Chapter 27: Pixels that stay put Frozen 9 October 2026 at 19:36 BST (game main at 8418d4e, clean; the frozen handoff and owner-decisions.md equal HEAD). Covers stage P9 (Signature E: ink and light edges, pixels fixed to the world): P9a edge ink and rim pass, P9b ink in the profile and tuning (with the review fix and the open cost question), P9c presentation post at art resolution, P9d camera snap to the art-pixel grid (with the review fix and the orchestrator ruling on shading parity), P9e smooth scroll by the residual and the integer-exact rect, and the P9 stage close with the repeated --verify, D31 and the M6 export. Owner answers: D31 live pan session "all looks good"; D30 "ink off" (8 October). Part codes: P9a edge ink, P9b ink tuning, P9c art-resolution post, P9d camera snap, P9e smooth scroll; D30 ink and rim default, D31 live pan session; M6 the milestone exports. 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 320-532 320: **Beautify P9 stage close (8 October 2026):** the stage-end overhead of stage P9 "Signature E: ink and light edges, pixels fixed to the world" (step `27-P9`). Opus 5.5 ran it on this Mac on the final P9 code: `main` d642dc3 (code 2c01eb0, P9e; d642dc3 only adds an owner-decisions note), clean apart from the orchestrator's `artifacts/beautify/beautify-stage.js` (`close/head.txt`). No game code changed. Evidence: `artifacts/beautify/27-P9/close/`. 321: - **Snapshots as they are** (`build/backup-feel/`): `27-P9` (pre-stage, code 8ccf61a), `27-P9-a` (P9a code, 26d33df), `27-P9-b` (final P9b code, 6b809b1; the pre-P9b state is only in `27-P9-a`), `27-P9-c-pre` / `27-P9-c` (P9c code 3aeb293), `27-P9-d-pre` / `27-P9-d` (P9d code d8dfd63, `SNAP_LIVE` false), `27-P9-e-pre` / `27-P9-e` (P9e code 2c01eb0, `SNAP_LIVE` true). Each `-pre` equals the previous part's folder on scripts shaders tests data scenes tools. No `-post` folders exist. 322: - **Before side:** `close/pre-tree` from `tools/feel_snapshot.sh tree 27-P9` (= `main` 8ccf61a, the P8-IV close code). It differed from live only in P9's 22 files, each of the 12 shared ones equal to 8ccf61a (`close/diff-pretree.txt`, `close/git-changed-since-8ccf61a.txt`; `project.godot` equal). 323: - **Power:** every run on AC, 100% charged, Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`); load averages 2.0-9.3 (1-min; the highest during the captures, 3.3-5.0 at the bench). 324: - **Full suite** (`close/suite.sh` a, b, v; 19:25-19:34): **46 suites, 23,994 checks, 0 failures** (`close/suite/suite_all.out`). `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. `test_creature_reactions` 5,583 checks and 0 failures in the run and the rerun; the known flaky check did not fire. 1440x900 not run (inherited stall). 325: - **Journey** (`artifacts/feel-journey/27-P9/`, six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): counts against 26-P8-IV (`27-P9/counts-vs-26-P8-IV.txt`) +1 draw call and rendered object on all six views (harbour-day 908 → 909, harbour-lane 750 → 751, harbour-night 907 → 908, quarter-day 1025 → 1026, quarter-night 1024 → 1025, styles-day 502 → 503): P9c's declared +1 (the post viewport's rect); P9a, P9b, P9d and P9e declared none. A single-view re-run of harbour-lane and quarter-night (`close/rerun/journey-live/`, `rerun/counts-journey-vs-rerun.txt`) gave identical counts. Frame >8 luma against 26-P8-IV 2.2-7.6% (`27-P9/frame-vs-26-P8-IV.txt`), within-tree re-run 2.0% (harbour-lane) and 6.3% (quarter-night): drops, rings, clouds and smoke at another TIME phase. `notes.json` has a `27-P9` entry (`close/journey-note.js`, backup `notes.json.bak-before-27-P9`); `index.html` and `journey-standalone.html` (110.37 MB) rebuilt. 326: - **Art capture and protected looks** (`close/caps.sh`, the P8-IV close's 101 views in groups g1-g3; 19:36-20:24; pre-1, after, pre-2, after-2, every view on try 1). Metrics `close/metrics/` (`metrics.sh`, `frame-table.js`, `prot.js` → `protected.md`, `elapsed.js`, `consistent.js` → `look/consistent/*.png`; regions = the P8-IV close's). Against the previous stage close `26-P8-IV/close/after`: `ab-p8ivclose-*`, `frame-p8ivclose-*`. Sheets `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P8-IV-close.png`, `noise-pre.png`, `noise-after.png`. 327: - **Counts** (`metrics/counts.txt`): +1 on all 101 views in both pairs (orbit views vary within a tree, as before). The previous close's after against this pre-1: identical on 100 views (quarter-orbit varies). 328: - **Frames:** 10 of 101 views above both noise legs (`frame-table.txt`: amsterdam-day/-night, harbour-2015, harbour-pipes, qr-day, qr-night, qt-day, quarter-r16-night, shops-night, signs-fog). Their consistent maps (looked at: qr-day, signs-fog, harbour-pipes, amsterdam-day) show only rain rings, water glints, cloud dither and chimney smoke; the live capture is 22-83 ms earlier in both pairs on those views (`elapsed-ms.txt`), since the P9 post is cheaper. No structural change. The pre-1 sitting was noisy on rainy views (mirror columns pre-1/pre-2 54%); pre-2 → after-2 is the cleaner pair. 329: - **Canal mirror columns:** p50 43.5 → 43.4 (43.6 → 43.5) quarter-night, 52.3 → 52.4 at 21:00; changed >8 in pre-2 → after-2 40.2% and 1.6% against noise 54.0%/0.8% and 49.2%/14.8%: within noise. 330: - **HDR neon cores:** fish 66.3 = 66.3, bar 70.9 → 70.6 (70.6 = 70.6), lane sign 140.2 = 140.2, tower windows 96.0 = 96.0; changed shares at noise. 331: - **05:45 and 20:45 skies:** changed ≤1.1% against noise ≤1.3%; p50 within 0.3 (quarter 20:45 horizon 151.1 = 151.1, zenith 97.7 = 97.7; 05:45 horizon 113.9 → 113.8, zenith 56.6 = 56.6). 332: - **Glade/Canal separation:** Glade night zenith `#001d45`, horizon, lawn and bank 0% changed; Harbour night zenith p50 53.9 = 53.9. 333: - **Approved tree style:** grove-day and grove-1930 crowns, conifers, feet and lawn 0.00% changed (p50 55.3, 55.6, 59.4; golden conifers 42.9); quarter-night undersides 27.7 = 27.7. 334: - **Whole-stage A/B bench** (`close/bench.sh` = the P8-IV close recipe; `bench-t1.out`, `-t13.out`, `-t19.5.out`, `bench/summary.txt`; 20:31-20:50): 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. The first 19:30 set was disturbed (dense live round 1 +1.28 ms, Harbour pre 8.2-8.3, harbour-live-1 needed 4 tries leaving no `performance.json` on tries 1-3) and was rerun on try 1 throughout; the first set is kept in `bench/first-t19.5/` with `NOTE.txt`. Raw medians pre → live (paired median; rounds live slower): 336: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 337: |---|---|---|---|---|---|---| 338: | Dense | 14.660 → 14.291 (−0.386, 0/3) | 16.444 → 16.179 (−0.314, 0/3) | 14.407 → 14.338 (−0.022, 1/3) | 16.260 → 16.271 (+0.052, 2/3) | 17.031 → 16.835 (−0.216, 0/3) | 18.757 → 18.624 (−0.030, 0/3) | 339: | Harbour | 6.931 → 6.707 (−0.184, 0/3) | 7.548 → 8.223 (+0.675, 2/3) | 6.599 → 6.657 (+0.041, 3/3) | 7.684 → 7.487 (−0.197, 1/3) | 7.492 → 7.261 (−0.235, 0/3) | 9.255 → 8.960 (−0.087, 1/3) | 341: - Draw calls +1 (dense 2587 → 2588, 2588 → 2589, 3439 → 3440; Harbour 917 → 918, 918 → 919, 830 → 831). The stage saves 0.2-0.4 ms p50 at night and in the evening (P9c's post; it skips the halo by day, so 13:00 is flat). Harbour p95 at 01:00 spreads 7.4-8.4 ms in both trees with one round each way, as in earlier closes; the plan's stop (dense p95 over about +0.5 ms) is not hit. The ink pass is off by default, so its cost (+0.41 ms dense with it on, P9b) is not in these numbers. Every run in both trees logs a missing `res://assets/ui/hologram/terrain_flowers.svg` (pre-existing). 342: - **Owner sheet** `artifacts/beautify/27-P9/close/sheet/` (`close/sheet.sh`, regions and sources in `close/sheet-regions.json`): `p9-01-ink-rim-night.png`, `-02-ink-rim-golden-day.png`, `-03-ink-rim-twilight-lane.png` (close/after against `close/ink-on`, the final code with `--edge-ink` at `EDGE_PROPOSED`, `close/inkcaps.sh`; +1 draw call with the pass on), `-04-post-art-resolution.png`, `-05-post-what-differs.png` (P9c frozen sets and map), `-06-snap-pan-frames.png` (P9d films), `-07-smooth-scroll-frames.png` (P9e films), `-08-doubled-column.png` (P9e crops at pixel size 3), `-09-existing-scenes.png` (pre-1 against after), and `sheet.md`. I looked at pages 01, 02, 04, 06, 07, 08 and 09. Pages 06-08 reuse the parts' films and crops (no new recordings). 343: - **Not done / open:** 344: - **D31, the live pan session on this Mac with the owner, was not run**; the camera snap is on by default in a person's editor (`CameraRig.SNAP_LIVE = true`) with P9d's accepted shading-parity change on each 1-px step (8-49% of pixels). Options after D31: a 2-art-px grid, parity-stable derivatives in `contact_shade.gdshader` and `surface_detail.gdshaderinc`, or `SNAP_LIVE = false`. 345: - **The repeated `--verify` was not done.** The close was asked to run the windowed `--verify` at 2880x1800 five times on the pre-tree and five times on the final code and report stalls per tree. On disk there is one run on the final code (`close/suite/verify.log`, `suite_all.out`: try 1, exit 0, 34 s, `PIXELPUNK_INTEGRATION []`, no stall) and none on the pre-tree (now in `the macOS Trash/`, `close/trash.txt`); 1440x900 was not run either. So 1 of 5 on the final code (1 passed, 0 stalled) and 0 of 5 on the pre-tree. The only P9 stall on disk is P9c's one-off first try (`c/suite/suite_all.out`: 293 s, no `PIXELPUNK_INTEGRATION`). 346: - **Repeated `--verify`, done afterwards by the orchestrator (8 October, evening):** the pre-stage tree was rebuilt from `build/backup-feel/27-P9` (`close/verify-repeat/pre-tree`) and `--verify` at 2880x1800 was run 5 times per tree, interleaved, one Godot at a time, on mains power (`close/verify-repeat/run.sh`, `results.txt`, `power.txt`, logs `pre-N.log` / `live-N.log`). Result: **0 stalls in 10 runs** (every run 34-35 s, 52 captures). **8 of 10 passed with `PIXELPUNK_INTEGRATION []`; pair 2 failed on both trees with the same line** (`Holding the raise brush did not build a hill: 2.05`, exit 1; `pre-2.log`, `live-2.log`). The identical value on the pre-stage tree shows it is not from P9; the two runs were back to back, so it looks like a machine-load timing effect on the raise-brush check (not investigated). The P9c stall (293 s) did not recur in 10 runs. 347: - **D31 and M6 (8 October 2026):** the owner ran the live pan session on this Mac on the final build (camera snap on, smooth scroll) and said "all looks good" (`owner-decisions.md`, "D31 live pan session done"). The M6 export then ran from `main` 32c9706 with the pinned Godot 4.6.3 (`artifacts/beautify/27-P9/m6/m6_export.sh`, `m6_export.out`; no errors in the logs): `the Desktop/PixelPunk-M6.app`, `PixelPunk-M6-macOS.zip` (66 MB, macOS universal, ad-hoc signed) and `PixelPunk-M6-Windows.zip` (43 MB, `PixelPunk.exe` plus README, notices and START HERE). The builds were not launched after export; Windows was not run (P6e kit still deferred, so Windows cost and correctness of P6 to P9 remain unmeasured). 