Chapter 29: Finer wires and a replanted harbour 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-II of stage P10 (P10e cables as 1-px lines, P10g Harbour replanting mock, P10m beacon as a tower param, P10k quay swags and rainy overcast mocks) and its stage close, then the merged P10-III parts: P10j speck guard (measured, not built), P10n art-pixel size follows the window (waiting for the owner), P10o pixel font and integer UI scale options (no download; waiting for the owner) and P10q, which applies the owner's P10-II answers. The P10-III stage close was still running at the freeze and is not covered. Owner answers: P10-II "All approved, beacon to be a placeable option" (9 October). 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 5-204 5: **Beautify P10q Apply the P10-II yeses: replant, swags, overcast, placeable beacon (9 October 2026):** Opus 5.5 on this Mac (plan section "P10q"; owner on the P10-II sheet: "all approved, beacon to be a placeable option"; step `28-P10`, sub-batch P10-III, last part before its close). **This part changes default looks**, each behind its own constant with a revert switch and tests that follow it: (1) **Replant** (`SampleScene.HARBOUR_MOCK = "replant"`; `""` brings the pre-P10g Harbour back byte for byte): P10g's option A is the default Harbour steps (starter, menu, the drag and release fixtures). One change against the sheet: the east-bank tree (from (-16,-11), 3.4 m) stood 2.8 m in front of the bank's lantern post at (14.6,-2.4) and its crown hid it (`look/east-cands-day.png`, `-night.png`, first panel); it now stands at `REPLANT_EAST` (16.5,-2.0), between the post and the bench, 1.5 m from the water, so the post shows by day and night and fewer reeds grow on the front edges (`sheet/p10q-5-east-bank-day-night.png`). Four other spots were captured and looked at (`probe/east-cap/`, `look/east-cBcD-x3.png`: (17.4,-2.8) and (17.6,-1.6) overlap the bench, (13.8,-6.8) leaves the front bare). `REPLANT_EAST_CLEAR = true` is its switch (false: the sheet's option A exactly, md5 `710cc572db9f0f7f7264a08520e4fc5a`); the beacon mock keeps P10g's spot. New default document md5 `8fb5346fbdd4442dab6c29c7926a4d64` (`probe/md5.log.out`). (2) **Swags** (`Composition.QUAY_SWAGS = true`): as P10k built them. Lamp 195 with no string is **accepted, no rule change** (a lone post between two swags is one of the dark gaps). Measured here: the west quay's lamp 243 is also left with no string (its one string 243:244 is a quay span over 244's limit); the P10k entry and sheet named only 195. Both are pinned in test_links. (3) **Overcast** (`WorldView.RAIN_OVERCAST = true`, strength A = P10k's constants, unchanged). **The rule, written down** (LightingProfile header): the profile is independent of rain; the view dims a rainy daylight as soft overcast, overcast = rain x (1 - windows), key x (1 - 0.4 oc), ambient x (1 + 0.25 oc), shadow blur x (1 + 2 oc); nights and dry days are exactly the profile; false restores "daylight is independent of rain". (4) **Placeable beacon** (`EditorUI.BEACON_OPTION = true`): a selected tower with a flat roof, and the tower tool, show a "Beacon" row (None / Lamp room and beam) under Roof, like the Roof, Windows and Stone rows. It writes P10m's optional `beacon` param through the ordinary parameter edit: one undo step, saved with the scene, version 2. Default off: choosing the tower tool starts at None; selecting a tower no longer inherits the last tower's value (`Main.select_id` erases `beacon` before the merge); a placed tower carries the key only if the row was turned on; no other kind ever carries it (`Main.object_candidate`); a moved beacon tower stays one. A pitched roof hides the row and the beacon (the key stays and is ignored, P10m). In no shipped scene or starter (test_beacon `shipped_checks`, 12 scenes: Harbour, courtyard, castle, styles, benchmark, the quarter, Amsterdam, the four demos and the blank canvas). **Light slots:** the beacon lamp is one source among the 16 engine light slots, and the editor refuses nothing and warns of nothing. The 16 best-ranked on-screen sources (nearest the view target for their energy) get a light; the others keep their lit glass without a pool. The beam is its own node and draws either way: with 24 extra lanterns, 205 sources, 16 lights, the beacon lamp held no slot and the beam was drawn (`suites/r1-test_beacon.log.out`). The beam is steady under Reduced and frozen under Off, as P10m built and test_beacon checks. Files: `scripts/sample_scene.gd`, `scripts/composition.gd`, `scripts/world_view.gd`, `scripts/lighting_profile.gd`, `scripts/editor_ui.gd`, `scripts/main.gd`, `tests/test_starters.gd`, `test_links.gd`, `test_time.gd`, `test_sun.gd`, `test_beacon.gd`, `test_ui.gd`, `test_foliage_merge.gd`, `test_release_spread.gd`, `test_warmup.gd`; no new class, suite or fixture. Snapshots: pre-part `build/backup-feel/28-P10-q-pre/`, saved before the first edit. `diff -rq` on scripts, shaders, tests, data, scenes, tools and `project.godot` is empty against `28-P10-o` and against the live tree, so nothing unexpected had changed (`q/diff-o-vs-q-pre.txt`, `q/diff-live-vs-q-pre-at-start.txt`; `main` 094397c, clean, `q/head.txt`, `q/status.txt`). Post-part `build/backup-feel/28-P10-q-post/`. Scratch trees: `artifacts/beautify/28-P10/q/pre-tree` (from `28-P10-q-pre`); `q/on-tree` (live code); `q/off-tree` (the four switches off); `q/east-off-tree` (`REPLANT_EAST_CLEAR` false); diffs in `q/switch-trees-diff.txt`, `q/off-tree-code.diff`, `q/east-off-tree-code.diff`. Evidence is under `artifacts/beautify/28-P10/q/` (below, `q/`); file backups in `q/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, drag run, bench and suite (`q/cap-power.txt`, `q/drag/power.txt`, `q/bench/power.txt`, `q/suites/power.txt`, `q/suite/suite_all.out`). 1-min load 2.38-2.92 during the captures, 2.35-2.94 during the drag runs and 2.37-3.81 during the bench. 6: - **Profiles and modes reached:** replant: every look that draws Harbour (canal day, golden, night and rain; Glade `--look=glade` on the same scene; photo mode and tilt-shift not captured), the drag and release fixtures. Swags: every scene with lamps along water (the quarter; Harbour has no quay span, 12 strings as before). Overcast: canal and Glade by day and at golden hour when it rains; nights, dry days and the profile data unchanged. Beacon: only scenes where a person turns it on. Reduced/Off motion: nothing new moves. 7: - **Tests:** test first (`q/testfirst/`, before any code change): test_starters 367 checks, 1 failure (no `REPLANT_EAST`/`REPLANT_EAST_CLEAR`; the rest of the new checks sit behind that guard); test_links 404, 2 (QUAY_SWAGS off; bare lamps `[195, 243]` against my first expectation `[195]`, which I then corrected to the measured pair); test_time 447, 1 (RAIN_OVERCAST off); test_ui 437, 1 (no `BEACON_OPTION`; the option checks then took the "not offered" path); test_sun 101/0 and test_beacon 242/0 (their new checks describe behaviour that already held). After: affected suites `q/suites/r1-*` (test_links 404/0, test_time 447/0, test_sun 101/0, test_beacon 242/0), `r2-` (test_starters 381/0, test_ui 461/0), `r3-`/`r4-` (test_release_spread 131/0, test_warmup 53/0, test_foliage_merge 92/0). New checks: test_starters `p10q_harbour_checks` (the switch, the pin, the default follows `HARBOUR_MOCK`, the default against the sheet's option A file differs only in the east tree, the post not hidden by any crown and the check sees P10g's spot hide it, no beacon in any Harbour); test_ui `p10q_beacon_checks` (above, 26 checks); test_beacon `shipped_checks`, `slot_checks`; test_links (switch on, `quay_swags_on()` follows it, bare lamps 195 and 243); test_time (switch on, the view follows the switch with no override). **Off position** (`q/switch.sh`, logs `q/switch//`): `off-tree` test_starters 379/1, test_links 395/1, test_time 447/1, test_ui 431/3, test_sun 101/0, test_beacon 242/0, test_foliage_merge 61/0, test_warmup 53/0. The only failures are the four pins on the shipped values ("is ON" / "replant"), plus test_ui's two chemistry-fixture checks, which fail the same way in `on-tree` (scratch trees have no `res://artifacts`). test_release_spread hits script errors on the missing fixtures in both trees, so a scratch-only copy that runs Harbour alone (`test_release_spread_harbour.gd`, in the scratch trees only) ran: 60 checks, 0 failures in both. `east-off-tree` test_starters 377/1 (the pin only). **Full suite** (`VERIFY_RES=2880x1800 tools/suite_all.sh q/suite`, 19:12-19:21): **47 suites, 26,356 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (25 s), `PIXELPUNK_INTEGRATION []`. test_creature_reactions 6,871/0, rerun 6,871/0, so the known flaky check did not fire. Its count rose from 5,583 although its file is unchanged: its loops run over Harbour's parts. 1440x900 `--verify` not run (inherited stall). 8: - **The stage gap, understood before changing expectations:** tree growth T3 builds at most one stage mesh a frame, and a group with no built stage to borrow waits undrawn (designed and tested: test_tree_grips "A sync builds at most one stage mesh", even on a forced sync). The replanted Harbour has six such plants (trees 85, 87, 90, 91 and conifers 137, 138), so after a fresh `sync_scene` they get no decoration slot until `tick_stage_builds` has run 7 frames (0-8.7 ms each) (`q/probe/stage_probe.log.out`; the pre-P10g Harbour waits for none). This is not a bug in state: in the app the start-up screen covers it, and from the menu the six plants appear over about 7 frames. Changing it would reverse T3's tested pacing, so it is left as is. The three tests that read a view without running frames now let it build the stages it waits for first. These six plants are what P10g's three "not just counts" failures were. 9: - **Changed expectations (old -> new):** test_starters: "Harbour steps is byte for byte the pre-P10g document" now reads `harbour("")`, not the default; "HARBOUR_MOCK stays \"\"" is removed and replaced by "is \"replant\"" (P10q); "harbour() follows HARBOUR_MOCK" compares the default with `harbour(HARBOUR_MOCK)`. test_beacon: the P10m pins `harbour` (cluster signature) and `harbour_doc` are now of `harbour("")`, which P10m measured. test_links: QUAY_SWAGS false -> true. The checks labelled "off" ask for override false instead of null (null now means on): the quay-row chain, flags, lawn, the lived-in signature, Harbour off, quarter off. "the override false is the switch's off" -> "no override is the switch's own setting". test_time: RAIN_OVERCAST false -> true; the overcast checks' "off" asks for `debug_overcast = false` (was null). view_checks: key = the hour's x (1 - 0.4 oc) and fill x (1 + 0.25 oc). twilight_checks: the anchor, preset, crossing and handover ambient each x (1 + 0.25 oc), with oc from `overcast_of`, which follows the switch (with the courtyard's rain 0.35, oc > 0 by day). test_sun: "The fill follows the turned sun's elevation" x (1 + 0.25 oc). test_foliage_merge: Harbour foliage MultiMeshes `<= 34` -> `<= 40` (32 -> 38: four new tree sizes and two conifers draw from their own stage meshes, `q/probe/mm_probe.log.out`); the slot check runs after the view has built its pending stages. test_release_spread "harbour animated, drained": both views build their pending stages before the compare. test_warmup `world_of`: the view builds its pending stages before the counts (the counts were 167/99 after the warm-up against fewer before it, because the scene's own stage builds landed during the warm-up's frames, not a warm-up leak). Counts: test_starters 367 -> 381, test_links 402 -> 404, test_time 440 -> 447, test_beacon 235 -> 242, test_ui 435 -> 461, test_foliage_merge 89 -> 92, test_release_spread 130 -> 131, test_creature_reactions 5,583 -> 6,871 (data-driven); suite 25,008 -> 26,356. 10: - **Expected journey count deltas** (from `q/metrics/counts-pre1-live1.txt`, same views and cameras as the journey set): harbour-day draw calls and rendered objects +24 (912 -> 936), harbour-night +24 (911 -> 935), harbour-lane +17 (753 -> 770); on all three: parts 136 -> 138, foliage plants 221 -> 235, foliage total 661 -> 675, light_sync_calls 1 -> 2, pool_resets 0 -> 1, pool_kept 0 -> 16. quarter-day, quarter-night and styles-day identical (the swags and the overcast change no integer count). P10g saw the same extra light sync on its options (not traced here either). **Drag baseline re-based** (`q/drag.sh`: `tools/feel_drag.sh`'s Harbour run without its pkill, 1600x1000, 3 runs each, pre and live alternating 18:49-18:51 (the last run started 18:50); folded by `tools/feel_drag_report.js` into `q/drag/pre/` and `q/drag/live/`). **`q/drag/live/` (README.md, baseline.json) is the new Harbour drag baseline.** Replay errors none; every run ends on one scene, digest f0459cd73da0 -> 6fcf1d5f0538. Step total median / p95 pre -> live: handle 17.9/52.5 -> 18.7/53.7; slider 10.9/51.3 -> 17.5/53.7 (slower in 3/3 runs, 10.0-11.7 -> 13.5-17.6; probably the shrubs and trees moved beside the main house the script resizes, not traced); terrain raise 127.2/149.6 -> 128.2/146.1; wall stroke 8.6/10.8 -> 8.8/12.8. The dense and cluster6 drag fixtures do not use Harbour and were not rerun. 11: - **Captures** (`q/cap.sh` stages a, b, c, `tools/art_capture.sh`, 1600x1000 = art-pixel factor 1, so P10n's default "follow" equals `--pixels-follow=off`; one sitting 18:51-18:58, every view on try 1): `pre-1`, `pre-2` (pre-part tree: before and its noise pair), `live-1`, `live-2`. Views: harbour-day, -golden, -night, -lane, lane-night, quarter-day, -night, -golden, styles-day, glade-day, harbour-day-rain9, quarter-day-rain9 (`--rain=0.9`), qn-front, qn-west (P10k's quay cameras). Sheets, all looked at: `q/sheet/p10q-1-harbour-day-golden.png`, `-2-harbour-night-rain.png`, `-3-quarter.png`, `-4-quarter-swags-rain.png`, `-5-east-bank-day-night.png`. They show the planted lawns, conifers and east bank; the quarter's front-quay garland thinned to swags; and a rainy day a little dimmer and greyer. Night Harbour looked darker in a downscaled tile; at 2x it is the same, and the warm lawn pool reads p50 71.1/71.0 pre and 70.9/71.1 live (`look/night-pool-pre1-live1.png`). The pre-1/live-1 mask shows only trees, rain and smoke (`masks/harbour-night-pre1-live1.png`). 12: - **Acceptance, measured:** (1) every change behind its own constant with a revert switch and a test that follows it: `HARBOUR_MOCK`, `REPLANT_EAST_CLEAR`, `QUAY_SWAGS`, `RAIN_OVERCAST`, `BEACON_OPTION`, with the off-position runs above. (2) Before/after sheet for Harbour (day, golden, night, rain) and the quarter: `q/sheet/`. Frame changed >8 luma, pre-1 -> live-1 against the noise pairs pre-1/pre-2 and live-1/live-2 (`q/metrics/frame-pre1-live1.txt`, `noise-*.txt`): harbour-day 4.84% (noise 1.29/1.71), golden 4.94% (1.51/1.90), night 4.13% (0.49/1.15), rain9 12.52% (1.01/0.88), lane 5.50% (1.73/1.73), glade-day 4.54% (0.27/0.38); quarter-night 1.69% (0.27/0.45), qn-front 2.58% (0.40/0.34), qn-west 1.52% (0.16/0.30), quarter-day-rain9 11.93% (0.19/0.57), quarter-golden 1.43% (3.15/4.01, within noise), quarter-day 0.74% (0.66/0.26), styles-day 0.81% (0.22/0.39). Overcast (`q/metrics/overcast-stats.txt`, P10k's regions): harbour-day-rain9 sky 202.8 -> 198.4, the same as P10k's mock; quarter-day-rain9 spread 141.4 -> 138.9 (mock 138.8), sky 177.5 -> 173.0; nights: sky equal (54.7, 57.4); quarter-night p90 88.1 -> 87.3 (fewer lit strings). (3) Bench, interleaved rounds: below. (4) Suite green: above. (5) Every changed expectation listed: above. 13: - **Bench** (`q/bench.sh`, P10k's recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, 1920x1080 (factor 1), `--disable-vsync`; pre-part tree and live interleaved, rounds 0-5, round 0 discarded, 19:00-19:12; `q/bench-summary.js`, `q/bench/summary.txt`), medians p50 / p95 in ms, pre -> live: Harbour 13:00 6.817 -> 6.940 (+0.123, slower 5/5) / 7.830 -> 7.871, draw calls 922 -> 948; Harbour 01:00 6.728 -> 6.850 (+0.122, 5/5) / 7.577 -> 7.925, 921 -> 947 (P10g's mock predicted +0.11 and +26); quarter 01:00 +0.004 (2/5) / +0.225, quarter 13:00 +0.015 (3/5) / +0.536, 1051/1052 draw calls unchanged; dense 01:00 -0.017 / -0.023, 2588 unchanged. The plan's dense p95 stop line is not hit; quarter p95 moves both ways within its 7.9-8.7 ms spread. 14: - **Review:** filled by the orchestrator. 15: - **Not done:** the optional cone-speck fix (P10j's harbour-day cones `hsp` 6-7 against P0d's 0) was not attempted; it stays open. Not done either: drag dense and cluster6 reruns (no Harbour); photo mode, tilt-shift and the perspective lens not captured; the journey folder itself (the orchestrator's, at the P10-III close); the extra light sync on the replanted Harbour not traced; the slower slider-scrub step in the Harbour drag not traced; `--verify` at 1440x900 (inherited stall); no commit, push, deploy, export or download. The scratch trees under `q/` are left in place. 16: - **Platforms:** Mac run; Windows not run. 18: **Beautify P10o Pixel body font and integer UI scale (9 October 2026):** Opus 5.5 on this Mac (plan section "P10o"; report polish backlog "Pixel body font and integer UI scale", life lens F6 as revised; step `28-P10`, sub-batch P10-III, third part). **This part produced options and measurements only. No font was downloaded or added, and no UI scale or font change was made.** The owner's 8 October instruction stops the part before any download, and owner decisions (3) and the 9 October note add that no download, UI-scale change or font change is made without a named file, source and size. The options sheet and the questions for the owner are in `o/sheet/README.md`. 19: - **Files:** none in the game. No script, shader, test, scene, `project.godot`, asset or `THIRD_PARTY_NOTICES.txt` changed. No new fixture, so `tools/repo_evidence.sh` is unchanged; its `*/*/sheet/*.png` and `*.md` globs already cover `o/sheet/`. 20: - **Snapshots, as on disk:** 21: - Pre-part `build/backup-feel/28-P10-o-pre/`, saved before anything else. Its `diff -rq` against `28-P10-n`, against `28-P10-n-post` and against the live tree is empty on scripts, shaders, tests, data, scenes, tools and `project.godot`. Nothing had changed that I did not expect (`o/diff-n-vs-o-pre.txt`, `o/diff-n-post-vs-o-pre.txt`, `o/diff-live-vs-o-pre.txt`; live `main` fa88e75, clean, `o/head.txt`). 22: - Post-part `build/backup-feel/28-P10-o-post/`. Only `docs/visual-handoff.md` differs from the pre snapshot (this entry; `o/diff-pre-vs-post.txt`). 