Chapter 33: Crags, lanterns and a castle on the loch Frozen 21 October 2026 at 01:11 BST (game main at 214f068). Line numbers refer to the frozen copies in research/update-2026-10-21-R5d/ (visual-handoff-full-at-freeze.md, decor-owner-decisions.md). Covers stage R8 Castle grounds of docs/castle-elements-plan.md: R8a, R8b, R8c (15 and 16 October), R8d, R8e, R8f (20 October), the stage close with milestone M-R4 and R8x (20 October), with the orchestrator rulings of 15, 16 and 20 October, the owner's push of 21 October about 01:00 and the owner's answers on the R8 sheet items A to M (21 October). The follow-up stage R9 is quoted only as the queue line under the A to M answers; nothing in it is built. Not covered: R5b to R5d (chapter 34), R5e and later. Excerpts are given in the order of the work. Local Trash folders are given as "the macOS Trash". SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 357-430 (Castle R8a Crag and cliff stairs) 357: **Castle R8a Crag and cliff stairs (15 October 2026):** Opus 5.5 ran this on the Mac, 15 October 14:19 to about 16:00, fix pass 1 from 16:05 to 16:50, and fix pass 2 from 17:00 to 17:30 (step `36-R8`, part a; `docs/castle-elements-plan.md` R8a, owner K6 and K8). Snapshots `build/backup-feel/36-R8/` and `36-R8-a-pre/` (= HEAD f90e005). Scratch pre-tree `a/pre-tree`; backups `a/backup/` (.gdignore'd). Evidence `artifacts/decor/36-R8/a/` (`a/` below). Design `a/design.md` (with a section on the orchestrator's look notes of about 15:16). Patch `a/part.diff` (2567 lines); `a/diff-live-vs-pre-at-end.txt`: `scripts/scene_state.gd`, `main.gd`, `world_view.gd`, `architecture.gd`, `cluster_architecture.gd`, `editor_ui.gd`, `adapt.gd`, `composition.gd`, `creatures.gd`, `shaders/masonry_body.gdshaderinc`, `data/catalog.tres`, `tools/test.sh`, `tests/test_decor.gd`, `tests/test_bikes.gd`, new `scripts/crag.gd` and `tests/test_crag.gd`. Save format unchanged (version 2; two new optional keys); every edit one undo step. Review: see the Review line below. Platforms: Mac run; Windows not run. 358: - **The kind.** `crag` (a Walls card in the tray, 70 to 71 cards). Drawn with its own tool as one closed freehand stroke (the wall-path stroke, simplified to at most 16 points); a click places the default 8-point crag. Keys `crag_pts` (the outline, local, 3 to 24 points; a crossing or tiny outline is refused) and `crag_h` (3 to 16 m over the water, default 8, omitted at default), both checked against the shared inspector keys (no clash with `height` or `points`). Edited by a knob on every vertex at the top's edge and a height arrow, and by the Height row (step 0.5); each drag or row change is one undo step. 359: - **The rock** (`scripts/crag.gd`, its own cluster, rebuilt only when its signature changes). Faces lean in by `CRAG_BATTER` 0.15 of the drop from the top (the foot 2.4 m out under a 16 m crag; `crag_base` is that footprint). Bands about 2.1 m tall, each stepped in or out, jittered, with a seeded tone. Moss sits on the ledges, in a rim under the turf and in patches. Rounded boulders lie at the waterline. The top is exactly the drawn outline. It draws in the masonry material through a new natural-rock branch (`vcol.a` 0.05 to 0.15: no joints). No other piece uses that alpha, and the existing views are within noise. 2818 triangles at 16 m with a 12-point outline; no MultiMesh kinds of its own (test ≤12). 360: - **Top as terrace.** `ground_height` and `land_at` answer the crag's top inside the outline; `terrain_land_at` keeps the terrain's own answer for brushes, so the heightfield and the 8 m limit are untouched. A tower on the top has no skirt, band or pillars (`Supports.plan` empty; `tower-top`). A bridge lands on it. A house over the edge or the battered foot is refused with its hint for the hand, and is soft when the crag is moved, shrunk or lowered under it (lesson 7). The foot follows the ground to the bed; a terrain stroke beside or under it refreshes it. Planting keeps off the top. The per-pass index means 500 queries take at most one walk. 361: - **Cliff stairs** (after the look notes). These wrap the faces: from the outline point nearest the foot, round the crag at 0.72 m per metre, 0.4 m out from the battered face, with a landing at each corner and a deck onto the top. The stair goes round the way whose deck is clear of houses. Rock-toned treads with short corbels and no piers. The shore keeps clear of the stair's low pieces. Shown from the waterline (`stair-water`) and from the bank. 362: - **Look notes** (`a/notes1/notes.diff`, 1122 lines; earlier evidence in `a/superseded/`). Accepted 15:36. The dusk warmth is the sunset key light and sky on the masonry material, as on the castle's walls at that hour. It is left as is. 363: - **Tests.** `tests/test_crag.gd` (after test_castle_hardening): tests first 16/16 failing (`a/testfirst/`), 104/0 before fix pass 1, 116/0 after it, and 200/0 after fix pass 2 (below). Changed expectations: tray pins 70 to 71 (test_decor ×4, test_bikes ×1). Affected suites 26, 0 failures (`a/suites/summary.txt`, `a/suites2/`). Full `tools/suite_all.sh` a+b+v: 57 suites, 31539 checks, 0 failures, verify 2880x1800 clean (`a/suite/suite_all.out`). 364: - **Unchanged without a crag.** R7a 37, R7c 66 and bridge 84 pins identical (`a/pins/final/`). The g0 ident 15/15 is identical (`a/ident/final/`). test_starters md5 pins pass. Castle starter, Harbour, quarter and Amsterdam pre against live are within the noise pair, with identical counts (`a/metrics/`). 365: - **Captures** (`a/caps.sh`, `a/cap/`, sheets `a/sheet/r8a-crags.png`, `r8a-close.png`, `r8a-castle.png`, `r8a-existing.png`; all on AC). The cue castle sits on a 12 m crag (`a/scenes/build.gd`; viaduct and outer towers removed). 366: - **Cost** (live only: crag scenes do not load on the pre-tree; `a/cost/release.out`, medians of 7, 3 runs). On the realistic scene, Height row change and undo take 42 / 32 ms (clusters 26 ms); a tower on the top 25 / 18 ms; a vertex drag step 0.8 ms (p95 23) and its release 30 ms. On the stress scene (6 crags, 40 buildings): 40 / 20, 44 / 28, 0.8 (p95 19) and 35 ms. Frame time with six 16 m crags is 2.52 ms p50 against 2.50 for the bank alone (`a/bench/bench.txt`, vsync off). Harbour handle p95 is 57.5 ms live against 56.2 ms pre, 3+3 (`a/drag/summary.txt`; P10q baseline 53.7). 367: - **Older builds.** A build before R8a refuses a scene that holds a crag with "Unknown or damaged building part." (the load check of `scene_state.gd`, `a/pre-tree` line 6336). The save format version stays 2, and a scene without a crag loads both ways. 368: - **Not done / open. These limits are carried to the R8 close:** 369: - Creatures do not walk crag tops, and cliff stairs carry no walkers (the same as tower wraps today). 370: - A house that a crag move leaves straddling the edge keeps standing. Its own Height row is refused until it is moved. 371: - A house placed later on the stair's deck turns the stair the other way. With both ways blocked, the deck runs under the house. 372: - A second crag beside a stair is not refused. 373: - Close up, the faceting reads as a regular sawtooth. Show this to the owner at the R8 close. 374: - A viaduct or a covered bridge cannot be set between two crag tops: its crowd check refuses the rock under its deck. So the fix pass 2 ride is untested for those two styles. 375: - Paths painted under a crag are hidden. The cliff stair ignores the stair body and rail options, and it does not check things at its foot. 376: - **Fix pass 1** (orchestrator rulings of 15 October about 16:03, `artifacts/decor/owner-decisions.md`; review `a/review/`). Backups `a/fix1/backup/` (.gdignore'd). Evidence `a/fix1/`. Patch `a/fix1/fix1.diff` (530 lines): `scene_state.gd`, `architecture.gd`, `composition.gd`, `creatures.gd`, `main.gd`, `adapt.gd`, `tests/test_crag.gd`. The false "hop limit" line in `a/design.md` is corrected. 377: - **1. Terrain chunks.** The terrain chunk is drawn from the heightfield alone. `Architecture.terrain_chunk`, `puddle_chunk`, `water_signature` and `hidden_tufts` raise `DioramaState.crag_off` while they run, so `land_at` and `ground_height` give the terrain's own answer. Their tufts use `Composition.footprint_index(state, false)`, which leaves crags out and keeps its own cache. The chunk fingerprint never read crags, and now nothing it draws depends on them. The crag draws its own sides and top. 378: - **2. Naps.** `Creatures._nap_keep_some` skips every site within a crag's foot (`_on_crag`: the top, its faces, and all that stands there) while tops are unwalked. 