348: - **D30** (ink and rim on by default, outline level) unanswered; default keys 0 stand. The P8-IV sheet verdict is also still open. 349: - **M6 final exports (both platforms) not built**; the optional P6e Windows kit not rerun. Windows not run for any of P9. 350: - No perspective, photo-mode, Reduced or Off captures at this close; the live margin's cost is unmeasured (automation does not use it). 351: - The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 352: - Scratch trees: `close/pre-tree` and the part trees `a/`-`e/pre-tree` were moved to `the macOS Trash/` with `mv` as `beautify-27-P9--pre-tree-205415` (`close/trash.txt`) once their numbers were recorded. `d/probe-tree` (a P9d probe tree, not a pre-tree) was left in place. 353: - Platforms: Mac run; Windows not run. 355: **Beautify P9e Smooth scroll by the residual; integer-exact rect (8 October 2026):** Opus 5.5 on this Mac (plan section "P9e"; report Signature E "scroll the finished frame by the remainder. Use a 1-art-px margin and an integer-exact rect" and Defect 9, the doubled row or column at odd window sizes (`main.gd:282, 489` in the report); step `27-P9`, fifth and last part; a measuring part, second review required). While the view snaps, the 3D view is drawn 1 art px larger on every side (`CameraRig.SNAP_MARGIN`) and the window shows it at exactly `pixels` window px per art px, moved by the P9d residual in whole window px, so pixels stay fixed to the world and a pan or ease scrolls at window resolution. At every window and pixel size the window now shows each art px at exactly `pixels` window px (the 3D view is `ceil(window / pixels)`), which removes the doubled row or column. **The snap is now ON in a person's editor by default** (`CameraRig.SNAP_LIVE = true`, my decision from the tests below: the plan says the snap stays off "until P9e lands"). Automated runs (`--smoke`, `--verify`, `--drag`, `--benchmark`, captures, recordings) and headless tests do not snap and have no margin, so they keep the old camera; only `--snap` turns it on there. Reverting is one line (`SNAP_LIVE = false`). With the snap on, the D31 live pan session now shows P9d's accepted shading-parity gap in the editor. Snapshots as they are on disk: the pre-part state is `build/backup-feel/27-P9-e-pre/` (the P9d code d8dfd63, `SNAP_LIVE = false`; `diff -rq` against `27-P9-d` on scripts shaders tests data scenes tools is empty; at part start the live tree equalled it, so nothing unexpected had changed), and `build/backup-feel/27-P9-e/` holds the P9e code (commit 2c01eb0, `SNAP_LIVE = true`; equal to the live tree on those folders). There is no `27-P9-e-post`. Pre-part scratch tree: `artifacts/beautify/27-P9/e/pre-tree`. Evidence: `artifacts/beautify/27-P9/e/`; file backups in `e/backup/`. 356: - **Code:** 357: - `scripts/camera_rig.gd`: `SNAP_LIVE = true`; `SNAP_MARGIN = 1`; `view_size(window, pixels, margin)` returns `ceil(window / pixels) + 2 margin`; `frame_scroll(residual, pixels)` returns `(round(-rx x pixels), round(ry x pixels))` window px. 358: - `scripts/world_view.gd`: `signal frame_moved`, emitted by `update_camera` when `snap_residual` changes (never while the snap is off). 359: - `scripts/main.gd`: 360: - `resize_viewport` now calls `layout_frame(window)`, which sets `frame_window`, `frame_margin` (1 while `world.pixel_snap`) and `viewport.size = CameraRig.view_size(...)`. 361: - `_place_frame` (on a layout and on `frame_moved`) places `upscale` at `frame_origin()` = (-margin, -margin) x pixels + the scroll, sized exactly `viewport.size x pixels`. 362: - New helpers `view_frame()` (the window's part of the view, in view px), `view_point()` (window to view) and `window_point()` (view to window). 363: - `upscale` moved into its own `CanvasLayer` "FrameLayer" (layer 0, under the editor's layer 1). Its scale is `visible / frame_window`, so one of its units is one window px. 364: - A 3D view sized by hand (tests) is shown whole, as before (`_fit_post`). 365: - `pan.txt` gains `scroll_x scroll_y` columns, and its header gains `margin` and `window`. 366: - `scripts/scenery_audio.gd`: optional `view_origin` on `update` and `ground_samples` (default zero: unchanged). 367: - **Two designs tried and dropped** (both runs kept in `e/probe/`): 368: - A rect in the editor's canvas (`e/probe/try1/`): a Control's position is rounded to whole UI units, so at 2880x1800 (UI scale 1.53) a +1 px scroll drew half a window px off. 0.68-0.89 million of 1.24 million checked pixels were wrong on 4 of 6 residuals. 369: - An AtlasTexture region on a still rect (`e/probe/try2-atlas/`, `e/probe/atlas/atlas.out`): exact at pixel sizes 2 and 4, but at pixel size 3 the GPU's interpolated region coordinates miss the texel on 17-58% of pixels. 370: - **Audit of every `viewport.size` use** (the plan's line numbers had moved; found with grep): 371: - Now through the window's part of the view (`view_frame` / `view_point` / `window_point`): `screen_point` (every pick: tools, editor drag, selection lift, grips, tray drops), `replay_press` and `test_click_at` (event position only; nothing reads it), `_screen_row` (lens rows), the ambience grid (`scenery_audio.update`), `dock_goal` (share of the shown rows), `_photo_subject_depth`'s 5x5 grid, `_middle_ground` and `photo_focus_point`'s middle. 372: - Kept on the whole view, because the camera frames all of it, margin included, so metres per art px = size / its rows: `_pan_camera`'s scale, WorldView's `art_pixel`, `_snap_grid`, `_apply_dock_offset`, the near-rain aspect, `_scale_handles`' per-pixel size, the light-culling bounds (`grow(40)`), the `foliage_lamps` cache key, and the `pan.txt` / record / performance logs (they report the real render size). 373: - Consequence: with the margin the snapped editor frames 2 rows more than a capture of the same camera, 0.4% closer at 1600x1000 (`art_pixel` 38/502 against 38/500). 374: - **Tests:** 375: - Test-first: `e/testfirst-test_camera.log(.out)` (338 checks, 7 failures: `SNAP_LIVE`, `snap_wanted`, `SNAP_MARGIN`, the statics and three source checks; the frame part returned early) and `e/testfirst-test_presentation.log(.out)` (44 checks, 2 failures). 376: - After: `test_camera` **452 checks, 0 failures** (`e/after4-test_camera.log.out`; 333 + 119) and `test_presentation` **44, 0** (`e/final-test_presentation.log.out`). 377: - What `frame_checks()` covers: 378: - `view_size` and `frame_scroll` values, and the old 718-over-1437 stretch modelled (a 3-px column). 379: - At windows 1437x901, 1600x1000, 1438x902, 1601x999, 2880x1724 and 1100x700, pixel sizes 1-4, snap on and off: the rect, read off the node and its layer, is exactly view x pixels, every interior art px is exactly `pixels` window px, and the window is covered. 380: - At 1437x901 and 1600x1000, pixel sizes 2 and 3, with and without the dock offset, over 12 sub-pixel targets (residual up to 0.5): 381: - the rect sits at (-p, -p) + scroll in whole window px; 382: - the eased target stays within 0.5 window px of one place (the snap alone: up to p/2); 383: - every window pixel of a drawn art px maps back to it; 384: - a ground pick from a point's drawn pixel lands in the same art px and within its footprint; 385: - the lens row, the window middle and the dock share agree. 386: - The scroll follows `update_camera` by itself, and the ambience grid takes the origin. 387: - With the snap off the frame is the old one, and a hand-set view is shown whole. 388: - `screen_point`, driven through `Input.parse_input_event`, reads the mouse through the rect. 389: - **Mutation checks** (`e/mutant/run.sh`; live files restored and checked with `cmp`): no scroll gives 23 failures, the old stretch mapping 25, floor instead of ceil 4. 390: - Full suite (`VERIFY_RES=2880x1800 tools/suite_all.sh e/suite` a, b, v; 11:32-11:41, AC 100%, Low Power Mode 0, `e/suite-power.txt`): **46 suites, 23,994 checks, 0 failures** (23,873 + 121; `e/suite/suite_all.out`). `--verify` 2880x1800 passed on try 1 (exit 0, 35 s), `PIXELPUNK_INTEGRATION []`. The `test_creature_reactions` rerun had 5,583 checks, 0 failures; the known flaky check did not fire. 1440x900 was not run (inherited stall). 391: - **Changed expectations (old -> new):** 392: - `test_camera` "the snap stays off in a person's editor until P9e lands": `SNAP_LIVE == false` -> `true`. `snap_wanted([], live)`: false -> true. 393: - `test_presentation`: "upscale anchored to the full window" -> anchored top-left and placed by Main (with no snap at a divisible size it still covers exactly the window). "The upscale is in the window's canvas layer" with the lens nodes in the same canvas, ordered by index -> the upscale has its own canvas layer under the editor's; the lens nodes stay in the editor's canvas, before the UI. 394: - New: `view_size` in `test_presentation`. 395: - **Expected journey count deltas: none.** Measured on 11 frozen views: identical draw calls and objects, pre against live (`e/metrics/frozen-cmp.txt`; for example quarter-night 1024/1100, harbour-night 907/983, photo-golden 1351/1427, tilt 1026/1102). The new canvas layer adds no draw call. 396: - **Captures** (`e/caps.sh`, engine `--time-scale 0`, 1600x1000 unless noted; pre-tree `fz-pre-1`, `fz-pre-2` and live `fz-post`; AC 100%, Low Power Mode 0, `*/capture-meta.txt`): 397: - **Snap off = byte-identical** (`cmp`, courtyard and editor) on quarter-night, quarter-day, quarter-night `--dock-offset`, quarter-night 1437x901 (the window is 1438x902 on this Retina Mac, so p2 cannot be odd here), quarter-night `--tilt-shift=0.6` (lens over the new layer, looked at), and photo-golden. The pre-tree noise pair is also byte-identical on these. 