23: - Scratch trees, all from `28-P10-o-pre`, each with one patch (`o/sheet/README.md`): `o/pre-tree` (none), `o/int-tree` (`Main.ui_scale_for` rounded), `o/floor-tree` (floored), `o/silk8-tree` and `o/silk16-tree` (rounded, plus Silkscreen as the body font at 8 or 16 px with antialiasing, hinting and subpixel positioning off). Each has a `.tools` link to the pinned engine. 24: - Evidence is under `artifacts/beautify/28-P10/o/` (below, `o/`). A backup of this file from before the entry is in `o/backup/`. 25: - **Findings:** 26: - The body font is `SystemFont ["Segoe UI","Arial"]` (`hologram_theme.gd:41-44`): Segoe UI on Windows, Arial on this Mac. 27: - The repo's only font is Silkscreen (OFL 1.1). It has no lowercase: all 26 lowercase glyphs are byte-identical to the capitals (`o/fonts/ttfinfo.js`, `silkscreen-info.json`). It is drawn on a grid of 8 px per em. 28: - This Mac has no other pixel font. Apple's fonts are not redistributable, and the Steam, Tiny Glade and Adobe fonts are neither pixel faces nor ours to ship. 29: - Since the 25 September UI_GROWTH rule, the UI scale is exactly 1 up to 1920x1080. The report's 1.11x and 1.2x predate that rule. Fractional scales start at 1920x1200 (1.067): 1.28 at 2560x1440, 1.42 at 2560x1600, 1.53 at this Mac's 2880x1724 fullscreen (`o/metrics/scale-table.txt`). 30: - **Options for the owner:** 31: - A: round, which is the P10n art factor: 2 at 2880x1724 (+31% against 1.53, above full growth), 1 at 2560x1440 (-22%). 32: - B: floor: 1 at 2880x1724 (-35%). It also drops P10n's art factor to 1 unless `pixel_factor` keeps its own sum. 33: - C: Godot's `scale_mode="integer"`, which ignores UI_GROWTH. Not mocked. 34: - Keep today. 35: - Three download candidates, each with its licence, URL, file and listed size: Pixel Operator (CC0), monogram (CC0) and Pixelify Sans (OFL). 36: - **Tests:** none written and no test-first log, because nothing was built. The constant, revert switch and following test come with the font in the follow-up part. `test_ui` on the scratch trees (`o/opt-tests.sh`, `o/opt-tests/*-test_ui.log.out`; AC, Low Power Mode 0, `o/opt-tests/power.txt`): 37: - pre 429 checks, 2 failures: the B2c and Q1a chemistry-fixture checks that every scratch tree fails, as in P10n. 38: - int 429, 3: plus "P10n: ui_scale_for ... 1.6 at 2880x1800". 39: - floor 429, 10: plus that check and 7 P10n factor checks. 40: - silk8 429, 3 and silk16 429, 3: the same as int. The 1100x700 "Control fits" checks pass, but they do not test text clipping. 41: - **Full suite and `--verify`: not run**, because the code is byte-equal to `28-P10-n-post`. Its suite ran in P10n: 47 suites, 25,008 checks, 0 failures; `--verify` 2880x1800 exit 0 (`artifacts/beautify/28-P10/n/suite/suite_all.out`). The known flaky check `test_creature_reactions` "The re-plan is applied after the edits" is recorded as known and did not run here. `--verify` at 1440x900 was not run (inherited stall). 42: - **Changed expectations:** none. **Expected journey count deltas:** none. 43: - **Captures** (`o/caps.sh`, `o/capset.sh`; frozen clock `--time-scale 0`; `harbour-day` with `--tool=window` and `--tool=scenery`; the P10n preference left at its default, on). Order: `cap/pre-1`, `live-1`, `int`, `floor`, `silk8`, `silk16`, `pre-2`, 18:03-18:09 BST, every view on try 1. Sizes: 1440x900, 1600x1000, 1920x1080, 2560x1440 and 2880x1800 for pre and live; 2560x1440 and 2880x1800 for int and floor; 1600x1000 and 2880x1800 for the Silkscreen trees. The 2880x1800 request gives a 2880x1724 window. Power: "Now drawing from 'AC Power'", "100%; charged", lowpowermode 0, "load averages: 2.26 2.76 2.66" at the start and "3.37 3.06 2.82" at the end (`cap/*/capture-meta.txt`). 44: - **Sheets** (`o/sheet/o-1` to `o-7`, all looked at): 45: - `o-1`, `o-6`: Silkscreen 8 is crisp but 5-px caps, unreadable at 1x. 46: - `o-6`, `o-7`: Silkscreen 16 is legible but caps only and about twice as wide. The Scenery labels are cut off ("MIRROR THE TOWN IN THE WAT"), and at 1600x1000 the top-right hint runs off the window. 47: - `o-2`, `o-5`: option A at 2880 is a bigger UI and dock over the island. 48: - `o-2`: option B at 2880 gives small text and a fine art pixel. 49: - `o-4`: the brand at x3. 50: - **Measured** (`o/crisp.js`, `o/metrics/crisp-table.txt`). The metrics: the partial share is partly covered over covered text pixels; "on grid" is the share of the Silkscreen heading's horizontal runs that are whole font pixels. 51: - Today at 1440x900, 1600x1000 and 1920x1080: brand and title partial 0, on grid 100%. 52: - Today at 2560x1440: brand 0.394 and title 0.377, on grid 0%. 53: - Today at 2880x1724: 0.161 and 0.224, on grid 0%. 54: - Option A at 2880: 0 and 100% (runs of 4, 8 and 12 px). Option A at 2560 and option B at both sizes: 0 and 100% at scale 1. 55: - Arial body text: partial 0.52-0.61 at every size today, 0.459 with A at 2880. 56: - Silkscreen 16 body: 0. Silkscreen 8 body: 0.154, which is the separator line inside the rect (`o/check/check-silk8.png`). 57: - **Noise and "unchanged"** (`o/metrics/noise.txt`). The whole `editor.png` changes on 0.84-0.99% of pixels between pre-1 and pre-2 (Harbour's CPU smoke and rain), and on 0.82-1.00% between pre-1 and live-1, which is within noise. The inspector text rect is 0 differing pixels in both pairs at all 10 views. The crispness numbers are identical in pre-1, pre-2 and live-1. 58: - **Draw calls** (`o/metrics/counts.txt`): 59: - Pre, live, int and floor are identical at each size: 911/987 at 1440x900 and 1600x1000, 921/997 at 1920x1080 and 2560x1440, 915/991 at 2880. 60: - The Silkscreen trees draw 1 call fewer and 45 more rendered objects (910/1032, 914/1036). Not traced, because it is a mock. 61: - The floor tree draws 2880 at `pixel_factor` 1 instead of 2. 62: - **Bench:** none, because no game code changed. 63: - **Review:** filled by the orchestrator. 64: - **Not done, and why:** 65: - The font download, its import with antialiasing off, the `THIRD_PARTY_NOTICES.txt` entry, the theme and integer-scale code, and the acceptance captures. All of these wait for the owner's reply. The candidates are listed with their URLs, files, sizes and licences, read from the pages on 9 October 2026 and not verified by a download. 66: - Option C was not mocked. 67: - 3840x2160 and 5120x2880 were computed only; this display gives at most 2880x1724. 68: - Windows font rendering was not checked. 69: - No commit, push or export. The scratch trees are left in place. 70: - **Platforms:** Mac run; Windows not run. 72: **Beautify P10n Art-pixel density follows the UI scale (preference) (9 October 2026):** Opus 5.5 on this Mac (plan section "P10n"; report polish backlog "Art-pixel density follows the UI scale in fullscreen (`effective_pixels`), as a preference toggle after an A/B", pixel lens F4 as its skeptic revised it; step `28-P10`, sub-batch P10-III, second part; a look change approved as built under the owner's standing approval if it clears review). In a big or fullscreen window each art pixel is now drawn at the pixel size x the window's factor, so the town keeps a 1440x900 window's art-pixel density while the UI grows: `Main.pixel_factor(window, follow)` is the editor's own UI scale (`Main.ui_scale_for`, the same UI_GROWTH 0.8 sum `resize_viewport` uses) rounded, at least 1: 1 for every window up to 2560x1440 (1440x900, 1600x1000, 1920x1080/1200 included), 2 at 2880x1724-2880x1800 and 3840x2160, 3 at 5120x2880-5120x3200. `Main.effective_pixels()` (pixel size x `frame_factor`) replaces the pixel size in every P9e frame sum: the 3D view (`CameraRig.view_size`), the window rect, the 1-art-px margin (now `effective` window px), the residual scroll, `frame_origin`, `view_frame`, `view_point` and `window_point` (so every pick), and in the lens: `tilt_radius`, `lens_grid` and `lens_bokeh` take an optional `factor` (default 1, so the static calls and their old tests are unchanged), the photo depth-of-field radius uses `effective_pixels()`, and `layout_frame` re-applies the lens when the factor changes. It is the person's preference, never the document's: `UserPrefs` "display"/"pixels_follow", default `UserPrefs.PIXELS_FOLLOW_DEFAULT = true` (the plan default "on after the owner sees a fullscreen A/B"; the revert switch: false restores the old frame for everyone who has not chosen and for automation); an older preference file loads unchanged (no key = the default, its other keys kept); a stored non-bool reads as the default; an interactive launch takes the file, an automated run (smoke, verify, bench, drag, tests) the default unless `--pixels-follow=on|off`. The Scenery panel has a "Pixel size follows window size" switch under Pixel size (no edit, no undo, nothing in the saved scene, version stays 2; it re-lays the frame out for the same window). `performance.json` gains `pixel_factor` and `art_pixels`. My choice, for the owner: the factor follows the UI scale with UI_GROWTH (the part's title); the lens packet's literal formula `round(min(w/1440, h/900))` ignores UI_GROWTH and would give 2 already at 2560x1440 and 2560x1600. Files: `scripts/main.gd`, `scripts/user_prefs.gd`, `scripts/editor_ui.gd`, `tests/test_camera.gd`, `tests/test_tilt.gd`, `tests/test_ui.gd` (316 lines in, 37 out). No shader, scene or document change; no new fixture (the sheets are covered by `tools/repo_evidence.sh`'s existing `*/*/sheet/*.png` glob, so it is unchanged). Snapshots as on disk: pre-part `build/backup-feel/28-P10-n-pre/`, saved before the first edit; `diff -rq` against `28-P10-j` and against the live tree on scripts, shaders, tests, data, scenes, tools and `project.godot` is empty (`n/diff-j-vs-n-pre.txt`, `n/diff-live-vs-n-pre.txt`; live `main` af82114, clean, `n/head.txt`), so nothing had changed that I did not expect. Post-part `build/backup-feel/28-P10-n-post/` (equal to the live tree, `n/diff-live-vs-n-post.txt`). Scratch trees: `n/pre-tree` (from `28-P10-n-pre`) and `n/off-tree` (the live code with `PIXELS_FOLLOW_DEFAULT = false`). Evidence under `artifacts/beautify/28-P10/n/` (below `n/`); file backups in `n/backup/`. 73: - **Profiles and modes reached** (2880x1800; nothing changes where the factor is 1, below): Canal and Harbour at Balanced 2 -> 4 window px per art px (view 1440x862 -> 720x431 at 2880x1724), Chunky 3 -> 6, Retro 4 -> 8. **Glade (Fine, pixel size 1) 1 -> 2** (view 2880x1724 -> 1440x862), and its lens now takes P10h's art-grid branch (grid 2) where Fine used the window-pixel blur. **Tilt-shift and photo depth of field:** the blur is the same in art px (0.6 x 7.8 = 4.68 art px), so in window px it doubles (9.36 -> 18.72 px at Balanced); fullscreen now blurs the same share of the frame as a 1440x900 window (before, fullscreen had half of it; `n/sheet/n-7`, `n-8`). Photo bokeh follows the grid. **Ambient motion:** the setting is untouched; motion that steps in whole art pixels (crowns, boats, rain drops) steps 4 window px instead of 2 in fullscreen Balanced. Grips and other art-px-sized overlays grow 2x in window px with the UI. `world_view.gd`'s `story_fine` still reads the pixel-size setting, not the factor (unchanged). 74: - **Tests:** test first (`n/testfirst/`): `n-testfirst-test_tilt.log.out` (parse errors: `tilt_radius`, `lens_grid`, `lens_bokeh` called with the new factor argument), `n-testfirst-test_camera.log.out` 534 checks, 2 failures (no `pixels_follow`; 0 of 8 grown-pixel snap runs), `n-testfirst-test_ui.log.out` 401, 2 (no `UserPrefs.pixels_follow*`, no `Main.ui_scale_for`/`pixel_factor`/`effective_pixels`); an earlier test-first run of test_camera died on my own loop-variable parse error and was overwritten. New: test_tilt radius, grid and bokeh at factors 1-3 (pixel sizes 1, 2, 4); test_ui `p10n_follow_checks` (preference file: no file, an older file, stored values, storing keeps the other keys, interactive and automated choice, `--pixels-follow`; the factor table at 11 window sizes = round(UI scale); layouts at 1440x900, 1600x1000, 2880x1724, 2880x1800, pixel sizes 1-3, follow on and off; Balanced at 2880x1800 = 720x450; lens radius and grid at 2880x1800 for Balanced and Fine, on and off; the Scenery switch, re-layout, file written only for a person's editor, no document change, no undo step, version 2, rebuilt panel); test_camera P9e frame checks at 7 windows (5120x3200 added) x pixel sizes 1-4 x snap x follow on/off, and the snapped margin, scroll, picks, lens row, middle and dock at 2880x1724 and 5120x3200 as well as 1437x901 and 1600x1000. After (`n/after/`): test_tilt **258, 0** (`a1-test_tilt.log.out`), test_camera **534, 0** (`a2-test_camera.log.out`), test_ui **435, 0** (`a3-test_ui.log.out`). Revert switch off (`n/off.sh`, `n/off/`, `n/off-tree`): test_tilt 258/0, test_camera 534/0, test_ui 429 checks, 2 failures, both fixture checks that read `res://artifacts/visual-audit/chemistry.json`, which scratch trees do not have (B2c, Q1a). Full suite (`VERIFY_RES=2880x1800 tools/suite_all.sh n/suite` a, b, v; 17:37-17:46; AC, 100% charged, Low Power Mode 0, `n/suite-power.txt`): **47 suites, 25,008 checks, 0 failures** (`n/suite/suite_all.out`). `--verify` 2880x1800 exit 0 on try 1 (25 s), `PIXELPUNK_INTEGRATION []`, and it ran at factor 2 (its integration captures are whole 4x4 blocks: `adapt-b-levitate-clean.png` 100% of 4x4 blocks uniform, 4.68% in the P10-II close's); the only ERROR line is the inherited missing `terrain_flowers.svg` (also in the P10-II close). test_creature_reactions 5,583/0 and rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). 75: - **Changed expectations (old -> new):** test_camera P9e "the rect is exactly view x pixels, every art px is `pixels` window px" -> `pixels` is now `Main.effective_pixels()` (at 2880x1724 with the preference on: 2 x the pixel size), checked with the preference on and off, over 7 windows instead of 6; the snapped margin/scroll/pick loop's `pixels` likewise, over 4 windows instead of 2, and its labels read "px %d (%d window px per art px)". `main.gd`'s `--verify` tilt check: radius `0.3 x TILT_REACH x pixels` -> `x effective_pixels()`, the chunky step `x chunky` -> `x chunky x factor`, the grid `lens_grid(env)` -> `lens_grid(env, factor)`. Counts: test_tilt 212 -> 258, test_ui 399 -> 435, test_camera 452 -> 534; suite total 24,844 -> 25,008. **Expected journey count deltas: none** (the journey is 1600x1000, factor 1; draw calls and objects identical in every capture below). 76: - **Captures** (`n/caps.sh`, frozen clock `--time-scale 0`, the P9c-P9e recipe; AC, 100%, Low Power Mode 0, 1-min load 2.0-3.6, `*/capture-meta.txt`). **Set A, factor 1** (`n/capA.sh`, 17:10-17:15; pre-1, live-1, pre-2, live-2; quarter-night, quarter-day, photo-golden, harbour-night, glade-day and quarter-night with the tilt-shift at 0.6, each at 1600x1000 and 1440x900): `n/metrics/capA-cmp.txt`. **Set B, 2880x1800** (`n/capB.sh`, 17:16-17:22; pre-1, live-1 (default on), off-1 (`--pixels-follow=off`), pre-2, live-2, off-2; the same views plus quarter-day and harbour-day): `n/metrics/capB-cmp.txt`, `capB-grid.txt` (`n/grid.js`: share of b x b window-px blocks that are uniform), `capB-counts.txt` (`n/counts.js`). Sheets (`n/sheet.js`, `n/sheet/README.md`), all looked at: `n-1-quarter-night-full.png` (same framing, twice as chunky), `n-2-quarter-night-1to1.png` (Balanced at 4 window px; thin balcony rails, festoon cords and cobble joints thin out or drop, as in a 1440x900 window), `n-3-harbour-day-full.png` (with the editor UI: the UI identical), `n-4-glade-day-1to1.png` (Fine at 2 window px), `n-5-tilt-1to1.png` and `n-6-photo-golden-1to1.png` (the blur twice as wide in window px), `n-7-photo-golden-1440x900-live.png` with `n-8-photo-golden-2880-full.png` (the fullscreen photo now reads as the 1440x900 one). 77: - **Acceptance, measured:** 78: - **1440x900 and 1600x1000 byte-identical: met.** `cmp`, courtyard and editor, pre against live in both pairs: identical on quarter-night, quarter-day, photo-golden and the tilt-shift view at both sizes; the pre and live noise pairs are identical too. Harbour-night and Glade (CPU smoke) are within their noise pairs (pixels changed, any): harbour-night 1600x1000 pre/live 1.217% and 1.438% against noise 1.062% (pre) and 1.423% (live); 1440x900 1.030% and 1.279% against 1.223% and 1.052%; glade-day 0.755% and 0.789% against 0.773% and 0.801%; 1440x900 0.823% and 0.781% against 0.799% and 0.790%. Views, draw calls and objects identical (for example quarter-night 800x500, 1030/1106; photo-golden 1360/1436; harbour-night 910/986). 79: - **At 2880x1800 the art pixel grows with the UI: met.** The window is 2880x1724 (UI scale 1.53, factor 2): view 1440x862 -> 720x431 (Glade 2880x1724 -> 1440x862). Uniform 4x4 blocks pre -> live: quarter-night 8.66% -> 100%, quarter-day 12.05% -> 100%, harbour-day 11.65% -> 100%, harbour-night 19.46% -> 100%, tilt 6.80% -> 100%, photo-golden 3.93% -> 100%; Glade's 2x2 blocks 23.7% -> 100%. Pixels changed pre-1 -> live-1 (courtyard, any / > 8 levels): quarter-night 85.4% / 44.6%, quarter-day 79.2% / 38.9%, harbour-day 69.3% / 20.0%, harbour-night 62.1% / 21.2%, glade-day 64.0% / 11.9%, tilt 92.2% / 40.4%, photo-golden 95.4% / 46.7%. Draw calls and objects identical pre/live/off (quarter-night 1046/1122, harbour-day 915/991, photo-golden 1366/1442). **With the preference off the old frame:** pre against off byte-identical on the four artwork views in both pairs, Harbour and Glade within noise (harbour-day 0.900% and 0.897% against noise 0.892% pre and 0.892% off; harbour-night 1.064% and 1.080% against 1.068% and 1.100%; glade-day 0.750% and 0.742% against 0.754% and 0.734%). 80: - **Frame time falls: met.** Bench below: dense p50 -18.5%, Harbour p50 -33.3%. 81: - **The preference lives in user prefs, not the document: met** (tests above; no key in the environment or the saved scene). 82: - **The P9e margin is included, picking exact at every scale: met.** Headless (test_camera above): at 2880x1724 and 5120x3200, pixel sizes 2 and 3, with and without the dock offset, the snapped view has the 1-art-px margin, the rect is at (-e, -e) + the scroll, every window pixel of a drawn art px maps back to it and every pick lands in the art px under it. Windowed in a person's editor (snap and margin on; `n/probe/pick_probe.gd`, a copy of P9e's probe reading `effective_pixels()`; `n/probe/run.sh`; `n/probe/pick-*.txt`; AC, Low Power Mode 0, `n/probe/power.txt`): 2880x1800 (2880x1724) at pixel sizes 1, 2 and 3 (2, 4 and 6 window px per art px), 1600x1000 at 2, and 2880x1800 at 2 with the preference off: **0 of 7,447,680 checked window pixels (2,400,000 at 1600x1000) differ from the post texel `view_point` names**, over 6 residuals each; ground picks 0 of 270 outside the art px under the pointer; grips 10/10; object picks 136/136 equal to the pick through the rect's own geometry; the photo-mode screenshot 0 differ. The OS mouse did not reach `get_mouse_position` in a windowed run (as in P9e). 