and their variants), `tests/test_look.gd`, `tests/test_weathering.gd`, `tools/art_regions.json` (new `p10b_lawn_back`, `p10b_lawn_right`, `p10b_lawn_wide`, `p10b_lawn_left` in `_harbour`, `p10b_grove` in `grove-day`; nothing moved). No new fixture, so nothing for `tools/repo_evidence.sh`. Snapshots: pre-part saved by me as `build/backup-feel/28-P10-b/` before the first edit (it was missing; `diff -rq` against `28-P10-a` and against the live tree on scripts shaders tests data scenes tools and `project.godot` empty: nothing unexpected, `b/diff-live-vs-28-P10-b.txt`); that pre-part tree is now `build/backup-feel/28-P10-b-pre/` (equal to `28-P10-a-post-fix2`), and `28-P10-b/` now holds the final P10b code, the post-commit snapshot (a3c2948; equal to `28-P10-b-post`) (checked with `diff -rq` 9 October after the stage close); post-part `build/backup-feel/28-P10-b-post/` (equal to the live tree at the time). Scratch tree `artifacts/beautify/28-P10/b/pre-tree` (equal to `28-P10-b-pre`). Evidence `artifacts/beautify/28-P10/b/`; file backups `b/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, the bench and both suites (`b/cap-power.txt`, `b/bench/power.txt`, `b/suite-power.txt`, `b/suite2-power.txt`). 273: - **Tests:** test first on the unchanged code `b/testfirst/b-testfirst-test_look.log.out` (first draft of the checks: 700 checks, 23 failures) and the final test_look on a copy of the pre-part tree `b/testfirst/b-testfirst-final-test_look.log.out` (703 checks, 27 failures, all P10b; `b/testfirst-pre.sh`). After: test_look 703/0 (`b/suites/test_look.log.out`). New checks: the uniform and its default, every `LAWN_*` constant, `lawn_tone`/`lawn_step`, the branch inside the meadow between the drift and the mix, the dither, the octave after the old turf and before the blooms, the 0.17/0.09 clumps on their old fade, `lawn_bands` read nowhere outside the meadow (not the bank, earth or sand); values from GDScript ports: the drift takes 3 tones and keeps its mean (within 0.01), the straw 3 steps and none where it is 0, the octave holds the lawn's mean (within 1.5%), the fades at footprints 0.2/0.41; canal 1 or 0 (the switch), Glade 0; the editor's grass, earth and sand (and build variants) carry canal's value at start, after undo and by index, 0 in Glade. Full suite on the final code (`b/suite2/suite_all.out`, `VERIFY_RES=2880x1800`, 03:03-03:12): **46 suites, 24,203 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and its rerun 5,583/0, the known flaky check did not fire. The first full run (`b/suite/`, 02:54-03:03) failed one check in test_weathering (below; log kept as `b/testfirst/b-suite-test_weathering-before-exemption.log.out`). 1440x900 `--verify` not run (inherited stall). 274: - **Changed expectations (old -> new):** test_weathering P3d "no constant of its own carries the tide's 0.15 or 0.30, except `BANK_ROCK`" -> "except `BANK_ROCK` and `LAWN_TURF`" (the plan's 0.30 m turf clump, not a tide number); test_weathering 508 -> 513 checks (the new `LAWN_*` floats pass the guard); test_look 656 -> 703 checks. 275: - **Captures** (`b/cap.sh` -> `tools/art_capture.sh`, 1600x1000; final sitting 02:40-02:46 interleaved pre-5, live-5, pre-6, live-6, 13 views each, every view try 1; small-step orbit 02:47-02:49 o1-pre, o1-live, o2-pre, o2-live): harbour-day, -golden, -night, -lane, glade-day, grove-day, styles-day, quarter-day, quarter-night and an orbit by day at yaws 0.70/0.74/0.82/0.86 (and 0.77/0.79) around harbour-day's 0.78. Sets 1-4 and t1-t3 are earlier code (tuning; not used for the numbers below). Sheets, all looked at: `b/sheet/p10b-1-day.png` (Harbour lawns and the grove, before | after, colour and greyscale), `p10b-2-golden-night-glade.png`, `p10b-3-banks.png`, `p10b-4-orbit-strip.png` (the right-hand lawn patch at 6x in seven yaws, registered: pre-5, live-5, live-6). What they show: the lawn has three green tones with stepped, dithered borders and a crisp 2-4 px turf mottle instead of airbrushed drifts; at the grove's zoom the turf reads as a fine scale pattern over the bands; golden and night keep their value; Glade's lawns and every bank look the same. Draw calls and every integer count identical pre -> live in all 13 views of both pairs (`b/metrics/final2/counts.txt`). 276: - **Acceptance, measured** (`b/lawn.js`, `b/zones.js`, `b/crawl.js`, `b/maskdiff.js`; tables in `b/metrics/final2/` and `b/metrics/crawl*.txt`): (1) **3-4 discrete tone zones on harbour-day: met.** Three drift tones (by construction, and tested); along the linear-light lush-to-sunny axis the sunlit lawn's 2x2-art-px blocks go from 2 broad peaks (pre-5/-6) to 6 sharp ones in both live sets (the 3 tones, two of them also in half and whole straw); the same in orbit-74 and -82. (2) **tufts of at least 2 px: met for the octave.** Bright local runs on the harbour-day lawn (art grid, hp > +5 over a 9x9 mean): 19.8 -> 54.5 per 1000 lawn px, median 2 -> 3 px, runs of 2 px or more 56% -> 68% (one-pixel runs 44% -> 32%; they include rain drops and specks, which the pre lawn has too); golden 2 -> 3 px; grove 1 -> 9 px. (3) **no crawl over two orbit rounds: met.** Registered patches (texture correlation r and tuft IoU after the best whole-pixel shift; same-yaw noise pairs r 0.93-1.00: pre-5/pre-6 on the upper patch (1088 524 1152 564) in harbour-day 0.925, every other pair 0.99-1.00, `b/metrics/crawl.txt`): at +-0.01 yaw the live lawn keeps r 0.64-0.88, IoU 0.64-0.79 in both rounds, as well as the unchanged slate roof beside it (r 0.72-0.73, IoU 0.44-0.48) and close to the pre lawn's smooth drifts (r 0.57-0.91); at +-0.04 r 0.58-0.70, at +-0.08 0.42-0.57, falling smoothly with the angle (the rotation reshapes 2-3 px tufts). Round 2 repeats round 1 closely, not exactly: live-5 -> live-6 at +-0.04 0.598 and 0.579 equal on the upper patch, 0.627/0.695 -> 0.617/0.687 on the lower; at +-0.08 0.493 -> 0.479, 0.430 -> 0.418, 0.452 -> 0.445 (0.572 equal); the small-step orbit (`b/metrics/crawl-small.txt`) o1-live -> o2-live lawn-back 0.884 -> 0.828 and 0.852 -> 0.805, lawn-right 0.638 -> 0.673 and 0.710 = 0.710; only the unchanged roof reference repeats in r to the third decimal (0.733, 0.724). The strip shows the same tufts frame to frame. (4) **Glade and the P3 bank branch unchanged: met.** Glade lawn pixels (lawn mask, whole frame) pre -> live 0.20%/0.28% changed > 2 luma against its noise pairs 0.23%/0.29%, mean 92.09 -> 92.08. Bank region (`bank`, Harbour) day 78.2 -> 78.2, golden 68.6 -> 68.6, night 59.1 -> 59.5 / 59.1 (live's own pair moves it 59.5 -> 59.1), lane 74.2 -> 74.2; the changed pixels of the region's 2,304 (`b/metrics/final2/ab-5.txt`, `ab-6.txt`: day 20 in both pairs (0.868%, 6 of them by more than 8 luma), golden 28 (1.215%), night 38 in pair 5 (1.649%) and 6 in pair 6 (0.26%)) are the lawn edge row the region's top clips and their 1-2 luma ink neighbours. Value held (lawn mask, linear Y): day -0.25%, golden +0.8%, night +2.1% (58.84 -> 59.33 luma), grove -0.1%, styles +0.4%. 277: - **Bench** (`b/bench.sh`, the P10a recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, `--time=1`, 1920x1080, `--disable-vsync`; pre-tree against live, interleaved, round 0 discarded, 02:51-02:54; `b/bench.out`): medians of rounds 1-3, dense p50 14.356 -> 14.344 ms, p95 16.215 -> 16.246; Harbour p50 6.707 -> 6.720, p95 7.586 -> 7.331 (single runs 7.24-8.27). Draw calls 2,588 and 918 in every run. Within noise. 