398: - Harbour and Glade (CPU smoke) are within their noise pairs, lane slightly above one of them (courtyard 3.167% against 2.909%, editor 3.067% against 2.813%; the code path is identical at 1600x1000 pixel size 2 with the snap off, so this is single-noise-pair smoke variance, second review) (`e/metrics/frozen-diff-harbour.txt`): harbour-night pre-to-post 1.16/1.34% against noise 1.27%; lane 2.81-3.17% against 2.81-2.91%; Glade 0.785-0.808% against 0.788%. 399: - **Odd size (Defect 9)**, the artwork scene at pixel size 3 (`e/scenes/artwork-p3.json`), measured by `e/gridphase.js` (`e/metrics/gridphase.txt`: interior column and row runs whose length is not a multiple of p): 400: - pre 1600x1000: 1 doubled column (x 798) and 1 doubled row (y 498); 401: - pre 1438x902: 1 column and 2 rows; 402: - **live: 0 and 0**. The view grew from 533x333 to 534x334 (and 479x300 to 480x301), so the frame moves by half an art px; that is expected. 403: - Crops `e/maps/p3-pre-crop-x798.png` and `p3-post-crop-x798.png` (looked at), and `fz-post/quarter-night+p3/courtyard.png` (looked at). 404: - **Pan filmstrips** (`--snap --pan=6,2.5`, 24 frames; pre-tree = P9d, the snap with a still rect; `e/film-pre/`, `e/film-post/`): analysed by `e/scroll.js` (best whole-window-px shift of frame k against frame 0; stride 3, because stride 2 aliases the 2x2 blocks; `e/metrics/scroll-summary.txt`). 405: - **Live:** the measured shift equals the one `pan.txt` predicts on 24/24 frames in all 4 films. The drawn target sits within **0.40-0.47 window px** of the continuous pan, and frames move by at most **1** window px per frame. 406: - **Pre (P9d):** within 0.80-0.95 px, with jumps of 2 px. 407: - Sheet `e/sheet/qn-pre-post-f0-7.png` (looked at; `README.txt`). 408: - **Drag replay** (`e/drag.sh`, `data/drag/baseline.json`, 1600x1000, frozen clock; pre, post, pre, post): no errors; the 25 steps (names, types, selections) are equal; all 13 document digests are identical across the four runs; draw calls 899 = 899; `drag-after.png` pre-to-post 21-25k pixels against a noise pair of 24k (`e/metrics/drag-compare.txt`; `drag/post-1/drag-after.png` looked at). 409: - **Picking probe** (windowed live editor: no automation argument, so the snap and margin are on, with the dock offset; `e/probe/pick_probe.gd`, `run2.sh`, logs `e/probe/pick--p.txt`, final six in `final6/`; AC, Low Power Mode 0, `power.txt`): 410: - Windows 1437x901 (1438x902), 1600x1000 and 2880x1800 (2880x1724, UI scale 1.53), each at pixel sizes 2 and 3, six residuals each. 411: - **Every 2nd window pixel equals the post texel `view_point` names: 0 differ out of 1.9-7.4 million per case.** Ground picks: 0 of 270 per case outside the art px under the pointer. Grips of a selected building: 10/10 picked identically and are the grip itself. 412: - Object picks (136 part centres; 1437x901 p2, 1600x1000 p3, 2880 p2): 136/136 equal to the pick through the rect's own geometry. 413: - Photo mode (perspective lens, residual 0, rect at (-p, -p)) and its screenshot (`grab_viewport`, full window size): 0 differ. 414: - A mouse moved through the engine's input did not reach `get_mouse_position` in a windowed run (the real pointer wins). `screen_point` is covered by the headless check and the source check. 415: - **Acceptance:** 416: - **Picking exact to the art pixel at odd sizes: met.** Headless at 1437x901 and 1600x1000, pixel sizes 2 and 3, with a residual, with and without the dock offset. Windowed in the live editor at six window and pixel cases, with 0 mismatches. 417: - **No doubled row at 1437 px: met.** Headless at 1437 and 1601 px; in captures at pixel size 3: 1-2 -> 0. At pixel size 2 this Mac cannot make an odd window. 418: - **Pans scroll smoothly with pixels locked to the world: met for position.** Shifts are exactly as predicted on 96/96 frames, the target within 0.47 window px, steps of 1 window px. P9d's shading-parity changes on grid steps remain (accepted, D31). 419: - **`--verify` unchanged: met** (pass, integration []). **`--drag` unchanged: met** (digests identical). 420: - **Automation byte-identical with the snap off: met** on every deterministic view, and within noise on Harbour. 421: - **Bench:** none. The plan has none for P9e, and automated runs and benches take the old path (identical frames and counts). The live margin draws 2 more rows and columns (about 0.6% more pixels at 1600x1000); its cost was not measured. 422: - **Not done:** 423: - The D31 live pan session (owner, at the P9 close). 424: - Windows. 425: - `--verify` at 1440x900 (inherited stall). 426: - Odd p2 windows on screen: impossible on this Retina Mac; covered headless only. 427: - The vignette now spans the margin and moves with the scroll (at most 1 window px at the frame edge; not measured). 428: - The OS mouse path was not exercised windowed. 429: - The shading parity (P9d ruling) is unchanged and is now live. 430: - Review: filled by the orchestrator. 431: - Platforms: Mac run; Windows not run. 433: **Beautify P9d Camera snap to the art-pixel grid (8 October 2026):** Opus 5.5 on this Mac (plan section "P9d"; report Signature E "Pixel stability: snap the orthographic camera to the art-pixel grid", gated to a live camera, and the pixel lens's finding F3; step `27-P9`, fourth part). Under the orthographic camera the drawn look-at point is now the eased target moved, in the frame's own plane, onto the nearest whole art pixel, and the remainder is kept as `WorldView.snap_residual` (art pixels, each axis in [-0.5, 0.5]) for P9e to scroll the finished frame by. **The snap is OFF by default**: `CameraRig.SNAP_LIVE` is false until P9e lands (a person's editor does not snap yet; without the scroll a pan would step a pixel at a time), and only the test flag `--snap` turns it on. With it off every camera is exactly what it was. The perspective lens and photo mode's free camera are never snapped. No export may happen between P9d and P9e (plan). Snapshots as they are on disk: the pre-part state is `build/backup-feel/27-P9-d-pre/` (the P9c code 3aeb293; `diff -rq` on scripts shaders tests data scenes tools against `27-P9-c` is empty; at part start the live tree equalled it, so nothing unexpected had changed), and `build/backup-feel/27-P9-d/` holds the P9d code (commit d8dfd63, `SNAP_LIVE = false`). There is no `27-P9-d-post`. Pre-part scratch tree: `artifacts/beautify/27-P9/d/pre-tree`. Evidence: `artifacts/beautify/27-P9/d/`; file backups in `d/backup/`. 434: - **Code:** 435: - `scripts/camera_rig.gd`: `SNAP_LIVE = false`; `snap_wanted(args, live)` (`--snap`, or live and `SNAP_LIVE`); `snap_target(target, basis, mpp)` returns `{target, residual}` (the point whose right/up coordinates are whole art pixels; the depth along `basis.z` is kept); `parse_pan(text)` (`--pan=[,]`, art px/s along the frame's right and up, clamped to 240; P8e's one-number form is x only). 436: - `scripts/world_view.gd`: `pixel_snap`, `snap_residual`; `update_camera()`'s orthographic branch looks at the snapped point when `pixel_snap` is on, with mpp = `view_camera.size` / the viewport's rows, the same as `art_pixel`. The dock offset (P5e) is already whole art pixels and is untouched, so the grid holds with a non-zero `v_offset`. The near rain box follows the drawn look-at point (the report's challenge note). The veil depth, `art_pixel` and every other camera input are unchanged. 437: - `scripts/main.gd`: `world.pixel_snap = CameraRig.snap_wanted(...)` after `camera_live`. `record_pan` becomes a Vector2 (the `--record` pan writes x exactly as before and adds y only when it is non-zero). New pan filmstrip: a `--smoke` run with `--pan` and without `--record` takes `PAN_FILM_FRAMES` 24 consecutive frames from frame 100, moving the goal by pan x `PAN_FILM_STEP` (1/30 s) before each one, a fixed step so `--time-scale 0` freezes everything but the pan. It saves them as `pan_NN.png` and writes `pan.txt` (per frame: target, r/u, residual, the drawn look-at point's grid coordinates). 438: - `tests/test_camera.gd`: `snap_checks()` (new). 439: - **Tests:** test-first `d/testfirst-test_camera.log(.out)` (266 checks, 5 failures: no `snap_target`, `SNAP_LIVE`, `snap_wanted`, pan filmstrip; the editor part returned early). `d/testfirst0-parse-test_camera.log` is my own first try, a parse error in the new test (static `call` on the class name). After: `test_camera` **328 checks, 0 failures** (`d/after1-test_camera.log.out`; 261 + 67). The new checks cover: the maths over 240 views (residual within 0.5 on each axis, worst on-grid error under 0.002 px, depth kept, target = snapped + residual x mpp); `--snap` and `--pan` parsing; snap off: the camera position equals the old expression exactly on 12 views and targets, residual 0. Snap on, at 3 zooms with and without the dock offset (non-zero, whole-pixel `v_offset`): the eased target is drawn at centre + (rx, -ry) to within 0.01 px and within half an art pixel; static probes move by whole pixels between sub-pixel targets; `art_pixel` is unchanged; a view at rest does not jitter. In a live editor, frame by frame: the F-focus tween (a pan, and one with a zoom), a trackpad pan, and the same pan with the snap off (it must show fractional moves, which it does). Also: the perspective lens and the free camera are unchanged with the snap on. **Mutation check:** with `look = target` forced (no snap), 17 checks fail (`d/mutant1-test_camera.log.out`). Full suite (`VERIFY_RES=2880x1800 tools/suite_all.sh d/suite` a, b, v; 10:25-10:34, AC 100%, Low Power Mode 0, `d/suite-power.txt`): **46 suites, 23,868 checks, 0 failures** (23,801 + 67; `d/suite/suite_all.out`). `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. creature_reactions rerun 5,583, 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 440: - **Changed expectations:** none (new checks only; `record_pan` changed type, but no test read it). **Expected journey count deltas: none** (snap off by default). Measured on 8 frozen views, pre-tree against live: identical draw calls and rendered objects (`d/metrics/counts.txt`, for example quarter-night 1024/1100, harbour-night 907/983, photo-golden 1351/1427). The pan films have the same counts with and without `--snap`. 