83: - **Bench** (`n/bench.sh`, the P10h/P10i recipe at 2880x1800: smoke, `--churn=100000`, `--motion-live`, rain 0.45, `--disable-vsync`; dense fixture at 13:00 and Harbour at 21:00; pre-part tree, live (on) and live `--pixels-follow=off` interleaved; warm-up round 0 discarded; `n/bench1.out` rounds 1-3, `n/bench2.out` rounds 4-6, `n/bench/summary.txt`; 17:23-17:35, AC, 100%, Low Power Mode 0, 1-min load 2.2-2.5, `n/bench/power.txt`), pooled 6 rounds: **dense p50 17.308 -> 14.101 ms (-18.5%), p95 19.203 -> 15.814 (-17.6%); Harbour p50 9.868 -> 6.582 (-33.3%), p95 10.399 -> 7.536 (-27.5%)**; with the preference off dense p50 +0.2%, p95 -0.7%, Harbour p50 +0.2%, p95 +0.4% (the old cost). Both sets agree (dense on p50 -18.6% and -18.5%; Harbour -33.2% and -33.3%). Draw calls 2577 and 885 in all three. Not benched at 1600x1000 (the same code path and frames as before). 84: - **Review:** filled by the orchestrator. 85: - **Not done, and why:** 86: - The owner's fullscreen A/B and the default (the sheet is `n/sheet/`, its questions in `README.md`): keep on; exempt Glade/Fine (pixel size 1) so the miniature stays fine in fullscreen (one line in `pixel_factor`'s caller); the UI_GROWTH factor against the report's formula. 87: - A real macOS fullscreen space was not entered: the captures, probe and bench ask for a 2880x1800 window, which this Mac gives as 2880x1724. Window resizes go through the same `resize_viewport`, untested by hand. 88: - Detail loss in fullscreen (rails, 1-px cords and cable lines, joints) is inherent: fullscreen now shows what a 1440x900 window shows. 89: - Motion in fullscreen not filmed (stills only); the doubled step of whole-art-pixel motion is not judged. 90: - `--verify` at 1440x900 (inherited stall); Windows not run; no commit, push or export. 91: - Scratch trees `n/pre-tree` and `n/off-tree` are left in place. 92: - **Platforms:** Mac run; Windows not run. 94: **Beautify P10j Speck guard (conditional) (9 October 2026):** Opus 5.5 on this Mac (plan section "P10j"; report polish backlog "Speck guard in the outline, 8-neighbour version", pixel lens F6 as its skeptic revised it; step `28-P10`, sub-batch P10-III, first part). **Measured only. The guard was not built because the roof speck lattice it was meant for is gone.** In four captures from one sitting, the harbour-day slate roof (`roof_slate`, the roof planes of the report's hd-roofs crop) has 0 isolated dark art pixels on every metric. The report had 1.9% (the kit reads 2.449% on the audit capture). The guard's own trigger (darker than all 8 neighbours by more than 0.06 luma) finds 0 cells there. No game code, shader, test, scene or expectation changed. Files: none in the game. The evidence scripts are `artifacts/beautify/28-P10/j/cap.sh`, `specks.js` (counts on the art grid), `mark.js` (marked crops) and `sim.js` (a CPU estimate of what the guard would change in a capture). Live `main` a05fb93, working tree clean (`j/head.txt`). Snapshots: pre-part `build/backup-feel/28-P10-j-pre/`, saved before anything else. Its `diff -rq` against `28-P10-S2` and against the live tree on scripts, shaders, tests, data, scenes, tools and `project.godot` is empty, so nothing had changed that I did not expect (`j/diff-S2-vs-j-pre.txt`, `j/diff-live-vs-j-pre.txt`). Post-part `build/backup-feel/28-P10-j-post/`: its code equals the pre snapshot, and only `docs/visual-handoff.md` differs (this entry; `j/diff-pre-vs-post.txt`). Scratch tree `artifacts/beautify/28-P10/j/pre-tree` (from `28-P10-j-pre`). Evidence is under `artifacts/beautify/28-P10/j/` (below, `j/`). A backup of this file from before the entry is in `j/backup/`. 95: - **Target.** P10j gives no number, only "well over target". I read the target as P0d's acceptance line: on pitched roofs and cones in harbour-day and harbour-night, no pixel more than 0.10 luma below both horizontal neighbours (`hsp`, the kit's `art.hspecks`, target 0). I also read it as the report's aim that the slate lattice fades. Metrics (`j/specks.js`, art grid, borders excluded): 96: - `s4`: darker than all 4 neighbours by more than 0.04 (the report's 1.9%). 97: - `s8_06`: darker than all 8 neighbours by more than 0.06 (the guard's trigger). 98: - **Captures** (`j/cap.sh`, `tools/art_capture.sh`, 1600x1000, interleaved `cap/pre-1`, `live-1`, `pre-2`, `live-2` from 16:44:56 to 16:47:36 BST, every view on try 1): harbour-day, harbour-night, quarter-day, quarter-night, lane-day and glade-day. Pre and live run byte-identical code, so the live sets are a second "before" pair and the cross pairs measure the sitting. The integer counts, draw calls included, are identical (`metrics/counts-pre1-live1.txt`). Power (`j/cap-power.txt`): start "Now drawing from 'AC Power'", "100%; charged", lowpowermode 0, "load averages: 2.60 2.59 2.34"; end the same with "load averages: 3.34 2.88 2.49". 99: - **Numbers** (`j/measure.txt`, `measure.json`; pre-1 / pre-2 / live-1 / live-2): 100: - harbour-day `roof_slate` (1,225 cells): `s4` 0 / 0 / 0 / 0 (0.000%; target "well under 1.9%": met), `hsp` 0 on all four (target 0: met), `s8_06` 0 on all four. In the audit capture the same script gives `s4` 30 (2.449%), `hsp` 29 and `s8_06` 29. 101: - harbour-day `roof_crop` (the whole hd-roofs crop): `s4` 0.403 / 0.403 / 0.412 / 0.403% (audit 0.832%). `s8_06` is 10 / 10 / 9 / 9 cells, 0.074-0.082% (audit 76, 0.626%). The `s4` cells left sit on the eave and ridge silhouette steps, the railing, the window bars, the chimney and the tree edge, not inside the slate (`look/roofcrop-live1-x3.png`; the audit side is `look/roofcrop-audit-x3.png`, where the yellow lattice fills both planes; plain crop on the left, marked on the right: red `s4`, yellow `s8_06`). Sheet: `j/sheet/p10j-roofs-audit-vs-live.png` (audit left, `live-1` right, colour and greyscale; `roof_crop`, `roof_slate` and `cones`). 102: - harbour-day `cones`: `s4` 13 (1.660%) on all four (audit 25, 3.193%). **`hsp` is 7 / 6 / 6 / 6 against P0d's 0: not met.** It was 0 after P0d's look-review fix, 7 at the P8-I close and 6 at the P9 close (`j/history-stages.txt`; the closes between P0 and P8-I have no PNGs left). I did not trace when it came back. By eye the marked cells are low-contrast dark shingles in the per-tile value grain, not black pepper and not a lattice (`look/cone1-live1-x6.png`, `cone2-live1-x6.png`). The guard would touch 4 of them. The CPU estimate puts `hsp` at 6 -> 3 (`metrics/guard-sim.txt`). 103: - harbour-night: `roof_slate` 0 and `cones` 0 on all four. `roof_crop` `s4` 0.214-0.280%, `s8_06` 1-3. 104: - glade-day `roof_slate` (pixel size 1): `s4` 33 (0.640%) and `s8_06` 20 (0.388%) on all four (audit 75, 1.454%, and 46). 105: - lane-day `roof_lane` (the same canal slate at lane zoom, 4,408 cells): `s4` 64-65 (1.45-1.48%), `hsp` 0, `s8_06` 21 on all four. **This is the one region where the guard's own trigger went up: `s8_06` 2 in the audit and in the P0 baseline -> 21 now (about tenfold), `s4` 0.749% (audit) and 0.794% (P0) -> 1.45%** (`j/fix/lane-remeasure.txt`; `hsp` 1 -> 0). It appeared by the P8-I close (`pre-1` there already reads 66 and 23) and has stayed at 21-24 at every close since (`j/review/lane-history.txt`, `lane-history-closes.txt`). I did not trace which stage between P0 and P8-I brought it: their close PNGs have been pruned. By eye, on the shaded slope (rect 1460,500,1580,590, x5: `j/fix/lane-roofB-audit-x5.png` against `lane-roofB-live1-x5.png`; plain left, marked right) the audit showed dark joint lines between the slate courses, with 25 `s4` and 2 `s8_06` cells. Now the lines are mostly gone and single dark art pixels are scattered over the slope: 49 `s4` and 19 `s8_06`. The contrast is low (`hsp` 0). The lit slope (rect 1180,440,1300,510, `lane-roofA-audit-x5.png` against `lane-roofA-live1-x5.png`) keeps its course lines: `s4` 8 -> 15, `s8_06` 0 -> 2. The decision not to build the guard follows the part note's hd-roofs crop. Whether this lane-zoom pepper is worth the guard, or a look at the slate's shaded-side value, is the owner's call. `cone_lane`: `s4` 10, `hsp` 7-8, `s8_06` 7 (audit and P0 28, 12-13, 15). 106: - Whole frames, `s4` with `s8_06` in brackets: harbour-day 0.852-0.872% (0.145-0.148%), audit 0.592% (0.150%); quarter-day 0.709-0.721% (0.145-0.147%), audit 0.617% (0.134%); quarter-night 0.861-0.893% (0.076-0.080%), audit 0.465% (0.074%). The 4-neighbour share rose after the audit, but the 8-neighbour share did not. In the audit capture the 2x2 blocks were not uniform (2.3% of blocks on harbour-day, because presentation ran at window resolution before P9c); now they are (0.000%), so the art-grid average no longer softens one-pixel steps. That is the likely reason; I did not trace it further. On whole frames the guard would touch about 0.15% of art pixels by day and 0.08% at night. The estimate for quarter-day `hsp` is 4,254 -> 4,005 (`metrics/guard-sim.txt`). The estimate mixes luma only and cannot undo the outline's extra darkening, so it understates the guard. 107: - History, the same script on earlier captures (`j/history.txt`): `roof_slate` `s4` was 2.449% in the audit and the P0 baseline, then 0.000% at the P9 close and the P10-II close. P0d's look-review fix removed it: the instanced tiles' side faces are shaded with the top normal (see the P0d entry). 108: - **Whole-frame noise** (`metrics/noise-pre.txt`, `noise-live.txt`, `frame-pre1-live1.txt`, `frame-pre2-live2.txt`; >8 luma). harbour-day: noise 5.32% and 0.62%, cross 5.85% and 1.45%. quarter-day: noise 0.52% and 0.88%, cross 3.32% and 3.06%. quarter-night: noise 0.30% and 4.01%, cross 4.59% and 0.85%. glade-day: noise 0.31% and 0.31%, cross 0.36% and 0.47%. The code is byte-identical, so the cross pairs above noise on quarter-day come from the sitting and the capture clock (the P10-II close saw the same between trees), not from code. The roof and cone region counts agree across all four sets within 0-6 cells. 109: - **Tests:** none, and no test-first log, because nothing was built. The guard's constant, revert switch and following test were only to be written with the guard. **Full suite and `--verify` not run:** the code is byte-equal to `28-P10-S2`, whose full suite ran at the P10-II close (47 suites, 24,844 checks, 0 failures; `--verify` 2880x1800 exit 0; `artifacts/beautify/28-P10-II/close/suite/`). The known flaky check `test_creature_reactions` "The re-plan is applied after the edits" is recorded as known and did not run here. **Bench:** none, because no code changed. 110: - **Changed expectations:** none. **Expected journey count deltas:** none; every count stays identical. 111: - **Review:** filled by the orchestrator. 112: - **Review fix (9 October 2026):** the review found that the entry gave `roof_lane` with no audit or P0 comparison, although it is the one region where the guard's trigger rose (2 -> 21) and the shaded slope now shows scattered dark art pixels where it had joint lines. Fixed in the entry text only: the lane-day bullet under Numbers now gives the audit and P0 values, when the change appeared and the look crops (`j/fix/`, made with `j/specks.js` and `j/mark.js` on the existing captures; no new captures and no Godot run). No code changed: `diff -rq` of scripts, shaders, tests, data, scenes, tools and `project.godot` against `build/backup-feel/28-P10-j-pre/` is still empty, so the suites were not re-run. 113: - **Not done, and why:** 114: - The 8-neighbour guard in `presentation.gdshader`, with its constant, revert switch, tests, interleaved bench and per-profile and per-mode numbers (photo mode and tilt-shift included). The plan's condition is not met: `roof_slate` is at 0, not over target. 115: - The rail and cable checks on qn-plaza and hd-fence-wires. They are only needed if the guard is built. 116: - The cones' 6 `hsp` against P0d's 0 is left to the orchestrator and owner. If wanted, the cheaper fix is in the cone tiles' per-tile value; the guard would halve it. 117: - The lane-zoom slate pepper (`roof_lane` `s8_06` 2 -> 21 since the audit, there by P8-I and not traced) is left to the owner: build the guard for it, look at the slate's shaded side, or accept it. 118: - The 1440x900 `--verify` stall is inherited and was not run. 119: - No commit, push or export. 120: - The scratch tree `j/pre-tree` is left in place. 121: - Platforms: Mac run; Windows not run. 123: **Beautify P10-II (parts e, g, m, k of P10) stage close (9 October 2026):** the stage-end overhead of sub-stage P10-II "Polish backlog" (step `28-P10-II` in the plan's overhead table). Opus 5.5 ran it on this Mac on the final P10-II code, `main` 434302e (P10k), working tree clean (`close/head.txt`, `close/status.txt`). No game code changed. Evidence: `artifacts/beautify/28-P10-II/close/`. 124: - **Before side:** `close/pre-tree` from `tools/feel_snapshot.sh tree 28-P10-S1` (= `main` 862ccbe, the P10-I close code). The snapshot differs from live only in the stage's files (`close/diff-S1-vs-live.txt`); of the 39 code files changed since 862ccbe, the 25 that exist there are byte-equal to it in the snapshot and 14 are new (`close/git-changed-since-pre.txt`, `close/s1-vs-862ccbe.txt`). 125: - **Power:** every run on AC, 100% charged, Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`); 1-min load 2.0-5.3 during the suite and captures, 1.5-2.7 during the bench. 126: - **Full suite** (`close/suite.sh` a, b, v; 15:12-15:21): **47 suites, 24,844 checks, 0 failures** (`close/suite/suite_all.out`; P10-I close 46 / 24,306, the new suite is `test_beacon`). `--verify` 2880x1800 exit 0 on try 1 (34 s), `PIXELPUNK_INTEGRATION []`. `test_creature_reactions` 5,583/0 in the run and the rerun; the known flaky check did not fire. 1440x900 not run (inherited stall). 127: - **Journey** (`artifacts/feel-journey/28-P10-II/`, six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1): against 28-P10-I (`28-P10-II/counts-vs-28-P10-I.txt`) draw calls and rendered objects +3 harbour-day (909 -> 912), +2 harbour-lane, +3 harbour-night, +6 quarter-day (1026 -> 1032), +6 quarter-night, styles-day identical; every other integer count identical. That is exactly P10e's declared delta (the cable line surfaces; its own `e/metrics/counts-pre1-live1.txt` has the same numbers); P10g, m and k declared none. Re-run of quarter-night and harbour-day in the same tree (`close/rerun/`): counts identical, frame >8 0.2% and 1.0%. Frame >8 luma against 28-P10-I (`28-P10-II/frame-vs-28-P10-I.txt`): harbour-day 4.8%, harbour-lane 2.6%, harbour-night 4.8%, quarter-day 3.8%, quarter-night 8.1%, styles-day 3.1%; diff maps `close/look/journey-diff-*.png` (looked at: quarter-night, harbour-day) show the continuous cable lines plus rain rings, clouds and chimney smoke at another phase. `notes.json` has a `28-P10-II` entry (`close/journey-note.js`, backup `notes.json.bak-before-28-P10-II`); `index.html` and `journey-standalone.html` (111.36 MB) rebuilt. 128: - **Art capture and protected looks** (`close/caps.sh`, `capall.sh`: the P10-I close's 101 views in groups g1-g3 plus g4 = P10e's far-night, far-day, q-yaw-a, q-yaw-c, tilt-night; 106 views; 15:24-16:09; pre-1, after, pre-2, after-2, every view on try 1). Regions `close/regions.json` = the P10-I close's (`tools/art_regions.json` unchanged by the stage). Metrics `close/metrics/` (`metrics.sh`, `frame-table.js`, `prot.js` → `protected.md`, `elapsed.js`, `consistent.js` → `consistent.txt`, `look/consistent/*.png`). Against the previous stage close `28-P10-I/close/after` (the pre-stage code, another sitting): `ab-p10iclose-*`, `frame-p10iclose-*`. Sheets `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P10-I-close.png`, `noise-pre.png`, `noise-after.png` (looked at: 1). 129: - **Counts** (`metrics/counts.txt`): draw calls and rendered objects rise on every view that has string lights or cords, +1 to +9 (quarter-2100, -2200, -golden, -0610, photo-golden +9), identical in both cross pairs; all other integer counts identical apart from qn-orbit's light admits, which vary within each tree as before. The previous close's `after` against `pre-1`: identical apart from the two orbit views. 130: - **Frames:** 88 of 106 views above both noise legs (`frame-table.txt`); not above: grove-day, grove-1930, shops, beached, banks, field-*, sheet-day/night, chemistry-day, harbour-r2, orbit-g3. Systematic change (all four cross pairs, neither noise pair; `consistent.txt`): grove-day and grove-1930 0.000%, glade-day 0.49%, glade-night 0.32%, harbour-day 3.9%, quarter-night 4.1%, up to 6.4% (courtyard-day). The live capture is a median 149 ms later after start than the pre one in 90 of 106 views (`elapsed-ms.txt`, P10e saw the same), so TIME-driven clouds, rain rings, sparkle and smoke land at another phase in a code-correlated way; the maps looked at (styles-day, quarter-night) show the cable lines plus that phase. 131: - **Canal mirror columns:** p50 43.6 → 43.9 quarter-night, 52.3 → 52.8 at 21:00, hex unchanged; changed >8 30.3%/27.3% and 33.3%/30.5% against noise 10.2%/3.5% and 9.5%/1.0%. Four-way crops `close/look/mirror-q2100-4.png` and `mirror-qdry-4.png` (looked at): the columns keep their shape in both trees; rings (rain) and sparkle (dry) sit elsewhere. Attributed to the capture clock as P10e did; not proven that no reflected cable pixel contributes. 132: - **HDR neon cores:** fish sign 67.3 → 67.9, bar 70.8 → 71.0, lane sign 140.2 → 140.3, lane fish 135.9 = 135.9, tower windows 95.8 = 95.8 (0.0%); the 1.6-2.7% of changed pixels in the sign boxes are the bunting cord and festoon lines in front of them (`protected-1.png`), the letters unchanged. 133: - **05:45 and 20:45 skies:** changed ≤1.06% against noise up to 0.55%, p50 equal (quarter 05:45 zenith 56.6, horizon 113.9 → 113.8; 20:45 zenith 97.7, horizon 151.2 → 151.1; Harbour 05:45 horizon 115.8, 20:45 154.8). 134: - **Glade/Canal separation:** Glade night zenith `#001d45`, horizon, lawn and bank 0.00% changed, Glade night water 0.4-0.6%; Harbour night zenith p50 53.9 = 53.9 (23% of its pixels change between trees: cloud phase, 0.1-0.6% within each tree; the P10-I close showed the same 22.7% between sittings of one code). 