278: - **Journey count deltas expected:** none (no node, mesh or draw call; counts identical). Pixels change on every canal meadow (Harbour views, styles-day, the quarter's few lawn cells); Glade none. 279: - **Not done / open:** the octave's 0.30 m clumps at the grove's zoom read as a fine, slightly regular scale pattern (left for the sheet; `LAWN_TURF_FADE` or `LAWN_TURF_CROWN` are the constants); harbour-night lawn +2.1% in linear Y (within the night's value noise by eye, not tuned); no perspective or photo-mode capture; the quantised zones also apply at night and golden hour (by design, albedo only). Review: filled by the orchestrator. Platforms: Mac run; Windows not run. 281: **Beautify P10a Canal stone toward violet-grey, calmer upper storeys (9 October 2026):** Opus 5.5 on this Mac (plan section "P10a"; report polish backlog "Canal stone toward violet-grey (`STONES`) with calmer upper storeys (`wall_calm`)" and the surfaces lens finding F5; step `28-P10`, sub-batch P10-I, part a; a look change, approved as built under the owner's standing approval if it clears review). The canal stone moves from grey 858c92 with a tan 918a80 and a teal 76898b variant to violet-grey 888a9d, mauve-grey 918891 and blue-grey 7c8698 (the report's 85879a, 8d858d and 7a8496, lifted 1.3-2.5% per channel by the review fix below; `QUAY_FACE` follows `STONE.darkened(0.05)`: 7e858b → 818395); the per-block casts are tightened from warm (1.08,0.98,0.88) / cool (0.93,0.99,1.06) to (1.04,0.99,0.95) / (0.96,0.98,1.05); and from 2.5 m to 4.5 m over a wall's own ground (`weather_height`) a masonry wall gives up part of its texture by the look's new `wall_calm` (canal 0.6, Glade 0.8): per-block spread 0.22 → 0.13, odd blocks 12% → 5%, mortar floor 0.52 → 0.68, lip ring x0.6, the joint normaliser following the floor by the joints' share at each pixel (`joint_share`, from the second review fix; first `/(1-(1-floor)*0.1875)`; exactly the old /0.91 at calm 0). Paving, setts and anything under 2.5 m keep their full texture. **Reverting:** `Architecture.STONE_VIOLET = false` (palette and quay face; the pinned tests follow it), `BLOCK_WARM`/`BLOCK_COOL` in `masonry_body.gdshaderinc` back to the old casts written beside them, `wall_calm` 0 in `LookProfile.RENDER`. Files: `scripts/architecture.gd` (`STONE_VIOLET`, `STONE`, `STONES_VIOLET`/`STONES_OLD`/`STONES`, `QUAY_FACE`), `shaders/masonry_body.gdshaderinc` (`wall_calm`, `WALL_CALM_LOW/HIGH`, `BLOCK_WARM/COOL`, `JOINT_FLOOR`, `joint_share`, `block_shade(..., calm, ...)`), `scripts/look_profile.gd` (`wall_calm` key), `scripts/world_view.gd` (`sync_profile` sets `wall_calm` on `stone`, `stone_uv` and their variants), `tests/test_look.gd`, `tests/test_building_tools.gd`, `tests/test_meet.gd`, `tests/test_dusk.gd`, `tools/art_regions.json` (new `p10a_*` regions in `_quarter`, `_lane`, `_harbour`; nothing moved). Snapshots: pre-part saved by me as `build/backup-feel/28-P10-a/` before the first edit (it was missing; `diff -rq` against `28-P10-p` and against the live tree on scripts shaders tests data scenes tools and `project.godot` empty: nothing unexpected, `a/diff-live-vs-28-P10-a.txt`); that pre-part tree is now `build/backup-feel/28-P10-a-pre/` (equal to `28-P10-p`), and `28-P10-a/` now holds the final P10a code, the post-commit snapshot (1486774; equal to `28-P10-a-post-fix2`) (checked with `diff -rq` 9 October after the stage close); post-part `build/backup-feel/28-P10-a-post/` (the code as first reviewed). Scratch trees under `artifacts/beautify/28-P10/a/`: `pre-tree` (equal to `28-P10-a-pre`), `calm0-tree` (P10a code with `wall_calm` 0: the palette alone), `old-tree` (P10a code with `STONE_VIOLET` false). Evidence: `artifacts/beautify/28-P10/a/`; file backups in `a/backup/`. 282: - **Review fix (9 October 2026, 01:28-01:47, Opus 5.5 on this Mac; evidence `a/fix/`).** The review found two must-fix defects. (1) Acceptance line "same value within ±3%" was not met on the rendered stone: the report's hexes match the old stone only in sRGB luma, and in linear light (where the renderer works) they are 3.2-5.4% darker, so walls the calm does not reach rendered 4-7% darker (fish market −6.3%, quay −6.6 to −7.3%). I lifted the three constants to hold the old linear luminance: 85879a → **888a9d** (linear Y +0.11% against 858c92), 8d858d → **918891** (−0.32% against 918a80), 7a8496 → **7c8698** (−0.02% against 76898b); hue and saturation unchanged (234°/219°/300°, s 0.134/0.184/0.062); `QUAY_FACE` = `STONE.darkened(0.05)` = 818395 (`Color(0.5067, 0.5141, 0.5849)`). Still one switch, `STONE_VIOLET`. (2) test_meet and test_dusk load `artifacts/beautify/28-P10/a/stairs-pre-violet.json` and `geo-pre-violet.json`, which `tools/repo_evidence.sh` did not carry (a fresh checkout would fail both suites): both are now in its fixture list. Files: `scripts/architecture.gd`, `tools/repo_evidence.sh`, `tests/test_look.gd`, `tests/test_building_tools.gd`, the two JSON fixtures (regenerated by `a/stairs_violet.gd` and `a/geo_violet.gd`: the same 20 of 42 and 253 of 297 rows as before change, now to the lifted colours; old copies in `a/fix/backup/`). Tests first: test_look now also checks each constant's linear luminance within 1% of the old and that `repo_evidence.sh` lists both fixtures; on the unlifted code 648 checks, 10 failures (`a/fix/testfirst/test_look.log.out`); after 648/0. test_building_tools on the stale violet pins 930/15 (`a/fix/testfirst/test_building_tools.log.out`), the same 15 canal rows re-pinned from it → 930/0; test_meet 1397/0, test_dusk 461/0, test_masonry 501/0 (`a/fix/suites/`). Full suite (`a/fix/suite/suite_all.out`, `VERIFY_RES=2880x1800`, 01:32-01:41, AC 100%, Low Power Mode 0, `a/fix/suite-power.txt`): **46 suites, 24,143 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; creature rerun 5,583/0, the known flake did not fire. Captures (`a/fix/cap.sh`, same sitting 01:41-01:46, every view try 1, AC 100%, Low Power Mode 0): pre-3, live-3, pre-4, live-4, calm0-2 (`calm0-tree` given the lifted `architecture.gd`); metrics `a/fix/metrics/regions.txt`, `frame-change.txt`, `counts.txt` (draw calls and every integer count identical, pre-3 → live-3 and → calm0-2, all 8 views). Sheets (I looked at all four): `a/fix/sheet/p10a-fix-1-quarter.png`, `-2-house.png`, `-3-walls.png`, `-4-calm.png` (the magenta lines in mosaic crops are the kit's tile separators, not outlines); the hue reads as before, the stone a touch lighter. **Re-measured (pre-3 → live-3; noise pre-3/pre-4 ≤ 1.8% of the crop mean, quay up to 1.8% from the water):** the palette alone (`calm0-2`) now holds value on every crop, −0.5% to +0.5% (quarter-day walls 53.7 → 53.7, tower_lo 39.5 → 39.5, night walls 35.0 → 35.1, golden walls 67.2 → 67.0); uncalmed walls **met**: fish market 73.4 → 73.1 (−0.4%), quarter-day quay 61.5/60.4 → 60.5, quarter-night quay 33.1 → 33.1, golden quay 50.4/51.3 → 50.5, lane walls 84.8 → 85.1 (+0.4%). **But the calmed upper walls now read lighter than pre by more than 3%:** Harbour house face day 82.5 → 85.1 (+3.1%, 17% of it calmed), quarter-day walls 53.7 → 55.6 (+3.5%), tower_lo 39.5 → 42.5 (+7.6%), left_lo +3.8%; golden walls 67.2 → 69.6 (+3.6%), tower_lo +7.7%, stack_hi +3.5%; night walls 35.0 → 36.2 (+3.4%), tower_lo +7.1%, right_hi +4.6%. That lift is the calm alone (calm0-2 → live-3, the review's separate JOINT_SHARE finding: the normaliser assumes 18.75% joint area, more at distance where joints widen to the art-pixel floor); before the fix it was hidden by the palette's −5%. It was not in the must-fix list, so I left `JOINT_SHARE` alone; it is the open item for the value line. Texture calm unchanged in kind (sd/mean calm0-2 → live-3: quarter-day walls 0.706 → 0.642, tower_hi 0.568 → 0.490; fish market 0.624 → 0.624, quay 0.479 → 0.479). Hue unchanged: quarter-day walls 143° → 232°, lane 191° → 234°, quarter-night walls 228° → 236° (s 46% → 57%, not purple). Bench not re-run: three colour constants only, no node, mesh or draw-call change (counts identical). Snapshot after the fix: `build/backup-feel/28-P10-a-post-fix/` (equal to the live tree on scripts shaders tests data scenes tools and `project.godot`; `28-P10-a-post` is the reviewed code, not overwritten). The numbers in the bullets below are from the first implementation (sets pre-1/live-1/calm0-1) unless they say otherwise. 