441: - **Captures** (`d/caps.sh`, engine `--time-scale 0` as in P9c, 1600x1000, pixel size 2; AC 100%, Low Power Mode 0, `d/*/capture-meta.txt`, `d/film-power.txt`): 442: - **Snap off = byte-identical** (`d/metrics/frozen-diff.txt`). The pre-tree sets are `fz-pre-1`, `fz-pre-2` and `fz-pre-3` (quarter-day and harbour-night only); the live sets are `fz-post` and `fz-post-2` (quarter-day and harbour-night). Views: quarter-night, quarter-day, quarter-night with `--dock-offset`, quarter-night at 1437x901, photo-golden (perspective lens), harbour-night, lane-night, glade-day; `courtyard.png` and `editor.png`. **Pre and live are byte-identical (`cmp`) on every artwork view**, with a noise pair also byte-identical. Two exceptions, one traced and one open: 443: - **Traced:** in `fz-post/quarter-day/editor.png`, a hover box on the tower (0.92% of pixels; `d/maps/qd-editor-pre1-vs-post-crop.png`, looked at) is a pointer-position effect: the recapture `fz-post-2` is byte-identical to the pre-tree, and `fz-post` differs from `fz-post-2` by exactly that box. 444: - **Open:** `fz-pre-1/harbour-night` is an outlier of the pre-tree run: 911 draw calls against 907 everywhere else, a different camera framing and zoom (the island framed further left), and 78.96% of its pixels differ from `fz-pre-2/harbour-night` (courtyard; 43.23% over 8 levels; `d/metrics/frozen-diff.txt`; `d/maps/hn-pre1-vs-pre2-half.png`, looked at). No cause was found. It is in the pre-tree (P9c code), so it is inherited nondeterminism in the automated Harbour framing, not a P9d effect; it stays open. 445: - Without those: Harbour, lane and Glade differ pre/live by 1.31-1.49% (Harbour), 2.2-3.1% (lane) and 0.75-0.79% (Glade) against noise pairs of 1.42-1.53%, 2.9% and 0.77%. That is the CPU smoke and the variance P9c already recorded. 446: - **Pan filmstrips** (`d/film/+snap+pan`, `+pan`; `--pan=6,2.5`: 0.2 art px right and 0.083 up per frame; 24 frames): quarter-night, quarter-day, harbour-night and harbour-day, with and without `--snap`, plus `quarter-day+snap+panx` (x only). Analysed by `d/panshift.js` (art-pixel frames; the best whole shift in [-3, 3]^2 per pair; each compared pixel is classed world-locked, screen-locked or neither), output in `d/metrics/pan-*.txt` and `d/metrics/summary-pan.txt`. Sheets: `d/sheet/qn-snap-vs-nosnap-f0-8.png` and `d/sheet/hd-snap-vs-nosnap-f0-8.png` (looked at: without the snap, lamp heads and the table change shape every frame; with it they hold, then step one pixel). 447: - **Acceptance:** 448: - **On-screen target within half an art pixel of the eased target, residual in [-0.5, 0.5], with a non-zero `v_offset`, the F-focus tween and a trackpad pan: met** (tests above; worst placement error under 0.01 px). 449: - **Snap off: the camera exactly as before, captures byte-identical: met** (position equality in the test; frozen captures above). 450: - **Perspective and free cameras untouched: met** (test). 451: - **Static pixels stay or move by whole art pixels: met for positions, not in full for shading.** 452: - **Positions:** with the snap, the best whole shift equals the one `pan.txt` predicts on 23/23 pairs in all four films. Pairs where the grid did not step (16 or 17 of 23) are **byte-identical** on the quarter (0.00%), and at most 0.14% changed on Harbour (frozen CPU smoke). Without the snap, the median pair changes 33.6-60.9% of pixels across the four films, 7.4-11.3% by more than 8 levels; the largest single pair is 62.8% (quarter-night) (`d/metrics/summary-pan.txt`). 453: - **Shading:** on the pairs where the grid steps one pixel (6-7 per film), content does not translate exactly. **"Neither" class** (pixels that match neither the world-shifted nor the unshifted previous frame; `classes.t0/t8.neither` in `d/metrics/pan-*+snap+pan.txt`), over 8 levels in brackets: quarter-night 43-49% (2.3-5.2%), quarter-day 26-37% (3.2-7.8%), harbour-night 14-19% (1.1-1.4%), harbour-day 8-20% (1.7-1.9%). **Total mismatch against the predicted world shift** (`100 - at_predicted.eq`, over 8 levels `at_predicted.gt8`; screen-locked pixels count here too): quarter-night 51-62% (7.3-11.0%), quarter-day 28-50% (4.2-10.1%; worst pair 4-5, a y step: best_eq 49.9%, best_gt8 10.06%), harbour-night 21-34% (5.4-5.6%), harbour-day 10-39% (2.5-3.4%). y steps are worse than x steps (quarter-day x 28% (4.2%), y 50% (10.1%)). 454: - **Probed cause** (a scratch probe tree `d/probe-tree` with test-only switches; nothing in the live code): it is not the sun or moon shadow map (switching it off changes nothing), nor the fog, cloud-shadow or veil passes. The contact-shade pass accounts for most x-step changes over 8 levels (quarter-day 3.2% to 0.3%). The rest follows **screen parity**. On the x-only film, odd shifts change 28-31% of pixels (4.2-4.7% over 8 levels), even shifts only 5.2-5.3% (0.7%), and 4 px 8.8-9.0% (1.1%) (`d/metrics/pairstat-panx.txt`, by `d/pairstat.js`; film `d/film/quarter-day+snap+panx`). The maps (`d/film/map-qd-snap-2-3.png`, `d/probe/map-qd-ns-nc-4-5.png`, `-2-3.png`, looked at) show a diamond moiré on lit surfaces for x steps and brick-course banding for y steps, while foliage translates exactly. My reading, not proven: derivatives over the GPU's screen-fixed 2x2 quads (fwidth and dFdx on non-linear terms such as the masonry lines). Only the foliage reads FRAGCOORD, and it is clean. At night, lamp-lit areas also change by 1-8 levels on a step. 455: - A **2-art-pixel grid** (probe only, `--probe-snap-step=2`, `d/metrics/probe-*step2.txt`) cuts quarter-day x steps to 0.10% over 8 levels and y steps to 1.39%, but leaves quarter-night at 1.1-3.2%. It doubles the step and the residual range, which is a P9e design choice, not taken here. 456: - **Orchestrator ruling (8 October 2026): the shading-parity gap is ACCEPTED AS RECORDED for now.** The gap: 8-49% of pixels change on a 1-px grid step ("neither" class above), from the contact-shade normals (dFdx/dFdy) and the `surface_detail` bumps depending on the 2x2 quad parity; odd shifts 28-31% against even 5%. **P9e continues with a 1-px grid**, and the owner judges the result in the D31 live pan session. Options kept for after D31: (1) a 2-art-pixel grid (quarter-day x steps 0.10% over 8 levels, night 1.1-3.2%; it doubles the step and the residual range), or (2) quad-stable derivatives in `contact_shade` and `surface_detail`. 457: - **Bench:** none; the plan has no bench for this part. With the snap off nothing in the frame changes. With it on, the cost is one `basis_of` and two dot products per camera update. 458: - **Not done:** 459: - The shading parity above is not fixed: it is outside the camera. Accepted as recorded by the orchestrator ruling above; revisit after D31. 460: - No live editor run: the snap is off live until P9e, and the D31 live pan session is at the P9 close. 461: - No recording with `--record --pan`: Movie Maker under `--time-scale 0` gives a zero delta. 462: - `--verify` at 1440x900 not run (inherited stall); the stage-close repeat of `--verify` on both trees is left to the close. 463: - `--time-scale 0` and the pan filmstrip are measuring tools, not part of the P0a kit. 464: - **Review fix 1 (8 October 2026):** the second `_place_rain` call (a ceiling change in a foliage pass, `_refresh_foliage`) passed the eased `view_camera.target`, so with the snap on the near rain box did not follow the drawn point. It now passes `drawn_look()` (new; `view_camera.target` unchanged when the snap is off, the perspective lens or the free camera), and `update_camera` shares the grid maths through `_snap_grid()`. Test-first in `snap_checks()`: `d/fix1-testfirst-test_camera.log.out` (331 checks, 1 failure: rain origin off the drawn point); after `d/fix1-after-test_camera.log.out` **333 checks, 0 failures**. Full suite (`VERIFY_RES=2880x1800 tools/suite_all.sh d/fix1-suite`, 10:43-10:52, AC 100%, Low Power Mode 0, `d/fix1-suite-power.txt`): **46 suites, 23,873 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1, `PIXELPUNK_INTEGRATION []`; creature_reactions rerun 5,583, 0 failures. Backups `d/backup/fix1/`. Sheet row labels: `d/sheet/README.txt`. 465: - Review: filled by the orchestrator. 466: - Platforms: Mac run; Windows not run. 468: **Beautify P9c Presentation post at art resolution (8 October 2026):** Opus 5.5 on this Mac (plan section "P9c"; report Signature E "Presentation post at art resolution instead of window resolution" and the pixel lens's missed finding M0; step `27-P9`, third part). The presentation post (`shaders/presentation.gdshader`: luma outline, P1c toe grade, P5f pixel halo, vignette) now runs once per art pixel in its own SubViewport the size of the 3D viewport; the window TextureRect only scales that up, nearest, with no material. Before, every window pixel repeated its art pixel's 5 outline and 12 halo taps (4x at pixel size 2). The only intended pixel change is the vignette, now taken at each art pixel's centre: at most one 8-bit level, and none at pixel size 1. Snapshots as they are on disk: the pre-part state is `build/backup-feel/27-P9-c-pre/` (the final P9b code 6b809b1; `diff -rq` on scripts shaders tests data scenes tools against `27-P9-b` is empty; at part start the live tree equalled it, so nothing changed that I did not expect), and `build/backup-feel/27-P9-c/` holds the P9c code (commit 3aeb293). There is no `27-P9-c-post`. Pre-part scratch tree: `artifacts/beautify/27-P9/c/pre-tree`. Evidence: `artifacts/beautify/27-P9/c/`. File backups: `c/backup/`. 469: - **Code:** 470: - `scripts/main.gd`: new `post_view` (SubViewport "PostView": `disable_3d`, opaque, UPDATE_ALWAYS) holds the 3D `viewport` as its child. A child viewport is drawn before its parent, so the post always reads this frame. `presentation` (same name, same material, so `main.presentation.material` and `apply_presentation` are unchanged) now sits in `post_view` and reads `viewport.get_texture()`. New `upscale` TextureRect (window canvas, full rect, nearest, stretch, no material) shows `post_view.get_texture()`. New `_fit_post()` keeps `post_view.size == viewport.size`; it is called from `resize_viewport` and from `viewport.size_changed`, so it also follows tests that set `viewport.size` directly. `_over_ui` now names `upscale` instead of `presentation`. The tilt-shift lens nodes, the UI and their order in the window canvas are unchanged, so the lens-radius self-tests (`lens_radius`, `lens_focus`, `run_tilt_checks`) needed no change. Captures, filmstrips, photo mode and Movie Maker all still read the root viewport. 471: - `shaders/presentation.gdshader`: a header comment only (no maths changed). 