135: - **Approved tree style:** grove-day and grove-1930 crowns, conifers, tree feet, lawn and frame 0.00% changed; quarter-night undersides 27.7 = 27.7; Harbour, quarter and Glade crowns 0.1-1.2% (cables crossing them). 136: - **Whole-stage A/B bench** (`close/bench.sh` = the P10-I close recipe; `bench-t1.out`, `-t13.out`, `-t19.5.out`, `bench/summary.txt`; 16:18-16:30, 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): 138: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 139: |---|---|---|---|---|---|---| 140: | Dense | 14.315 → 14.287 (−0.029, 1/3) | 16.176 → 16.087 (−0.089, 0/3) | 14.361 → 14.374 (−0.013, 1/3) | 16.286 → 16.217 (−0.044, 1/3) | 16.234 → 16.058 (−0.144, 0/3) | 18.039 → 17.848 (−0.180, 0/3) | 141: | Harbour | 6.728 → 6.739 (+0.005, 2/3) | 7.777 → 8.471 (+0.525, 3/3) | 6.804 → 6.807 (+0.000, 1/3) | 7.399 → 7.472 (+0.127, 3/3) | 6.682 → 6.694 (+0.000, 1/3) | 7.471 → 7.478 (+0.229, 2/3) | 143: - Draw calls: dense unchanged (2588 / 2589 / 3440: the dense fixture has no cable the stage redraws), Harbour 918 → 921, 919 → 922, 831 → 833. p50 within ±0.03 ms except dense 19:30, where live is 0.14 ms faster in all three rounds (pre ran 0.14 ms slower than the P10-I close's pre at that hour; not chased). Harbour p95 spreads 7.3-8.6 ms in both trees and swings both ways (P10-I close: −0.57 to +0.77); the plan's stop line (dense p95 over about +0.5 ms) is not hit. 144: - **Owner sheet** `artifacts/beautify/28-P10-II/close/sheet/` (`close/sheet.sh`, regions `close/sheet-regions.json` from `close/sheet-regions.js`; before left, after right, colour plus greyscale on the P10e crops): `p10ii-01-e-cables-quarter.png`, `p10ii-02-e-cables-harbour.png`, `p10ii-03-e-far-bulbs.png`, `p10ii-04-existing-scenes.png` (close pre-1 | after), and the parts' own mock pages copied unchanged because their switches are off and no shipped scene changed: `p10ii-05-g-mock-day.png`, `-06-g-mock-night.png`, `-07-g-mock-pier.png` (`28-P10/g/sheet`), `p10ii-08-m-beacon.png` (`28-P10/m/sheet`), `p10ii-09-k-swags-quarter.png`, `-10-k-swags-closeups.png`, `-11-k-overcast-heavy.png`, `-12-k-overcast-default.png` (`28-P10/k/sheet`); `sheet.md` (one paragraph per part with its measured numbers, then the bench and the suite). The close agent looked at pages 01 (first version, regions then narrowed to 2x), 02, 03, 04 and the P10g day page. After the handoff check (9 October): all 12 pages looked at; pages 05-12 are byte-identical to their sources (`cmp`); `sheet.md` amended (backup `close/sheet.md.pre-docfix`): page 04's caption ("the only code change"), P10g's p95 rise, the east-bank lantern post hidden by A's new tree and the reeds there (crops `close/look/p10g-east-bank-pre-1.png`, `-replant-1.png`), the default's above-noise spots, and the test expectations a yes changes; P10k's lamp 195 left with no string, wall-hung strings not thinned, and the corrected noise and quarter-night wording. Page 08's title (copied unchanged from `28-P10/m/sheet`) is cut at the right edge ("... AND WITH"); left as copied. Handoff corrections of the parts' entries: backup `close/visual-handoff.md.pre-docfix`. 145: - **Not done / open:** 146: - The owner's answers on the P10g sheet (replanting A or B, beacon placement and colour) and the P10k sheet (swags, overcast A or B, water under overcast); until then Harbour, the beacon placement and both P10k switches stay off, and nothing in the journey or drag baseline was re-based. 147: - P10e's pale laundry and bunting cords read as continuous near-white lines by day and at the far zoom (`p10ii-03`, far-night east crop: one bright cord by the boathouse roof); the cable night colour is not tuned toward the report's 1a1826. Taste, left for the owner. 148: - The canal mirror columns' above-noise pixel change is attributed to the capture clock, not proven (a fixed-step pair, as P10m used, would settle it). 149: - The harbour-night front-water dashes question from the P10-I close is not re-examined here. 150: - No photo-mode, perspective, hover or Reduced/Off views in the close capture beyond photo-golden and tilt-night; P10m's beam and the P10k mocks are shown from the parts' captures. 151: - No export (none asked); the handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 152: - Scratch trees: `close/pre-tree` was moved to `the macOS Trash/p10ii-close-pre-tree-163250` with `mv` once its numbers were recorded (`close/trash.txt`). The part scratch trees under `artifacts/beautify/28-P10/` (`e/pre-tree`, `e/fix/revert-tree`, `e/review/revert-tree`, `e/review/revert-tree2`, `g/pre-tree`, `g/switch-tree`, `m/pre-tree`, `m/off-tree`, `k/pre-tree`, `k/on-tree`, `k/off-tree`, `k/strong-tree`) were left in place (outside this close's folder). 153: - Platforms: Mac run; Windows not run. 155: **Beautify P10k Owner-call tunings: quay string lights, rainy overcast (9 October 2026):** Opus 5.5 on this Mac (plan section "P10k"; report polish backlog "String lights: thin the quay rows into separate swags; keep lawn and plaza spans" and "Rainy days read as overcast", architecture lens F5 as its skeptic revised it, light lens F9; step `28-P10`, sub-batch P10-II, part k, the last before the P10-II close). **Stopped at the mock, as the plan's owner gate and the owner's 8 October note require: both tunings are built behind a switch that is OFF, so no scene, starter, fixture or capture changes; neither was turned on.** (1) **Quay swags** (`Composition.QUAY_SWAGS = false`): on, a string of lights (never bunting) whose two ends and middle each lie within `QUAY_NEAR` 1.5 m of a side between a land cell and a water cell running along it (|cos| > `QUAY_COS` 0.8; `Composition.quay_span`) is a quay span; each lamp keeps at most `QUAY_LINKS` 1 quay span, and a quay span left dark still takes its place in its lamps' `ANCHOR_LINKS` counts, so no lamp takes a new partner: on, the strings are exactly the switch-off ones less some quay spans. `LIGHT_MIN`, `ANCHOR_LINKS`, lawn, plaza, stair and cross-water strings are unchanged. With the swags on, the lived-in cache keys on the terrain too (painting water beside a row thins it); off, the signature is `hash(state.objects)` as before. Deviation from the lens revision: 1.5 m, not 1.2 m (the quarter's west quay lamps stand 1.4 m from the water), and wall lanterns count as well as standing lanterns (the quarter's front quay chain runs through wall lamps on the quay-facing fronts; standing-only would have changed nothing there). (2) **Rainy overcast** (`WorldView.RAIN_OVERCAST = false`): on, the view (never the profile, whose data stays rain-free) takes overcast = rain x daylight, daylight = 1 - windows (`WorldView.overcast_amount`; 0 at night and 0 when dry): key x (1 - `OVERCAST_KEY` 0.4 x oc), sun shadow blur x (1 + `OVERCAST_BLUR` 2 x oc), ambient x (1 + `OVERCAST_AMBIENT` 0.25 x oc), and the panel greys by rain x mix(0.22, `OVERCAST_SKY` 0.35, daylight) toward cloud x mix(0.85, `OVERCAST_CLOUD` 0.70, daylight) (`backdrop.gdshader`, new uniform `overcast` = daylight with the switch on, 0 off: exactly the old rain x 0.22 toward cloud x 0.85). The report's numbers, except the panel: the report's flat rain x 0.35 would have greyed the night sky too, so the weight follows daylight. All in `sync_profile` (a running clock carries it; `sync_environment` sets the same blur). `LightingProfile`'s header keeps "Daylight is independent of rain" with a note naming the switch. Test and mock only, never saved: `--quay-swags=on|off` (`Composition.quay_swags_override`) and `--overcast=on|off` (`WorldView.debug_overcast`); null follows the switch. **Turning either on** is one constant; the changed tests are listed below. Files: `scripts/composition.gd`, `scripts/world_view.gd`, `scripts/main.gd` (the two flags), `shaders/backdrop.gdshader`, `scripts/lighting_profile.gd` (header note), `tests/test_links.gd`, `tests/test_time.gd`, `tests/test_water.gd`. No new suite, class or fixture (so `tools/repo_evidence.sh` unchanged). Snapshots: pre-part `build/backup-feel/28-P10-k-pre/` (saved before the first edit; `diff -rq` against `28-P10-m` and the live tree on scripts shaders tests data scenes tools and `project.godot` empty, so nothing unexpected had changed: `k/diff-live-vs-28-P10-m.txt`, `k/diff-k-pre-vs-live-and-m.txt`); post-part `build/backup-feel/28-P10-k-post/`. Scratch trees `artifacts/beautify/28-P10/k/pre-tree` (from `28-P10-k-pre`), `k/strong-tree` (live code with the stronger overcast B, mock only), `k/on-tree` (live code, both switches true) and `k/off-tree` (live code unchanged), diffs `k/switch-trees-diff.txt`. Evidence under `artifacts/beautify/28-P10/k/` (below `k/`); file backups `k/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, bench and suite (`k/cap-power.txt`, `k/cap-fixed/power.txt`, `k/bench/power.txt`, `k/suite-power.txt`); 1-min load 1.7-3.6. 156: - **Profiles and modes reached:** switched off (as shipped), none. Swags on: every look that draws a scene with lamps along water (canal and Glade alike, by night and day; the strings are geometry). Measured: the quarter (artwork.json) 28 -> 24 strings (dropped `lights:162:195`, `167:195`, `196:197` on the front quay and `243:244` on the west quay, none added); Harbour steps 12 -> 12 strings and 2 bunting, same keys (no Harbour string runs along its water); dense 100 -> 100, styles 3 -> 3, lane, grove and chemistry 0 (`k/probe/dense_probe.log.out`, `k/suites/r2-test_links.log.out`). Overcast on: canal by day, at golden hour (daylight 0.34) and slightly at 05:45 and 20:45 (daylight 0.15); nights none; dry days none; Glade only when its rain is raised (it starts at 0); photo mode and tilt-shift follow the same light (not captured). Ambient motion: neither moves anything. 157: - **Tests:** test first `k/testfirst/k-testfirst-test_time.log.out` (382 checks, 5 failures: no switch, constants, shader uniform or flag) and `k/testfirst/k-testfirst-test_links.log.out` (381 checks, 3 failures: no switch, constants or flag; the rest sit behind that guard). New `overcast_checks` in test_time (switch off by default; `overcast_amount` off, on, night, dry, clamps; for 7 hour/rain cases the view with the switch off is exactly the hour's key, ambient, blur and panel 0, and with `debug_overcast` on each value follows the formula while the profile stays the hour's; a running clock into the night takes it off; the rain is never changed; the shader's uniform and the old greying at 0; the flag) and `p10k_checks` in test_links (switch off; constants; LIGHT_MIN and ANCHOR_LINKS unchanged; the flag; a quay row of five lamps 4 strings off, 2 separate swags on, no lamp with two quay spans; post lanterns thin the same; flags keep their bunting; a lawn row and a row 2.2 m back keep their chain; a string across a channel stays; the swags derived, undo exact, nothing stored; the cache follows painted water on, `hash(objects)` off; Harbour identical; the quarter thins, only quay spans dropped, nothing added, bunting unchanged). Affected suites `k/suites/r1-test_time.log.out` 440/0, `k/suites/r2-test_links.log.out` 402/0. **Full suite** (`k/suite/suite_all.out`, `VERIFY_RES=2880x1800`, 14:23-14:32): **47 suites, 24,844 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). 158: - **Changed expectations (old -> new):** test_water `far_quay_checks` finds the rain greying line by its text: `sky = mix(sky, cloud_color.rgb * 0.85, rain * 0.22);` -> `sky = mix(sky, cloud_color.rgb * mix(0.85, OVERCAST_CLOUD, overcast), rain * mix(0.22, OVERCAST_SKY, overcast));` (same order checks, same meaning at overcast 0). Counts test_time 377 -> 440, test_links 377 -> 402; suite total 24,756 -> 24,844, 47 suites. **Journey count deltas expected: none** (both switches off; every capture's integer counts identical, below). 159: - **What a yes would change** (`k/on-tree`, both switches true, against `k/off-tree`, every suite of `tools/test.sh` once in each; both trees lack `res://artifacts`, so fixture failures are in both and left out; per-suite logs `k/switch/on-tree/`, `k/switch/off-tree/`, list `k/switch-on-only-failures.txt`): test_links 4 (the switch-default checks, by design); test_time 92: the switch-off checks (by design) and, with the courtyard's rain 0.35, view_checks "the key light is the hour's colour and strength" and "the fill drops as the sun gets low" at 05:45, 06:45, 10:18, 16:30, 20:45, and twilight_checks' "the ambient is the fill plus x of skylight" / "exactly what it was" in canal and Glade: these are the rain-independence rule's checks the plan says to update deliberately on a yes; test_sun 1 ("The fill follows the turned sun's elevation", the same rule). Nothing else differs. 160: - **Captures** (`k/cap.sh` stages a, b, c, all `tools/art_capture.sh` 1600x1000, one sitting 14:02-14:13, every view try 1; `k/cap/strong-1` 14:21-14:22 from `k/strong-tree`, after the fixed-step runs of 14:14-14:19 (`k/cap/strong-1/capture.tsv`, `k/cap-power.txt`; Low Power Mode 0 is recorded at the start of the strong set only, its closing block has no lowpowermode line)): `pre-1`, `pre-2` (pre-part tree: before and noise pair), `live-1` (live, both off), `swags-1` (`--quay-swags=on`), `overcast-1` (`--overcast=on`), 19 views each: quarter-night, quarter-dry, quarter-day, quarter-golden, quarter-0545, quarter-2045, harbour-night, harbour-day, harbour-golden, lane-night, lane-day, glade-day, qn-front (quarter at night, camera 0.845,0.315,11 at (9,3.5,12)), qn-west (at (-11,3.5,3)), harbour-day-rain9, harbour-day-dry, harbour-night-rain9, quarter-day-dry, quarter-day-rain9 (`--rain=0.9` / `0`). Fixed-step pairs (`k/fixed.sh`, `--fixed-fps 30`, so TIME is the same on the capture frame): `k/cap-fixed/pre`, `live-1`, `live-2`, `swags`, `overcast`, 8 views. Owner sheet `k/sheet/` (all 8 pages looked at) with `p10k-sheet.md` (4 questions): `p10k-1-swags-quarter`, `-2-swags-closeups`, `-3-swags-front-1to1`, `-4-swags-harbour`, `-5-overcast-heavy`, `-6-overcast-default`, `-7-overcast-1to1`, `-8-overcast-unchanged`; option B is key 0.7, blur 4, ambient 0.5, sky 0.5 toward cloud x0.6 (a mock only). Also looked at: `k/look/sw-qn-front.png`, `sw-qn-west.png`, `oc-harbour-rain9-1to1.png`, `oc-shadow-3x.png`, `oc-quarter-rain9-3.png`, masks `k/masks/quarter-day-rain9-pre1-live1.png`, `qn-front-pre1-live1.png` (rain drops, rings and water glints only). 161: - **Acceptance, measured** (the plan's lines apply "after the mock and a yes"; measured on the mocks now): (0) **switched off nothing changes**: `pre-1` vs `live-1` integer counts identical on all 19 views (`k/metrics/counts-pre1-live1.txt`); changed >8 0.02-4.39% against the noise pair's 0.05-6.69% (`k/metrics/summary-frame.txt`); above the pre-1/pre-2 pair on six views: clearly on quarter-day-rain9 (1.40% vs 0.28%) and qn-front (0.75% vs 0.27%), where the masks show rain and water phase only, and by small margins on glade-day (0.411% vs 0.398%), harbour-day-dry (0.364% vs 0.315%), harbour-day-rain9 (1.047% vs 1.036%) and harbour-night-rain9 (1.03% vs 0.859%), for which no mask was made (not attributed); fixed-step `pre` vs `live-1` 0.37-1.34% against `live-1`/`live-2` 0.40-1.48% (`k/metrics/summary-fixed.txt`). (1) **"quay rows become separate swags while the Harbour lawn swags stay"**: by test (above) and by picture (sheet pages 1-4); fixed-step swags vs live: Harbour views within noise (harbour-night 0.72% vs 0.75%, lane-night 1.31% vs 1.48%, harbour-day 0.72% vs 0.79%), the quarter above it (qn-front 2.39% vs 0.43%, quarter-night 1.16% vs 0.40%); integer counts identical on all views (the strings are one mesh). (2) **"a rainy day reads as soft overcast while dry days and nights are unchanged"**: `k/metrics/overcast-stats.txt`, `overcast-stats-strong.txt` (town region luma, sky band): harbour-day-rain9 town p50 105.7 -> 101.4 (B 98.4), spread p90-p10 195.3 -> 189.2 (B 181.4), sky 202.8 -> 198.4 (B 191.6); quarter-day-rain9 spread 141.4 -> 138.8 (B 135.2), sky 177.5 -> 173.0 (B 166.5); at the default rain harbour-day spread 193.5 -> 191.3, quarter-day 141.0 -> 139.5; golden 177.3 -> 177.0 and 149.4 -> 149.1; dry days and harbour-night identical to the decimal (harbour-day-dry, quarter-day-dry, harbour-night), quarter-night within 0.1 (live-1 -> overcast-1 p90 88.1 -> 88.0, spread 75.9 -> 75.8, inside the pre-1/pre-2 spread 75.8/76.0; p50 and sky equal); fixed-step overcast vs live: nights within noise (harbour-night 0.74% vs 0.75%, quarter-night 0.42% vs 0.40%, qn-front 0.50% vs 0.43%, lane-night 1.29% vs 1.48%), quarter-0545 31% of pixels move but only 0.39% by more than 8 (noise 0.42%), harbour-day-rain9 8.9% and quarter-day-rain9 10.2% over 8. Judged by eye: A is subtle (a little dimmer and greyer, shadows a little softer), B reads more overcast; the water still mirrors the hour's bright sky (not part of the report's proposal; an owner question). (3) **"the rain-independence rule and its test_time checks updated deliberately if the overcast is taken"**: not taken; the checks that would change are listed above. 162: - **Bench** (`k/bench.sh`, P10g/P10m recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, 1920x1080, `--disable-vsync`; rounds 0-5 interleaved, round 0 discarded, 14:47-14:58; `k/bench-summary.js`, `k/bench/summary.txt`): Harbour pre -> live p50 -0.003 ms (13:00) and -0.003 ms (01:00), p95 -0.85 / +0.34 (bimodal p95, noise); dense 01:00 p50 -0.013 ms, p95 +0.04; draw calls 922 / 921 / 2588 unchanged. Mocks against the live default: Harbour 13:00 `--overcast=on` p50 -0.020 ms, p95 +0.95 (noise), draw calls 922 both; quarter 01:00 `--quay-swags=on` p50 -0.012 ms, p95 -0.05, draw calls 1051 both. 163: - **Open for the owner / known:** (a) the owner sheet's four questions: swags yes/no; overcast yes/no; A or B; the water's sky under overcast. (b) The swags thin the quarter's front quay to [186-162] [195 alone] [167-196] [197 alone on the quay, still strung up the steps; 198 keeps its strings to 29 and 132]: separate swags, but one quay lamp (195) ends up with no string at all. (c) With the overcast on, Glade with its rain raised also turns overcast; whether the shadow blur shows in GL Compatibility at x1.6-x2.6 is hard to see on the stills (the contrast drop is what reads). (d) `quay_span` reads water from the terrain only: with the Water switch off, a quay row still thins. 164: - **Review:** filled by the orchestrator. 