283: - **Second review fix (9 October 2026, 01:48-02:13, Opus 5.5 on this Mac; evidence `a/fix2/`).** The recheck found one must-fix: the calm by itself lifted the calmed upper walls above the pre-part tree by more than the 3% the value line allows (pre-3 → live-3: quarter-day tower_lo +7.6%, golden tower_lo +7.7%, night tower_lo +7.1%, walls +3.4-3.6%, Harbour-day house face +3.1%), because the normaliser took a fixed joint share (`JOINT_SHARE` 0.1875) while at play distance the joints widen to the art-pixel floor and cover a third to nine-tenths of a face. **Fix:** `masonry_body.gdshaderinc` measures the share per pixel: `joint_share(m, px)` integrates `bevel_joint`'s fill (smoothstep(m-aa, m+aa+0.035, edge), edge even on [0, 0.5], the two axes independent; about 0.18 up close, 0.50 at 10 cm pixels, 0.90 at 30 cm), and `block_shade` multiplies the old `/0.91` by `held = (1-(1-JOINT_FLOOR)*share)/(1-(1-joint_floor)*share)`, which is exactly 1 at calm 0 (the old wall untouched) and holds the calm wall's linear mean to the full wall's; `JOINT_SHARE` is gone, `JOINT_FLOOR = 0.52` named. No new uniform, node or draw call; the revert switch is still `wall_calm` 0. Files: `shaders/masonry_body.gdshaderinc`, `tests/test_look.gd` (backups `a/fix2/backup/`). **Tests first:** on the first-fix shader test_look 656 checks, 6 failures (`a/fix2/testfirst/test_look.log.out`: no fixed share left, `joint_share` present and following `bevel_joint`, `JOINT_FLOOR`, `held`, `/ 0.91 * held`); the 4 numeric checks passed there (a GDScript port of `joint_share` against the sampled fill within 0.002 for 1-30 cm pixels; the calm wall's sampled mean held within 0.2%; the old fixed share lifting it by more than 3% at distance, the defect; up close the share near 0.1875). After: test_look 656/0, test_masonry 501/0 (`a/fix2/suites/`). Full suite (`a/fix2/suite/suite_all.out`, `VERIFY_RES=2880x1800`, 02:00-02:09, AC 100%, Low Power Mode 0, `a/fix2/suite-power.txt`): **46 suites, 24,151 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; creature rerun 5,583/0, the known flake did not fire. **Captures** (`a/fix2/cap.sh`, one sitting 01:52-01:57, interleaved pre-5, live-5, pre-6, live-6, calm0-3, every view try 1, AC 100%, Low Power Mode 0, `a/fix2/cap-power-*.txt`; `calm0-tree` given the new shader); metrics `a/fix2/metrics/regions.txt`, `value-summary.txt` (`a/fix2/summary.js`), `oddblocks.txt` (`a/fix2/oddblocks.js`), `frame-change.txt`, `counts.txt` (draw calls and every integer count identical pre-5 → live-5 and → calm0-3, all 8 views). **Re-measured value, pre-5 → live-5 (noise pre-5/pre-6 0.0% on every tower, stack, left and quay crop except quarter-golden tower_hi −1.1% and left_lo +0.4%; ≤1.6% elsewhere):** walls quarter-day 53.8 → 54.2 (+0.7%), golden 67.4 → 67.7 (+0.4%), night 35.0 → 35.3 (+0.9%); Harbour-day house face 82.6 → 82.5 (−0.1%); left_lo +2.0/+2.0/+1.9%; right_hi +1.3/+0.5/+1.9%; stack_hi 0.0/+0.6/0.0%; tower_hi −1.0/−1.5/−1.5%; lane walls −0.2%, market −0.4%, quays +0.2/+0.2/0.0%; Glade house face −1.1%. **Still over 3%: tower_lo +3.5% (day), +3.7% (golden), +3.4% (night)**, down from +7.6/+7.7/+7.1%. Why (`oddblocks.txt`): the 34x40 px crop holds about ten bricks; half of its +3.6% is 3.5% of its pixels moving by more than 12 luma (one dark odd block going back to a plain block, its joint with it: the plan's odd blocks 12% → 5%), the rest the per-block spread pulling its few bricks toward the mean. The whole tower column (x 552-592, y 205-500, both tower crops and the wall between) reads calm0-3 → live-5 0.0% / +0.1% / −0.2% in the three lights, and tower_hi on the same tower goes −1.0 to −1.5%: a crop-sampling effect of which blocks the calm reshuffles, not a lift. Texture calm kept (sd/mean calm0-3 → live-5): quarter-day walls 0.706 → 0.653, tower_hi 0.568 → 0.500; night walls 0.567 → 0.510; lane parapet 0.444 → 0.395; market 0.624 → 0.624, quay 0.479 → 0.479. Hue unchanged (quarter-day walls 144° → 232°, night 228° → 236°). **Sheets** (I looked at all seven): `a/fix2/sheet/p10a-fix2-1-pre-live-1.png`/`-2.png` (pre-5 | live-5, the calmed crops), `p10a-fix2-2-calm-1.png`/`-2.png` (calm0-3 | live-5, the calm alone), `p10a-fix2-3-fix1-fix2-1.png`/`-2.png` (first fix live-3 | this fix live-5, different sittings: the upper joints a little darker again, the lift gone), `p10a-fix2-4-tower.png` (the round tower, pre-5 | calm0-3 | live-5 | live-3). **Bench** (`a/fix2/bench.sh`, the recipe above; pre-tree against live, interleaved, rounds 1-3, round 0 discarded; AC 100%, Low Power Mode 0, `a/fix2/bench/power.txt`, stdout `a/fix2/bench.out`): medians dense p50 14.287 → 14.338 ms, p95 16.219 → 16.217; Harbour p50 6.685 → 6.692, p95 8.201 → 8.491 (pre 7.33-8.46, live 7.40-8.49 single runs); draw calls 2,588 and 918 in every run. Within noise. Snapshot after this fix: `build/backup-feel/28-P10-a-post-fix2/` (equal to the live tree on scripts shaders tests data scenes tools and `project.godot`; earlier snapshots untouched). **Open for the orchestrator:** whether tower_lo's +3.4-3.7% (a ten-brick crop, block sampling, the tower as a whole at 0.0%) passes the ±3% line or the crop should be widened; holding each small crop exactly would need the odd-block thinning or the spread changed, a design change beyond this fix. 284: - **Tests** (`a/testfirst/`, `a/suites/`): test-first on the unchanged code `a/testfirst/a-testfirst-test_look.log.out`: 642 checks, 33 failures (the new `p10a_stone_checks` and `wall_calm_wired`). After: test_look 642/0. New checks: the three violet-grey constants and `STONE`, each within 3% of the old sRGB luma and violet-grey (blue the largest channel or tied with red, green never largest, saturation < 0.2; my first bound of 0.15 failed on the plan's own 7a8496, s 0.187, and was widened), `STONE_VIOLET` on, `QUAY_FACE` = `STONE.darkened(0.05)`; the shader's casts, the calm's height band, spread, odd-block threshold, mortar floor, lip and normaliser, paving and setts at calm 0; `wall_calm` canal 0.6 / Glade 0.8; the editor's materials (and build variants) at 0.6 / 0.8 at start, after choosing Glade, undo, redo and both index choices. Full suite on the final code (`a/suite2/suite_all.out`, `VERIFY_RES=2880x1800`, 01:03-01:11, AC 100%, Low Power Mode 0): **46 suites, 24,137 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`; `test_creature_reactions` rerun 5,583/0; the known flaky check did not fire. The first full run (`a/suite-first/`) failed 15 checks in test_building_tools, 1 in test_meet and (part b, `a/suite/`) 1 in test_dusk: geometry pins that hash vertex colours (below). 1440x900 `--verify` not run (inherited stall). 285: - **Changed expectations (old → new):** second review fix: test_look's fixed-share normaliser checks (`JOINT_SHARE = 0.1875`, `/ (1.0 - (1.0 - joint_floor)*JOINT_SHARE)`, 1-(1-0.52)*0.1875 = 0.91) → the per-pixel `joint_share`, `JOINT_FLOOR`, `held` and `/ 0.91 * held` checks plus 4 numeric ones; `mix(0.52,0.68,calm)` → `mix(JOINT_FLOOR,0.68,calm)`; test_look 648 → 656 checks. First review fix: test_look `P10A_NEW` 85879a/8d858d/7a8496 → 888a9d/918891/7c8698, plus 6 new checks (linear luminance within 1%, the two fixtures in `repo_evidence.sh`); `P7F_PRE_VIOLET`, `TRIM_REST_VIOLET`, `P7G_PRE_REST_VIOLET` canal rows and the two violet JSON fixtures re-pinned to the lifted colours. First implementation: the geometry pins that include vertex colours now pick a violet set when `Architecture.STONE_VIOLET` is true and the old set when false: test_building_tools `P7F_PRE`, `TRIM_REST`, `P7G_PRE_REST` → `*_VIOLET` copies (canal12, canal12a, canal4, canal7, cl_canal changed; the Victorian rows identical) via `p10a_pins`; test_meet Q1c stair fingerprints `artifacts/feel/06-B2q/c-probe/stairs-pre.json` → `artifacts/beautify/28-P10/a/stairs-pre-violet.json` (20 of 42 rows, written by `a/stairs_violet.gd` with the test's own fingerprint); test_dusk P7b `GEO_PRE` → `GEO_PRE_VIOLET` = `artifacts/beautify/28-P10/a/geo-pre-violet.json` (253 of 297 hashes, silhouette counts kept, `a/geo_violet.gd`). Proof that only colours changed: on `old-tree` (`STONE_VIOLET` false) the old pins pass, test_building_tools 930/0, test_meet 1397/0 with the old tests (`a/oldpins/`) and again with the new ones (`a/oldpins2/`), test_dusk 461/0 (`a/oldpins/`). The orchestrator must carry the two JSON files with the evidence (`tools/repo_evidence.sh`), as it does `stairs-pre.json`. 