472: - `tests/test_presentation.gd` (new; added to `tools/test.sh` and `tools/Test.ps1`). `tools/art_metrics.js`: the art-grid comment brought up to date (no code change). 473: - **Draw order, probed** (`c/lagprobe/`, a windowed probe script: the 3D background alternates red and blue each frame, and the window is read after each `frame_post_draw`): nested (P9c's layout) 40/40 frames show this frame; post added after the 3D view as a sibling 40/40; post added before it 0/40 (39 show the previous frame). That last case is why the 3D view is nested. 474: - **Tests:** test-first `c/testfirst-test_presentation.log(.out)` (11 checks, 3 failures: no `post_view`, no `upscale`, no art-resolution note; it stopped before the node checks). After: `test_presentation` **42 checks, 0 failures**. The checks cover: the vignette twin (worst change under 1 level at 1600x1000, 2880x1800 and 1440x900; frame centre 1); the shader text; nesting and draw order; post and upscale properties; lens and UI order; sizes after `resize_viewport`, after direct `viewport.size` writes (720x450, 718x451, 1440x900) and at pixel size 1; uniforms still set; captures read the root viewport; the hover test. `test_look` 601, `test_camera` 261 and `test_tilt` 176 checks, all 0 failures (`c/after1-*`). Full suite (`VERIFY_RES=2880x1800 tools/suite_all.sh c/suite` a, b, v; 09:33-09:50, AC 100%, Low Power Mode 0, `c/suite-power.txt`): **46 suites, 23,801 checks, 0 failures** (23,759 + 42; `c/suite/suite_all.out`). creature_reactions rerun 5,583 checks, 0 failures; the known flaky check did not fire. **`--verify` 2880x1800:** the first try ran 293 s, exited 0, and **stalled** (no `PIXELPUNK_INTEGRATION` line, 0 captures, with the owner using the laptop). I could not reproduce it. Four standalone runs (`c/verify1.sh`) then passed, each with 52 captures and `PIXELPUNK_INTEGRATION []` in its log: the pre-tree once (`c/verify/pre-2880`) and the live tree three times (`c/verify/dbg1` with temporary progress prints, since removed; `live-a`, `live-b`); their run times were not saved to disk. The suite's own v part also passed again (`suite_all.out`: try 1 exit 0, 34 s, `PIXELPUNK_INTEGRATION []`). Recorded as a one-off. 1440x900 not run (inherited stall). 475: - **Changed expectations:** none of the existing assertions changed (new checks only). **Expected journey count deltas: +1 draw call and +1 rendered object on every view** (the post viewport's rect). Measured on all 12 frozen views and the 4 timed views (`c/metrics/counts.txt`, `c/metrics/timed-frame.txt`), for example quarter-night 1023 -> 1024 / 1099 -> 1100 and dense 2587 -> 2588. 476: - **Captures.** Frozen-time sets, all from one sitting (`c/caps.sh`, 09:18-09:23, timed sets 09:24-09:27; 1600x1000 unless noted, AC 100%, Low Power Mode 0, `c/fz-*/capture-meta.txt`): the engine's `--time-scale 0` stops shader TIME and the clocks, so the same tree twice is byte-identical on the artwork scene. Sets: `fz-pre-1`, `fz-pre-2` (pre-tree) and `fz-post` (live). Views: quarter-night, quarter-day, quarter-2045, quarter-night+edge (`--edge-ink=0.55,0.22`), quarter-day+glade (pixel size 1), quarter-night@1437x901 (window 1438x902, art 719x451), photo-golden (perspective lens with DoF), harbour-night, harbour-golden, harbour-night@2880x1800 (window 2880x1724), lane-night, glade-day. Diffs by `c/pxdiff.js` go to `c/metrics/frozen-diff.txt` and `frozen-diff-glade.txt`; maps are in `c/maps/`. Timed sets (`tools/art_capture.sh --set noise`): `before-1`, `before-2` (pre-tree), `after-1`, `after-2`, with `c/metrics/timed-frame.txt`. Recording check: `c/record-qn/` (2 s Movie Maker clip, 60 frames, contact sheet looked at). Looked at: `maps/quarter-night-courtyard.png` (the 1-level pixels lie along the frame edge and scatter thinly inside), `maps/harbour-night-courtyard.png` (the Harbour noise is the CPU smoke plumes and one tree), `fz-post/quarter-night@1437x901/courtyard.png`, `fz-post/photo-golden/courtyard.png` (lens and DoF intact), `maps/timed-hn-b1-a1.png`, `record-qn/contact.png`, and `suite/integration/tilt-inspector.png` and `integration_photo.png`. 477: - **Acceptance:** 478: - **Frames identical apart from the vignette: met.** On the deterministic artwork views, pre -> post changes 1.11-2.02% of pixels, **all by exactly 1 level (max channel difference 1; 0 pixels over 1)**, against a noise pair of 0.000%. The views: quarter-night 1.114%, quarter-day 2.021%, quarter-2045 1.490%, edge ink on 1.115%, 1437x901 1.239%, photo-golden 1.254%. Where the vignette is 1 nothing changed (1 pixel at 20:45, at the band's edge). At **pixel size 1 (Glade look on the quarter), pre -> post is 0 pixels**. On the Harbour scene the frozen noise pair is not zero (CPU smoke and one tree vary): pre -> post over 1 level is 0.84% against noise 0.82% (harbour-night), 1.00 vs 0.85% (harbour-golden), 2.19 vs 2.21% (lane-night), 0.85 vs 0.83% (harbour-night 2880), and glade-day 0.60 vs 0.58%. Everything above that is the 1-level vignette. Timed captures (TIME running): harbour-night over 8 luma is 6.7-8.7% before -> after against noise legs of 3.9% (pre) and 1.2% (post); quarter-night 7.4-9.2% against 9.0/3.7%. The map shows only water, cloud, rain and smoke phase, so this reads as the capture frame landing at a different TIME. The frozen sets carry the claim. 479: - **Fullscreen frame time falls: met** (`c/bench.sh`, `c/bench-summary.txt`, `c/bench-r01.out`, `c/bench-r23.out`; pre-tree vs live interleaved, round 0 discarded, rounds 1-3, vsync off, rain 0.45, `--motion-live --light-pin=20`, 01:00 with the halo's ring taps live, every run on try 1, 491 samples; 09:27-09:33, AC 100%, Low Power Mode 0, `c/bench/power.txt`). Paired post - pre p50: **Harbour 1600x1000 -0.188/-0.137/-0.159 ms** (6.34 -> 6.18); **Harbour 2880x1800 (window 2880x1724) -0.525/-0.531/-0.543 ms** (9.96 -> 9.43, about -5.4%); dense 1920x1080 -0.192/-0.251/-0.281 ms (14.55 -> 14.31). p95: Harbour 2880 -0.49/-0.54/-0.21; dense -0.27/-0.12/-0.24; Harbour 1600 +0.30/+0.19/-0.32 (noise). By day the halo is skipped, so the saving there is smaller; that was not benched. GPU-only time is not measured (no timer query in this path). Windows iGPU not run. 480: - **Captures still work: met** (smoke, filmstrip and integration captures, photo action, Movie Maker; see above). 481: - **Lens nodes and lens-radius self-tests:** the node paths did not change, so no update was needed; `--verify` passes them, and `test_tilt` 176 / 0. 482: - **Chain pieces since the plan:** P5f halo and P1c toe (in the post, measured above); P8 veil, cloud shadows and edge ink (3D side, under the post; ink on is identical within 1 level); P8a `art_pixel` globals (from the 3D viewport, unchanged; `test_camera` green); Reduced/Off (untouched; the timed and record runs use the default). Not separately exercised: a Reduced or Off capture, and the 1440x900 `--verify`. 483: - **Not done:** a day-hour bench; Windows (the P6e kit was not rerun); the doubled row or column at odd window sizes is unchanged (P9e's job; the 1437x901 pair shows the same frame as before). `--time-scale 0` is a measuring trick I used here, not part of the P0a kit. 484: - Review: filled by the orchestrator. 485: - Platforms: Mac run; Windows not run. 487: **Beautify P9b Ink in the profile, Glade scaling, tuning, owner sheet (8 October 2026):** Opus 5.5 on this Mac (plan section "P9b"; step `27-P9`, second part). The edge ink and rim pass (P9a) now takes its strengths from the lighting profile: every anchor dictionary ends with `edge_ink`, `edge_rim` and `edge_rim_color`, **defaults 0, 0** (the pass stays hidden at no cost until the owner answers D30), Glade scales them x0.3 and x0.4, and the proposed "on" values are a named constant, `LightingProfile.EDGE_PROPOSED`, shown on the owner sheet through `--edge-ink`. The tuning pass took P9a's open items: the pass moved from priority 108 to -2 (below the glass, the lamp halos, glows and rain drops, the placement fill and the grips), ink and rim fade with the distance fog, the rim follows the key light's handover, railing openings are no longer ringed, and lone rim specks are dropped. Snapshots as they are on disk: `build/backup-feel/27-P9-b/` now holds the final P9b code (commit 6b809b1; `diff -rq` against `27-P9-a` lists 9 differences on scripts shaders tests data scenes tools). The pre-P9b state (`main` 26d33df, the P9a code) is only in `build/backup-feel/27-P9-a/`; at part start the live tree equalled it: nothing unexpected. There is no `27-P9-b-post`. Pre-part scratch tree `artifacts/beautify/27-P9/b/pre-tree` (`feel_snapshot.sh tree 27-P9-b`, i.e. the P9a code). Evidence: `artifacts/beautify/27-P9/b/`; file backups in `b/backup/`. 488: - **Code:** 489: - `scripts/lighting_profile.gd`: `edge_ink` 0.0, `edge_rim` 0.0, `edge_rim_color` (the anchor's key lifted 30% toward white, written out as floats) at the end of all nine dictionaries (3 PROFILES + 6 authored; the ten ANCHORS share them, 01:00 and 04:00 both being Night: the plan's "ten anchor dictionaries" is ten anchors, nine dictionaries). **Named `edge_*`, not the plan's `ink`/`rim`/`rim_color`, because `rim` is already the fill light's colour.** `EDGE_PROPOSED` = [ink, rim] per anchor hour: 01:00 and 04:00 0.55/0.22, 05:45 0.5/0.2, 06:45 0.45/0.22, 09:00, 13:00 and 16:30 0.35/0.2, 19:30 0.45/0.24, 20:45 0.5/0.22, 22:00 0.55/0.22 (P9a's 0.35 night rim was too strong; night ink kept at the lens's 0.55). 490: - `scripts/look_profile.gd`: `GLADE_INK` 0.3, `GLADE_EDGE_RIM` 0.4; Glade's rim colour follows its warmed key (lerp toward `WARM_KEY` lifted 30%, by `GLADE_KEY_WARMTH`). 491: - `scripts/edge_ink.gd`, `shaders/edge_ink.gdshader`: `PRIORITY` 108 -> **-2**; `amounts(profile, override)`; `fade(begin, end, curve)` and `fade_at` (the engine's depth fog: smoothstep over `fog_depth_begin`-`_end` of the distance from the eye, to the curve, times the hour's `fog`); `light(toward_key, lit)` (`key_lit`, the rim x SunPath.handover); `tint` takes `edge_rim_color`; `BOX_REACH` 3 and `boxed()` (only on ink candidates: no far-side ink where nearer things lie within 3 art px on both sides along either axis, a railing opening); `RIM_WEAK` 0.5, `LINE_MIN` 2 and `rim_lit()` (only on rim candidates the 5x5 depth grid, corners not taken, is read; a rim needs 2 other creases of half the floor within two art px, so lone specks go; `LINE_MIN` 0 is a one-line switch that skips the grid and its cost). The CPU twin's `pixel()` now takes the two-out taps and the opening flag (sliver parity). 