165: - **Not done:** turning either tuning on (owner gate: stops at the mock); the rain-independence rule and its test_time checks (only on a yes); photo mode, tilt-shift and the perspective lens not captured; `--verify` at 1440x900 (inherited stall); GPU-only time not measured. Platforms: Mac run; Windows not run. 167: **Beautify P10m Beacon as a tower param (9 October 2026):** Opus 5.5 on this Mac (plan section "P10m"; report polish backlog "Beacon as a tower param (not a new kind)"; step `28-P10`, sub-batch P10-II, part m; restart of a run stopped part-way at the owner's request, whose uncommitted draft was inspected, kept and finished here). A round tower with the optional param `beacon: true` and a flat roof that nothing covers carries a lamp room on its roof (`Architecture.beacon_room`: a stone drum, a painted iron gallery and rail on eight posts, "warm" lit glass between six mullions with one lamp emitter (energy 2.0, range 6 m, fixture order 1, steady: no blink), a lintel ring, a cap in the roof colour and a finial; radius 0.46 of the tower's, 0.42-0.85 m, at least 0.22 m inside the shell; all of it under `DioramaState.BEACON_TOP` 2.2 m over the wall head, appended after everything the tower draws, no rng drawn), and the view draws a beam from the lamp (`BeaconBeam`, `shaders/beam.gdshader`): two soft cones back to back, 15 m each, 0.12 to 1.5 m, tipped 0.05 rad down, added light (unshaded, `blend_add`, no depth written, depth-tested, banded in 4 alpha steps, render priority 116: after the water (-10) and cloud shadows (115), before the rain veil (117) and fog (120); reads no screen or depth copy). **When:** `amount_of(lamps, rain)` = max(smoothstep(0.6, 0.95, lamps), 0.6 x smoothstep(0.45, 0.8, rain)): night 1.0, 05:45 about 0.95, 20:45 about 0.61, golden hour (lamps 0.55) and every day at rain up to 0.45 nothing (node hidden: no draw call); `gain_of(lamps)` scales the alpha, `DAY_GAIN` 2.0 by day falling to 1 at night (new in this run: the draft's day-rain beam was nearly invisible, `look/g2-rain.png` against `draft-run/d-rain.png`). **Sweep:** one turn per 16 s of the light clock under Ambient motion Full only (two beams, so a point is passed 0.125 times a second, D27's limit is 3); Reduced and Off hold it where it stands; every automated still holds the clock, so it rests at `REST_YAW` 0.35 rad. **Where:** a beacon tower's roof gets no derived roof pieces (`Composition.roof_enabled` false); the hover lift box reaches `BEACON_TOP` higher (`Main.lift_boxes_of`, `Architecture.beacon_lift`); the object's own boxes are unchanged. Merged clusters draw the lamp room at the end of `ClusterArchitecture.build_buildings`; drag previews (massing) do not draw it. The beam nodes exist only in a scene with a lit beacon, are moved on every drag step (`sync_live`, new in this run; the draft left the beam where the tower was until release) and are rebuilt only when the number of beacons changes. The shader warm-up draws the beam's material (`ShaderWarmup.entries_of`, `WorldView.beam_material_of`; new in this run) so the first beacon night does not compile on screen (an automated run, which skips the warm-up, ends about 56 ms later on its first beam frame). Param in the scene JSON only (no editor control, D26 style); only true or false validate; save version stays 2. **The param is OFF everywhere: no starter, fixture or Harbour object carries it, and Harbour steps is unchanged (P10g stopped at its mock; placing the beacon waits for the owner's answer to the P10g sheet).** **Reverting:** `DioramaState.BEACON_ON = false` ignores the param everywhere (no lamp room, beam, lift or roof change): checked in a scratch tree with the switch off, `test_beacon` 146 checks, 0 failures (`m/off/test_beacon-off-final.log.out`), and captures `m/cap/off-tower-1` (beacon-tower.json) against the live plain twin `after-plain-1` (`m/metrics/offswitch.txt`, 3 views): integer counts identical; changed >8 pier-day 2.36%, pier-night 2.24%, pier-rain 1.82%, against the noise pairs before-tower-1/2 2.22%, 1.03%, 1.66% (`m/metrics/oldbuild.txt`) and after-tower-1/2 1.26%, 0.95%, 1.40% (`m/metrics/beam.txt`): pier-night is over twice the noise; no mask was made of this pair, so the pixel difference is not attributed (the integer counts are the evidence that nothing extra is drawn). Test-only: `--beam=off` draws the beams at amount 0 (node, draw call and first-draw compile kept; a capture then differs by the beam's light only). Files: `scripts/scene_state.gd` (`BEACON_PARAM`, `BEACON_ON`, `BEACON_TOP`, `has_beacon`, validation), `scripts/architecture.gd` (`building()` hook, `beacon_*` constants and functions, `beacon_room`), `scripts/cluster_architecture.gd`, `scripts/composition.gd` (`roof_enabled`), `scripts/main.gd` (`--beam=off`, `lift_boxes_of`), `scripts/world_view.gd` (beam state, `beam_material_of`, `_sync_beams`, `_shine_beams`, `_sweep_beams`, calls in `_process`, `_lights_commit`, `sync_live`, `sync_environment`, `sync_profile`), `scripts/shader_warmup.gd`, new `scripts/beacon_beam.gd`, `shaders/beam.gdshader`, `tests/test_beacon.gd`, `tools/beacon_scene.gd` (and their `.uid`), `tools/test.sh`, `tools/Test.ps1` (the new suite), `tools/repo_evidence.sh` (lists `artifacts/beautify/28-P10/m/scenes/beacon-*.json`). Snapshots: pre-part `build/backup-feel/28-P10-m-pre/` (equals `28-P10-g`; at the restart the live tree differed from it by exactly the stopped run's draft list and nothing else, `m/diff-live-vs-m-pre-at-start.txt`); post-part `build/backup-feel/28-P10-m-post/`. Scratch trees `artifacts/beautify/28-P10/m/pre-tree` (from `28-P10-m-pre`, built before `BeaconBeam` was registered) and `m/off-tree` (live code, switch off). Evidence under `artifacts/beautify/28-P10/m/` (below `m/`); file backups `m/backup/`; the stopped run's unreviewed logs and captures moved to `m/draft-run/` (not cited). Power: AC, 100% charged, Low Power Mode 0 for every capture set, recording, bench and suite (`m/cap-power.txt`, `m/cap-fixed/power.txt`, `m/rec-power.txt`, `m/bench/power.txt`, `m/suite-power.txt`); 1-min load 1.84-2.55 during the captures (`m/cap-power.txt`) and 1.86-2.50 during the bench (`m/bench/power.txt`); the fixed-step, recording and suite power files record no load. 168: - **Profiles and modes reached:** only scenes whose tower carries `beacon: true` (none shipped). There: canal look (day, golden, 05:45, 20:45, night, rain by day and night) and Glade (`--look=glade`: the beam shows at night, +1 draw call), photo mode and tilt-shift draw it too (not captured; it writes no depth, so photo bokeh blurs it with what is behind). Ambient motion Full sweeps, Reduced and Off freeze; the lamp never blinks. 169: - **Tests:** test first `m/testfirst/m-testfirst-test_beacon.log.out` (the stopped run's, 40 checks, 15 failures: no API) and `m/testfirst/m-testfirst2-test_beacon.log.out` (this run's new checks against the draft: 230 checks, 6 failures: no `gain_of`, no gain uniform, the warm-up not drawing the beam, the gain not set by day or night, the beam not following a drag). New suite `test_beacon` 235 checks, 0 failures (`m/logs2/test_beacon-2.log.out`): the 13 pins (11 tower fixtures in 5 styles, flat, pitched, small, arched, merged, covered, plus Harbour's clusters and document md5 `45390862e758ec662cc67855410785e5`) equal the pre-part tree's (`m/testfirst/pins-pre-tree-3.log.out`, same PIN lines); validation (only bools), save and reopen byte for byte, version 2, undo and redo of the param; `beacon: false` draws as absent; pitched or covered draws as without; the lamp room only appended, one lamp emitter at `beacon_lamp`, on the roof under `BEACON_TOP`, inside the shell; lift, boxes, roof pieces; `amount_of`, `gain_of`, `advance`, `sweeps`, the mesh (one surface); the shader's render modes, no screen/depth read, no TIME; the view: no node in Harbour, one beam at the lamp, hidden by day and at golden hour, shown in heavy rain by day with the gain, night gain 1, `--beam=off`, the warm-up, a drag step moves the same node, Full turns it, Reduced and Off hold it, the automated hold rests it at `REST_YAW`, taking the param away removes it. Full suite (`tools/suite_all.sh`, `VERIFY_RES=2880x1800`): 47 suites, 24,756 checks, 0 failures, `--verify` 2880x1800 exit 0 with no integration failure (`m/suite/suite_all.out`); the known flaky `test_creature_reactions` check did not fail (its rerun 5583/0). 1440x900 `--verify` not run (inherited stall, rule 2). 170: - **Older builds:** the pre-part tree (an older build) opens `beacon-tower.json`: validates (''), keeps the key on save, draws every cluster with the same signature as `beacon-plain.json` (`m/oldbuild/oldbuild.log.out`); captures `before-tower-1` against live `after-plain-1` (`m/metrics/oldbuild.txt`): integer counts identical on all 9 views; changed >8 0.41-2.36%, e.g. pier-night 2.357%, pier-2045 2.25%, harbour-night 1.053%, glade-night 0.407%, against the before-tower-1/2 noise pair's 1.03%, 1.815%, 0.734%, 0.326% (after-tower-1/2 pier-night 0.95%, `m/metrics/beam.txt`): above noise, pier-night over twice it. Only the mask `m/masks/pier-night-before1-plain1.png` attributes the difference (smoke, cloud, rain and sparkles, no tower or lamp pixels), and only for pier-night; the other views are not masked. 171: - **Changed expectations (old -> new):** none changed. New suite `test_beacon` (235); `test_warmup` 53 unchanged. Suite total 24,521 -> 24,756, 46 -> 47 suites. **Journey count deltas expected: none** (no journey scene has a beacon; Harbour views' integer counts identical, below). 172: - **Captures** (`m/cap.sh` stages a, b1, b2, b3, all `tools/art_capture.sh` 1600x1000, one sitting 12:55-13:08, every view try 1; fixtures `m/scenes/beacon-tower.json` = P10g's "beacon" option with the mock lamp room and its lanterns taken off and the param set, and its twin `beacon-plain.json`, written and read back by `tools/beacon_scene.gd`, `m/logs2/beacon_scene.log.out`; that log's md5 values `4acf3a7ff726a166a35adf0fc82a962f` (tower) and `3104ff7d3e6868c370e79d387d019c40` (plain) are of `JSON.stringify(doc)`, not of the file bytes, whose md5 is `60725fb62b5c09a03cde829de0b254e7` and `435d085b77eef13ffd9ba43171da8c46`): Harbour `before-harbour-1/2` (pre-part tree) and `after-harbour-1`; beacon scene `before-tower-1/2` (pre-part tree), `after-plain-1`, `after-tower-1/2`, `after-nobeam-1` (`--beam=off`), `after-hiddenbeam-1` (the earlier hide-the-beam flag, kept), `off-tower-1`; views pier-day, pier-golden, pier-0545, pier-2045, pier-night, pier-rain (rain 0.9), pier-nightrain, harbour-night, glade-night (pier camera 0.78,0.5,20 at (12,4,-4)). Sheet `m/sheet/p10m-plain-vs-beacon.png` (plain twin left, beacon right; looked at). Looked at: `after-tower-1` pier-night and harbour-night (a lighthouse: lit lantern inside the square parapet rail, beam over the water), pier-rain (`look/g2-rain.png`: a pale warm shaft over the grey water), `look/g2-lamp-night-x3.png` (the lamp room 3x), masks below. Fixed-step captures `m/fixed.sh` (`--fixed-fps 30`, so TIME is the same on the capture frame): `m/cap-fixed/on-1`, `off-1`, `on-2`. Recordings `m/rec/full`, `reduced`, `off` (`tools/art_record.sh`, pier camera at night, 8/4/4 s; `contact.png` looked at: the beam turns under Full, stands still under Reduced and Off). 173: - **Acceptance, measured:** (1) "a tower without `beacon` draws exactly as before": the 13 pins equal the pre-part tree's (byte-level batch signatures); Harbour default `before-harbour-1` vs `after-harbour-1` integer counts identical on 6 views (draw calls 921/922 at the bench), changed >8 0.56-2.48% against three-capture noise 0.43-4.55% (`m/metrics/harbour.txt`, `harbour-b2-a1.txt`). (2) "older builds ignore the param": above. (3) "sweeps at night and in rain, not by day": beam hidden (`after-hiddenbeam-1`) vs shown (`m/metrics/drawcalls.txt`): +1 draw call at night, 05:45, 20:45, rain by day and night and Glade night; integer counts identical at pier-day and pier-golden (no draw), and fixed-step beam band changed 34.4% (pier-night), 24.8% (pier-rain), 16% (harbour-night) against 0.35-1.05% noise (`m/metrics/mirror-fixed.txt`). (4) "never in the mirror": fixed-step, the water below and in front of the beam changes 0.33-0.43% (pier-night), 0.36-0.59% (pier-rain), 0.43-0.52% (harbour-night, water from y 520) with the beam on vs off, against 0.33-0.46% for on vs on; masks `m/masks/fixed-*.png` show only rain streaks in the water. (The sitting's free-running captures showed 2-9% water differences on vs off: rain-ring phase, the beam runs ending 56 ms later in engine time, `m/metrics/water-pairs.txt`; the fixed-step pair removes it.) (5) "freezes under reduced motion": headless (Full turns a quarter turn in 4 s; 0.5 and 0 hold to 1e-6) and recorded (`m/metrics/sweep.txt`: the beam's sky band against second 0, Full 4.7-9.4% rising, Reduced 2.6-2.9%, Off 2.4-2.6% flat = rain and stars). 174: - **Counts (lamp room alone, `after-plain-1` -> `after-hiddenbeam-1`, `m/metrics/drawcalls.txt`):** draw calls +4 (pier views) and +6 (harbour-night, glade-night), emitter sources, reflection streaks and paving streaks +1 each. Whole scene: plain twin 972/973 -> beacon 979 draw calls at the bench (13:00 and 01:00). 175: - **Bench** (`m/bench.sh`, P10g's recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, 1920x1080, `--disable-vsync`; 5 interleaved rounds after a warm-up, medians, `m/bench/summary.txt`): Harbour live vs pre-part p50 -0.003 ms (13:00), -0.017 ms (01:00), p95 -0.66 / -0.02 ms, draw calls 922 / 921 both; dense fixture 01:00 p50 -0.049 ms, p95 -0.04, 2588 both; the beacon scene, tower vs plain twin, p50 +0.026 ms (13:00, beam hidden), -0.046 ms (01:00, beam shown), within noise. 176: - **Open for the owner / known:** (a) the beacon lamp is one more lamp in the 16-light budget: in the pier-night view it takes the slot of the tree house's pink light, whose pool on the bay window goes (`look/blob-pair.png`, mean luma -9 there); an ordinary lamp's energy (1.3) does not change that (`m/iter/e13`); not changed. (b) a user can stand parts on a beacon roof (it stays a flat roof) and they may cross the lamp room; perching birds not checked. (c) a covered beacon tower also loses its derived roof pieces (`roof_enabled` has no state). (d) a beacon tower made with a build animation (paste, undo) shows its beam at the light pass, before the tower has risen. (e) the colour and strength of the beam are the draft's tuning plus the day gain; the P10g sheet's questions (placement, tower colour) are still the owner's. 177: - **Review:** filled by the orchestrator. 178: - **Not done:** Harbour placement (owner gate: the P10g sheet is unanswered, so no `sample_scene.gd` change); photo mode, tilt-shift and the perspective lens not captured; `--verify` at 1440x900 (inherited stall); GPU-only time not measured. Platforms: Mac run; Windows not run. 180: **Beautify P10g Harbour replanting, with the beacon mock (9 October 2026):** Opus 5.5 on this Mac (plan section "P10g"; report polish backlog "Harbour replanting" and "Beacon as a tower param", trees lens F6, architecture lens F7 revised; step `28-P10`, sub-batch P10-II, part g). **Stopped at the mock, as the plan's owner gate and the owner's 8 October note require: Harbour steps (the default scene and the starter) is unchanged** (its document is byte for byte the pre-part one, md5 `45390862e758ec662cc67855410785e5`; every capture's integer counts identical, pre-part tree against live). The replanting and the beacon exist as two options behind one switch: `SampleScene.HARBOUR_MOCK` (`""`, the default, builds Harbour as before; `"replant"`; `"beacon"`), and `SampleScene.harbour(mock)` builds an option for the sheet without the switch. The options are applied after the whole Harbour build (`_harbour_mock`), so every other object keeps its id and bytes. **"replant"**: five trees the default camera cannot see (the ones at (-13.5,-9.5), (-15.5,-4.5), (-12.5,-12.5), (-16.5,-7.5), (-16,-11)) move to the garden's far corner (5.6 m), the front-left corner outside the garden wall (4.0 m), the front lawn by the stairs (4.4 m), the right lawn behind the road (4.8 m) and the east bank's front corner (3.4 m), as `grow_h`; two conifers (8.8 m, 6.2 m) stand behind the gate tower (the right round tower); the five hidden shrubs and their flower patches move into two groups of three (garden west wall; the canal house's stair corner); the two seen shrubs stay. The front-left tree is near the quay, so the existing shore rule (Composition, reeds within TREE_REEDS of a tree, lily pads near them) grows reeds and lily pads there. **"beacon"**: "replant" plus the east pier (cells x 14..18, z -8..-4) painted stone, a 2.5 m round canal tower (variant 2) 8.5 m to a flat roof at (16.5,-6.5), on its roof a 1.5 m round tower 2.0 m tall with a pitched roof as the lamp room, two wall lanterns on it and four post lanterns round it; the bench and bollard move onto the pier and the pier tree to (17.6,-2.4). It is a placement mock from existing parts: no beam, no `beacon` param (that is P10m, after the owner places it). Save format unchanged (no new key; the options are version 2 documents that older builds open). New `tools/harbour_mock.gd` writes the three documents (`harbour-before.json`, `harbour-replant.json`, `harbour-beacon.json`) to `artifacts/beautify/28-P10/beacon-mock/` (the plan's folder; listed in `tools/repo_evidence.sh`), validated and read back byte for byte. **Reverting / switching:** nothing to revert for the default; on the owner's yes, `HARBOUR_MOCK = "replant"` or `"beacon"` makes it the default, and the expectations below then need refreshing. Files: `scripts/sample_scene.gd` (constants `HARBOUR_MOCK`, `HARBOUR_MOCKS`, `REPLANT_TREES`, `REPLANT_CONIFERS`, `REPLANT_SHRUBS`, `BEACON_PIER`, `BEACON_TOWER`, `BEACON_MOVES`; `_move_object`, `_harbour_mock`; `harbour(mock := HARBOUR_MOCK)`, one line before the starter atmosphere), `tests/test_starters.gd`, new `tools/harbour_mock.gd` (and its `.uid`), `tools/repo_evidence.sh`. Snapshots: pre-part `build/backup-feel/28-P10-g-pre/` (saved before the first edit; `diff -rq` against `28-P10-e` and the live tree on scripts shaders tests data scenes tools and `project.godot` both empty, so nothing unexpected changed: `g/diff-live-vs-28-P10-e.txt`, `g/diff-pre-vs-live-and-e.txt`); post-part `build/backup-feel/28-P10-g-post/` (equal to the live tree, `g/diff-live-vs-post.txt`). Scratch trees `artifacts/beautify/28-P10/g/pre-tree` (from `28-P10-g-pre`) and `g/switch-tree` (live code, switch set per run). Evidence under `artifacts/beautify/28-P10/g/` (below `g/`); backup `g/backup/sample_scene.gd`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, bench and suite (`g/cap-power.txt`, `g/bench/power.txt`, `g/suite-power.txt`); 1-min load 1.74-2.60 during the captures (`g/cap-power.txt`), 1.77-2.61 during the bench (`g/bench/power.txt`) and 3.08 at the start of the suite (`g/suite-power.txt`). 