286: - **Captures** (`a/cap.sh` → `tools/art_capture.sh`, 1600x1000, same sitting 00:38-00:45, interleaved pre-1, live-1, pre-2, live-2, calm0-1, every view on try 1; AC 100%, Low Power Mode 0, `a/cap-power-*.txt`): quarter-night, -day, -golden, harbour-day, -night, -lane, glade-day, styles-day in `a/cap///`. Owner sheets (crops, before | after, colour then greyscale, no outlines): `a/sheet/p10a-1-quarter.png` (quarter day, night, golden), `p10a-2-house.png` (Harbour stone house: lane, day, night, Glade day), `p10a-3-walls-1.png`/`-2.png` (wall crops), `p10a-4-calm.png` (the calm alone: palette with `wall_calm` 0 | 0.6). Calm maps (`a/maps/calm-quarter-day.png`, `calm-harbour-lane.png`; magenta = lighter by > 8 luma): the calm acts only on upper walls (lighter joints, fewer odd blocks); street-level walls, quays and paving untouched. I looked at all four sheets, both calm maps, `maps/noise-live-quarter-night.png` and the whole frames. What they show: day walls go from tan, teal and green-grey to mauve and lavender-grey, the round tower from teal to blue; at night the walls are a slightly bluer indigo than before, not purple; at golden hour the stone reads rose-mauve instead of orange-tan; Glade's house goes from warm grey to a low-saturation mauve. Draw calls and every integer count identical in all 8 views (`a/metrics/counts.txt`). 287: - **Acceptance, measured** (`a/measure.js` → `a/metrics/regions.txt`/`.json`; `a/metrics/frame-change.txt`): (1) **hue violet-grey: met.** Building wall crops (`p10a_walls`, mean colour hue / saturation-weighted pixel hue): quarter-day 144° → 230°, Harbour lane 191° → 233° (the reference's walls are 228°, its tower 275°); per crop by day 219-273°. (2) **same value within ±3%** (first implementation; superseded by the review fix above, which lifted the values: now met on the palette and the uncalmed walls, not on the calmed upper walls): sRGB luma of the constants: 85879a −2.2%, 8d858d −2.4%, 7a8496 −1.6%. Rendered (mean luma of the crop, pre-1 → live-1): quarter walls day 53.8 → 52.8 (−1.9%), golden 67.2 → 66.2 (−1.5%), night 35.0 → 34.7 (−0.9%), Harbour house face day 82.5 → 82.2; but walls the calm does not lift lose 4-7%: lane walls 84.8 → 80.8 (−4.7%), fish market (street level) 73.5 → 68.9 (−6.3%), quarter quay 60.4 → 56.3 (−6.8%), palette alone (`calm0-1`) −4 to −6% everywhere. Reason: the report's hexes match in sRGB luma, but the renderer works in linear light, where they are 3.2-5.4% darker (linear Y 858c92 → 85879a 0.2582 → 0.2465, 918a80 → 8d858d 0.2576 → 0.2436, 76898b → 7a8496 0.2361 → 0.2284). I did not push the values (stop rule); a lift of about 2% per channel would hold the rendered value; owner call. (3) **upper storeys calmer: met.** Relative texture contrast (luma sd/mean, the report's measure) with the calm against the same palette without it (`calm0-1` → `live-1`): quarter-day tower top 0.579 → 0.498, middle stack upper 0.689 → 0.625, right block upper 0.589 → 0.492, wall over the arcade 0.543 → 0.486; lane parapet 0.448 → 0.389, front band 0.588 → 0.544; quarter-night tower top 0.460 → 0.397, stack 0.588 → 0.516. Street level unchanged: fish market 0.634 → 0.634 (0% of pixels changed), quay 0.494 → 0.490. Against pre-1 the five quarter-day wall crops go 0.704 → 0.653 together (the report: 0.85 by day, reference 0.50). (4) **quarter-night walls not purple: met.** Night wall crops hue 228° → 236° (each crop 226-243°, all blue-indigo; purple would be 260-320°), saturation 46% → 58% (the reference's night walls 47%); Harbour-night house (with its lit windows) 307° → 292°. Noise pairs (`frame-change.txt`, % of frame > 8): pre-1/pre-2 0.3-6.0%, live-1/live-2 0.3-11.7% (live quarter-night: clouds, water and the neon bay on TIME, walls identical, `maps/noise-live-quarter-night.png`); pre → live 2.4-29.5% (styles-day 5.9%, inside its own 6.0% noise). 288: - **Bench** (`a/bench.sh`, the P10p recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, `--time=1`, 1920x1080, `--disable-vsync`; pre-tree against live, interleaved, round 0 discarded; AC 100%, Low Power Mode 0, `a/bench/power.txt`). Three runs: run 1 rounds 1-3 (`a/bench/bench-run1-stdout.txt`: its run folders were overwritten by run 2, started by mistake when my `tee` target did not exist), run 2 rounds 1-3 (`bench-run2.out`, the folders on disk), run 3 rounds 4-6 (`bench-run3.out`). Medians of the nine rounds: dense p50 14.292 → 14.325 ms, p95 16.159 → 16.194; Harbour p50 6.679 → 6.689, p95 7.753 → 7.773 (single runs 7.17-8.58 pre, 7.44-9.53 live; in run 2 alone the live median p95 was +0.75 ms (7.753 → 8.506), in run 3 −0.75 (8.319 → 7.569; paired per round +0.02, +0.27, −1.01, median +0.02), `bench-run2.out`, `bench-run3.out`). Draw calls 2,588 and 918 in every run. Within noise. 289: - **Journey count deltas expected:** none (no node, mesh or draw call added; counts identical in all 8 captured views). Pixels change on every canal masonry wall, quay face and stone part in every view. 290: - **Not done / open:** the calmed upper walls' lift (3.4-7.7% after the first review fix) is fixed by the second review fix on every crop but tower_lo, which still reads +3.4-3.7% (block sampling in a ten-brick crop, the whole tower 0.0%; the orchestrator's call whether that passes); the tighter casts apply to every masonry surface (Victorian and medieval stone, paving, setts), not only canal walls, as the report's line 51 change does; `wall_calm` also calms the Victorian "stone" walls (it is a look key, not a style key); golden hour now reads rose-mauve on the stone (left for the sheet); the round tower is blue-grey (variant 2), bluer than the reference's 275° tower; no perspective or photo-mode capture; walls without a weathering pair read their own world height for the calm. Review: filled by the orchestrator. Platforms: Mac run; Windows not run. 292: **Beautify P10p Hover highlight, arrows only when selected, no boxes, double-click to move (8 October 2026):** Opus 5.5 on this Mac (plan section "P10p", owner request of 8 October 2026; step `28-P10`, part p; a measuring part, second review required). Hover in Select mode is now a soft lift of the part under the pointer (0.6 of the old selection lift, `WorldView.HOVER_LIFT`, fading in and out in about 0.07 s) with no wireframe; a selected part shows only its arrows (no box, no lift; the 25 September selection lift is superseded); no wireframe box is drawn round any object in any tool (placement shows only the translucent preview object, the wall-path and first-landing markers are a flat ring on the ground, parts hanging in the air breathe the lift instead of a wireframe); a double-click on a piece with the Select tool starts the Move command on it. Files: `scripts/world_view.gd` (`show_bounds`, `show_cage`, `fade_selection`, `selection_contains` removed; `show_hover`/`show_hover_parts`, `release_hover`, `fade_hover`, `_write_lift`, `lift_contains`, `hover_world_box`, `show_ground_ring`, `_levitation_box`, `_shimmer_shown`; `show_selection` records