492: - `scripts/world_view.gd`: the strengths come from `EdgeInk.amounts(profile, debug_edge_ink)` in `sync_environment` and in `sync_profile` (a running clock blends them); `edge_ink.light(..., lit)`; `edge_ink.fade(...)` in `_apply_depths`. `scripts/main.gd`: `--outline-scale=` (0-2, never saved) scales the luma outline for the D30 sheet's outline levels. 493: - **Tuning tried and dropped (all on the measuring mask, `b/r2-*`, `b/r3-*`, `b/probe1`):** a "crease on both sides across" rule (r2) and a 3x3 "two neighbours 90 degrees apart" rule (r3) both ate one pixel of every 2-px dash on the 2:1 quay lip (`b/look/r3-quay-qn.png`); the 5x5 count (r4, kept) leaves the dashes as P9a had them (`b/look/r4-quay-qn.png`). A weak-crease fill (a crease of half to 1 x the floor lit when the larger of its pair) changed nothing measurable (r4 vs `b/r5-mask`: frame rim 2.33 vs 2.34%) and was dropped after the first bench. The probe `b/look/probe1-terr.png` (blue = weak, yellow = strong but rejected) shows why the dashes stay: their gap pixels carry no crease on either screen axis. 494: - **Tests:** test-first `b/testfirst-test_edge_ink.log(.out)` (70 checks, 57 failures: no keys, no `amounts`/`fade_at`/`boxed`, no proposed values) and `b/testfirst-test_look`, `-test_time` (green: their new expectations only apply once the keys exist). After: `test_edge_ink` 86 -> **204 checks, 0 failures** (`b/after6-test_edge_ink.log.out`), `test_look` 601, `test_time` 377, `test_warmup` 45, all 0 failures (`b/after4-test_look.log.out`, `b/after4-test_time.log.out`, `b/after4-test_warmup.log.out`; `b/after4-test_edge_ink.log.out`, 210 checks and 1 failure, is an intermediate run). Full suite on the final code (`b/suite.sh` a, b, v; 05:40-05:49, AC 100%, Low Power Mode 0, `b/suite-power.txt`): **45 suites, 23,759 checks, 0 failures** (`b/suite/suite_all.out`; 23,641 + 118; the run before the weak fill was dropped, 23,765, is in `b/suite-r4/`). `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`. creature_reactions rerun 5,583, 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 495: - **Changed expectations:** `test_edge_ink`: priority 108 -> -2; the tautological "0 or 2" check -> "a 0.064 m crease takes the rim"; the tap-count check drops its total (the line and opening searches add conditional taps); the stale "near side" comment fixed. `test_time`: P1c's toe, black, shadow_tint are found after `outline` instead of being the last three keys. `test_look`: its independent Glade mirror adds the three keys. **Expected journey count deltas: none** (keys 0: `b/metrics/counts.txt`, pre-tree -> live identical on all 11 views). With `--edge-ink` on: +1 draw call and rendered object on every view, the same as P9a's pass (pre-on -> on identical). 496: - **Captures** (`b/caps.sh` 04:56-05:06 and `b/caps2.sh` 05:36-05:40, the `on` sets recaptured on the final code; the first `on` sets are in `b/old-r4/`; 1600x1000, 100% charged and Low Power Mode 0 in `b/cap-power.txt`, which has no power-source line, so AC is implied, not recorded): `pre-1`, `pre-2` (pre-tree, default), `off-1`, `off-2` (live, default), `on` (live, EDGE_PROPOSED at each view's hour via `--edge-ink`), `pre-on` (P9a code, same values), `grey-off`, `grey-on`, `outline-05`, `outline-0` (on, luma outline x0.5 and x0), `mask`; views quarter-0545, -2045, -2100, quarter-night, -golden, -day, harbour-night, lane-night, grove-day, photo-golden (perspective lens), glade-day (Glade, 0.105/0.08 passed by hand because the override bypasses Glade's scaling). Plus `on-ink075`, `on-ink090` (night ink 0.75 and 0.9). 497: - **Owner sheet for D30** (`b/sheet/`): `d30-off-vs-on.png` (quarter night, golden, day, harbour-night, lane-night, grove-day; colour and greyscale), `d30-twilight-lens-glade.png` (05:45, 20:45, 21:00, the perspective lens, Glade), `d30-outline-1-vs-05.png`, `d30-outline-1-vs-0.png` (the three outline levels with ink on), 2x crops `crop2x-qn-terraces(-grey).png` (off | on | on with outline 0), `crop2x-hn-island.png`, `crop2x-qd-tower.png`, `crop2x-glade.png`, `crop1x-photo-golden.png`. Looked at: the outline sheet, the terraces, island, tower, Glade and lens crops. By eye: the ink makes the lit Harbour island read noticeably darker at night (`crop2x-hn-island.png`), Glade barely changes, the lens path inks the in-focus blossom and railings with no artefacts. 498: - **Acceptance:** 499: - **05:45 and 20:45 unchanged within noise:** met (keys 0, pass hidden). Frame >8 luma pre -> live 5.08/1.67% (05:45), 2.43/1.00% (20:45), 2.23/0.78% (21:00) against noise pre 5.58/3.03/4.10% and live 2.69/0.33/2.80% (`b/metrics/frame-default.txt`; every view's before -> after legs sit within the larger noise leg; glade-day 0.38/0.39 against 0.37/0.35, TIME phase). Counts identical. 500: - **Night iron and ink at or below 0.05 (luma 0-1, after P1c's toe):** iron met, ink line not met at the proposed 0.55. `b/inkluma.js` on the pixels the mask marks as ink: quarter-night iron_rail p50 0.039 (64% at or below 0.05; off 0.077); whole-frame ink p50 0.072 quarter, 0.101 harbour, 0.105 lane-night. At ink 0.75: 0.057/0.077/0.080 (iron 84%); at 0.9: 0.048/0.058/0.060 (iron 97%). The lamp halos and drops now draw over the ink (priority -2) and far ink fades with the fog, so the p90s stay 0.16-0.28. Not pushed: 0.9 reads as a solid black outline; the night ink level is a D30 question. 501: - **The glass decision:** pass below the glass. At -2 the glass and window panes (-1) draw over whatever ink and rim lie behind them; not isolated in a capture (no glass balustrade in these views). 502: - **P9a's open items:** (1) rim specks: rim pixels in 1-2 px components 44.2 -> 32.7% (quarter-night frame), 23.4 -> 10.1% (harbour), 43.9 -> 32.0% (quarter-day); rim in the crown rects 1.23 -> 0.86% (quarter), 3.35 -> 1.94% (harbour); cones 5.78 -> 4.00% (`b/metrics/mask-regions.txt`, mask in art pixels, `b/m.js`). **Not met in full:** what remains inside crowns is still specks (82%), the shingle cones keep short vertical rim ticks by day (`b/sheet/crop2x-qd-tower.png`), and the 2:1 dashes are unchanged. (2) railings: ink 13.33 -> 6.03% (quarter railings), iron_rail 16.57 -> 13.06% (`b/look/final-roofrail-qn.png`, P9a | P9b masks: the openings' inner rings are gone; the remaining ink is the near side against the sky and the frame's outer edge). Ink speck share rose 2.9 -> 5.2% (single pixels the opening rule leaves). (3) depth fade: in the shader and tested (range, curve, density follow the fog); its pixel effect was not isolated. (4) handover: the rim x `SunPath.handover` (tested at 13:00, 20:45, 21:00, 06:10, 01:00). (5) draw order: BAR sign neon core (pixels with off luma >= 200) 245.0 -> 244.8, 4 of 762 changed by >8; FISH MARKET core 230.9 -> 232.4, 140 of 1664 changed (P9a code 136) (`b/core.js`), the same under P9a's 108 and P9b's -2, so what still moves is the rim and ink on the signs' own geometry, not draw order. The bar_sign region p50 70.6 -> 76.3 (P9a code 76.9) comes from the box's non-core pixels. Grips: below them by priority (tested), not captured. (6) perspective lens captured (`photo-golden`, `b/sheet/crop1x-photo-golden.png`): ink on the in-focus blossom and railings, no artefacts, DoF over it. (7) housekeeping done (pixel parity, comment, tautology); GPU cost: the bench below is wall-frame time only; no GPU timer query exists in the GL Compatibility path, so per-pass GPU time is not measured. 503: - Sky and water untouched: 0.00% ink and rim in zenith, horizon_band, sky_right and harbour water_open in the masks. The colour frames' sky regions change with the pass on (`tools/art_metrics.js ab off-1 on`, share changed by >8): quarter-night zenith 38.3%, horizon_band 6.3%, sky_right 13.2%; harbour-night zenith 22.9%, horizon_band 1.7% (off-1 -> off-2 noise >8: 0.15-0.95%). In P9a code (`ab pre-1 pre-on`) harbour-night zenith is 22.7% but quarter-night zenith only 2.0%. (`b/metrics/protected.txt` prints the any-change share, not the >8 share.) The cause is cloud dither and drops at a later capture time, not the pass, whose mask draws nothing there (`b/look/zenith-off-on.png`). 504: - **Bench** (`b/bench.sh`, form of P9a's: preon = pre-tree with `--edge-ink=0.55,0.22` (P9a's pass), off = live default, on = live with the same override; interleaved, round 0 discarded, rounds 1-3, 1920x1080, vsync off, rain 0.45, `--motion-live --light-pin=20`, 01:00, every run on try 1, 491 samples; AC 100%, Low Power Mode 0, `b/bench*/power.txt`; summaries by `b/bsum.js`). Final code (`b/bench2-t1.out`, 05:26-05:32): dense p50 off 14.595, P9a pass 14.644, P9b pass 15.114; **paired on - off +0.474/+0.621/+0.478 ms p50, p95 +0.52/+0.82/+0.52**; on - P9a pass +0.33 to +0.49. Harbour on - off p50 +0.15/+0.17/+0.15. Draw calls +1 with the pass on (2587 -> 2588, 917 -> 918), none by default. Attribution (`b/bench-nogrid.out`, the rim grid bypassed, rounds 1-2): dense on - off +0.12/+0.14, so **the line rule's 5x5 grid costs about 0.35-0.4 ms on dense**. With the pass on, **P9a's cost line (at most 0.4 ms on dense) is exceeded**; with the default keys 0 nothing is drawn and nothing is spent. Not optimised in this part (time limit); `LINE_MIN` 0 removes the grid and its cost and gives back P9a's specks. GPU-only time is not measured (no timer query in this path). 505: - **Not done:** making the rim grid cheap enough for the 0.4 ms line (a two-stage test reading the 3x3 ring first, or a half-resolution crease map); the 2:1 dashes (needs the diagonal taps on every pixel; P9a measured that doubling crown rims); the cone ticks and the remaining crown specks; no 13:00 bench; grips not captured; Windows cost unmeasured; the night-ink level left to D30 (0.55 proposed, 0.75 and 0.9 measured). `build/backup-feel/27-P9-b/` holds the final P9b code (6b809b1); the pre-P9b state (26d33df) is only in `27-P9-a`. 506: - Review: filled by the orchestrator. 507: - **Review fix 1 (8 October 2026, Opus 5.5 on this Mac; evidence `b/fix*`, `b/bench3`-`b/bench6`, shader backups `b/backup/fix1/`).