379: - **3. Crag index.** `crowd_dirty(replaced)` keeps the crag index unless the replaced object is a crag or a stair, or none is named. The memo of stair plans and landings is dropped either way. 380: - **4. Landings.** `crag_error` refuses a crag whose rock holds a bridge's ground end. The rock is its foot, `crag_base`, and an end counts when it lies more than `CRAG_BURY` 0.25 m under the top. The line names the landing (`CRAG_LANDING_ERRORS`), for example "Keep the crag clear of the bridge landing.", with walkway and jetty forms. The other way, a bridge whose ground end is set in a crag's rock is refused with `CRAG_FOOT_ERROR` ("Set the landing on the crag's top or clear of its foot."). An end set on the top stands on the ground there, so it seats. 381: - Both are soft in `soft_error`. `Main.adapt_all` refuses them only for the part the open edit placed or changed (`_part_changed`, which also compares the ends). `validate_candidate` returns them for the moving part, the crag knob refuses live with the hint, and Adapt's literal path does not let them through. So a stroke, a move, a knob or a Height row that would bury an end is refused, and an edit elsewhere still goes on. 382: - Every scene in `a/scenes/` holds no buried end (`a/fix1/probe/landings.gd`). The scenes rebuilt by `a/fix1/scenes/build.gd` are byte-identical to `a/scenes/`. 383: - **Tests first** (`a/fix1/testfirst/test_crag.log.out`, on the unfixed scripts swapped back in): 8 of 116 checks fail. 384: - The chunk test differs at `drawn` and `undone` in 3 chunks. 385: - The cat's nap list includes `roof:2@15.4`. 386: - The walk test counts 6 walks. 387: - Four landing checks fail: drawn, hint, moved and grown over a landing, and an end set in the foot. 388: - Now 116/0 (`a/fix1/suites/test_crag.log`). The new checks are `r8a_fix1_checks`: 389: - draw a crag over a bank, move it away, undo and undo again, with every chunk's terrain and puddle groups hashed equal to the never-touched chunk; 390: - no nap within a crag's foot, while a ground house keeps its nap; 391: - one walk for `adapt_all` after a Height row with 6 houses on the top, all 6 settled; 392: - drawn, moved and grown over a landing are refused with the hint; 393: - an end in the foot is refused; 394: - a bridge between two crag tops seats; 395: - an edit elsewhere commits. 396: - **Suites.** Affected suites (`a/fix1/suites/summary.txt`): 28 suites, 0 failures, with test_creature_reactions passing on try 1. Full `tools/suite_all.sh` a, b, then v with `VERIFY_RES=2880x1800` (`a/fix1/suite/suite_all.out`, 16:30-16:44, on AC at 100%): 57 suites, 31563 checks, 0 failures. Verify is clean, and the creature_reactions rerun passes. 397: - **Pins and ident** against `a/pre-tree` (`a/fix1/pins/`, `a/fix1/ident/`): R7a 37, R7c 66 and bridge 84 pins all show diff 0, and ident 15/15 shows diff 0. 398: - **Costs** (`a/fix1/cost/release.out`, medians of 7, 3 runs, live). 399: - Crag index walks for one Height row: 1 (was 3) on the realistic scene and 1 (was 8) on the stress scene. 400: - After fix pass 1, raising the realistic scene's main crag 1 m was refused, because its bridge lands on the top (`a/fix1/probe/raise.log.out`). Its run that raised 1 m is kept in `a/fix1/cost/raise-refused/` (about 17 / 10 ms, a refused edit). So `a/fix1/probe/release.gd` lowered that crag 1 m instead: 41.0 / 31.9 ms fixed, against 41.4 / 32.1 ms on the unfixed scripts (`a/fix1/cost/unfixed/`). Fix pass 2 lifts that refusal (below). 401: - Stress Height row: 39.3 / 20.0 ms (was 40.5 / 20.1). 402: - Vertex release: 29.8 ms realistic and 34.7 ms stress (were 30 and 35). Drag steps are unchanged (0.8 ms, p95 23 and 19). 403: - The clusters phase (the crag's own rebuild, about 26 ms) is most of the cost. 404: - **Captures** (`a/fix1/caps.sh`, on AC). The 16 crag views were taken twice (after, after-2), plus the new bank-crag view. Bank-crag (where the 8 m crag meets the bank, low from the water) was also taken on the unfixed scripts (`a/fix1/cap/unfixed`). 405: - Metrics are in `a/fix1/metrics.out`. Against the accepted set `a/cap/after`, nine views move beyond their noise pair in pixels changed by more than 8 (corrected after the recheck, `a/review2/diff/`): castle-crag-high 0.52% (noise 0.03%), castle-crag-low 0.20% (0.02%), crag8-close 0.18% (0.10%), tower-top 0.14% (0.01%), crag16-close 0.11% (0.002%), stair-water 0.11% (0.03%), crags-high 0.11% (0.04%), shore-moss 0.013% (0.004%) and stair-bank 0.013% (0.007%). The dusk and low views over moving water stay within it. 406: - The castle-crag-high change is stray path-edging lines on the courtyard top, now gone. Most likely the terrain chunk drew them from the crag-aware land answer before fix pass 1 (`a/review2/diff/castle-crag-high-crop.png`). They are not the cat or the terrain under the rock. The sheet `a/fix1/sheet/r8a-fix1-changed.png` is too small to show it. 407: - Draw calls rise by 2 to 6 in 13 of the 16 views, which fits the terrain now drawn under and beside the rock. Castle-crag and castle-crag-low are identical, and castle-crag-high falls by 2. 408: - Bank-crag, unfixed against fixed: 0.62% of pixels change by more than 8, against a noise pair of 0.58%. The sheet `a/fix1/sheet/r8a-fix1-bank-crag.png` shows no visible change. 409: - The existing views hold no crag, so they were not retaken (the pins and ident are unchanged). 410: - **Fix pass 2** (orchestrator ruling of 15 October about 16:57, `artifacts/decor/owner-decisions.md`; recheck `a/review2/`): what stands on a crag's top rides with it. Backups `a/fix2/backup/` (.gdignore'd). Evidence `a/fix2/`. Patch `a/fix2/fix2.diff` (319 lines): `scene_state.gd`, `main.gd`, `tests/test_crag.gd`. 411: - **The carry.** `DioramaState.crag_carry(before, after)` moves every connection end with no parent (bridges, walkways, jetties, a cliff stair's head) to the new top. It applies to an end that sat on the old top (inside the old outline, within `CRAG_SEAT` 0.2 m of its top) and is still inside the new outline. Each moved connection is replaced by a copy, so the change stays inside the edit's one undo step. 412: - `Main.adapt_all` runs it first (`_carry_crag_tops`) for every crag the open edit changed, compared with `history.pending`. `crag_buries` then never sees a carried end. 413: - Each style's own rules then judge the result. A stone, timber, japanese, modern or rope bridge whose ends differ by 0.5 or 1 m takes the walkway form and stands. A jetty stands. A lift bridge is refused with "A lift bridge needs both landings at one height." and its own hint. 414: - Lowering carries the ends down, so nothing is left levitating. A cliff stair's head always rides, so its plan is never left empty. 415: - An outline edit that leaves an end outside the new top does not carry it. It keeps the soft lost-landing state it had, or the fix pass 1 refusal when the end is in the rock. 416: - **Live.** The Height row's scrub (`preview_parameter`) and the crag's height arrow (`drag_crag_handle`) carry the ends as they go (`_crag_live_carry`; the copies join the drag's rollback point). Each tick checks the crag and every carried connection. A release that would be refused shows its hint during the scrub, and the scrub snaps back. The arrow shows the hint and goes on. 417: - **Tests first** (`a/fix2/testfirst/test_crag.log.out`, before the change): 29 of 189 checks fail. Raise 9 and 8.5 m were refused with "The bridge lands here", lower 7 left the end levitating, for each of the 7 styles; the walkway, both previews and the cliff stair at 13, 11, 16 and 8 m also failed (plan empty, head at 12.15). 418: - **Now 200/0** (`a/fix2/suites/test_crag.log`; `r8a_fix2_checks`). The checks cover: 419: - for each style, raise 1 and 0.5 m and lower 1 m with one end on the top, each accepted change undone in one step, and an edit elsewhere afterwards; 420: - the walkway, raised and lowered; 421: - the preview for stone (it rides) and for the lift bridge (it hints); 422: - a cliff stair at 13, 11, 16 and 8 m from 12, with undo; 423: - an outline edit that narrows past the end, and one that widens round it. 424: - The viaduct and the covered bridge are not placed between two crags: their crowd check refuses the rock. The test records this (`R8A_RIDE`). 425: - **The recheck's probe rerun** on this code (`a/fix2/probe/review2/probe.log.out`): every style raises and lowers with "" except the lift bridge, which refuses with its own line. The walkway raises and lowers. On the realistic scene the Height row now commits 13 m, and lowering to 11 m leaves the bridge standing (not levitating). 