only; `show_candidate`, `show_levitation`, `tick_shimmer` rewritten), `scripts/main.gd` (`_update_hover`, `_outline_hover`, new `lift_boxes_of`, `_update_selection`, `_preview_body` rings, `_double_click_move`, the press handler and motion disarm, `--hover=` bench pin, `test_double_click_at`, the `--verify` A1b hover and A3a ghost checks, a P10p double-click check), `shaders/selection.gdshaderinc` (hover rows first, `(3,0).w` counts them, new global `selection_breath`; the nine including shaders unchanged), `project.godot` (`[shader_globals] selection_breath`), `scripts/shader_warmup.gd` (comment), `tests/test_owner_0924.gd`, `tests/test_ui.gd`, `tests/test_building_tools.gd`. Snapshots: pre-part `build/backup-feel/28-P10/` (saved by me before the first edit, the orchestrator's was missing; `diff -rq` against `27-P9-post` on scripts shaders tests data scenes tools and `project.godot` empty, so nothing unexpected had changed). There is no `28-P10-post` folder: the code as first reviewed is `build/backup-feel/28-P10-p-wip/` (its `scripts/main.gd` and `tests/test_owner_0924.gd` are byte-equal to `p/fix1/backup/`), and `build/backup-feel/28-P10-p/` holds the final P10p code (follow-up commit 3401cc9; `diff -rq` equal to `28-P10-a-pre`); the two differ only in `scripts/main.gd` and `tests/test_owner_0924.gd` (fix round 1 and the follow-ups; checked 9 October after the stage close). Fix round 1 on its own (eb7ea0b) has no snapshot. Pre-part scratch tree `artifacts/beautify/28-P10/p/pre-tree`. Evidence: `artifacts/beautify/28-P10/p/`; file backups in `p/backup/`. 293: - **Audit of every line mesh (`PRIMITIVE_LINES` and `LINE_STRIP`) in `scripts/` before → after:** 295: | Where (pre-part `world_view.gd`) | What it drew | After | 296: |---|---|---| 297: | `_build_grid` :913 `PRIMITIVE_LINES` | terrain-tool grid on the ground | kept (lines on the ground) | 298: | `show_bounds` :4333 `PRIMITIVE_LINES` | wireframe box: non-building placement ghost, wall-path cursor marker, connection first-landing marker (selected path already redirected to the lift) | **removed**; preview object only, markers are `show_ground_ring` (flat triangle annulus, `ArrayMesh`) | 299: | `show_cage` :4406 `PRIMITIVE_LINES` | building/tower placement cage | **removed**; preview object only | 300: | `_outline` :4839 `PRIMITIVE_LINES`, used by `show_hover` | hover wireframe box | **removed**; hover lift | 301: | `_outline` used by `show_levitation` | breathing wireframe boxes on floating parts (A3b) | **removed**; breathing lift (`selection_breath`, 0.10-0.45, at most `LEVITATION_MAX` 8 parts) | 302: | `_outline` used by `show_guide` | magnet guide line / cross on the ground or ridge | kept (not a box) | 303: | `show_composition` :4906 `PRIMITIVE_LINES` | picked group's cell edges on the ground | kept (not a box) | 304: | `show_wall_stroke` `LINE_STRIP` | control polyline of a wall being drawn, on the ground | kept | 305: | `show_terrain_brush` `LINE_STRIP` | brush circle on the ground | kept | 307: Other boxes looked at: `handle_mesh` falls back to a `BoxMesh` for an unknown grip shape (no caller uses one), `_detail_mesh` / `mesh_batch.gd` `BoxMesh` are brick and tile geometry, `hologram_button.gd` `draw_rect` is 2D UI. After the part `world_view.gd` holds exactly 3 `Mesh.PRIMITIVE_LINES` (a test reads the source). 308: - **Tests** (`p/suites/`, `p/testfirst/`): test-first logs on the unchanged code: `p-testfirst-test_owner_0924.log.out` (parse error, `HOVER_LIFT` missing), `p-testfirst-test_building_tools.log.out` (930 checks, 3 failures), `p-testfirst-test_ui.log.out` (390 checks, 2 failures, then script errors: `shimmer_breath`, `hover_world_box` missing). After: test_owner_0924 87 → 168 checks (145 at the first review, +23 in fix round 1), test_ui 397 → 399, test_building_tools 930 = 930, all 0 failures (`p/fix1/suites/`). New checks (`p10p_hover_checks`, `p10p_furniture_checks`, `p10p_double_click_checks`): hover lifts the part under the pointer and nothing 0.2 m outside or on ground beside, no line node visible, fade in under 0.2 s, texel (3,0).a counts hover rows, one texture kept, fade out over open ground leaves no rows; selected shows the grips and lifts nothing; hover over the selected part lifts it, over a grip stays off; clean capture and photo mode drop it at once and the capture puts it back; a freehand wall is lifted along its run and not 2 m beside; `show_bounds`/`show_cage` gone; placement of a building and 12+ other kinds shows a preview object and no lines; wall-path and bridge markers are a flat triangle ring (< 0.2 m tall); a floating bridge gets lift rows, breathes (0.10-0.45), no wireframe, dropped by a clean capture, stops when set down; double-click: first click selects, second starts the Move (Move button's line), the next click is one undo step, undo restores, Esc cancels, nothing on empty ground, on a grip, after a body drag, in photo mode, with a placing tool, and on a freehand wall the Move button's "Reshape a freehand wall…" line. Full suite on the final code after fix round 1 (`p/fix1/suite/suite_all.out`, `VERIFY_RES=2880x1800`, 23:37-23:46 on 8 October, AC 100%, Low Power Mode 0): **46 suites, 24,077 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (35 s), `PIXELPUNK_INTEGRATION []`; `test_creature_reactions` rerun 5,583/0. Before the fix (`p/suite2/`): 46 suites, 24,054 checks, 0 failures, `--verify` exit 0 (36 s). The first full run (`p/suite/`) found the `--verify` A3a check still reading the cage colour (`no cage colour to read for adapt-cage-roof.png`, `-wall.png`); fixed to read the preview object's colour and assert no cage, then the whole suite was rerun (`suite2`). Known flaky check: `test_creature_reactions` "The re-plan is applied after the edits" fired once in the first run (rerun 5,583/0); not in `suite2`. 1440x900 `--verify` not run (inherited stall; owner at the laptop). 309: - **Changed expectations (old → new):** test_owner_0924 `selection_lift_checks`: "The lift covers the building", "reaches the building's corner", "stops 3 cm outside", "No other building's middle is lifted", the texture row and bounds checks, "fades in over a few frames", "A clean capture has no lift", "A freehand wall is lifted along all of its line", "Nothing a metre beside the wall is lifted" → selecting lifts nothing, shows ≥ 8 grips, no rows, the recorded box still holds the part, a selected freehand wall is not lifted (the lift checks moved to the hover in `p10p_hover_checks`); globals list + `selection_breath`. test_building_tools: "The selection is rewritten into the texture it already has" (via `show_bounds`) → the hover lift is rewritten into the same texture, no mesh; "A part under the pointer shows a hover outline" / "The hover outline redraws into the mesh it already has" → a hover lift with no mesh, rewritten in place (one row). test_ui: A3a review cage colour read from `ghost.mesh` → `show_candidate` draws no cage, colour read from the preview object; A3b "the shimmer moves" reads `shimmer_breath` (was the wireframe's alpha; still 0-0.6); B2c hover box read from `hover_world_box()` (was the outline mesh) and an unslid detail's box is `bounds.grow(0.03)` (was 0.05); Q1a selection box checks unchanged (now a record, nothing lifted). `--verify`: A1b hover checks read the lift (rows, no mesh, waits for the fade-out); A3a ghost colour read from the preview object. Grip checks in test_owner_0924 that read `selection.visible` unchanged (`selection` stays the editor's switch, draws nothing). Fix round 1: the P10p "not on a grip" check now fails when the house has no y grip (it was skipped silently); new `p10p_double_click_face_checks` (window-sized 1600x1000 root, so the 3D view is 800x500): a double-click from nothing selected on the middle of each wall and the roof of the 4 x 4 x 5.2 m house starts the Move (5 points, 4+ carry an arrow once selected), Esc leaves it in place, and with the house already selected a double-click on the arrow over a wall middle does not move it. 310: - **Captures, first set** (superseded for the numbers by fix round 1 below; `p/sheet_capture.gd`, run by `p/sheet.sh` on both trees with `--script`, 1600x1000, rain 0, clock off, the editor's own test pointer through `screen_point`, interface hidden; same sitting, interleaved pre-1, live-1, pre-2, live-2, every run on try 1; power AC 100%, Low Power Mode 0, `p/sheet-power-*.txt`): hover over a building (Harbour 3), a tree (88), a lamp (lantern 74), a freehand wall (16) and a pipe (quarter 173), day and night, each with a no-hover frame of the same view; selected (building 3) with the pointer away and on it; placement of a building, a planter, the wall-path marker and a bridge's first landing; a floating bridge. `lines.json` per run: before, every hover, placement and floating frame has 1 visible line node; after, 0 in all 27 frames. Picking: the pointer picked the wanted part in all 10 hover frames in both trees (same ids). Owner sheets (crops, before | after): `p/owner/p10p-1-hover-day.png`, `-2-hover-night.png`, `-3-selected.png`, `-4-placement-floating.png`; lift maps (no-hover → hover, magenta brighter) in `p/sheet/live-1/maps/`. I looked at sheets 1, 3 and 4, the maps of the building, tree, lamp and wall, and `suite2/integration/adapt-cage-roof.png` (translucent amber ghost, no cage). What they show: before, a pale wireframe round the hovered part, a lifted selected part, a cage or box round placements and markers, a wireframe round the floating bridge; after, the hovered part lifted inside its box (building p50 luma of the crop 51.8 → 58.6), no lines anywhere, the selected part at its own colour with only its arrows, placement as the preview object, rings on the ground, the floating bridge lifted. The lift is box-based as the selection lift was: anything drawn inside the hovered part's box lifts with it (the bench under tree 88, a sliver of roof behind building 3). 311: - **Acceptance, measured:** (1) no wireframe round any object in any tool or state: 0 line nodes in all 27 after frames against 1 in every hover/placement/floating before frame; audit table above; target met. Numbers from the fix round 1 set b captures (`p/fix1/sheetb/`, see below). (2) hover is a soft lift only: within-tree no-hover → hover changed > 8 (% of frame / brightened / mean Δluma of changed) building day 11.45-11.48 / 10.2-10.3 / +14.8 to +15.2, night 10.9-11.2; tree 3.6-3.8 day, 4.3-5.4 night; lamp 1.1 day, 1.5-1.9 night; wall 2.9-3.1 day, 3.2-4.0 night; pipe 1.35-1.40 day, 1.38-1.44 night (live-1, -2, -3; `p/fix1/metrics-b/summary.txt`); target met. (3) selected shows only arrows: `day-selected` pre-k → live-k 9.27 / 9.49 / 9.26% against within-tree noise 1.66-2.36% (the old lift gone); headless: no rows, ≥ 8 grips; target met. (4) double-click moves a piece: headless checks above, now including face middles at a window size, and `--verify` (P10p check: Move started by a double-click at window resolution, next click one undo step); target met. (5) unchanged with nothing hovered: `p/fix1/metrics/none_excl_b.txt` and `none9b.txt`, three frames per tree, % changed > 8. All 60 pairs pre-j → live-1 and pre-j → live-2 over the ten no-hover views sit inside the pre tree's own noise (each view's largest such pair against its largest pre-pre pair: day building 1.82/1.94, tree 0.41/0.57, lamp 0.08/0.11, wall 1.41/1.43, pipe 0.08/0.09; night building 0.83/0.89, tree 1.32/1.49, lamp 0.19/0.20, wall 1.06/1.32, pipe 0.07/0.11). live-3 is above the pre noise on 9 of 10 views (up to 2.09 against 0.89 on night building), but it differs from live-1 and live-2, the same code, by as much (night building 2.05/2.03): looked at (`p/fix1/look/b-nb-live1-live3-s.png`), the changed pixels are cloud edges, chimney smoke, water glints and stars, all on the shader's TIME, which no hold controls. Verdict: within noise; the 9-pair medians are at or below the 6-pair noise medians on 9 of 10 views (night tree 1.23 against 1.02, inside its noise range 0.19-1.49). 312: - **Bench, first run** (its hover column is void, see fix round 1; `p/bench.sh`, the P9 close recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, `--time=1`, 1920x1080, `--disable-vsync` before `--path`; pre-tree, live with nothing hovered, live with `--hover=` (dense building 265, Harbour building 3; the pin's own log line checked in `bench/pincheck/run.out`), interleaved, round 0 discarded, medians of rounds 1-3; `bench/bench-a.out`, `-b.out`, `power.txt`: AC, 100%, Low Power Mode 0 throughout). Dense p50 pre 15.634 → live 15.594 → hover 15.637 ms, p95 17.352 → 17.156 → 17.344; Harbour p50 8.027 → 8.116 → 8.189, p95 9.603 → 9.542 → 9.561 (single runs spread 7.86-8.34). Draw calls 2,588 and 918 in all three. Nothing hovered: within noise (target met); the "hover" column (+0.04 dense, +0.07 Harbour) is void: none of its eight runs printed the pin's line (only `bench/pincheck`, run after the bench, did), so it measured nothing hovered. Floating parts were not benched separately (no fixture flag); they use the same rows and shader path as the hover plus one global per frame. 313: - **Second review and fix round 1 (8-9 October 2026):** the second review found four must-fix items, all handled. (a) A double-click on the middle of a small building's wall or roof began a resize, not the Move: the pair's first click selected the part, its arrows appeared under the pointer and the second press found one (reviewer probe `p/review2/grip_share.gd`: 0 of 5 face middles moved). Fix in `scripts/main.gd` `_double_click_move` (no argument now): "on a grip" is decided by the pair's first press, which only arms the pair when it found no grip, so a grip the first click's own selection put there no longer counts; a grip that was under the first press still disarms it and resizes. Test first: `p/fix1/testfirst/test_owner_0924.log.out` (168 checks, 5 failures, the five face middles), after `p/fix1/suites/` 168/0; the reviewer's probe rerun on the fix (`p/fix1/grip_share.log.out`) moves at all 5 face middles and the interior point (the probe itself still ends on its 170 s timeout after the Harbour samples, as in the review). (b, c) The first bench's hover column was unpinned (above). Rerun on the final code (`p/fix1/bench.sh`, same recipe, every run's pin line counted: pin=1 on all hover runs, pin=0 on all others; `p/fix1/bench/bench-a.out`, `-b.out`, `-c.out`, `power.txt`: AC, 100%, Low Power Mode 0 throughout, load 3.7-4.9). Medians of rounds 1-3: dense p50 pre 15.672 → live 15.589 → hover 15.596 ms, p95 17.280 → 17.346 → 17.193; Harbour p50 8.105 → 8.287 → 8.164, p95 9.599 → 9.657 → 9.586. Draw calls 2,588 and 918 in every run. Harbour live sat 0.08-0.31 ms above pre in rounds 1-3, so four more Harbour rounds were run (`bench-c.out`, rounds 4-7): over rounds 1-7 the medians are pre 8.105, live 8.188 (+0.08), hover 8.164; single runs pre 7.88-8.25, live 6.84-8.34 (round 7 fell to 6.8 for live and hover together). Nothing hovered: dense −0.08, Harbour +0.08 over seven rounds, inside the run spread; hovering one building costs nothing measurable against live (dense +0.01, Harbour −0.02 to −0.12). The reviewer's own pinned rerun (`p/review/second/bench/`) agrees (dense live 15.601 → hover 15.648, Harbour 8.102 → 8.113). (d) Acceptance (5)'s quoted numbers could not be reproduced from the PNGs on disk, which probably came from an earlier capture set that a later