** The review's one must-fix: with the pass on, the line rule's 5x5 grid broke P9a's "cost at most 0.4 ms on dense" (+0.47/+0.62/+0.48). Only `shaders/edge_ink.gdshader` changed: `rim_line()` now takes fragment()'s centre and four neighbour depths and first reads the four 3x3 corners, which allow four of the grid's crease tests (x test of the pixels above and below, y test of those left and right); when `LINE_MIN` (2) of them hold it returns true, otherwise it counts the full grid as before, reusing the nine taps. The answer is the full grid's in every case (distinct grid pixels, the count only rises). Tried and dropped: a lazy loop reading each tap on demand (`b/bench3-t1.out`: dense +0.954/+0.881/+0.783, worse) and a second short stage over the corners' x tests (`b/bench5-t1.out`: +0.55/+0.36/+0.49, no gain). Bench, same form as above (01:00, AC 100%, Low Power Mode 0, `b/bench4/power.txt`, `b/bench6/power.txt`): final code `b/bench6-t1.out` (06:20-06:24) dense on - off p50 **+0.366/+0.383/+0.488**, p95 +0.49/+0.35/+0.57; the same code at 06:03 (`b/bench4-t1.out`) +0.350/+0.437/+0.427; six rounds mean **+0.41**, median +0.405 (was +0.52). Harbour +0.12/+0.12/+0.08. Draw calls unchanged (2587 -> 2588 with the pass on). **The 0.4 ms line is at the line, not under it: 3 of 6 rounds above 0.4.** Recorded, not pushed; with keys 0 nothing is spent; `LINE_MIN` 0 still brings it to about +0.13 with P9a's specks back; a half-resolution crease map is the next step. Mask parity (`b/fix2-mask/`, quarter-night, harbour-night, quarter-day; `b/cls_diff.js`): pixels whose ink/rim class changed against `b/mask` 0.012/0.015/0.003%, against the noise pair `b/r5-mask` vs `b/mask` 0.018/0.013/0.003%; rim speck shares 32.7/10.2/32.0% as before (looked at `fix2-mask/quarter-night/courtyard.png`). Tests: `test_edge_ink` 204 checks, 0 failures (`b/fix2-test_edge_ink.log.out`; no expectation changed: the rule's text checks still hold). Full suite on the final code (`b/fix-suite.sh` a, b, v, 06:12-06:20, `b/fix-suite-power.txt`): **45 suites, 23,759 checks, 0 failures** (`b/fix-suite/suite_all.out`), `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`, creature_reactions rerun 5,583/0. Journey count deltas: none (keys 0). Windows not run. 508: - **Review fix 2 (8 October 2026, 06:27, Opus 5.5 on this Mac): not fixed, left open for the owner at D30.** The recheck kept one must-fix: with the pass on, dense on - off p50 is still +0.41 mean, +0.405 median over the six rounds above (`b/bench4-t1.out`, `b/bench6-t1.out`), 3 of 6 rounds above 0.4. The round-to-round "off" p50 moves about 0.11 ms, so the 0.01 ms excess is inside bench noise: **P9a's "at most 0.4 ms on dense" line is neither met nor failed with the pass on.** No shipped state pays it: the default keys are 0, so the pass draws nothing. Nothing was changed. This fix round started at 06:27, and the owner's limit allows no Godot run to start after 06:30, so no change could be benched in time. Both remedies are also outside an implementer's fix. `LINE_MIN` 0 (about +0.13 ms) brings back P9a's rim specks and undoes tuning item 1. A half-resolution crease map is a design change beyond this part. **Owner decision with D30:** turn the pass on at this cost (about 0.4 ms on dense at 1920x1080, GPU-only time unmeasured, Windows unmeasured), or keep it off until a half-resolution crease map is built and benched, or take `LINE_MIN` 0 and accept the specks. The tree is unchanged since the full suite above (shader last written 06:11:48; suite 06:12-06:20, 45 suites, 23,759 checks, 0 failures). No new tests, captures or benches. Journey count deltas: none. 509: - Platforms: Mac run; Windows not run. 511: **Beautify P9a Edge ink and rim pass (depth, art resolution) (8 October 2026):** Opus 5.5 on this Mac (plan section "P9a"; report Signature E "edge_ink.gd cloned from ContactShade at priority 108, premultiplied blend, 5 depth taps at art resolution, second differences" and the pixel lens's revised finding "Ink and light edges"; step `27-P9`, first part; a measuring part, second review required). A new full-screen pass reads the depth buffer at each art pixel and its four neighbours and draws a dark ink line (toward `0c0a18`) at silhouettes whose depth jump is over max(8 mpp, 0.8 m), and a lit rim (the hour's key colour lifted 30% toward white) on convex creases from max(1 mpp, 0.06 m) to 6 mpp: copings, sills, quay lips, stair nosings, ridges. **It is OFF by default** (ink and rim 0, node hidden, no draw call) until P9b puts the keys into the lighting profile and the owner answers D30; `--edge-ink=` turns it on for captures and benches (never saved). Snapshot: `build/backup-feel/27-P9/` (made by the orchestrator, = `main` 8ccf61a, the P8-IV close code); at part start `diff -rq` on scripts shaders tests data scenes tools and `project.godot` against it was empty: nothing unexpected. Pre-part scratch tree `artifacts/beautify/27-P9/a/pre-tree` (`feel_snapshot.sh tree 27-P9`). Evidence: `artifacts/beautify/27-P9/a/`; file backups in `a/backup/`. 512: - **Code:** 513: - `scripts/edge_ink.gd` (new; no `class_name`, preloaded like ContactShade): camera-pinned 2x2 quad, `PRIORITY` 108 (before ContactShade 110, CloudShadow 115, RainVeil 117, HeightFog 120), `INK_COLOR` 0c0a18, `INK_MPP` 8, `INK_MIN` 0.8, `RIM_MPP` 1, `RIM_MIN` 0.06, `RIM_TOP_MPP` 6, `RIM_LIFT` 0.3; `configure(ink, rim, show_mask)` (visible only when ink or rim > 0), `tint(profile)`, `light(toward_key)`; CPU twin `ink_threshold`, `rim_range`, `classify`, `pixel`, `parse_override`. 514: - `shaders/edge_ink.gdshader` (new): unshaded, `blend_premul_alpha`, depth test and write off, no TIME. Taps one art pixel apart (`1.0 / VIEWPORT_SIZE`; the view renders at art resolution). mpp = the global `art_pixel` (P8a), x depth under the perspective lens (`art_pixel_slope`). The centre pixel is skipped on the sky panel and under the water plane (ContactShade's tests). 515: - `scripts/world_view.gd`: `edge_ink` made beside the contact shade; `configure` in `sync_environment` (0,0 unless `debug_edge_ink`), `tint` and `light(key_light.transform.basis.z)` in `sync_profile`, `backdrop_depth` in `_apply_depths`. `scripts/main.gd`: `--edge-ink=` and test-only `--edge-ink-mask` (ink pure red, rim pure green, opaque: the measuring view). `tools/test.sh`, `tools/Test.ps1`: `test_edge_ink` added. 516: - **Deviations from the lens design, all measured on probes in `a/probe*/`, `a/an/`:** (1) **Ink goes on the far side of a silhouette** (the pixel behind the edge), the near side only where the far side is sky or water: the lens's near-side ink inked every festoon bulb and lamp head whole (`a/an/probe-qn-map.png`, `probe-lamps.png`). (2) A one-pixel gap between two nearer things, a one-pixel sliver against the sky, and **the background of a one-pixel sliver** are not inked: without the last rule every terrace railing panel was inked solid (`a/probem1` vs `probem2`, `an/probe-rail.png`). That rule takes one extra tap two art pixels out, only on far-side candidates (short-circuit, `textureLod`), so the pass is "5 taps plus up to 4 conditional", not 5. (3) The lens's planar-prediction ink test is kept as "jump over t AND second difference over t" (the revised finding). (4) A diagonal-crease rule (both axes' second differences summed) was tried to mend dashed rims on 2:1 copings and reverted: it did not close the dashes and doubled crown rims 1.24% -> 2.66% (`probem3`). 517: - **Tests:** test-first `a/testfirst-test_edge_ink.log(.out)` (11 checks, 11 failures: no script, shader, functions or `--edge-ink`), `a/testfirst-test_warmup.log(.out)` (45, 1). After: `test_edge_ink` 86 checks, 0 failures (`a/after6-test_edge_ink.log.out`): priority and order, constants, thresholds by mpp, the twin on flat faces, grazing ground, far/near side, gaps, slivers, sliver backgrounds, crown shell gaps, proud brick 0.025 and tile lap 0.035 m (no rim), coping and sill creases (rim), parsing and clamping, the shader contract (render modes, depth texture, `art_pixel` globals, tap count, sky/water skip, constants equal the twin's, no TIME, nothing drawn at 0,0, mask off by default), and the view (camera child, hidden by default at night, by day and in Glade, override on, rim colour = key lifted 30%, key direction, `--edge-ink=0,0` hides it, backdrop depth like the fog, no environment key, saved scene unchanged). `test_warmup` 44 -> 45 (the quad is warmed). Two of my own new checks were corrected while running (`a/after1-*`: "no TIME" read comments; `base_height` is never set on the contact shade, so the check compares shader defaults). 518: - **Full suite** (`a/suite.sh` a, b, v; 04:09:30-04:18:13, AC 100%, Low Power Mode 0, `a/suite-power.txt`): **45 suites, 23,641 checks, 0 failures** (`a/suite/suite_all.out`; P8-IV close 23,554 + 86 + 1). `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. creature_reactions rerun 5,583 checks, 0 failures; the known flaky check did not fire. 1440x900 not run (inherited stall). 519: - **Changed expectations:** none of the existing assertions changed (only new checks). **Expected journey count deltas: none** (the pass is hidden by default; `a/metrics/counts.txt`: before -> after identical on all three views in both pairs). With `--edge-ink` on: +1 draw call and rendered object (quarter-night 1024 -> 1025, quarter-day 1025 -> 1026, harbour-night 907 -> 908). 520: - **Captures** (`a/caps.sh 1 2`, 03:50-03:53, one sitting, 1600x1000, AC 100%, Low Power Mode 0, `a/cap-power.txt`): `before-N` (pre-tree), `after-N` (live, default = off), `on-N` (live, `--edge-ink` night 0.55,0.35 / day 0.35,0.22, the lens's values), `mask-N` (the measuring mask); views quarter-night, quarter-day, harbour-night. Sheet `a/sheet/off-vs-on.png` (colour and greyscale); crops for the reviewer `a/sheet/crop*` (1x crowns colour and grey; 2x brick, tiles, copings, quay; after | on | mask) and `a/look/` (4x crowns, cones, slate, brick, water). All looked at. 521: - **Pass off = unchanged:** counts identical; frame >8 luma before -> after 6.49/1.83% quarter-night, 1.17/1.08% quarter-day, 3.41/0.81% harbour-night against noise pairs before-1/2 5.90/0.28/1.28% and after-1/2 0.76/0.20/3.91% (`a/metrics/frame.txt`). quarter-day sits above both legs; its map `a/an/qd-b1-a1-map.png` (looked at) shows only rain drops and water rings (TIME phase), no structure. 