426: - **Suites.** Affected suites (`a/fix2/suites/summary.txt`): 28 suites, 0 failures. Full `tools/suite_all.sh` a, b, then v with `VERIFY_RES=2880x1800` (`a/fix2/suite/suite_all.out`, 17:13-17:27, on AC at 100%): 57 suites, 31647 checks, 0 failures. Verify is clean, and the creature_reactions rerun passes. Pins against `a/pre-tree` (`a/fix2/pins/`): R7a 37, R7c 66 and bridge 84 all show diff 0. Ident 15 shows diff 0 (`a/fix2/ident/`). 427: - **Cost** (`a/fix2/cost/release.out`, `a/probe/release.gd` as first measured, 3 runs, medians of 7). The realistic Height row is a real raise of 1 m again: 42.7 / 38.0 ms, with 1 index walk. The crag rebuild is 27.7 ms of clusters, and the bridge and stair rebuild about 1.8 ms of connections. Before fix pass 1 it was 41.8 / 31.8 ms and left the bridge buried. Stress: 40.7 / 21.1 ms. Vertex release: 29.7 and 34.8 ms. 428: - **Captures** (`a/fix2/caps.sh`; `a/fix2/probe/raise_scene.gd` sets the realistic scene's main crag to 14 m and to 10 m through `change_parameter` → `a/fix2/scenes/realistic-14.json`, `-10.json`). In both the bridge and the stair stand: `raise_scene.log.out` shows the bridge ends at 14.15 / 12.15 and 10.15 / 12.15 in walkway form, and the stair head on the top with its plan. Sheet `a/fix2/sheet/r8a-fix2-raised.png` (default and low cameras; noise pair `a/fix2/metrics/noise.txt`, 0.03-0.28% over 8). At 14 and 10 m the stone bridge takes the game's walkway form, a timber deck on long legs down to the water, which is the existing rule for a bridge whose landings differ. Show this to the owner at the R8 close. 429: - **Review:** two rounds by fresh Opus 5.5 reviewers, each followed by a fix pass, and the orchestrator's check of the last; the orchestrator's look notes of about 15:16 (a battered mass with stepped, toned bands, visible moss, grey rock at dusk, a boulder shore, stairs wrapping the faces, the cue castle on a 10 to 14 m crag) were folded in before the bench and accepted on the remade sheets at 15:36. (1) Review (`a/review/`: `probe.log.out`, `drag/summary.txt`, `power.txt`): the look, the top as a terrace with no skirt, the per-pass index, the shader's rock branch (no existing masonry vertex in its alpha range, lowest 0.247), ground queries for planting, waterside pieces, tufts and gargoyle streams, the 16-point cap, self-crossing refusal and the pins held; the Harbour handle re-run, 3+3 interleaved on AC, gave p95 54.83 live against 55.72 pre (noise). Must-fix: terrain chunks drew crag-aware land without the crag in their fingerprint (stale chunks, edit-order holes); the cat's first nap site was an unreachable roof on a crag top. The orchestrator's rulings of about 16:03 (`artifacts/decor/owner-decisions.md`, reversible) added keeping the crag index across `crowd_dirty`, refusing a crag over a ground landing both ways, and the entry lines; closed by fix pass 1 (above). (2) Recheck (`a/review2/`: `probe.gd`, `probe2.gd`, `pins/`, `ident/`, `suites/`, `diff/`): the chunk, nap and index fixes sound (chunks byte-identical drawn, moved and undone; no nap site at crag-top level; one walk a commit); must-fix: raising a crag with a bridge end on its top was refused for every style with a hint naming a buried landing, lowering left the bridge levitating, and a cliff stair went soft after any Height change (there since the first build). The orchestrator's ruling of about 16:57 (what stands on a top rides with it) set the remedy; closed by fix pass 2 (above), checked by the orchestrator against `a/fix2/suite/suite_all.out` (57 / 31,647 / 0), the pins and ident (diff 0), the failing test-first log (`a/fix2/testfirst/`, 29 of 189), the re-run recheck probe (`a/fix2/probe/review2/`: every style raises and lowers with no error and no levitation, the lift bridge refused with its own line and hint, the realistic Height row commits 13 m, the stair keeps its plan) and the sheet `a/fix2/sheet/r8a-fix2-raised.png`. Accepted as own-edit content cost (the class of the R7 rulings): a crag Height row 42.7 / 38.0 ms (change / undo, realistic; 27.7 ms of it the crag's own rebuild), a vertex release 29.7 to 34.8 ms; the Harbour carries no crag. Notes, not defects (carried to the R8 close, reversible): creatures do not walk crag tops and cliff stairs carry no walkers (as tower wraps); a house left straddling the edge keeps standing with its Height row refused until moved; a later house on the stair deck turns the stair, and with both ways blocked the deck runs under it; a second crag beside a stair is not refused; viaduct and covered bridges cannot span two crags (their crowd check refuses the rock); the faceting reads as a regular sawtooth close up and the rock warms at dusk as the castle walls do; a stone bridge between crags at different heights draws in the walkway form, a deck on long legs (show the owner). **Closed.** SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 242-356 (Castle R8b Boathouse and lantern boats) 242: **Castle R8b Boathouse and lantern boats (15 October 2026):** Opus 5.5 ran this on the Mac, 15 October about 17:30 to 19:25, fix pass 1 from 19:35 to 20:00, and fix pass 2 from 20:12 to 20:36 (step `36-R8`, part b; `docs/castle-elements-plan.md` R8b, owner K10). The orchestrator's look notes (about 18:37) were folded in, and their sheets were accepted about 19:06. Snapshot `build/backup-feel/36-R8-b-pre/` (= HEAD 0440347); scratch pre-tree `b/pre-tree`; backups `b/backup/` and `b/notes1/backup/` (.gdignore'd). Evidence `artifacts/decor/36-R8/b/` (`b/` below). Design `b/design.md`, with sections for the cost pass and the look notes. Patch `b/part.diff` (2134 lines) covers `scripts/scene_state.gd`, `main.gd`, `world_view.gd`, `architecture.gd`, `editor_ui.gd`, `adapt.gd`, `composition.gd`, `creatures.gd`, `magnets.gd`, `build_audio.gd`, `data/catalog.tres`, `tools/test.sh`, `tests/test_decor.gd`, `tests/test_bikes.gd`, `tests/test_ui.gd`, and the new `scripts/fire.gd` and `tests/test_boathouse.gd`. Save format unchanged: version 2, with four new optional keys. Every edit is one undo step. Review: see the Review line below. Platforms: Mac run; Windows not run. 243: - **The kind.** `boathouse` is a Buildings card in Architecture (tray 71 to 72), 5 by 7 m. 244: - Rows: Walls (stone or timber), Roof (gable or hipped), Water door (round or pointed), and Stone (the variant). The keys are `boathouse_walls`, `boathouse_roof` and `boathouse_arch`, each omitted at its first value and each checked against the shared inspector keys. Each row is one undo step. 245: - Magnet (`Magnets.boathouse`): released within 3 m of a quay's lip or a crag's foot (`crag_base`), the house goes to the nearer one. Its back sits on the edge, it faces out, and its floor is on the quay or on the rock shelf. At a quay it slides along the lip until clear. Ctrl places it as drawn. 246: - Refusals: the house is refused off the water, in the rock, over a jetty's deck, or over a hull. A jetty drawn through it, a boat moored across its walls, or a crag drawn over it is refused too. Each has its own hint. All are soft for edits elsewhere and are refused for the part the edit changed (`R8B_SOFT`, `_part_changed`). 247: - **Following a crag.** Correction to the brief: raising a crag moves its foot outward (the batter is 0.15 of the height; +4 m moves the foot about 0.6 m). `crag_carry` therefore re-seats each boathouse that sits on the old foot, through `foot_seat`; the tests cover a 0.45 to 0.9 m move. This applies to the Height row, a vertex drag and a crag move, both live and on commit. 248: - **Jetty and boats.** A jetty runs along the house's side, clear of its walls both ways, with `BOATHOUSE_CLEAR` 0.05. 249: - A boat up to 3.7 m long released in the channel is berthed there (`Magnets.boathouse_berth`): on the channel's axis, bow to the water door. It takes no mooring line or buoy. 250: - Creatures do not nap or walk in the footprint. The R4c organic pieces keep 1 m clear. 251: - **Look** (`Architecture.boathouse`; no MultiMesh; materials stone, wood, tile, plain, metal, warm and pool): 252: - Inside: walkways on posts, a back landing, and a plank slip with rails into the channel. 253: - Walls: stone on a plinth with a cap course, or timber boarding on piles standing 0.3 m clear of the water. 254: - Openings: a land door in the back wall at a quay, or else a side door onto a landing. A voussoir ring or arch board round the water door. 255: - Roof: coursed slate or tile with 0.45 m eaves, fascia, bargeboards and a ridge cap. There is always a gable over the water door: a lit window under the gable roof, a louvre under the hipped one. 256: - Torches: two door torches. 