run overwrote, and the night building frame showed a window lamp lit in every live frame and no pre one. The harness ran the light and creature clocks free (it starts the editor with `--script`, which no automation mode holds). `p/fix1/sheet_capture.gd` now sets `light_hold` and `creature_hold` right after `main._ready` (set b, `p/fix1/sheetb/`, pre-1, live-1, pre-2, live-2, pre-3, live-3 interleaved, every run on try 1, 42 s each, AC 100%, Low Power Mode 0, `p/fix1/sheet-power-000606.txt`, `-000816.txt`); a first try that set them only after 240 start-up frames is kept as set a (`p/fix1/sheet/`, `sheet_capture_a.gd`, `metrics/`). With the clocks held the night window lamp is lit in all six frames of both trees (cell (2,2) 0.00-0.01% in every pair, `p/fix1/look/b3-window-night-building-none.png`). One small day-time glow in building 3's tower window (about 100 pixels) still varies from run to run: set b live 2 of 3, pre 0 of 3; set a live 2 of 3, pre 0 of 3; the first set pre 2 of 2, live 1 of 2 (`p/fix1/look/b3-window-sheetb.png`, `b3-window-6.png`, `b3-window-orig.png`). It shows in both trees, no lift row is live in those frames (`lines.json`: hover, selection and shimmer off) and with no rows the lift is 0, so it is not the lift; I did not find what drives it. Line nodes in set b: pre 15 of 27 frames per run, live 0 of 27; the pointer picked the wanted part in 10 of 10 hover frames in all six runs. The owner sheets (`p/owner/`) are still from the first set. Fix round 1 changed only `scripts/main.gd` and `tests/test_owner_0924.gd` against the reviewed code, `build/backup-feel/28-P10-p-wip/` (left as it was; the files before the fix are in `p/fix1/backup/`, equal to that snapshot's). 314: - **Follow-up fixes (9 October 2026):** five small review items, Opus 5.5 on this Mac, test-first (`scripts/main.gd`, `tests/test_owner_0924.gd`; backups in `p/fix2/backup/`). (1) Double-click to move is the Select (arrow) tool's only: new `_arrow_tool()` (not placing and not the Scenery tool; a part selected for editing counts, since `select_id` makes its kind the tool with placing off) is required when the pair's first press arms and again in `_double_click_move`, so the Scenery tool's selecting click and any placing tool never start a Move. (2) The pointer moving onto the interface releases the hover lift through `world.release_hover()` (it fades out as over open ground) from `_process`, where `update_preview` is skipped over the interface; new `_release_hover()` also forgets the last hover answer so a pointer back on the same pixel lifts the part again. (3) `_capture_furniture` records a hover that is fading out as gone, so a clean capture taken in the ~72 ms fade-out no longer brings the lift back on the part the pointer left. (4) On a freehand wall `_double_click_move` still says the Move button's line but returns `moving_id != 0` (false there), so the second press goes on to take hold of the wall's body like any press; its release, a click, re-selects the wall and shows the wall's own selected line ("Drag the yellow nodes to reshape ..."), which replaces the reshape toast (the changed expectation: the toast is read at the second press, the selected line after the release). (5) The "not after a drag" check moved to the new window-sized (1600x1000) `p10p_follow_up_checks`: in the 32x32 view of `p10p_double_click_checks` the 6 m drag was about 4 px, under `BODY_DRAG_SLOP`, so it never committed or disarmed. It now drags well past the slop and back to the press point (on the selected house, clear of grips) before the flagged press; with the motion disarm removed as a probe it fails (`p/fix2/testfirst/owner_nodisarm.log.out`, moving 1). Test-first on the unchanged code: 6 of the new checks fail, items 1-4 (`p/fix2/testfirst/owner_before2.log.out`); after: test_owner_0924 168 → **187 checks, 0 failures** (`owner_after.log.out`). Full suite (`tools/suite_all.sh` a, b, v with `VERIFY_RES=2880x1800`, `p/fix2/suite/suite_all.out`, on AC): **46 suites, 24,096 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`; `test_creature_reactions` rerun 5,583/0. No captures or bench rerun (input and capture bookkeeping only; nothing drawn changes while the pointer is on a part). 315: - **Journey count deltas expected:** none (journey and art captures are clean captures with no hover, selection or placement; no node added; draw calls equal in the bench). 316: - **Not done / open:** the day-time tower-window glow in building 3 that varies run to run in both trees (fix round 1, item d) has no identified cause; owner sheets not rebuilt from set b; floating-part cost not measured on its own (reasoned, not measured, against the 0.1 ms line); no perspective or photo-mode capture of the hover; the real OS double-click (macOS click count) not exercised by hand, only events with the engine's `double_click` flag; the old pick quirk stays: Harbour blossom 93 stands inside the tea house roof unseen, and its pick box above the roof takes the pointer before the roof (hovering that roof lifts the hidden tree's box), as the old outline did; the first two windowed sheet runs before any captures hung until my 240 s limit and were killed by my own runner (pid kill, no pkill), cause unknown, every run after was on try 1; `--hover` prints one line per pinned run. Review: filled by the orchestrator (second review required). Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy = HEAD 8418d4e, 9 October 2026, 19:36 BST), lines 330-374 330: ## New part P10p, owner request (8 October 2026, in chat) 332: "lets add another part to the polish, I am still seeing boxes around objects. An object should just be highlighted when hovered over, that is it. the geometric boxes should never be seen" and "even when selected it should just have the movement arrows. each piece should be able to double click to be able to move it when using the arrow tool". Added as P10p in docs/beautify-plan.md and run now (the owner asked for it; P10 is otherwise optional). Read by the orchestrator as: hover = a soft highlight (lift), no wireframe; selected = the movement arrows only (this supersedes the 25 September selection lift); no geometric box drawn around any object in any tool or state (hover, placement, cages, floating-part shimmer, ground markers); double-click with the Select tool starts the Move command on that piece. Defaults: hover strength 0.6 of the old selection lift; floating parts breathe the lift (or the cue is dropped and reported). 334: ## "run the rest of p10 after p10p" (owner, 8 October 2026, in chat) 336: Read by the orchestrator as: run every remaining P10 part (a, b, e, f, g, h, i, j, k, m, n, o) one workflow per part after P10p, with the plan defaults for unanswered decisions, in three sub-batches each closed with the usual sheet: P10-I (p, a, b, f, i, h; pre-tree `28-P10`), P10-II (e, g, m, k; pre `28-P10-I`), P10-III (j, n, o; pre `28-P10-II`). Limits that the plan or the standing rules put on it, which the instruction does not lift: (1) P10g stops at a mock and an owner sheet; the default Harbour scene is not changed until the owner says yes; (2) P10k builds each tuning behind a switch that is off and stops at its mock (it reverses "daylight is independent of rain"); (3) P10o makes no font download and no UI-scale or font change: the owner's explicit permission, filename, source and size are needed first, so it produces the options sheet and asks; (4) P10j builds the guard only if the roof-speck measurement is still over target. No push, no journal deploy, no new export without a request. D32 default ("P10 not run") is overridden by this instruction. 