522: - **Acceptance, measured on the mask (`a/an/maskcount.js`, `a/metrics/mask-regions.txt`; identical in both rounds):** 523: - **No ink inside crowns:** quarter `crowns` ink 1.93%, rim 1.24%; harbour `crowns` ink 1.76%, rim 3.31%. In the crops (`look/qn-crown1-6.png`, `hn-crowns-a.png`, `sheet/crop1x-qn-crowns*.png`) the ink in those rects is the crown's outer silhouette where a rect reaches the edge; no shell outlines inside. **Not met for the rim:** single-pixel rim specks sit inside crowns (lobe creases), 1.2-3.3% of the crown rects. For P9b. 524: - **No ink or rim on brick relief and tile laps (greyscale):** met. quarter `land_wall` ink 1.16% rim 2.82%, all on the tower's window reveals, sills and string courses (`look/qd-brick-grey.png`, `sheet/crop2x-qd-brick-grey.png`): no brick is outlined. Harbour `roof_slate` ink 0.11%, rim 1.04% (ridge and hip lines only, `look/hn-roof-grey.png`). **Harbour `cones` rim 5.83%:** the shingle cones' polygon facet edges take rim lines (`look/hn-cones-grey.png`; the quarter's tower cone too). Not tile laps, but not wanted either. For P9b. 525: - **Rims on copings, sills and quay lips:** met in kind: every parapet coping, sill, quay lip, stair nosing and ridge in the three views carries the rim (`sheet/crop2x-qn-copings.png`, `crop2x-qn-quay.png`). **On 2:1 diagonal edges the rim breaks into dashes** (one or two px on, one off); not mended in this part (see deviation 4). 526: - **Water and sky untouched:** met. 0.00% ink and rim in `zenith`, `horizon_band`, `skyline`, `sky_right`, `mirror_seam`, harbour `water_open`; `mirror_columns` 0.66/0.90% and `far_bay` 1.12/1.25% are the boat and the tower standing in those rects (`look/qn-mirrorcols.png`, `look/hn-farbay.png`): no water pixel is touched, the neon columns are intact. 527: - **Cost at most 0.4 ms on dense:** met (`a/bench.sh`, `bench-t1.out`, `bench-t13.out`, `bench-summary.txt`; pre-tree / live off / live on, interleaved, round 0 discarded, rounds 1-3, 1920x1080, vsync off, rain 0.45, `--motion-live --light-pin=20`, every run on try 1, 491 samples; 03:55-04:06, AC 100%, Low Power Mode 0, `a/bench/power.txt`). Dense 01:00 p50 14.537 off -> 14.596 on (paired +0.059, 2/3), p95 +0.128 (3/3); 13:00 p50 +0.045 (2/3), p95 +0.084 (2/3). Harbour p50 +0.038 (3/3) and +0.061 (3/3). Draw calls +1 (2587 -> 2588, 917 -> 918). Harbour p95 spreads 7.1-8.7 ms in every tree (off - pre +0.99 at 01:00 although the hidden pass draws nothing); a rotated-order rerun (`bench-order-t1.out`, off/on/pre, rounds 1-3) gave Harbour p95 pre 7.49/8.51/8.43, off 7.40/7.96/8.69, on 7.64/7.49/7.52: the slow p95s fell on both pre and off (the highest, 8.688 ms, on off round 3), so order and noise, not the pass. 528: - **For the owner sheet (P9b / D30), with the pass on:** the rim is strong at night at 0.35 (bright dashes on dark stone, `sheet/crop2x-qn-quay.png`); protected neon regions move: `bar_sign` p50 70.6 -> 84.4 and `fish_sign` 66.2 -> 65 (the cabinets' edges take rim and ink), `lamp` p50 54.4 -> 50.7 and 44.2 -> 41.9, `blossom_lamp` -1.2 to -3.1 (`a/metrics/ab-a1-on1.txt`). Glass balustrades were not isolated in these views (the plan's risk stays open for P9b). 529: - **Not done:** no rim fix for 2:1 dashes, crown specks or cone facets (left for P9b's tuning, not pushed here); Glade x0.3/x0.4 and the profile keys are P9b; no lane, grove or golden capture; no perspective-lens capture (the `art_pixel_slope` path is code-read only); Windows cost unmeasured. Post-part snapshot `build/backup-feel/27-P9-a/`. 530: - Review: filled by the orchestrator. 531: - Platforms: Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy = HEAD 8418d4e, 9 October 2026, 19:36 BST), lines 302-328 302: ## P8-IV close finished; P9 started (8 October 2026, 04:00, under the standing approval) 304: P8-IV close (main 6f706a9 with code 8ccf61a, snapshot `26-P8-IV`, sheet `artifacts/beautify/26-P8-IV/close/sheet/`, 8 pages sent in chat): 44 suites, 23,554 checks, 0 failures; +1 draw call on rainy views (the veil), stage cost -0.01 to +0.04 ms p50. Treated as approved as built under the standing approval until 08:00; the owner has not seen the sheet. Open for the owner: veil reads as sky and far-water rain only at Quarter zoom, lit rain faint in plain frames, the 300-drop near box (about 2.2x P8d's drops at light rain, model estimate), 05:45/20:45 horizon changed shares above noise, rain-ring phase differences, Glade with rain turned up gets the veil. P8 is done. P9a (edge ink and rim, second review required; pre-part snapshot `27-P9` = code 8ccf61a) is launched with the rule that the whole part must finish by 06:45; ink and rim stay OFF by default (D30 is the owner's gate after P9b). 306: ## P9a and P9b done; stopped for the morning (8 October 2026, 06:35, under the standing approval) 308: P9a (main 26d33df, snapshot `27-P9-a`) and P9b (main 6b809b1, snapshot `27-P9-b`) are committed locally. Edge ink and rim exist as a pass at render priority -2 (below glass), keys `edge_ink`, `edge_rim`, `edge_rim_color` in all ten anchors with default 0 (hidden, no cost; draw calls unchanged by default); proposed on-values are in `EDGE_PROPOSED`. D30 is the owner's decision (sheet sent in chat: `artifacts/beautify/27-P9/b/sheet/`): (a) ink and rim on by default, (b) which luma-outline level (x1, x0.5, x0), (c) the cost: with the pass on, dense p50 is about +0.41 ms (line 0.4; borderline, inside bench noise; Harbour +0.1), remedies are a half-resolution crease map (a P9 follow-up) or LINE_MIN 0 (+0.13 ms, rim specks come back). Not met and left to D30: night ink line at 0.55 is 0.07-0.10 luma against 0.05 (ink 0.9 reaches 0.05-0.06 but looks solid black). Open: 2:1 dashes, cone ticks, leftover crown specks (82% of crown rim), grips only covered by the priority test, GPU-only time and Windows unmeasured. Default if no answer: keys 0 (hidden). The orchestrator started no P9c: it cannot finish before 07:30. The standing approval ends at 08:00; no Godot runs. P9c (art-resolution post), P9d (snap), P9e (smooth scroll, second review), P9 close and D31 (live pan session on the Mac) remain. 310: ## Owner: "Continue please up until the close" (8 October 2026, about 09:10) 312: Read as: run P9c, P9d and P9e in order (each with its review, fix pass and handoff check, committed locally with a snapshot) and stop before the P9 stage close, which needs the D31 live pan session on this Mac (owner present) and the M6 exports. The 08:00 laptop limit no longer applies (the owner is back and asked for work to continue, so windowed Godot runs are allowed). Open items P8-IV sheet and D30 are unanswered and do not block P9c to P9e; the defaults stand (D30: keys 0). Parts that clear review are carried forward as built with their open items listed for the P9 sheet; no push, no deploy. 314: ## Orchestrator ruling on P9d (8 October 2026, 10:55, owner not asked; owner said "continue up until the close") 316: P9d (camera snap, off by default; main = the P9d commit, snapshot `27-P9-d`) meets placement and snap-off byte-identity but only partly meets "static pixels stay or move by whole art pixels": positions move by whole pixels (23 of 23 pairs), but shading changes on 1-px grid steps (8-49% of pixels, 1.1-7.8% by more than 8 levels), from the contact shade's dFdx/dFdy normals and the surface bumps depending on 2x2 pixel-quad parity (odd shifts 28-31%, even 5%). Ruling: accept as recorded and continue to P9e with a 1-px grid (no design change before the owner sees it); the owner judges it in the D31 live pan session. Options after D31: a 2-art-px grid (day x steps 0.10% over 8 levels, night 1.1-3.2%; doubles the step and the residual range) or quad-stable derivatives in `contact_shade.gdshader` and `surface_detail.gdshaderinc`. 318: ## Orchestrator note after P9e (8 October 2026, 12:05, owner not asked) 320: P9e (smooth scroll by the residual, integer-exact rect; snapshot `27-P9-e` holds the P9e code, commit 2c01eb0) passed two reviews with no must-fix. The camera snap is now ON by default in a person's live editor (`CameraRig.SNAP_LIVE = true`, as the plan's "off by default until P9e lands" says); automation (smoke, captures, tests) keeps the old camera, so byte comparisons hold; reverting is one line. A person therefore sees P9d's accepted shading-parity shimmer (8-49% of pixels change on each 1-px grid step) in the live editor; the D31 live pan session is where the owner judges it, and the options after D31 stay: a 2-art-px grid or quad-stable derivatives, or `SNAP_LIVE = false`. Open and left as built: frame_margin is also applied under the perspective lens and photo mode's free camera (about 0.4% less world shown there), the 3D view is ceil(window/pixels) so the extra pixels are cropped at the right and bottom at sizes the pixel size does not divide, smooth-pan filmstrips only at pixel size 2, the real mouse path only covered headless, margin cost and vignette movement not measured, Windows not run. The P9 stage close and D31 wait for the owner; the owner asked me to continue "up until the close", so I stop here. 322: ## D31 live pan session done (owner, 8 October 2026, in chat) 324: The owner ran the live editor on this Mac (final build, camera snap on, smooth scroll) for the pan session and said "all looks good". Read by the orchestrator as: D31 answered, the camera snap stays ON in the live editor as built (`SNAP_LIVE = true`, 1-px grid), with P9d's shading shimmer, the 1-art-px margin and the open P9e items accepted; no 2-art-px grid, no derivative fix. The owner did not mention picking or odd window sizes separately, so those rest on the automated probes. The P9 stage is therefore finished. M6 export started afterwards (script `artifacts/beautify/27-P9/m6/m6_export.sh`, outputs on the Desktop under M6 names). D30 (ink and rim default off) and the P8-IV sheet verdict remain unanswered. 326: ## D30 answered and P8-IV approved (owner, 8 October 2026, in chat) 328: "ink off p8iv approved". Read by the orchestrator as: (1) D30: ink and rim stay OFF by default (keys `edge_ink`, `edge_rim` stay 0 in all anchors; the pass stays hidden and costs nothing; `--edge-ink` and `EDGE_PROPOSED` remain for tests and later use; no outline level chosen, the open ink items 2:1 dashes, cone ticks, crown specks and the night ink line are not pursued); (2) the P8-IV sheet is approved as built (rain veil, lit rain with the 300-drop near box, the open items on that sheet left as built). No owner decision from the beautify plan is pending now except the optional P10 polish backlog, journal publishing and pushing to GitHub, none of which has been asked for.