181: - **Profiles and modes reached:** the default reaches nothing (unchanged everywhere). The options are the Harbour document, so after a yes they reach every look that draws Harbour: canal (harbour-day/golden/night, lane zoom), Glade (`--look=glade`, same scene; sheet page 3), photo mode and tilt-shift on it, and the drag and release baselines whose fixture is Harbour. Reduced/Off motion: nothing new moves (the mock lanterns are ordinary lanterns). 182: - **Tests:** test first `g/testfirst/g-testfirst-test_starters.log.out` (225 checks, 2 failures: the constants and `tools/harbour_mock.gd` missing; the rest of the new checks sit behind that guard). New `harbour_mock_checks` and `beacon_mock_checks` in test_starters: Harbour byte for byte against the pre-part md5; `HARBOUR_MOCK` is `""`; `harbour()` follows the switch; each option valid, version 2, saved and reopened byte for byte, built the same every time, same camera and environment; only trees, shrubs, flower patches (and for the beacon the bench and bollard) move, nothing removed, only conifers (and the beacon tower family) added; 4-5 moved trees were hidden from the default camera (a ray from the crown toward the orthographic camera meets a building's box) and every moved tree is now seen, `grow_h` 3.2-6.5, on a land cell, 0.5 m clear of buildings, 1.8 m from other trees, 1 m from ground parts; two conifers behind the gate tower, within 3.5 m across the view, clear of its wall, tops seen over the walls; eight shrubs all seen, at least six with two mates within 1.6 m, each moved shrub with its patch; perches (ridge on its roof, lamp on its lamp, ridge/lamp/nest kinds present) and ivy seeds (stored nothing, every seed at a wall base on the ground) valid on both options; the pier paved and the only terrain change, heights unchanged, the tower 2.5 x 8.5 m flat canal, a lamp among its parts, taller than the gate tower, its lamp room seen, bench and bollard on the paving, nothing inside the tower. Affected suite `g/suites/r2-test_starters.log.out` 367/0. **Full suite** (`g/suite/suite_all.out`, `VERIFY_RES=2880x1800`, 11:11-11:20): **46 suites, 24,521 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Log errors in captures and suites: the pre-existing `terrain_flowers.svg` not found, and the engine's exit line "ERROR: 2 resources still in use at exit" (16 lines in `g/suite`, 20 in `g/cap`, in pre-part and live sets alike); no other error. 183: - **Changed expectations (old -> new):** none changed; test_starters 222 -> 367 checks (new P10g checks only). Suite total 24,376 -> 24,521. **Journey count deltas expected: none** (Harbour unchanged: `g/metrics/counts-pre1-live1.txt` identical on all nine views; the scene-file path `harbour-before.json` identical too, `g/metrics/counts-live1-json1.txt`). 184: - **What a yes would change** (switch on in `g/switch-tree`, every suite once, against the same tree with the switch at a non-option value; scratch-tree failures from missing `res://artifacts` fixtures and suite order are in both and left out: per-suite logs `g/switch/beacon/`, `g/switch/off/` (`beacon/summary.txt` holds only the second half of the beacon run; the first half is in its logs), lists `g/switch/beacon-only-failures.txt`, `g/switch/replant-only-failures.txt`). With "beacon": test_starters (the pre-P10g md5 and `HARBOUR_MOCK` checks, by design), test_roofs "crownless" `["roof:9"] -> ["roof:9","roof:139"]` (the beacon tower's flat roof), test_nature "every chunk unchanged or under a footprint that lost vertices" (chunk `harbour_2_-1` changed by the pier paving), test_release_spread "harbour animated, drained: identical to a release built whole (differs in decorations)", test_foliage_merge "harbour: every grown plant has a decoration slot", test_warmup "The view has the children and decorations it had (170, 100)". With "replant" the last three fail too (167, 99). Those three are not just counts: they say the conifers (or the moved trees) take a different decoration path in the release spread and the merged foliage slots; they must be understood (expectation or a conifer gap) before a yes, together with the plan's re-based drag baseline and journey. Not investigated further (the part stops at the mock). 185: - **Captures** (`g/cap.sh` stages a, b, c, `g/cap2.sh`; all through `tools/art_capture.sh`, 1600x1000, one sitting 10:31-10:43, every view try 1): pre-1 and pre-2 (pre-part tree, Harbour, no scene: the before and its noise pair), live-1 (live, no scene), json-1 (live, `harbour-before.json`), replant-1, beacon-1, nine views each: harbour-day/night/golden, glade-day/night, harbour-lane, lane-night, pier-day and pier-night (camera 0.78,0.5,20 at (12,4,-4)). The lane and pier views of the scene-file sets were first drawn with the file's own camera (a `--camera` before `--scene` loses, the P10e lesson) and were recaptured with the camera after the scene (`g/cap2.sh`). Plant-mask sets at rain 0, day and night: `m-before`, `m-before-2` (noise), `m-replant`, `m-beacon` and each `-bare` (the same document without trees, blossoms, conifers and shrubs). Unchanged default, measured against the noise pair (`g/metrics/noise-pre1-pre2.txt`, `pre1-live1.txt`, `pre2-live1.txt`): pre-1/live-1 changed >8 on 0.39-3.17% of pixels per view against pre-1/pre-2 0.50-4.89%; the scene-file path (`live1-json1.txt`) 0.30-2.61%. Per view the ranges hide three views above the noise pair: pier-night 2.326% vs 0.91%, harbour-night 1.741% vs 1.084%, pier-day 1.612% vs 1.56% (pre-2/live-1 2.445%, 1.102%, 2.027%). The integer counts are identical on all nine views and the default document's md5 is identical, so nothing extra is drawn; no mask was made of these pairs, so the per-view pixel difference is not attributed (presumably TIME-driven rain, cloud and smoke phase, as in the other parts, but not shown here). Sheet `g/sheet/` (all looked at): `p10g-1-day` (whole island, day and golden, before / A / B), `p10g-2-night` (night, lane at night), `p10g-3-glade`, `p10g-4-pier` (pier and right lawn closer: B's paved pier, lit tower at night), `p10g-5-left` and `p10g-6-right` (1:1 details), notes and owner questions `p10g-sheet.md`. Tuning rounds `g/iter/r1`, `r2` (looked at `g/iter/look/`): the second front-lawn tree was moved off the stair line (it overlapped the first on screen) and the conifers spread so both show; the lamp room raised to 2.0 m. 186: - **Acceptance, measured:** (1) one owner mock sheet decides replanting and beacon placement: `g/sheet/` with `p10g-sheet.md` (four questions). (2) "After a yes the front and right lawns read as planted in harbour-day and harbour-night", measured on the options now (`g/plant.js`, `g/metrics/plant-share.txt`, masks `g/metrics/plantmask-*.png`, looked at `plantmask-replant-mday-crop.png`; plant pixel = option against its plant-free twin over 8 in any channel, so crowns, shrubs and their shadows, plus tree reflections where a region holds water): garden and front-left lawns (`p10b_lawn_left`) **7.08% -> 22.28%** by day, **6.48% -> 20.58%** at night; back and right lawn (`p10b_lawn_wide`) 7.04% -> 19.34% (6.54% -> 15.13%), with the beacon 14.43% (12.36%, the tower hides some); east bank (`lawn_right`) 9.76% -> 26.50% (8.35% -> 23.11%), beacon 31.57% (25.42%); noise 0.07/0.09/0.17% by day, 0.01-0.03% at night. (3) Every changed expectation listed: none now; the list for a yes is above. (4) Perches and ivy seeds still valid: by test, on both options. (5) Drag baseline and journey re-based once after the last Harbour-changing part: not applicable until a yes (nothing changed). 187: - **Counts (options against before, `g/metrics/counts-pre1-replant1.txt`, `counts-pre1-beacon1.txt`):** replant harbour-night draw calls 911 -> 935, foliage plants 221 -> 238, parts 136 -> 138; beacon 911 -> 973, parts 136 -> 146, emitter sources and reflection streaks 170 -> 190, paving streaks 25 -> 45, foliage MultiMeshes 6 -> 7 (one moss group). Both options also show light_sync_calls 1 -> 2 and pool_resets 0 -> 1 in a smoke: probes (`g/probe/sync/`) put it on the moved trees (moving only the trees gives 2; only `grow_h`, only a conifer or only the shrubs give 1); the smoke builds Harbour and then restores the scene file, and the clusters near the moved trees rebuild once more. Probably an artefact of the scene-file path (a default with the planting built in would sync once), not proven. 188: - **Bench** (`g/bench.sh`, P10e recipe: smoke, `--churn=100000`, `--motion-live`, rain 0.45, 1920x1080, `--disable-vsync`; rounds 0-5 interleaved, round 0 discarded, 10:44-10:54; `g/bench-summary.js`, `g/bench/summary.txt`): default Harbour pre -> live p50 6.825 -> 6.835 (13:00), 6.750 -> 6.742 (01:00), p95 7.882 -> 7.461 and 7.794 -> 7.725 (unchanged code; noise). Options against the live default: replant p50 **+0.110 / +0.109 ms** (slower 5/5), p95 +1.00 / +0.34; beacon p50 **+0.232 / +0.234 ms** (5/5), p95 +0.88 / +1.01; draw calls 922 -> 948 / 988. `g/bench2.sh` (10:54-10:58, 13:00, rounds 1-4, `g/bench/summary2.txt`): the scene-file path costs nothing (default 6.779, `harbour-before.json` 6.780), replant 6.885, beacon 7.041. Dense not benched (no code it runs changed). 189: - **Review:** filled by the orchestrator. 190: - **Not done:** the default Harbour, its expectations, drag baseline and journey (owner gate: wait for the sheet); the beacon's beam, lamp room geometry and `beacon` param (P10m); the three non-count failures a yes would bring (release spread, foliage merge slots, warm-up decorations) are listed, not investigated; the second light sync in the options' smoke is explained by probe only; no golden-hour 1:1 detail page; the beacon colour is the report's navy variant (an owner question); photo mode and tilt-shift not captured on the options. 191: - **Platforms:** Mac run; Windows not run. 193: **Beautify P10e Cables as 1-px line primitives (9 October 2026):** Opus 5.5 on this Mac (plan section "P10e"; report polish backlog "Cables as `PRIMITIVE_LINES`", pixel lens F7; step `28-P10`, sub-batch P10-II, part e; a look change, approved as built under the owner's standing approval if it clears review). Every cable is now a run of GL lines, one render pixel (one art pixel) wide and unbroken at every zoom, where the 0.018-0.028 m beams covered 0.3-0.5 art px and broke into dashes: the lived links' festoon cable, laundry cord and bunting cord (`Architecture._link_cable`), a string_lights fitting's cable (run wall/around, drawn in its cluster) and a string_lights decoration's swag, and every festoon bulb's hanger. `MeshBatch.cable()` appends a segment (two vertices, its direction as the normal) to the `cable` group; `build()` draws that group as one `Mesh.PRIMITIVE_LINES` surface; `merge()`, keyed pieces (the reveal, two vertices a piece) and `growth_log` carry it; UV2 stamping skips it (neither weathered nor dusk-seeded). The `cable` material is the plain body with `CABLE` defined (`shaders/cable.gdshader`, build variant `cable_build.gdshader`, in `BuildAnim.MATERIALS`): lit as plain, with the line's normal turned in the vertex shader to the part of the view direction square to the line (the camera side of a round cord), opaque in the default pass, selection-lifted with its part. Festoon bulbs (`MeshBatch.bulb()`, the same boxes) carry each corner's offset from the bulb's centre and the bulb's smallest side in CUSTOM1; the neon body grows a bulb about its centre to at least `BULB_MIN_ART` (1.0) art px (`art_pixel`, or its depth times `art_pixel_slope` under the lens), so nothing changes at the quarter and Harbour zooms (0.042 and 0.076 m an art px against 0.08 and 0.11 m bulbs) and a zoomed-out bulb no longer drops between pixel centres. Deviation from the report: the cable is lit (plain body), not an unshaded profile colour. **Reverting:** `MeshBatch.CABLE_LINES = false` (every cable the beam it was and every bulb the plain box: the stock scenes' P7b geometry fingerprints hold byte for byte, test_dusk 461/0 in a scratch tree with the switch off, see Review; a batch's `cable_beams` does the same, and `test_dusk` fingerprints with it; the `cable` material is still made and is warmed on the plain box it was warmed on before P10e, built as triangles); `BULB_MIN_ART = 0.0` in `shaders/neon_body.gdshaderinc` and `MeshBatch` (no growth). Test-only probe (never saved): `--cable-mask` draws every cable flat green, `--cable-mask=beams` also draws the cables as the old beams into the cable group as triangles (`MeshBatch.probe_beams`), so coverage is measured in one tree and sitting. Files: `scripts/mesh_batch.gd`, `scripts/architecture.gd` (the three string-light sites, `_link_cable`, `materials()`), `scripts/build_anim.gd`, `scripts/world_view.gd` (`debug_cable_mask`), `scripts/main.gd` (the probe flags), `scripts/shader_warmup.gd` (`_sample`: the cable warmed as lines, the warm/pink/cyan mesh piece as a bulb with CUSTOM1; with the switch off the cable on its old box), `shaders/plain_body.gdshaderinc`, `shaders/neon_body.gdshaderinc`, new `shaders/cable.gdshader`, `shaders/cable_build.gdshader` (and their `.uid`), tests `test_links.gd`, `test_warmup.gd`, `test_build_anim.gd`, `test_dusk.gd`, `test_foliage_merge.gd`. No new fixture or region (nothing for `tools/repo_evidence.sh`). Snapshots: pre-part `build/backup-feel/28-P10-e-pre/` (saved before the first edit; `diff -rq` against `28-P10-S1` on scripts shaders tests data scenes tools and `project.godot` empty, and the live tree equal to it: nothing unexpected, `e/diff-live-vs-S1.txt`); post-part `build/backup-feel/28-P10-e-post/` (equal to the live tree, `e/diff-live-vs-post.txt`). Scratch tree `artifacts/beautify/28-P10/e/pre-tree` (from `28-P10-e-pre`). Evidence `artifacts/beautify/28-P10/e/` (below `e/`); file backups `e/backup/`. Power: AC, 100% charged, Low Power Mode 0 for every capture set, bench run and suite (`e/cap-power.txt`, `e/bench/power.txt`, `e/suite2-power.txt`); 1-min load 3.44-4.81 during the bench (`e/bench/power.txt`). 194: - **Profiles and modes reached:** Canal (quarter) and Harbour at pixel size 2: every cable, the bulbs only when zoomed out past 0.08 m an art px (far zoom, below). Glade (pixel size 1): its cables were already about one window px; they are now hair-thin lines, lit slate-blue where the beams read near-black (`sheet/p10e-7`, top). Photo mode (perspective): lines one render px, bulbs grown by their depth; photo-golden changes on 1.6% of pixels >8 against noise 0.07%/0.01% (the cables; `sheet/p10e-7`, middle). Tilt-shift: the lens blurs the lines like everything else (tilt-night >8 3.3%/2.0% against noise 1.5%/0.3%). Reduced/Off motion: nothing moves in this part, so no change. 195: - **Tests:** test first `e/testfirst/e-testfirst-test_links.log.out` (366 checks, 32 failures, all P10e, no script error) and `e/testfirst/e-testfirst-test_warmup.log.out` (49, 4 failures). New `p10e_checks` and `p10e_world_checks` in test_links: the constants and switches, the shader contract (`BULB_MIN_ART` equal in GDScript and shader, CUSTOM1, `art_pixel`, the 0 gate, the CABLE wrappers and block), the CPU twin of the bulb growth (at least 1 art px at 0.042-0.32 m, unchanged at 0.042 and 0.076, doubled at 0.16 for the 0.08 m bulb, none at 0), a festoon's cable end to end on the sag curve plus its 0.07 m hanger lines and no beam left in metal, laundry and bunting cords as lines in d9d2c3 with the cloth unchanged, the `cable_beams` revert (old beam counts, plain bulbs, no CUSTOM1), `build()` (one lines surface, no UV2 or CUSTOM0, bulbs with CUSTOM1), `merge()` (offsets kept on their own vertices, rotated with their frame, zeros elsewhere), keyed reveal (CUSTOM0 on a lines surface), `growth_log` (one two-vertex piece, not grown), the probe, a wall fitting's and a decoration's cable and bulbs, and in WorldView the Harbour's links mesh (one lines surface, bulbs with CUSTOM1), its decoration lines, and a new house's fitting cable revealed in the cable build variant then handed back. test_warmup: the cable warmed on a mesh, both MultiMeshes and as its build variant, every warmed cable piece a lines surface, the three lit mesh pieces bulbs with CUSTOM1. Affected suites `e/suites/` (`r1`-`r3`). **First full suite** (`e/suite/suite_all.out`, 08:51-08:59): 3 failed checks in 2 suites from this part (test_foliage_merge 87 checks, 2 failures: the same check for harbour and quarter; test_tree_growth 1), all fixed: test_tree_growth "plain_body morphs the vertex before taking its world position" (the CABLE block sat above the morph; moved after `build_vertex`) and test_foliage_merge "grown surfaces, and only they, wear the grown materials" (changed expectation, below). **Final full suite** (`e/suite2/suite_all.out`, `VERIFY_RES=2880x1800`, 09:41-09:50): **46 suites, 24,372 checks, 0 failures**; `--verify` 2880x1800 exit 0 on try 1 (33 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions 5,583/0 and rerun 5,583/0 (the known flaky check did not fire). 1440x900 `--verify` not run (inherited stall). Log errors in captures and suite: the pre-existing `terrain_flowers.svg` not found, and the engine's exit line "ERROR: 2 resources still in use at exit" (20 lines in `e/suite2`, 48 in `e/cap`), both in both trees; no other error. 196: - **Changed expectations (old -> new):** test_links 325 -> 377 checks; the P8i1 cloth checks' `link_batch` now sets `cable_beams` on both sides, so "the cord stays plain" and "plain before = plain + cloth now" are held against the beam cord (the line cord is p10e_checks'). test_dusk P7b geometry fingerprints: the stock clusters are built with `cable_beams` (string-light fitting cables moved from metal beams to lines), so everything else is still held byte for byte (461 -> 461). test_build_anim: `cable` joins the A2b contract with `plain_body` as its body (515 -> 523, the per-material loop). test_foliage_merge: "a surface with CUSTOM1 wears a grown or bend material" now also allows a string_lights decoration's bulbs (CUSTOM1 offsets, settled neon material), and Harbour and the quarter must have such surfaces (87 -> 89). test_warmup 45 -> 53 (49 at implementation, +4 from the review fix). Suite total 24,306 -> 24,376 (24,372 at implementation). **Journey count deltas expected:** draw calls up by the new cable surfaces, measured +2 to +3 on harbour-night/day/lane and +6 on quarter-night/day (`e/metrics/counts-pre1-live1.txt`; styles-day not captured, it has no string lights); every other integer count unchanged apart from the orbit views' known within-tree variation. 