338: ## Standing approval until morning, and a clean-up (owner, 8 October 2026, in chat) 340: "approve all until morning, also lets clean up the folders again". Same scope as the earlier standing approvals: plan approvals and owner-sheet questions are taken as approved as built, with open items left as built and listed for the sheet, so the pipeline does not wait. It does NOT cover: pushing to GitHub, deploying the journal, a new export, or any download (so P10o still makes no font download: it needs a named file, source and size). Read for the P10 parts: look changes (P10a, P10b, P10f, P10i, P10h, P10e, P10m, P10n) are taken as approved as built when they clear review; P10g and P10k still build their changes OFF by default and stop at the mock, because turning on a new Harbour default or reversing "daylight is independent of rain" overnight is hard to review and is one switch to flip in the morning. The "morning" limit is not stated; the orchestrator launches nothing it does not expect to finish by 07:30 on 9 October and leaves no Godot running past 07:30. Clean-up: the frame folders (`frames/`) of earlier recordings are moved to the Trash again (not P10 scratch, not the running part). 342: Clean-up done (8 October 2026, evening): 8 PNG frame folders of the P8f recordings (1.6 GB) to `the macOS Trash/beautify-frames-0708b/`, 78 leftover project-copy trees (`pre-tree*`, `*-tree`, `diag-*`, 1.8 GB; each is reproducible with `tools/feel_snapshot.sh tree `) to `the macOS Trash/beautify-trees-0708/`, and the M6 staging copy (107 MB exe, duplicate of the zip) to `the macOS Trash/m6-stage-0708`. Kept: every clip, still, metric, log, sheet and note, everything under `artifacts/beautify/28-P10/`, and all snapshots under `build/backup-feel/`. The Trash is not emptied. `artifacts/` is 71 GB; what is left is evidence the handoff cites (the P0 to P9 stills, clips and metrics), so I did not touch it. 344: ## Raw captures of P0 to P7 trimmed to the Trash (owner, 8 October 2026, in chat) 346: "trim P0 to P7 raw captures to the Trash". 28,731 PNG files (29.6 GB) under `artifacts/beautify/18-P0` to `25-P7` were moved to `the macOS Trash/beautify-raw-P0-P7-0708/` with their relative paths kept (the list is `MOVED-FILES.txt` there), so putting them back is a move of that folder's contents over `artifacts/beautify/`. Kept: every PNG under `sheet*`, `look` and `crop*` paths (the pages sent to the owner and the 1:1 crops), all clips, metrics, logs, JSON and notes, and everything for P8 onward. Handoff entries for P0 to P7 that cite a raw capture path (`before-1/...`, `after-1/...`, `maps/...`) no longer resolve unless the Trash folder is restored; their numbers still stand on the metric files. `artifacts/` went from 71 GB to 41 GB. The Trash is not emptied. 348: Mistake found after the trim (8 October 2026): my commit 03a5ffb ("P9: D31 and M6 export recorded", local, not pushed) force-added `artifacts/beautify/27-P9/m6/` including its `stage/PixelPunk/PixelPunk.exe` (107 MB, a duplicate of the Windows zip). GitHub refuses files over 100 MB, so a later push of main would fail. The staging folder is now in the Trash, so git shows those four files as deleted in the working tree. Fixing it means either rewriting the unpushed commits after 03a5ffb to drop the exe, or a plain commit that untracks it (the blob would stay in history and block a push). Both were refused by the auto-mode git guard when I tried, so nothing was changed; the owner decides (and may allow a history rewrite of the unpushed commits, or squash them). Also noted: 30 tracked P0 baseline pictures that the repo keeps were moved by the trim and put back at once; the P0 to P7 trim therefore moved 28,701 untracked files, not 28,731. 350: History rewrite done (owner chose option 1, 8 October 2026, in chat): the 9 unpushed commits from 03a5ffb to ebd4126 were rewritten in a temporary clone (`git filter-branch --index-filter`, dropping `artifacts/beautify/27-P9/m6/stage`) and `main` was moved to the result with `git update-ref` (old value checked), so the working tree and the running part's uncommitted edits were not touched. Checked: the old and new trees differ only by the four staging files; the largest blob reachable from `main` is now 6.5 MB. The old history is kept on the branch `backup-before-m6-rewrite-0908` (it still holds the 107 MB exe: delete that branch, then run `git gc --prune=now`, before pushing if the exe should leave this machine's repo; the branch is not pushed unless asked). 352: ## P10f closed by the orchestrator (9 October 2026, 04:40) 353: - Committed adb5249 (local), snapshot 28-P10-f. Under the standing approval until morning, the look is taken as built. The owner A/B stays open: the mauve plaza, and whether Glade's canal roof decks (terrace, fish market; about +3 luma) should take it too. One switch reverts it all: `Architecture.PAVING_MAUVE = false`. 354: - P10i launched 04:41 (). P10h and the P10-I close wait for the morning unless time allows. 356: ## P10i closed by the orchestrator (9 October 2026, 05:57) 357: - Committed 9f0f36a (local), snapshot 28-P10-i. Look taken as built under the standing approval. Revert: `WATER_SUN_SHADOW` 0 and `STREAK_HDR_CORE` false. The review's must-fix (Harbour-day p95 over +5%) was fixed by cutting the shadow march from 32 to 24 steps; the shadow reaches about 36 m at Harbour. 358: - Open for the owner: Glade's water shadow is crisp next to blurred land shadows; the 20:45 view darkens on the far bay. 359: - P10h launched 05:58 (), finish limit 07:15. The P10-I close waits for the morning. 361: ## Laptop time lifted (9 October 2026, ~06:00) 362: - Owner: "I don't need the laptop today so your good to continue." The 07:30 limit no longer applies; P10h's own 07:15 limit stands for that run. Continue with the P10-I close, then P10-II (e, g, m, k) and P10-III (j, n, o). Still one Godot at a time. The earlier scope stays: no push, no journal deploy, no new export, no download (P10o), P10g and P10k stop at the mock. 364: ## Owner rulings on the P10f and P10i options (9 October 2026, ~06:10) 365: - "Keep all new." P10f stays as built: slabs and mauve plaza, and Glade's canal roof decks take the mauve (PAVING_MAUVE true). P10i stays as built: water sun shadow (Glade included, crisp edge) and the HDR streak core. These A/B items are closed; no softening part, no Quarter-only split. 367: ## P10-I closed by the orchestrator (9 October 2026, ~08:35) 368: - Close on main 862ccbe, entry and handoff corrections committed 9d2d8d5. Suite 46 suites, 24,306 checks, 0 failures; --verify 2880x1800 exit 0; journey draw calls and counts identical to 27-P9; bench dense +0.04 to +0.15 ms p50, Harbour +0.02 to +0.21. 369: - Owner sheet (10 pages) sent. Awaiting the owner on P10a (violet-grey stone), P10b (lawn bands, approved grove view lawn changed a lot), P10p (hover/arrows/double-click) and P10h (bokeh). P10f and P10i were already kept. 370: - Open: harbour-night cyan reflection dashes at the front lawn edge are missing in all four after captures; probably time-varying water dashes but not proven (needs 28-P10-i against 28-P10-h-pre). 371: - Snapshot naming: the file system is case-insensitive, so 28-P10-I is the same folder as 28-P10-i (P10i's final code). The stage snapshot is 28-P10-S1; the journey folder is artifacts/feel-journey/28-P10-I. 373: ## Owner rulings on the P10-I sheet (9 October 2026) 374: - "Keep all of P10a, P10b, P10p and P10h." Violet-grey canal stone and calmer storeys (P10a), lawn tone bands and turf octave (P10b), hover lift, arrows-only selection and double-click to move (P10p), and art-pixel tilt-shift and photo bokeh (P10h) stay as built. With P10f and P10i kept earlier, all of P10-I is approved. The open points on each part stay recorded as known behaviour, not work to do. The harbour-night dashes question stays open (not proven; no owner action).