257: - Sheets: `b/sheet/r8b-houses.png`, `r8b-close.png`, `r8b-flotilla.png` and `r8b-scenes.png`. 258: - **Shared torch** (`scripts/fire.gd`, `Fire.torch`, for R8e too). It draws a bracket, a shaft, flame cones and crossed pool cards at fixture order 1, plus one light of kind `torch` (energy 1.0, range 3). 259: - **Bow lantern** (`bow_lantern: true`, rowing style only; restyling the boat drops it; it changes nothing on other styles). 260: - Size: a pole and crook with a 0.2 by 0.28 m glass in a frame. It has two 0.8 by 0.9 glow cards and one light of kind `bow_lantern` (energy 1.0, range 3.0). It lights in the lamplighter's round and makes a water streak like a lamp's. 261: - Look: it reads at dusk, though it is still modest; the orchestrator ruled it be left as is. 262: - Ranking: `torch` and `bow_lantern` are in `WorldView.BELOW_LAMPS` with the fire bowls, so every town lamp keeps its light first. On the stress scene at dusk, 32 emitters share the 16-light pool with none starved. 263: - **Tests.** `tests/test_boathouse.gd` (after test_crag): 264: - Tests first: 17/17 failing (`b/testfirst/`). Now 124/0. 265: - The blocks cover data, water, quay, crag, follow, strand, jetty, boats, torches, lantern, boat pins (54 BOATPINs equal to the pre-tree), ranking, organic, creatures, undo, and walks (lesson 1: one index walk per pass with 6 houses and 20 boats). 266: - Changed expectations: tray 71 to 72 (test_decor ×4, test_bikes ×1); the test_ui tray filter "boa" now also finds "Boathouse". 267: - Affected suites after the notes (`b/suites/summary.txt`): 33 suites, 0 failures. 268: - Full `tools/suite_all.sh` a, b, then v at `VERIFY_RES=2880x1800` (`b/suite/`, 19:08 to 19:22): 58 suites, 31803 checks, 0 failures; verify clean; the creature_reactions rerun passes. 269: - **Unchanged without a boathouse.** 270: - Pins: R7a 37, R7c 66 and bridge 84 pins all show diff 0 (`b/pins/final/`). The g0 ident shows 15/15, diff 0 (`b/ident/final/`). The starters' md5 pins pass. All 54 boat pins are identical (`b/probe/`). 271: - Existing views: the Castle starter, Harbour, quarter, Amsterdam and five R1 boat views are within the noise pair, with identical counts (`b/metrics/README.txt`; diffs are only rain, glints and creatures). 272: - **Cost** (3+3 interleaved against the pre-tree; medians of 7 per run). 273: - **Harbour** (with a boathouse and 3 lantern boats; the pre-tree reads it without the boathouse) (`b/cost/summary.txt`): 275: | Edit | Pre | Live | Change | 276: |---|---|---|---| 277: | Lantern row | 27.28 ms | 27.34 ms | +0.06 ms | 278: | Boat placement | 26.38 ms | 26.38 ms | 0.00 ms | 279: | Boathouse placement | not loadable | 27.6 ms | live only | 281: - **Stress** (6 boathouses, 20 lantern boats; the pre-tree reads it without boathouses): two sittings. 282: - First (`b/cost/release.out`): the lantern row is +1.25 ms and the boat placement +1.63 ms against pre. 283: - Second (`b/cost/release2.out`): the boat placement is +3.03 ms against pre (13.74 against 10.71 ms) and +2.95 ms against the same scene without boathouses on the live tree (10.79 ms). 284: - Accepted by the ruling of 15 October about 19:34: `boathouse_error` is flat with scene size (0.52 to 0.59 ms for 6 houses, one index walk a pass), and the Harbour starter (the budget scene) has no boathouse. A memo keyed on terrain and position is carried as later hardening. 285: - Traced (lesson 11, `b/cost/summary2.txt`, `b/dev/`): on the same scene, live minus pre is +0.04 ms for the commit pass, +1.32 ms for the lantern row and +0.08 ms for the boat placement. The 5 boathouses present add about 1.0 ms to each commit pass, plus sync. 286: - A cost pass (`b/design.md`; first sitting in `b/superseded/cost/first/`) halved `boathouse_error`, from 0.91 to 0.47 ms a pass. 287: - **Frame time** at dusk, vsync off (`b/bench/bench.txt`): 289: | Scene | Frame p50 | Draw calls | 290: |---|---|---| 291: | 20 lantern boats, 6 houses | 2.92 ms | 321 | 292: | The same without lanterns | 2.86 ms | 281 | 293: | 5 lantern boats | 2.48 ms | 129 | 294: | 5 boats without lanterns | 2.45 ms | 101 | 296: - **Harbour handle drag**, 6+6 (`b/drag/summary.txt`): p95 55.4 ms live against 55.6 ms pre; release 9.62 against 9.60 ms. 297: - **Power.** Every capture, bench and suite sitting ran on AC at 100%; none ran on battery. 298: - **Older builds.** A build before R8b refuses a scene with a boathouse with "Unknown or damaged building part." A boat's `bow_lantern` key loads there and is ignored. 299: - **Not done / open:** 300: - R7g0's rubble Finish is not offered on the boathouse. 301: - No stair up from the water (the plan's element table); the land door at a quay or the side door's landing serves. 302: - With no multi-select (R5e), a flotilla is made by Alt-drag copies. 303: - The lantern's light and streak stand at the boat's rest pose, while the mesh bobs (at most 3.5 cm). 304: - On the stress scene, the boathouses' presence costs about 1 ms per commit pass. 305: - **Fix pass 1** (orchestrator rulings of 15 October about 19:34, `artifacts/decor/owner-decisions.md`; review `b/review/`), 19:35 to 20:00. Backups `b/fix1/backup/` (.gdignore'd); evidence `b/fix1/`; patch `b/fix1/fix1.diff` (311 lines: `scene_state.gd`, `main.gd`, `adapt.gd`, `tests/test_boathouse.gd`). No drawing changed, so nothing was recaptured. 306: - **1. Berthed boats ride with their boathouse.** `DioramaState.carry_berths(before, after, skip)` moves every boat lying wholly in the old house's channel (`berthed_in`) by the house's own move and turn; each boat keeps its height and is replaced by a copy. 307: - `crag_carry` calls it for each boathouse it re-seats, so the live Height scrub, the arrow drag, a crag move preview and the commit all carry the boats. 308: - `Main._carry_boathouse_berths` runs before every commit pass (after `_carry_crag_tops`). Every boathouse the open edit moved or turned (a move, a turn, any reseat) takes its berthed boats, except boats the edit itself changed; all of it is inside the edit's one undo step. 309: - `boathouse_error` no longer refuses a moved house for the boats berthed in it as it stands, since they ride. 310: - **2. Long boats.** A boat over 3.7 m whose middle is in a channel is refused with `BOAT_LONG_ERROR` "This boat is too long to berth here: boats up to 3.7 m fit the channel.", hint "Shorten it to 3.7 m or less, or moor it stern-in with its bow out through the water door." It is soft for edits elsewhere, like the wall line. Moored stern-in through the door, a long boat still stands. 311: - **3. Rock before water.** `boathouse_error` runs the rock check before the water check, so a house pushed into a crag gets the rock line. 312: - **4 and 5.** The older-build line now quotes the real message, and the stress cost lists both sittings with the ruling (above). 313: - **Tests first** (`b/fix1/testfirst/test_boathouse.log.out`, the unfixed `scene_state.gd`, `main.gd` and `adapt.gd` swapped back in): 18 of 167 checks fail. 314: - Five crag edits with a berthed boat are refused: raise to 12 m and to 14 m by the Height row, move 3 m east, move 1.5 m toward the house, and reshape narrower. 315: - The boat does not ride a hand move along the quay. 316: - The 4.4, 5.5 and 7.0 m boats get the berth hint. 317: - The house into the rock gets the water line. 318: - **Now** 172/0 (`r8b_fix1_checks`): each of the five crag edits commits in one undo step, the boat rides berthed on the channel's axis at its own height, and undo restores both; a hand move of 6 m carries the boat and undoes in one step; the long boats are refused with the limit and stand stern-in; the wall line is kept; the rock line holds. 319: - **Suites** (`b/fix1/suites/summary.txt`): test_boathouse, crag, decor, scene, building_tools, starters, ui, adapt, composition, bikes, creatures and dusk: 12 suites, 0 failures. 320: - **Pins against `b/pre-tree`:** R7a 37, R7c 66 and bridge 84 pins diff 0 (`b/fix1/pins/final/`); ident 15/15, diff 0 (`b/fix1/ident/final/`); 54 boat pins, diff 0 (`b/fix1/probe/`). 321: - **Full suite** (`tools/suite_all.sh` a, b, then v at `VERIFY_RES=2880x1800`, `b/fix1/suite/`, 19:45 to 19:59, on AC at 100%): 58 suites, 31854 checks, 0 failures; verify clean; the creature_reactions rerun passes. 322: - **Fix pass 2** (orchestrator rulings of 15 October about 20:11, `artifacts/decor/owner-decisions.md`; recheck `b/review2/`), 20:12 to about 20:36. Backups `b/fix2/backup/` (.gdignore'd); evidence `b/fix2/`; patch `b/fix2/fix2.diff` (229 lines: `scene_state.gd`, `main.gd`, `adapt.gd`, `tests/test_boathouse.gd`). No drawing changed, so nothing was recaptured. 