197: - **Captures** (`e/cap.sh`, `e/cap2.sh` for the two quarter yaw views, `e/capdry.sh`, all through `tools/art_capture.sh`, 1600x1000, one sitting 09:01-09:25, every view try 1 except one retry in `m-plain`): look sets pre-1, live-1, pre-2, live-2 (21 views: harbour-night/day/golden/lane, lane-night, quarter-night/day/golden/2100, glade-day/night, photo-golden, Harbour yaw 0.70 night and 0.86 day, quarter yaw 0.78 day and 0.91 night, quarter-orbit and qn-orbit `--orbit=0.6`, far-night and far-day (Harbour at camera size 64), tilt-night); mask sets on the live tree at rain 0, `m-plain`, `m-lines` (`--cable-mask`), `m-beams` (`--cable-mask=beams`), 14 views (Harbour day/night/golden, yaws 0.70 and 0.86, quarter day/night, yaws 0.78 and 0.91, lane, Glade, perspective, both `--orbit=0.6`); dry mirror sets `dry-pre-1/2`, `dry-live-1/2`. The quarter yaw views were first given the camera before `--scene` and mostly drew the scene's camera; they were recaptured with `--camera=` after it (a first recapture loop wrote 7 unusable runs into `cap/m-qa`, `cap/m-qc`, moved to the Trash (`the macOS Trash/p10e-bad-yaw`), noted in `e/cap-power.txt` line 99; not checked, the Trash is outside the sandbox). Sheet `e/sheet/`, all looked at: `p10e-1` (masks, beams left, lines right: dashes against one continuous line), `p10e-2` quarter-day plaza (pre: bulbs float with no cable; live: a thin grey-blue wire from bulb to bulb), `p10e-3` quarter-night plaza (the wire is dark and faint against the dark plaza), `p10e-4` Harbour garden, `p10e-5` fish-market bunting (the pale cord is now one continuous white-ish line by day, pale grey at night), `p10e-6` far-night (strings of bulbs read as dotted chains where pre lost most of them), `p10e-7` Glade, photo-golden, tilt-night. 198: - **Acceptance, measured:** (1) every cable continuous at Harbour and quarter zooms in stills and orbit frames (`e/continuity.js`, `e/metrics/continuity.txt`; cable pixels from the mask against the plain capture, 0 false positives on 12 of 14 views, 4 px on two): over the 14 mask views, 8-connected components **1,789 beams -> 416 lines**, fragments of at most 3 art px **1,144 -> 89**; e.g. Harbour day 127 -> 14 components (117 -> 1 fragments), quarter day 150 -> 45 (63 -> 9), quarter yaw 0.78 174 -> 48, Harbour yaw 0.86 135 -> 17, lane 92 -> 23, Glade 117 -> 20. The beams covered 41-44% of the lines' path at Harbour zoom, 75-78% at the quarter. The lines' remaining components and fragments are where a bulb, a lamp head or a facade is drawn in front of the cable (fragment maps `e/metrics/fragmap-.png`; looked at `m-hn` and `m-qd` close up, `metrics/frag-hn-x4.png`, `frag-qd-x4.png`: every red fragment sits between bulbs or behind geometry). The `--orbit=0.6` mask frames gave the same numbers as the matching stills at the quarter and nearly the same at Harbour (14 vs 15 components; the automated orbit barely moves these captures), so the orbit evidence is the four yaw stills. (2) Bulbs at least 1 art px: by the CPU twin in test_links; at the far zoom (0.128 m an art px) warm bulb pixels in the island rect 291 -> 340 art px, blobs 126 -> 141 (both pre captures identical, live-2 339/140); at the Harbour zoom unchanged within the noise pair (1,164/1,162 -> 1,126/1,145 art px, `e/metrics/bulbs.txt`). (3) The line surface passes through commit, build reveal, selection lift and growth paths, is skipped by UV2 stamping and is warmed: the tests above. (4) The mirror march is not broken: canal mirror columns p50 52.6 -> 52.8 (21:00) and 43.5 -> 43.6 (night), dry 50.7 -> 51.0 and 42.5 -> 42.2, means within 0.3 luma; but pixels changed >8 in the region sit above the within-tree noise (dry 21:00: 16.3%/11.5% between trees against 1.0%/5.0% within). The dry numbers were recomputed after the P10-II close's handoff check from the existing `e/cap/dry-*` captures (q2100-dry, qn-dry) with `tools/art_metrics.js ab` into `e/metrics/dry-mirror.txt` (regions `e/metrics/dry-regions.json`) and match; dry night (qn-dry) is noisier within a tree: 14.0%/22.8% between trees against 17.0%/5.3% within. The diff map (`e/metrics/mirror-q2100-dry-x3.png`, looked at) shows scattered single sparkle pixels with the columns' shape unchanged; the live captures run about 150 ms later after start than the pre ones (`elapsed_ms` 4,082-4,122 against 3,949-3,955), so TIME-driven sparkle and rain rings land at another phase (`e/metrics/mirror-q2100-x2.png`: with rain the frame difference is mostly the rings). Not proven that no reflected cable pixel contributes. (5) Glade at pixel size 1 gets hair-thin wires: yes (`sheet/p10e-7`). 199: - **Cable colour** (`e/colour.js`, `e/metrics/cable-colour.txt`, median of the drawn cable pixels incl. the pale cords): quarter day #1d2842 (luma p50 39.5), quarter night #352f45 (50.8), Harbour day #384565 (68.5), Harbour night #634a59 (79.4, the festoons lit by their own lamps). The report suggested 1a1826 at night; the lit cable is lighter. Not tuned. 200: - **Bench** (`e/bench.sh` = the P10i recipe plus the quarter at night; `e/bench2.sh` Harbour golden only, order alternating; `e/bench-summary.js`, `e/bench/summary.txt`; 09:25-09:41; round 0 discarded): medians over rounds 1-5, pre -> live: dense day p50 14.513 -> 14.553, **p95 16.335 -> 16.268**; Harbour day p50 6.964 -> 6.955, p95 8.390 -> 7.947; Harbour night 6.855 -> 6.868, p95 8.211 -> 8.295; quarter night 7.645 -> 7.680, p95 9.369 -> 9.385; Harbour golden p50 6.822 -> 6.835 and p95 7.642 -> 8.575 (5/5 slower), so 6 more alternating rounds: pooled over 11 rounds p50 **6.811 -> 6.828**, p95 **7.667 -> 8.144** (live slower 9/11; rounds 6-11 alone 7.889 -> 8.090). Harbour p95 is bimodal (7.5-8.9 in both trees). Draw calls dense 2589 = 2589, Harbour 919 -> 922 (golden 831 -> 833), quarter 1045 -> 1051. The plan's stop line (dense p95 over about +0.5 ms) is not hit; Harbour golden p95 +0.48 ms is reported for review. 201: - **Review:** filled by the orchestrator. The review's must-fix finding and what was done: the revert switch `MeshBatch.CABLE_LINES = false` broke the shader warm-up (the reviewer's scratch tree `e/review/revert-tree`, log `e/review/revert-test_warmup.log.out`: `_sample(batch, "cable")` drew the cable as a beam into "metal", so `_piece_mesh` read a missing `groups["cable"]` and `build({"cable": ...})` a missing `materials["metal"]`, a script error at start-up), and the entry's "byte for byte" claim was untested. Fixed (9 October, 10:00-10:15): `ShaderWarmup._sample` draws the cable as lines only with `CABLE_LINES` on and the batch's `cable_beams` off, otherwise the plain box in the cable group it drew before P10e; `MeshBatch.build()` draws the cable group as lines only under the same condition (with the switch off `cable()` writes nothing there, so only the warm-up's box can, and it is built as triangles); no change with the switch on. Test first: new `beams_sample_checks` in test_warmup (a `cable_beams` batch, which takes the switch-off branch, gets the sample in the cable group, not metal, and builds it settled and as its build variant as one triangle surface in the cable material) and the line and bulb checks now follow `MeshBatch.CABLE_LINES` (0 line pieces and 0 CUSTOM1 lamp pieces when off): failing log `e/fix/testfirst-test_warmup.console` (51 checks, 2 failures: the sample in `["metal"]`); after `e/fix/fix-test_warmup.console` 53/0. The switch itself, in a scratch tree of the live code with only `CABLE_LINES = false` (`e/fix/revert-tree`, runner `e/fix/run_revert.sh`): test_warmup 53/0 (`e/fix/revert-test_warmup.log.out`), test_dusk 461/0 with the P7b fixtures copied in (`e/fix/revert-test_dusk.log.out`: the stock scenes' geometry fingerprints, cables as beams, held), test_links 344 of 372 pass, its 28 failures all P10e checks of the line look and of the switch being on (`e/fix/revert-test_links.log.out`, expected, not made to follow the switch). Windowed start-up warm-up (`e/fix/warmcap.sh`, `--smoke --warmup --startup-report`, Harbour night, AC 100% Low Power Mode 0, `e/fix/cap-power.txt`): revert tree and live tree both reach `PIXELPUNK_STARTUP` (the first, build-variant phase done) with no script error (`e/fix/warm-revert/warm-hn.log`, `e/fix/warm-live/warm-hn.log`; only the pre-existing `terrain_flowers.svg`); `courtyard.png` of each looked at: the revert shows the old dashed beams, the live the continuous lines. Affected suites live (`e/fix/fix-*.console`): test_links 377/0, test_dusk 461/0, test_build_anim 523/0, test_foliage_merge 89/0. Full suite `e/fix/suite/suite_all.out` (`VERIFY_RES=2880x1800`, 10:02-10:10, AC 100% charged, `e/fix/suite-power.txt`, which has no lowpowermode line, so Low Power Mode is not recorded for this run): **46 suites, 24,376 checks, 0 failures**; `--verify` 2880x1800 exit 0 try 1 (34 s), `PIXELPUNK_INTEGRATION []`; test_creature_reactions rerun 5,583/0. No look change with the switch on, so no re-capture or re-bench. Post-fix snapshot `build/backup-feel/28-P10-e-post-fix/`. 202: - **Not done:** the cable's night colour is the lit plain body, not tuned toward 1a1826; the pale laundry and bunting cords now read as continuous near-white lines by day (taste, owner); no motion clip of an interactive orbit (the automated `--orbit=0.6` captures do not move here; yaw stills stand in); the mirror region's above-noise scatter is attributed to the capture clock, not proven; selection lift and hover not captured (shader path only, tested by text); no `--verify` check added for the cables. 203: - **Platforms:** Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen copy = HEAD 8418d4e, 9 October 2026, 19:36 BST), lines 376-386 376: ## P10-II closed by the orchestrator (9 October 2026, ~17:00) 377: - Close on main 434302e, entry and handoff corrections committed ae74643, snapshot 28-P10-S2, journey artifacts/feel-journey/28-P10-II. Suite 47 suites, 24,844 checks, 0 failures. Only P10e (cables as lines) changes the default look; P10g, P10m and P10k are off. 378: - Owner sheet (12 pages + sheet.md) sent. Waiting on the owner: P10e (pale cords, lit cable colour), P10g (replant / beacon / neither), P10m (beacon on any scene), P10k (swags yes/no, overcast yes/no). Nothing is turned on until a yes. 379: - P10-III started: P10j launched (conditional, measurement first). 381: ## Owner rulings on the P10-II sheet (9 October 2026, ~17:30) 382: - "All approved, beacon to be a placeable option." Orchestrator reading, recorded so it can be corrected: P10e stays as built; P10g option A (replanting) becomes the default Harbour; P10k quay swags ON and rainy overcast ON at the report's strength A (not the stronger mock B); P10m beacon stays out of every shipped scene and becomes a placeable option in the editor (default off). The rain-independence rule is updated deliberately. 383: - New part P10q (docs/beautify-plan.md, beautify-stage.js stage P10 part q) applies all of this. It runs after P10j, P10n and P10o (one Godot at a time) and before the P10-III close, so the close covers j, n, o, q. 385: ## Journal update and publish (9 October 2026, ~19:45) 386: - Owner: "update and publish the journal." Deploy to https://pixelpunk-journal.pages.dev is authorised for this edition only (the journal; nothing else is pushed or published). Workflow : chapter 26 (P8) wording fixes, chapter 27 (P9), and P10 chapter(s) from the parts' own entries; P10n and P10o written as open questions; no P10-III close result claimed (that close was still running). === Appended 12 October 2026 (journal update; earlier bytes above unchanged) === Frozen 12 October 2026 at 22:42 BST (game main at 36011bd; the lines quoted below are committed: the P10r entry in a184c16 and the P10-III close earlier). Covers the P10-III stage close (parts j, n, o, q), the owner's "All approved" on its sheet (P10n stays as built, default on), the owner's P10o answer (Pixelify Sans, UI scale rule A) and part P10r, which built it. Line numbers refer to the frozen copy research/update-2026-10-12-P10close/visual-handoff-R1c-to-P10III-close.md and to artifacts/beautify/owner-decisions.md at the freeze. "~/.Trash/" reads "the macOS Trash". SOURCE: docs/visual-handoff.md (frozen 12 October 2026), lines 117-177 (Beautify P10-III stage close) 117: **Beautify P10-III (parts j, n, o, q of P10) stage close (9 October 2026):** the stage-end overhead of sub-stage P10-III "Polish backlog" (step `28-P10-III` in the plan's overhead table). Opus 5.5 ran it on this Mac on the final P10-III code: `main` 8418d4e (code ae7349b, P10q; 8418d4e only adds an owner-decisions note), working tree clean (`close/head.txt`, `close/status.txt`); the live code equals `build/backup-feel/28-P10-q` (`close/live-vs-q.txt`). No game code changed. Evidence: `artifacts/beautify/28-P10-III/close/` (below, `close/`). 118: - **Before side:** `close/pre-tree` from `tools/feel_snapshot.sh tree 28-P10-S2` (= `main` 434302e, the P10-II close code). The snapshot differs from live only in the stage's 20 code files (`close/diff-S2-vs-live.txt`), and all 20 are byte-equal to 434302e in the snapshot (`close/git-changed-since-pre.txt`, `close/s2-vs-434302e.txt`). 119: - **Power:** every run on AC, 100% charged, Low Power Mode 0 (`close/suite-power.txt`, `cap-power.txt`, `bench/power.txt`). The 1-min load was 4.8-9.4 during the suite and captures and 6.0-7.6 during the bench (other apps), higher than at the P10-II close (1.5-5.3). 120: - **Full suite** (`close/suite.sh` a, b, v; 19:39-19:48): **47 suites, 26,356 checks, 0 failures** (`close/suite/suite_all.out`; equal to P10q's own full suite). `--verify` 2880x1800 exit 0 on try 1 (27 s), `PIXELPUNK_INTEGRATION []`; its only ERROR line is the inherited missing `terrain_flowers.svg`. `test_creature_reactions` 6,871/0 in the run and the rerun; the known flaky check did not fire. 1440x900 not run (inherited stall). 121: - **Journey** (`artifacts/feel-journey/28-P10-III/`, six views through `tools/art_capture.sh --set journey`, `close/journey.sh`, all on try 1). Against 28-P10-II (`28-P10-III/counts-vs-28-P10-II.txt`): 122: - harbour-day draw calls and rendered objects +24 (912 -> 936), harbour-night +24 (911 -> 935), harbour-lane +17 (753 -> 770). 123: - On all three: parts 136 -> 138, foliage plants 221 -> 235, foliage total 661 -> 675, light_sync_calls 1 -> 2, pool_resets 0 -> 1, pool_kept 0 -> 16. 124: - quarter-day, quarter-night and styles-day are identical. 125: - `performance.json` gains `pixel_factor` 1 and `art_pixels` (P10n). 126: - That is exactly P10q's declared delta; j, n and o declared none. 127: - Frame >8 luma against 28-P10-II (`frame-vs-28-P10-II.txt`): harbour-day 13.3%, harbour-lane 6.4%, harbour-night 12.6%, quarter-day 7.2%, quarter-night 12.6%, styles-day 8.3%. 128: - Re-run of quarter-day, styles-day and quarter-night in the same tree (`close/rerun/`): counts identical, frame >8 0.5%, 0.8% and 2.2%. The same code captured by P10q at 18:51 (`28-P10/q/cap/live-1`) differs from this journey by 6.7% (quarter-day) and 11.9% (quarter-night). 129: - So most of the journey change is rain, rings, clouds and water at another clock phase between sittings (maps `close/look/journey-diff-*.png`, `qlive1-vs-journey-*.png`; looked at: quarter-day, quarter-night). The systematic part (stage capture below) is the replant, the swags and the overcast. 130: - `notes.json` has a `28-P10-III` entry (`close/journey-note.js`, backup `notes.json.bak-before-28-P10-III`). `index.html` and `journey-standalone.html` (111.85 MB) were rebuilt. 131: - **Art capture and protected looks** (`close/caps.sh`, `capall.sh`, `runall.sh`; 19:58-20:57; pre-1, after, pre-2, after-2; 6 pre-1 views (glade-night, harbour-lane, glade-golden, glade-rain twice, towers-night, towers-golden) were retried by the kit after a run left no capture, every other view on try 1). The groups are the P10-II close's 106 views (g1-g4), plus g5 (harbour-day-rain9, quarter-day-rain9, harbour-golden-rain9, qn-front, qn-west, harbour-day-dry) and g6 (quarter-night and harbour-day at 2880x1800, into `pre-1-2880` etc.): 112 + 2 views. 132: - Regions `close/regions.json` = the P10-II close's (`tools/art_regions.json` is unchanged by the stage). 133: - Metrics `close/metrics/` (`metrics.sh`, `frame-table.js`, `prot.js` → `protected.md`, `elapsed.js`, `consistent.js` → `consistent.txt`, `look/consistent/*.png`). Against the previous stage close `28-P10-II/close/after` (the pre-stage code, another sitting): `ab-p10iiclose-*`, `frame-p10iiclose-*`. 134: - Sheets `close/protected/protected-1.png` to `-4.png`, `protected-5-vs-P10-II-close.png`, `noise-pre.png`, `noise-after.png` (looked at: 1). 135: - **Counts** (`metrics/counts.txt`): every view of the Harbour scene gets P10q's delta (draw calls +14 to +34 at the wide cameras; −3 to −27 at the close Harbour cameras r2, r7 and pipes, where trees now hide other parts). No other view changes an integer count, apart from the orbit views and quarter-night-px1, which vary within each tree as before. 136: - **Frames:** 47 of 112 views are above both noise legs (`frame-table.txt`). Systematic change (all four cross pairs, neither noise pair; `consistent.txt`): 137: - grove-day, grove-1930, shops and styles-mod 0.000%. 138: - Harbour views 2.4-5.2% (the replant). 139: - Rain-0.9 views 6.9-9.1% (overcast plus the replant). 140: - quarter-night 3.1%: the removed quay strings, their reflections, and cloud dither edges, which shift between trees but not within one (sky mean equal, 57.9). 141: - quarter-2045 0.8% and banks-golden 1.0% (the overcast's small twilight and golden share). 142: - styles-day 0.18%. 143: - **Canal mirror columns:** quarter-night p50 43.1 -> 43.2, changed >8 16.3% and 12.7% against noise 3.4% and 0.6%. Four-way crop `close/look/mirror-qn-4.png` (looked at): the columns keep their shape; the after pair has fewer scattered reflected quay bulbs (the swags). quarter-2100: 65.0% and 20.8% against noise 64.8% and 42.9% (within noise). 144: - **HDR neon cores:** fish sign 67.9 -> 68.0, bar 70.4 -> 71.0 (0.5% changed, = its noise), lane sign 140.4 = 140.4, lane fish 135.9 -> 135.8, tower windows 95.8 = 95.8; all within their noise pairs. 