323: - **1. Hand moves of a house with its own berthed boat.** `Main.validate_candidate` now carries the berths on its check copy for a moved boathouse (`check.carry_berths(old, new)` after the candidate replaces the house), so the drag's preview and release gate no longer refuse the house for its own boat. **The boat follows at release, not the ghost:** the drag ghost draws the house alone, and the boat moves when the edit commits (`_carry_boathouse_berths`), inside the same undo step. 324: - **2. Water first, then rock, then crag tops.** `_boathouse_wet(o, tops)` takes a flag; `boathouse_error` checks water on the terrain alone (crag tops off), then the rock, then water again with the crag tops. A house on the land beside a land crag gets the water line again; a water house pushed into a crag keeps the rock line. 325: - **3. Long boats skewed or off the axis.** The long-boat check in `boat_boathouse` no longer has the `local.z <= BOATHOUSE_D + 0.5` bound, so a long boat with its middle in the channel gets `BOAT_LONG_ERROR` however it lies. The hint is now "Shorten it to 3.7 m or less, or moor it stern-in, straight along the channel, with its bow out through the water door." 326: - **4. Walks.** `carry_berths` reads `boathouse_index().boats` instead of walking every object; `_carry_boathouse_berths` returns at once when the index has no boathouses and compares only the indexed houses and boats with the edit's pending state. 327: - **Tests first** (`b/fix2/testfirst/test_boathouse.log.out`, the unfixed scripts swapped back in): 13 of 200 checks fail. 328: - Four hand moves are refused for the house's own boat: 0.5, 1 and 2 m along the quay, and a 0.1 rad turn. 329: - Three land houses beside a land crag, at x -12, -10 and -8, get the rock line. 330: - Six long boats, 4.4, 5.5 and 7.0 m at 0.3 m off and 10° skew, get the berth hint. 331: - The walk test passes on the old code too: the unfixed code walked every object, not the index, so the index walk count cannot tell the two apart. The cost figures below show the difference. 332: - **Now** 200/0 (`r8b_fix2_checks`, `b/fix2/dev-test_boathouse.log.out`): 333: - each hand move passes `validate_candidate`, commits, carries the boat to the same channel-local place, and undoes in one step; 334: - with nothing moved, the carry makes no index walks; a moved house makes at most one; 335: - the land houses get the water line, and the house pushed into rock gets the rock line; 336: - the long boats get `BOAT_LONG_ERROR`, and its hint says "straight along the channel". 337: - **Recheck probe on the live tree** (`review2/probe2.gd`, `b/fix2/probe/probe2-*.log.out`; before is `b/review2/probe2-*.log.out`): 338: - **Cost** of `_carry_boathouse_berths` with nothing moved, median of 9: 339: | Scene | Before | After | 340: |---|---|---| 341: | Harbour starter | 0.067 ms | 0.001 ms | 342: | Harbour with a boathouse | 0.069 ms | 0.010 ms | 343: | Stress (6 houses, 20 boats) | 0.023 ms | 0.031 ms | 344: | Stress + 400 objects | 0.219 ms | 0.032 ms | 345: | Stress at 1600 objects | 0.788 ms | 0.031 ms | 346: - With one house turned, at 1600 objects: 1.152 ms before, 0.502 ms after. 347: - **Drag:** `validate_candidate` returns '' for moves of 0.25, 0.5, 1, 2, 4, 6 and 12 m, and the boat is carried at commit. A 2 m body drag moves the house and the boat in one step. 348: - **Rock:** as item 2. 349: - **Long:** all six skewed cases get the long-boat line and the new hint. Before, they got "Moor the boat in the boathouse's channel, or clear of its walls." 350: - **Edges:** the log is the same as the recheck's, apart from the new hint. 351: - **Limit carried to the R8 close.** A boat waiting just outside the water door is not berthed (`berthed_in` is false), so it does not ride. A crag raise or reshape that would move the house onto it is refused with the boat line ("Keep the boathouse's walls clear of the boats moored here."). This refusal is soft: the player moves the boat (`b/fix2/probe/probe2-edges.log.out`, "on-axis just outside door"). 352: - **Suites** (`b/fix2/suites/summary.txt`): test_boathouse 200, crag 200, decor 2717, scene 1736, building_tools 982, starters 397, ui 590, adapt 876, composition 160, bikes 499, creatures 250 and dusk 461: 12 suites, 0 failures. 353: - **Pins against `b/pre-tree`:** R7a 37, R7c 66 and bridge 84 pins diff 0 (`b/fix2/pins/`); ident 15/15, diff 0 (`b/fix2/ident/`); 54 boat pins, diff 0 (`b/fix2/probe/boats-summary.txt`). 354: - **Full suite** (`tools/suite_all.sh` a, b, then v at `VERIFY_RES=2880x1800`, `b/fix2/suite/`, 20:20 to 20:34, on AC at 100%, `b/fix2/suite-power.txt`): 58 suites, 31882 checks (28 more than fix pass 1, all in test_boathouse), 0 failures; verify clean; the creature_reactions rerun passes. 355: - **Review:** two rounds by fresh Opus 5.5 reviewers, each followed by a fix pass, and the orchestrator's check of the last; the orchestrator's look notes of about 18:37 (house roof materials with eaves, a gable over the arch, a stone water course, piles under timber, the slip visible, bigger bow lanterns with a water streak) were folded in before the bench and accepted on the remade sheets at 19:06. (1) Review (`b/review/`: `probe.gd`, `probe-crag.log.out`, `probe-refuse.log.out`, `pins/`, `suites/`): exact reseats on a crag foot and a quay for every crag edit without a boat, refusals both ways with their own hints, `Fire.torch` as the one torch for R8e, torch and bow lantern ranked below every town lamp, organic waterside pieces clear, no cat naps in or on a house, the look, the save keys and every pin held; the commit-pass cost of present boathouses (`boathouse_error`, 0.52 to 0.59 ms for 6 houses, one index walk a pass) is flat with scene size. Must-fix: a crag raise, move or reshape was refused for a boat berthed in a boathouse at its foot; a boat too long for the channel was told to release it in the channel. The orchestrator's rulings of about 19:34 (`artifacts/decor/owner-decisions.md`, reversible) set the remedies and accepted the stress cost (both sittings in the entry; a memo keyed on terrain and position carried as later hardening); closed by fix pass 1 (above). (2) Recheck (`b/review2/`: `probe2.gd` and its logs): the crag carries with berthed boats, two boats, lantern boats, turns and a deleted house sound, no boat lost, duplicated or left in the walls; must-fix: a hand move was still refused for the house's own berthed boat (the move check did not carry it), a house on land beside a land crag got the rock line, a skewed long boat got the berth hint; and `carry_berths` walked every object per moved house. The orchestrator's rulings of about 20:11 set the remedies; closed by fix pass 2 (above), checked by the orchestrator against `b/fix2/suite/suite_all.out` (58 / 31,882 / 0), `b/fix2/suites/summary.txt` (12 suites, 0 failures), `b/fix2/pins.run` (37, 66, 84 diff 0), the ident and the 54 boat pins (diff 0), the failing test-first log (`b/fix2/testfirst/`, 13 of 200) and the re-run probe (`b/fix2/probe/`: hand moves accepted with the boat carried at release, the land and rock lines right, every skewed long boat given the limit hint, the carry with nothing moved 0.03 ms at 1,600 objects against 0.79). Notes, not defects (carried to the R8 close, reversible): a boat waiting just outside the water door is not berthed, so a crag raise or reshape onto it is refused with the boat line (soft); a berthed boat follows a hand move at release, not in the drag's ghost; R7g0's rubble finish is not offered on the boathouse; no stair up from the water; a flotilla is made by copies until multi-select (R5e); the bow lantern's light holds the boat's rest pose through its bob (at most 3.5 cm); about 0.5 ms a commit pass with boathouses present. **Closed.** SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 177-241 (Castle R8c Glasshouse, owl tower, stone circle, pumpkins) 177: **Castle R8c Glasshouse, owl tower, stone circle, pumpkins (15 October 2026):** Opus 5.5 ran this on the Mac, 15 October about 20:41 to 22:30, fix pass 1 from 22:45 to 23:20, and fix pass 2 from 23:25 to 00:10 on 16 October (step `36-R8`, part c; `docs/castle-elements-plan.md` R8c). The orchestrator's look notes (about 21:38) were folded in, and the remade sheets were accepted about 22:08. Snapshot `build/backup-feel/36-R8-c-pre/` (= HEAD 03ed81f); scratch pre-tree `c/pre-tree`; backups `c/backup/` and `c/notes1/backup/` (.gdignore'd). Evidence `artifacts/decor/36-R8/c/` (`c/` below). Design `c/design.md`, with sections for the look pass and the look notes. Patch `c/part.diff` (2489 lines) covers `scripts/scene_state.gd`, `main.gd`, `world_view.gd`, `architecture.gd`, `editor_ui.gd`, `adapt.gd`, `composition.gd`, `creatures.gd`, `build_audio.gd`, `data/catalog.tres`, `tools/test.sh`, `tests/test_decor.gd`, `tests/test_bikes.gd` and the new `tests/test_grounds.gd`; the notes' own diff is `c/notes1/notes.diff`. Save format unchanged: version 2, with new optional keys only. Every edit is one undo step. Review: see the Review line below. Platforms: Mac run; Windows not run. 178: - **The kinds** (generic names, K8; tray 72 to 75): 179: - `glasshouse` (Architecture, Buildings, after the boathouse). Rows Form (`glasshouse_form` "dome"; absent gable), Length (`glasshouse_len` 3 to 12 m, step 0.5; absent 6) and Frame (the shared `variant`: white, the R2 ironwork green, black). Width `clamp(2.4 + 0.2 L, 3, 4.8)`, eave 2.4 m, on a 0.45 m brick plinth sunk 0.55 m. The dome form is a curved vault over the length with half-domed ends (3 m long: a round pavilion). Inside: a bed of shrubs, a staging bench of pots, hung bulbs and a small lamp at each shrub. One light of kind `glasshouse` (energy 1.2, range 5) in `WorldView.BELOW_LAMPS`; the glass is the existing `glass` at fixture order 1, so it lights in the lamplighter's round. Refused over water and on ground falling more than 0.5 m; on a terrace its footprint must fit. 180: - `stone_circle` (Architecture, Walls, after the crag). Rows Stones (`circle_stones` 5 to 13; absent 7), Radius (`circle_r` 3 to 10 m; absent 4), Centre (`circle_centre` "stone" altar slab or "sundial"; absent none) and Stone (grey, crag, dark). Stones seeded from the part's seed and id (lean, height, rough-hewn loft, moss at the foot); each foot stands on the ground under it, read each sync and never saved. Refused over water, off the canvas, or on a parent. 181: - `pumpkin` (Garden, Beds and pots, after the shrub). Row Size (`pumpkin_size` "small" 0.35 m, "large" 1.1 m; absent medium 0.6 m); a ribbed gourd with stem and leaf, tint and tilt from seed % 4. A patch is several pumpkins or Alt-drag copies: no patch kind. 182: - Tower row Perches (`perches: true`, written only when true; canal, medieval and victorian towers, alone, joined by curtain walls or merged with a hall since fix pass 1). The water and slope lines are soft for edits elsewhere and refused for the part the edit changed (`R8C_SOFT`). 183: - **The owl tower** (after look notes 1). Under each window of the top window floor (at most 4, a D-plan's on its round face) the window is a glassless dovecote opening (`Architecture.dovecote`: a dark mouth split into six pigeon holes) with a timber landing board on beams and struts, 0.8 m out. The window is still drawn into a spare batch, so the rng runs as before and the tower's other windows are unchanged (tested). 184: - Birds (`Creatures`, derived each re-plan, never saved): rock pigeons, 6 a tower, 18 for all towers together (1 to 3 towers 6 each, 4 towers 4, 6 towers 3), counted in a cap of their own outside the 30 bodies, as the swarms are, and drawn in the pigeons' multimesh (room 9 + 18). The ordinary pigeons keep their own cap. Two of each flock start on ledges; the rest circle over the cap. Every 8 to 20 s one lands on a free ledge by an approach point 1.4 m out, sits 10 to 35 s, and leaves the same way. A ledge whose flights are not clear is not used. Motion Off, or no clear circle: one bird a ledge, all sitting. Startled, a perched bird rejoins its circle. 185: - Correction (fix pass 1): before fix pass 1 a perched tower in a cluster drew nothing (merged with a hall) or drew its ledges with no birds (joined by a curtain wall). The Harbour test scene's perched canal tower is in a 2-building cluster, so it drew nothing, and the first Perches-row cost below was measured on a tower that drew nothing. Since fix pass 1 it draws 1 ledge (its other top windows face its neighbour) and brings 6 birds (`c/fix1/harbour_check.log.out`). 186: - **Look notes 1** (`c/notes1/`): the first towers showed no readable perches and a bird or two, and the white glasshouses read at dusk as opaque lanterns. Now: the dovecotes, boards and flocks above; the glass is thinner and warmest low down (lower walls a 0.42, upper 0.34, roof 0.28; the bulbs lower, a lamp at each shrub, the pool light at 1.1 m), so the bars and planting show. First sheets and the stopped first cost sitting are in `c/superseded/`. 187: - **Tests.** `tests/test_grounds.gd` (after test_boathouse): tests first 17/17 failing (`c/testfirst/`), now 164/0 (`c/notes1/dev/test_grounds.log.out`). Blocks: data, keys, glass ground, light, view, soft, perch drawing (dovecotes, other windows unchanged), perch birds (flock of 6, 2 sitting on their own ledges, the circle over the cap, caps, a minute of flight never inside a wall or roof, deterministic twice), motion Off, circle terrain and seed, pumpkins, undo, round trip, meshes, unchanged, walks (one crowd pass with 6 glasshouses, 6 perched towers, 4 circles and 20 pumpkins). 188: - Changed expectations: tray 72 to 75 and Garden 20 to 21 (test_decor ×4, test_bikes ×1). 189: - Affected suites after the notes (`c/notes1/suites/summary.txt`): 36 suites, 25413 checks, 0 failures. 190: - Full `tools/suite_all.sh` a, b, then v at `VERIFY_RES=2880x1800` (`c/suite/`, `c/suite-run.out`, 22:11 to 22:25): 59 suites, 32081 checks, 0 failures; verify clean; the creature_reactions rerun passes. 191: - **Unchanged without the new parts** (`c/pins/summary.txt`, after the notes): R7a 37, R7c 66 and bridge 84 pins diff 0; g0 ident 15/15 diff 0; R8b's 54 boat pins identical; 959 creature-plan pins (every starter, the R4 close's Courtyard, Castle and Amsterdam, day and night, at the start and after 20 s) identical pre and live. Starters' md5 pins pass in the suites. 192: - Existing views (`c/notes1/metrics/`): integer counts identical on all 10 (Castle starter, Harbour, quarter, Amsterdam, Courtyard and Harbour with birds; pre-1 vs n1-a). Change pre-2 vs n1-a is within the larger noise pair on every view but harbour-night (30.0 % against 29.5 % live noise; diffs are rain, smoke and creatures). 193: - **Captures and sheets** (`c/cap/n1-a`, `n1-b`, noise pair `c/notes1/metrics/noise-n1-new.txt`): `c/sheet/r8c-glass.png`, `r8c-towers.png` (default and close, each with a frame half a second on), `r8c-circles-pumpkins.png` (5 and 13 stones on a slope, 9 with a sundial on a crag top, 11 with an altar), `r8c-crag4.png` (all four in R8a's crag scene, day and dusk) and `r8c-existing.png`. The tower scene now perches three towers (medieval round, canal flat, D-plan) and leaves the victorian plain. 194: - **Cost** (`c/cost/summary.full.txt`; 3+3 interleaved against the pre-tree, medians of 7): 196: | Edit | Harbour pre | Harbour live | Stress pre | Stress live | 197: |---|---|---|---|---| 198: | Perches row (superseded: the Harbour tower drew nothing; see fix pass 1) | 75.59 ms | 75.79 ms | 51.81 ms | 55.72 ms | 199: | Pot placement | 28.25 ms | 26.91 ms | 14.00 ms | 17.30 ms | 200: | Glasshouse placement | not loadable | 30.40 ms | not loadable | 19.99 ms | 201: | Circle placement | not loadable | 29.78 ms | not loadable | 19.93 ms | 202: | Pumpkin placement | not loadable | 28.00 ms | not loadable | 16.90 ms | 204: - The pre-tree reads the scenes without the three kinds. On the stress scene the live side holds 27 more parts (33 objects against 6), so its +3.3 to +3.9 ms is their presence in each commit and sync pass (the sync's decorations phase is +1.7 ms of it), not a change to an existing edit. The Harbour (budget scene) is within noise. 205: - Frame at dusk, vsync off, creatures live (`c/bench/bench.txt`): 6 glasshouses and 6 perched towers 4.44 ms p50, 637 draw calls; the same towers plain without glasshouses 3.96 ms, 529 draw calls. 16 of 16 lights, none starved. 206: - Creatures (`c/cost/creature-summary.txt`): stress (18 tower birds) plan 10.53 ms against 7.73 ms plain (the plan job spreads it over frames), tick 0.086 against 0.042 ms; Harbour unchanged (11.56 against 11.52 ms, tick 0.061). 207: - Harbour handle drag, 3+3 (`c/drag/summary.txt`): p95 57.7 ms live against 56.0 ms pre (runs overlap: pre 60.4, 56.0, 55.5; live 57.7, 60.0, 54.5); release 9.57 against 9.59 ms. Baseline 53.7 ms. 208: - **Power.** Every capture, bench, cost and suite sitting ran on AC at 100%; none ran on battery. 209: - **Older builds** refuse a scene with a glasshouse, stone circle or pumpkin ("Unknown or damaged building part."); a tower's `perches` loads there and is ignored (drawn plain). 210: - **Not done / open:** 211: - No rubble Finish on the glasshouse; no patch kind for pumpkins. 212: - Corrected by fix pass 1: the line that stood here (a perched tower joined to a building or wall brings no birds) no longer holds. 213: - Tower birds are outside the 30-body cap (bounded at 18); with 6 perched towers the plan costs about 2.8 ms more, spread by the plan job. 214: - The 27 new parts on the stress scene cost about 3.3 ms per commit pass; no memo yet. 215: - **Fix pass 1** (orchestrator rulings of 15 October about 22:38, `artifacts/decor/owner-decisions.md`; review `c/review/`), 22:45 to 23:20. Backups `c/fix1/backup/` (.gdignore'd); evidence `c/fix1/`; patch `c/fix1/fix1.diff` (783 lines: `scene_state.gd`, `main.gd`, `world_view.gd`, `architecture.gd`, `cluster_architecture.gd`, `creatures.gd`, `composition.gd`, `adapt.gd`, `tests/test_grounds.gd`). `c/part.diff` is regenerated (3098 lines). 216: - **1. Clustered perched towers.** `_tower_alone` is gone: every perched tower offers its ledges. A tower sharing its cluster with another building (`DioramaState.tower_clustered`, walked out from the tower and stopping at the first other building) is drawn by `ClusterArchitecture.tower_windows`, which now takes its windows from `tower_slots`. Its perches are `tower_perch_slots`: the windows of its top floor whose 0.8 m board stands clear of the cluster's other bodies. The perched windows draw as dovecotes with boards, and the plain window is still drawn into a spare batch, so the rng and the other windows are unchanged. `Architecture.perch_points(o, state, clustered)` reads the same slots, so the birds' sites are what is drawn; the creature plan learns each tower's cluster from one `clusters()` walk a plan, and only when a perched tower is present. A tower joined only by curtain walls is still drawn by `Architecture.building` and now gets its flock. 217: - The Perches row never does nothing silently: when a tower has no window a perch can stand under (solid walls, or every top window covered by the hall it joins), the change is refused with `TOWER_PERCH_ERROR` "This tower has no open window for a perch." and the hint "Give the tower windows, or keep its upper floors clear of the hall it joins." Correction (fix pass 2): as built in fix pass 1 this check also ran after any row on a perched tower, so a tower whose windows another part's edit had covered refused every row of its own (stranded); see fix pass 2, item 1. 218: - The R7 cue castle with every tower perched (`c/fix1/scenes/castle-perched.json`, `scenes.log.out`): all 12 towers draw dovecotes and boards. 11 have 4 ledges and tower 139 has 1; birds are 2 a tower on 8 towers, 1 each on towers 139 and 140 and none on 141 and 142: 18 in all, the cap. 219: - **2. Covered perches.** No ledge site inside another building's box, a freehand wall's box, a crag (in its battered foot, below its top) or a glasshouse's box (`Creatures._perch_buried`). Motion on and Off, nothing sits or flies inside a crag, wall or house drawn over a ledge. 220: - **3. The glasshouse floor.** A ground glasshouse's footprint joins both footprint indexes (`Composition.footprint_index`, crags on and off), appended only when one is present, so other scenes key as before. No meadow tuft or planting stands inside its floor (`c/fix1/sheet/r8c-fix1-floor.png`: before n1-a, after f1-a). 221: - **4. The entry.** The Harbour lines are corrected above. The Perches row is re-measured on the Harbour tower now that it draws (`c/fix1/cost/summary.txt`, 3+3 interleaved, medians of 7): 76.99 ms live against 74.61 ms pre (+2.38: the cluster draws the dovecote and board, about +1.3 ms of clusters, and the row's check walks out the tower's cluster), undo 65.48 against 64.20. On the stress scene it is 55.62 against 52.06 ms. The other rows are within the earlier figures. A first sitting before the cluster walk replaced `clusters()` in the row check (+4.16 ms) is in `c/fix1/superseded/cost-first/`. 222: - **5. Crag-edge circles.** A stone circle whose stones (and centre piece) stand partly on a crag's top and partly off, or with any stone in the battered foot, is refused with `STONE_CIRCLE_CRAG_ERROR` "Stand the whole circle on the crag's top, or clear of its edge and foot." and the hint "Move the circle wholly onto the top or out past the foot of the rock, or change its radius." It is soft for strokes (`R8C_SOFT`), as a house over the edge is. 223: - **6. Rule 5.** Settled glasshouses and stone circles no longer draw one MultiMesh a look: each terrain chunk's glasshouses and circles are built into one shared mesh (`WorldView._sync_merged`, `MERGED_KINDS`). A chunk's mesh is rebuilt only when one of its parts changes its key or frame, and each part still lists its own lights under `decoration:`. A part growing in or sinking away is drawn on its own look's MultiMesh until its animation ends, then moves into the shared mesh. Test: 20 glasshouses of 19 lengths, both forms and three frames, and 20 circles of 5 to 13 stones, three centres and three tones: 0 MultiMeshes a kind once settled, all 40 in the shared meshes, 20 glasshouse lights listed. Correction (fix pass 2): the sentence that stood here (the glass's turn one per chunk, the shared mesh's origin) was false. The shared mesh stood at the world origin, so the settled glass took a turn of the world origin's sweep and its own seed, while its light kept the glasshouse's own origin: at (22, 20) the glass turned at 0.035 and its light at 0.896. See fix pass 2, item 2. 224: - **Tests first** (`c/fix1/testfirst/test_grounds.log.out`): 12 of 187 checks failed, plus a script error (no `merged_count`). Now `test_grounds` 189/0 (`c/fix1/dev/`). 225: - **Suites** (`c/fix1/suites/summary.txt`): test_grounds, test_creatures, test_creature_reactions, test_creature_life, test_decor, test_scene, test_building_tools, test_starters, test_crag, test_boathouse, test_composition and test_dusk: 12 suites, 14562 checks, 0 failures. Full `tools/suite_all.sh` a, b, then v at `VERIFY_RES=2880x1800` (`c/fix1/suite/`, `c/fix1/suite-run.out`, 23:01 to 23:16): 59 suites, 32112 checks, 0 failures; verify clean; the creature_reactions rerun passes. 226: - **Pins** (`c/fix1/pins/summary.txt`): R7a 37, R7c 66, bridge 84, ident 15, 54 boat pins and 959 creature-plan pins, all diff 0. 227: - **Captures** (`c/fix1/cap/f1-a`, `f1-b`, noise `c/fix1/metrics/noise-f1.txt`): the cue castle perched, default and half a second on, the clustered towers close and half a second on (`c/fix1/sheet/r8c-fix1-castle.png`), and gable-close-day. 228: - All fix-pass sittings ran on AC (`c/fix1/cap/power.txt`, `c/fix1/cost/power.txt`, the suite's POWER lines). 229: - **Fix pass 2** (orchestrator rulings of 15 October about 23:35, `artifacts/decor/owner-decisions.md`; recheck `c/review2/`), 23:25 to 00:10 (16 October). Backups `c/fix2/backup/` (.gdignore'd); evidence `c/fix2/`; patch `c/fix2/fix2.diff` (313 lines: `main.gd`, `world_view.gd`, `tests/test_grounds.gd`). `c/part.diff` is regenerated (3296 lines). 230: - **1. Stranding.** `change_parameter` records whether the tower had a perch slot before the row (`_perch_slots_open`, which walks only a perched tower). It then refuses the Perches row itself when the tower has no slot, or a row on the tower that takes away the last slot it had. A tower whose slots another part's edit took away is a soft state: it keeps its `perches` key, draws no ledges and gets no birds until a window opens, and its own rows work. Test (the recheck's case): a tower ringed by four halls takes its perches with 4 slots; raising each hall to 11 m takes them to 0 with no site and no bird; the tower's Variant row is then accepted, the Perches row (off and on again) is refused, and a Windows row that would leave an open tower no slot is refused. The recheck's probe (`c/fix2/cost/probe-row2.log.out`): after the four raises the Variant, Roof, Style and Perches-off rows are all accepted. Cost: the Variant row on the Harbour's perched tower 71.42 ms against 70.81 ms on the same tower unperched (medians of 7, `probe-hcost2.log.out`). 231: - **2. Dusk.