145: - **05:45 and 20:45 skies:** p50 within 0.3 except the quarter 20:45 horizon at 151.2 -> 150.3 (151.0 -> 150.3 in the second pair; 0.5% and 0.2% changed against noise 0.7% and 0.0%): the overcast at daylight 0.15, as P10k described. Harbour 05:45 and 20:45 are equal. 146: - **Glade/Canal separation:** Glade night zenith `#001d45` 0.00% changed and Harbour night zenith 53.9 = 53.9; Glade night water 33.8 = 33.8. Glade and Harbour banks and crowns change where the replant puts trees (bank 15-20%, crowns 52-60%, in both pairs, 0-1% noise). 147: - **Approved tree style:** grove-day crowns, conifers, feet, lawn and frame 0.00% changed; grove-1930 0.00%; quarter-night undersides 27.7 = 27.7. 148: - **Whole-stage A/B bench** (`close/bench.sh` = the P10-II close recipe; `bench-t1.out`, `-t13.out`, `-t19.5.out`, `bench/summary.txt`; 20:57-21:10, 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. Medians pre → live (paired median; rounds live slower): 150: | Scene | 01:00 p50 | 01:00 p95 | 13:00 p50 | 13:00 p95 | 19:30 p50 | 19:30 p95 | 151: |---|---|---|---|---|---|---| 152: | Dense | 16.922 → 16.992 (+0.041, 2/3) | 18.613 → 18.677 (+0.076, 2/3) | 17.045 → 17.090 (+0.045, 2/3) | 18.649 → 18.967 (+0.249, 3/3) | 18.470 → 18.415 (−0.104, 1/3) | 20.329 → 20.218 (−0.124, 0/3) | 153: | Harbour | 8.227 → 8.270 (+0.024, 3/3) | 8.870 → 8.898 (+0.061, 3/3) | 8.266 → 8.406 (+0.080, 3/3) | 8.855 → 9.072 (+0.165, 3/3) | 7.967 → 8.151 (+0.157, 3/3) | 8.711 → 8.801 (+0.008, 2/3) | 155: - Draw calls: dense unchanged (2588 / 2589 / 3440); Harbour 921 → 947, 922 → 948, 833 → 857. 156: - Harbour is slower by 0.02-0.16 ms p50 in every round, as P10q measured (+0.12) and P10g's mock predicted (+0.11). 157: - Absolute times are about 2.5 ms (dense) and 1.4 ms (Harbour) above the P10-II close's, under the higher machine load. Both sides ran under that load, so the A/B stands, but the absolute numbers are not comparable with earlier closes. 158: - The plan's stop line (dense p95 over about +0.5 ms) is not hit. 159: - **Owner sheet** `artifacts/beautify/28-P10-III/close/sheet/` (`close/sheet.sh`, regions `close/sheet-regions.json` from `close/sheet-regions.js`; before left, after right, colour plus greyscale on the crops; `sheet.md`: one paragraph per part with its measured numbers, then the bench and the suite). 160: - From the close captures: 161: - `p10iii-01-q-harbour-whole.png`, `-02-q-harbour-closeups.png`, `-03-q-quay-swags.png`, `-04-q-overcast-whole.png`, `-05-q-overcast-sky.png`, `-08-existing-scenes.png` (pre-1 | after). 162: - `-06-n-fullscreen-1to1.png`, `-07-n-fullscreen-whole.png` (pre-1-2880 | after-2880). 163: - Copied unchanged from the parts: 164: - `-09-j-roofs-audit-vs-now.png` (`28-P10/j/sheet`). 165: - `-10-o-font-options.png`, `-11-o-integer-scale-options.png` (`28-P10/o/sheet`). 166: - `-12-q-placeable-beacon.png` (P10m's `28-P10/m/sheet` page: the beacon a person can now place). 167: - Looked at: 01, 02, 03 (its west crop was moved after the first look), 05, 06, 08. 168: - Handoff check (after the close): all 12 pages looked at; 09-12 are byte-identical to their sources (`cmp`). Captions in `sheet.md` amended for 06 (its title says 1:1, but the 640x520 crops are drawn at 1/1.2), 08 (which views carry the replant; styles-day is within its noise), 09, 10 and 11 (P10o's pages have no labels, so `sheet.md` now names their panels) and 12 (its title is cut off at the right in P10m's original). No PNG was rebuilt. Backup of `sheet.md` before the change: `28-P10-III/handoff-fix/sheet.md.before-handoff-fix`. 169: - Also looked at: `close/look/hd-east-pre1-after.png`, `hd-west-pre1-after.png`, `qn-front-pre1-after.png`, `qn-west-crop.png`, `diff-qn-west.png`, `hn-pool-pre1-after-x2.png`, `diff-styles-day.png`. The Harbour night lamp pools by the tower are unchanged (rect means equal, `close/rectmean.js`). The darker look in the scaled sheet is the new trees in the right half. 170: - **Not done / open:** 171: - The owner's answers on P10n (keep the default on; exempt Glade/Fine) and P10o (font, integer scale, download permission with a named file); the P10j cone shingles (`hsp` 6-7 against P0d's 0) and the lane-zoom slate pepper. 172: - P10q's open items stay open: the slower slider step in the Harbour drag, the extra light sync on the replanted Harbour, and the photo-mode and tilt-shift views of the new Harbour. 173: - The quarter-night cloud-edge shift between trees (code-correlated, not traced; sky mean unchanged). 174: - The P10n fullscreen was captured as a 2880x1800 window, not a real macOS fullscreen space. 175: - No export (the plan's "exports after the last sub-batch" was not asked for). The handoff check, the progress board, `tools/repo_evidence.sh` and the commit are the orchestrator's. 176: - 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/` (`j/pre-tree`, `n/pre-tree`, `n/off-tree`, `o/*-tree`, `q/pre-tree`, `q/on-tree`, `q/off-tree`, `q/east-off-tree`) were left in place (outside this close's folder). 177: - Platforms: Mac run; Windows not run. SOURCE: docs/visual-handoff.md (frozen 12 October 2026), lines 55-106 (Beautify P10r) 55: **Beautify P10r Pixelify Sans body font and rounded UI scale (9 October 2026):** Opus 5.5 on this Mac, 22:29 on 9 October to 00:15 on 10 October (plan section "P10r"; owner, 9 October 2026, on the P10o options: font Pixelify Sans (OFL), scale rule A; step `28-P10`, part r). The body text is now Pixelify Sans at 11 px, drawn with every pixel on or off; Silkscreen stays for headings; the UI scale is rounded to a whole number. Each change sits behind its own constant with a revert switch, and the tests follow both constants. **Concurrency, read first:** the Decor R1a implementer (docs/decor-orchestration-handoff.md) worked in the same live tree from 22:30, from snapshot `29-R1`, which is identical to my `28-P10-r-pre`. It edited `main.gd` and `editor_ui.gd` (as I did) and several other scripts and tests. Its unfinished edits broke compilation at 22:44 (`BOAT_HEAVE_MIN`; `r/suites/test_ui.log.out`). For that reason every P10r measurement below comes from `r/post-tree`, which is the pre-part tree plus P10r only. The live `main.gd` and `editor_ui.gd` hold both parts' hunks; P10r's hunks are exactly those in `r/part.diff`. Evidence is under `artifacts/beautify/28-P10/r/` (`r/` below). 56: - **Download** (`r/download.txt`). Exactly two files were downloaded. `https://raw.githubusercontent.com/google/fonts/main/ofl/pixelifysans/PixelifySans%5Bwght%5D.ttf`: HTTP 200, **79,160 bytes (matches)**, sha256 `9ba86cd010a4de309d263ceff8e8044092c9db7efda869620cb9ff1c4389e8a5`. `.../ofl/pixelifysans/OFL.txt`: 4,488 bytes, sha256 `b66ba46f…73a7`. Its licence body is identical to Silkscreen's `OFL.txt`. 57: - **Files:** 58: - `scripts/hologram_theme.gd`: `PIXEL_BODY`, `BODY_FONT`, `BODY_SIZE` 11, `SYSTEM_BODY_SIZE` 14, `body_font()` and `body_size()`. The theme takes the body font. 59: - `scripts/editor_ui.gd`: `label()` passes sizes through `body_size()`. The Silkscreen tool title is pinned at its old 12 px. New `_nearest_window`: every Window the editor opens gets a nearest canvas filter. 60: - `scripts/main.gd`: `UI_SCALE_ROUND`; `ui_scale_for` rounds. 61: - `tests/test_ui.gd`, `THIRD_PARTY_NOTICES.txt` (Pixelify credit; "Fonts use installed system fonts" corrected). 62: - New: `assets/ui/hologram/fonts/PixelifySans[wght].ttf`, its `.import` (antialiasing 0, hinting 0, subpixel 0, `keep_rounding_remainders=false`, oversampling 1.0) and `PixelifySans-OFL.txt`. 63: - **Snapshots:** 64: - Pre: `build/backup-feel/28-P10-r-pre/`. It equals `28-P10-S3` except R5a's uncommitted `main.gd` and four tests (`r/diff-S3-vs-r-pre.txt`; the hunks match R5a's `part.diff`, `r/diff-R5a-pre-vs-r-pre.diff`). It equals `29-R1`. 65: - Post: `build/backup-feel/28-P10-r-post/`, copied from `r/post-tree` with the live docs. It is not taken from the live tree, which also holds R1a. 66: - Scratch trees: `r/pre-tree`, `r/post-tree` and `r/off-tree` (both switches false). Post and off link `artifacts` and carry README.md and `animalartreference.png`, so the fixtures resolve. 67: - **Body size 11 (my choice; the P10o sheet named no Pixelify size).** The face is drawn on a pixel grid of about 91 of 1000 units, so 11 px is its design size (`r/fonts/grid.js`, `glyphs.js`). Measured in the glyph cache: H is 5x7 px, x is 5x5 and l is 1x7. The next crisp size, 22, doubles everything and was not tried. 68: - **Two defects found by looking, then fixed:** 69: - Remainder carry: the 6.45-px advances spaced letters 1 or 2 px apart and closed some word gaps (`r/look/probe-1600-window.png`). Each advance is now rounded on its own, so letters sit 1 px apart and words 3 px (`r/look/probe2-1600-window.png`). 70: - At scale 2 the window's oversampling rasterised the font at 22 px with 3-px stems (runs 3/7/9). Oversampling is now 1 with a nearest filter, so scale 2 is the 1x text doubled (`r/look/probe-editor-os0-os1-zoom.png`). Popups had blurred without the filter (`r/look/probe-popups-v2-zoom.png`, middle). 71: - **Tests** (`r/testfirst/`, `r/suites/`): 72: - Test-first, 481 checks, 2 failures: the constants were missing. The spacing test-first: 497 checks, 2 failures. The new clipping helper, run on P10o's Silkscreen-16 mock, fails 8 clipping checks there: the hint at both sizes and six panel cases (`r/testfirst/r-check-on-silk16-mock-test_ui.log.out`; the run's other 9 failures come from that tree lacking later parts). So it detects the P10o findings, and it passes on the unchanged code. 73: - After: post-tree test_ui **499/0**, test_camera 544/0, test_tilt 258/0. Off-tree (both switches false): test_ui 493/0, test_camera 544/0, test_tilt 258/0. Live (with R1a in it) test_ui 499/0. 74: - test_ui 465 -> 499 (+34). New checks: the file's size and sha, the licence, the notices, the font settings and import, whole-pixel advances, every glyph pixel on or off and the H/x/l shapes, the scale table at 11 windows, every body label at 11, no clipped text in six panels or the hint at 1100x700 and 1600x1000, and the window filter. 75: - **Full suite** (`r/suite/suite_all.out`; `VERIFY_RES=2880x1800`, on `r/post-tree`; AC, 100%, Low Power Mode 0): **47 suites, 26,430 checks, 1 failure**. The failure is the known flake test_creature_reactions "The re-plan is applied after the edits"; its rerun gave 6,871/0. The count is 26,396 + 34. `--verify` 2880x1800 exit 0 on try 1 (25 s), `PIXELPUNK_INTEGRATION []`, and its only ERROR is the inherited `terrain_flowers.svg`. A first run without the reference sheet in the tree failed 2 checks (`r/suite-try1-no-animalref/`). The 1440x900 run was not made (inherited stall). Part b took 24 min under load. 76: - **Changed expectations (old -> new):** 77: - test_ui's P10n check "ui_scale_for 1.6 at 2880x1800" -> 2.0 with `UI_SCALE_ROUND`. 78: - UI scale: 1.067 -> 1 at 1920x1200, 1.28 -> 1 at 2560x1440 (-22%), 1.42 -> 1 at 2560x1600, 1.53 -> 2 at 2880x1724 (+31%), 1.92 -> 2 at 3840x2160, 2.56 -> 3 at 5120x2880. Up to 1920x1080 it stays 1. 79: - The P10n factor is unchanged everywhere. 80: - Body text: Arial (Segoe UI on Windows) at 10-14 px -> Pixelify 11 px. 81: - **Expected journey count deltas (1600x1000):** draw calls -1, rendered objects +44 per view (harbour-day, every size: 935/1011 -> 934/1055). Not traced; P10o's mock showed the same -1/+45. courtyard.png is unchanged; editor.png changes. 82: - **Captures** (`r/capset.sh`, frozen clock, harbour-day `+window` and `+scenery` at 1440x900, 1600x1000, 1920x1080, 2560x1440 and 2880x1800; P10n at its default, on, so 2880 is factor 2 in both trees). Order pre-3, post-3, off-3, pre-4, post-4, 23:03-23:10; AC, 100%, LPM 0, load 3.7-4.7 (`cap/*/capture-meta.txt`). The P10o recipe is copied. `r/metrics/cmp.txt`: 83: - The noise pairs differ by 0.82-0.97% of pixels. 84: - Off against pre: 0.82-0.98%, within noise. The revert switches restore the old look. 85: - courtyard.png pre against post: 0.81-0.96%, within noise. 86: - editor.png changes on 2.1-5.6% of pixels at scale 1, 8.9-13.1% at 2560 and 14.2-19.0% at 2880. 87: - Sheets, all looked at (`r/sheet/README.md`): the panels at four sizes, x4 text, half-size frames at 2560 and 2880, the hint and the popups. 88: - **Acceptance, measured:** 89: - **Body text crisp at 1440x900, 1920x1080, 2560x1440 and 2880x1800: met** (`r/metrics/crisp-table.txt`, P10o's metric). The body's partly covered share is 0.61/0.61/0.57/0.52 -> **0/0/0/0** in both post runs, and runs on the font grid are 100%: 2/6/8 px at 2880. Headings are 0 and on grid everywhere; at 2560 they were 0.39 and off grid. The menu text in popups is 0.56 -> 0 at 1600 and 0.63 -> 0 at 2880 (`r/metrics/popup-crisp.txt`). 90: - **Every window-pixel pick exact: met.** The P10n probe on the post tree (`r/pick/`): 1440x900, 1920x1080, 2560x1440 and 2880x1800 at pixel size 2, and 2880 at pixel size 3. 0 of 25,479,360 window pixels differ. Ground picks: 0 land outside. Grips 10/10; object picks 138/138. 91: - **test_ui at 1100x700 and the P10o clipping findings: solved.** The Scenery labels and the top-right hint are whole (tests, `r/sheet/r-hint-1600x1000-x2.png`). 92: - **Bench: within noise** (`r/bench/summary.txt`; P10n recipe, pre against post, interleaved, round 0 discarded; AC, LPM 0, load 3.9-5.3, `r/bench/power.txt`). 93: - 2880x1800: dense p50 15.270 -> 15.439 (+1.1%), Harbour 7.788 -> 7.906 (+1.5%). 94: - 1600x1000, 6 rounds: dense 14.560 -> 14.637 (+0.5%), Harbour 7.892 -> 7.806 (-1.1%). Rounds 1-3 alone gave +3.0% and rounds 4-6 -0.7%. 95: - Draw calls -1. 96: - **Review:** filled by the orchestrator. 97: - **Not done, and why:** 98: - Body text is smaller than Arial 14 (x-height 5 px). At 2560x1440 the whole UI is 22% smaller (rule A). Both are owner look calls. 99: - Tooltips were not captured; the filter covers them by code only. 100: - The help dialog's long lines run off a 1600 window before and after. 101: - The Silkscreen tool title at 12 px is off its grid (unchanged). 102: - Pixelify lacks 10 of the 11 terrain-card signs (▰ ✿ ▬ ∿ ⋰ ❀ ↟ ↡ ∽ ═; ≈ is present) and the recovery toast's →. These draw in the fallback system face (`r/fonts/coverage.txt`). 103: - 3840x2160 and above: computed only. 104: - Windows font rendering unverified. 105: - No commit, push or export. The scratch trees are left in place. 106: - **Platforms:** Mac run; Windows not run. SOURCE: artifacts/beautify/owner-decisions.md (frozen 12 October 2026, equal to HEAD), lines 388-409 388: ## Journal published (9 October 2026, 20:01) 389: - Deployed to https://pixelpunk-journal.pages.dev (deployment 2cb957d8-aee7-4c42-bc80-d1bea8698f3a, 296 files, 67 uploaded). Live check 34 of 34, 0 errors. Chapters 26 (P8, corrected), 27 (P9), 28 (P10-I), 29 (P10-II and P10q); P10n and P10o written as open questions. Two fact-check must-fixes (chapter 27 dek, chapter 29 overcast wording) were fixed before deploy. 390: - The P10-III close result is not in the journal yet; the next update starts there (journal/HANDOFF.md). 392: ## P10-III closed by the orchestrator (9 October 2026, ~22:20) 393: - Close on main 8418d4e (code ae7349b), entry and handoff corrections committed with this note, snapshot 28-P10-S3, journey artifacts/feel-journey/28-P10-III. Suite 47 suites, 26,356 checks, 0 failures; --verify 2880x1800 exit 0. Draw calls changed exactly as P10q declared (Harbour day and night +24, lane +17). Bench Harbour p50 +0.02 to +0.16 ms, dense within the stop line. 394: - Owner sheet (12 pages + sheet.md) sent. Waiting on the owner: P10n (keep the default on? exempt Glade's Fine pixel size? factor formula) and P10o (font file and permission to download, scale option A/B/C, body size). P10 is otherwise complete. 395: - Open and not traced: Harbour drag slider step 10.9 -> 17.5 ms, six plants pop in after a fresh sync, lamps 195 and 243 without swag string, cone specks 6-7 (P0d had 0), lane-day roof_lane 2 -> 21, harbour-night cyan dashes (unproven), quarter-night cloud edges differ between code versions. 396: - AGENTS.md has another contributor's uncommitted edit (a pointer to docs/decor-controls-ui-plan.md); left out of every commit. 398: ## Owner ruling on the P10-III sheet (9 October 2026) 399: - "All approved." Orchestrator reading: P10n stays as built (pixel size follows the window, default ON, Glade's Fine pixel size not exempt, factor formula as built with the UI's 0.8 growth); P10q's applied changes stand; P10j needs nothing (measure only). All of P10 is approved. 400: - P10o is NOT answered by "all approved": it needs a font file the owner names and explicit permission to download it, plus the scale option and body size. Nothing downloaded or changed. Asked the owner in chat. 402: ## P10o answered (9 October 2026) 403: - Owner picked: font Pixelify Sans (OFL), UI scale rule A (round to whole numbers). Download permission is therefore given for exactly `PixelifySans[wght].ttf` (79,160 bytes) and its `OFL.txt`, both from github.com/google/fonts, folder ofl/pixelifysans. Body size not asked: the P10o sheet's recommendation applies. 404: - New part P10r (docs/beautify-plan.md; beautify-stage.js stage P10 part r) builds it. 406: ## P10r committed by the orchestrator (10 October 2026, ~01:00) 407: - Commit a184c16: Pixelify Sans body font (11 px, Silkscreen kept for headings) and the rounded UI scale, each behind a constant (HologramTheme.PIXEL_BODY, Main.UI_SCALE_ROUND). Downloaded exactly the two permitted files (font 79,160 bytes, sha256 9ba86cd0...e8a5, and OFL.txt; URLs in artifacts/beautify/28-P10/r/download.txt). Body text partly covered share 0.52-0.61 -> 0 at 1440x900, 1920x1080, 2560x1440, 2880x1800; 0 of 25,479,360 window-pixel picks differ. 408: - Commit scope: another contributor (the Decor session, R1a and R5a) has uncommitted edits in the same working tree (main.gd, editor_ui.gd, tests, README, HANDOFF.md, AGENTS.md ...). Only the P10r hunks were staged and committed, composed against HEAD; the composed tree was tested separately (test_ui 495/0; R5a's own checks are not in it). The Decor edits are left untouched in the working tree. 28-P10-r-post is a copy of the P10r scratch tree and does not contain Decor. 409: - Owner look calls recorded as open: body text is smaller than the old Arial 14, and the UI is 22% smaller at 2560x1440 under rule A. Not done: tooltips not captured, help dialog long lines run off a 1600-wide window (before and after), 10 terrain-card signs fall back to the system font. Review point: the nearest-filter hook runs on every node_added (cost on rebuild/release not measured).