** Each chunk's shared mesh now stands at `Masonry.chunk_center`, and each part's piece is built relative to it. A light's `dusk_origin` is the chunk's centre, its position is moved back to world space, and its seed is the hash of the piece's frame, which is what the glass's vertices carry. Settled, a glasshouse's glass and its light take the same turn. The sweep (where the town's round reaches) is per chunk, at the chunk's centre; the seed is per part, so two glasshouses in one chunk still light apart. While a glasshouse grows in it is drawn on its look's MultiMesh at its own origin, and its light takes that origin, so glass and light also agree then. Test (four glasshouses at (2, 2), (6, 9), (22, 20) and (-24, 20)): glass equal to light within 0.001 growing and settled, all four at turns of their own, every chunk node at its centre. 232: - The turn moves when a part settles, because the sweep moves from the part's origin to its chunk's centre. In the test the turns go 0.272 to 0.525, 0.529 to 0.495, 0.835 to 0.834 and 0.843 to 0.996. At dusk a glasshouse placed during the round can therefore light or go out about a second after it is placed, glass and light together. 233: - Captures (`c/fix2/cap/f2-a`, `f2-b`, noise `c/fix2/metrics/noise-f2.txt`, sheet `c/fix2/sheet/r8c-fix2-dusk-mid.png`; turns from `c/fix2/cap/hours.log.out`): the glass scene at 19:10 (progress 0.555 against turns 0.523 to 0.582) shows one glasshouse lit, two lighting and one out. The stress scene at 21:00 (progress 0.832 against turns 0.809 to 0.928) shows three lit or lighting and three out. 234: - **3. Cost.** Each settled part's piece is kept by id (`WorldView.merged_pieces`, keyed by its look key, frame and chunk centre), so a chunk's rebuild draws only the parts that changed and joins the rest. A part that leaves the list (deleted, growing in, sinking away) drops its piece, so the memo holds at most one piece per settled part. The recheck's probe (`c/fix2/cost/probe-merged.log.out`, medians of 5): a Length row on one of 20 glasshouses in a chunk releases in 13.37 ms (decorations phase 7.44) against 10.14 ms (3.67) for a lone one, where it was 30.96 against 9.40 ms. A rebuild of the 20-part chunk mesh takes 3.78 ms, where it was 22.98. Moving, deleting and undoing still keep the chunk keying, and picking still works. Test: the row draws exactly 1 piece each time, 6.07 against 2.72 ms decorations phase, and the memo holds 21 pieces and 20 once the lone one is deleted. 235: - **4. The entry.** The fix pass 1 dusk sentence and the stranding sentence are corrected above. Two of the cue castle's towers, 136 and 137, have 2 ledges each that are drawn but get no perch site, because each lies inside a neighbouring tower's box (136's inside 137's and 133's; 137's inside 133's and 136's; `c/fix2/cost/probe-row2.log.out`). No bird sits on a ledge that cannot be seen, and each of those towers keeps 2 ledges with sites. The hall-ring soft state is item 1. 236: - **Tests first** (`c/fix2/testfirst/test_grounds.log`): 6 of 201 checks failed (the Variant row refused; glass and light apart when settled, -1 meaning no piece memo; chunk nodes at the origin; no memo; 24.93 against 2.60 ms). Now `test_grounds` 201/0 (`c/fix2/dev/test_grounds.log`; its NOTE lines carry the numbers above). 237: - **Suites** (`c/fix2/suites/summary.txt`): test_grounds, test_creatures, test_creature_reactions, test_dusk, test_decor, test_scene, test_building_tools, test_starters and test_warmup: 9 suites, 13668 checks, 0 failures (creature_reactions passed on its first run). Full `tools/suite_all.sh` a, b, then v at `VERIFY_RES=2880x1800` (`c/fix2/suite/`, `c/fix2/suite-run.out`, 23:41 to 00:08): 59 suites, 32124 checks, 0 failures. The first two v runs (23:54 and 00:01) stalled in `--verify`: exit 0 after 216 s on `--quit-after`, with no PIXELPUNK_INTEGRATION line and 2 captures. Both stopped in the roof checks, after `scenery-inspector.png` and `roof_saves/profile.json` and before `roof-inspector.png`; their logs are in `c/fix2/verify-stall1/` and `verify-stall2/`, with no script error. An instrumented run that stopped at 6000 frames (`c/fix2/verify-debug/dbg1/`, prints since removed) got past the roof checks. A plain run (`verify-debug/plain2/`, 28 s, PIXELPUNK_INTEGRATION []) passed, and then the suite's third v run (`c/fix2/suite-run-v3.out`, 29 s) passed with verify clean, `PIXELPUNK_INTEGRATION []`. All of these ran the same code. The stall is intermittent and its cause is not found; the roof checks edit only a building, whose rows return before any perch walk. That v run's creature_reactions rerun hit the known flake "The re-plan is applied after the edits" (1 of 6871). The one rerun allowed passes 6871/0 (`c/fix2/reactions-rerun/`), as does the first run in part a. 238: - **Pins** (`c/fix2/pins/summary.txt`): R7a 37, R7c 66, bridge 84, ident 15, 54 boat pins and 959 creature-plan pins, all diff 0. 239: - Power: the captures, the cost probes and the full suite recorded AC (`c/fix2/cap/power.txt`, `c/fix2/cost/power.txt`, the suite's POWER lines). The test-first, suites and pins sittings logged `pmset -g` (settings, Low Power Mode 0) but not the source; a check at 23:41, between them and the captures, read AC at 100% (`c/fix2/power.txt`). 240: - **Review:** two rounds by fresh Opus 5.5 reviewers, each followed by a fix pass, and the orchestrator's check of the last; the orchestrator's look notes of about 21:38 were folded in before the bench and accepted on the remade sheets at about 22:08. (1) Review (`c/review/`: `probe.gd` and its `probe-*.log.out`, `older.log.out`, `pins/`, `suites/`): the cost flat with scene size and within 0.6 to 2 ms of the same slots filled with benches and lamps, the save keys and an older build, and every pin held. Must-fix: perched towers joined by walls or merged with a hall drew nothing (the Perches row did nothing silently); birds sat on covered perches; tufts and planting grew in a glasshouse floor; a stone circle across a crag edge was not refused; rule 5 broke with many circles and glasshouse lengths; the entry's Harbour perch lines were wrong. The orchestrator's rulings of about 22:27 (tower birds capped at 18 outside the 30 bodies) and 22:38 (`artifacts/decor/owner-decisions.md`, reversible) set the remedies; closed by fix pass 1 (above). (2) Recheck (`c/review2/`): clustered and covered perches, the glasshouse floor, the chunk keying and the crag-edge circles hold; the Perches row's +2.38 ms accepted as own-edit content cost. Must-fix: a hall ring raised round a perched tower stranded its rows; settled glasshouses left the lamplighter's round (the shared chunk mesh swept as one); one part's edit rebuilt its chunk neighbours. The orchestrator's rulings of about 23:27 set the remedies; closed by fix pass 2 (above), checked by the orchestrator against `c/fix2/suite/suite_all.out` and `c/fix2/suite-run-v3.out` (59 / 32,124 / 0, verify `PIXELPUNK_INTEGRATION []`), `c/fix2/suites/summary.txt` (9 suites, 0 failures), `c/fix2/pins/summary.txt` (37, 66, 84, ident 15, 54 boat pins, 959 creature-plan pins, all diff 0), the failing test-first log (`c/fix2/testfirst/`, 6 of 201), the cost probes (`c/fix2/cost/`: a Length row on one of 20 in a chunk 13.37 ms against 30.96, the 20-part chunk rebuild 3.78 against 22.98 ms) and the sheet `c/fix2/sheet/r8c-fix2-dusk-mid.png`. The orchestrator's ruling of 16 October about 00:15 accepted the two departures below. Notes, not defects (carried to the R8 close, reversible): a glasshouse's turn in the lamplighter's round shifts when it settles (the sweep is per chunk once settled), so one placed mid-round at dusk can light or go dark about a second after placing, glass and light together; the full suite's verify stalled twice in the roof checks (`c/fix2/verify-stall1/`, `verify-stall2/`) with the same code that then passed twice, cause not found (watch it in R8d's suites); tower birds are capped at 18 in all, so some towers get none; two cue towers' ledges are covered by a neighbour's box and get no perch sites; a tower whose slots another edit took away keeps its `perches` key and draws nothing until a window opens; `circle_feet` still needs a memo; no rubble glasshouse and no pumpkin-patch kind; glasshouses in one chunk share one mesh. **Closed.** SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 150-176 (Castle R8d Arcade, fountain, gate toppers) 150: **Castle R8d Arcade, fountain, gate toppers (20 October 2026):** Opus 5.5 ran this on the Mac, 20 October about 09:40 to 11:10, and fix pass 1 from 11:20 to 12:15 (step `36-R8`, part d; `docs/castle-elements-plan.md` R8d). Snapshot `build/backup-feel/36-R8-d-pre/` (= HEAD a2d33b8); scratch pre-tree `d/pre-tree`; backups `d/backup/pre/` (.gdignore'd). Evidence `artifacts/decor/36-R8/d/` (`d/` below). Design `d/design.md`. Patch `d/part.diff` (2500 lines) covers `scripts/scene_state.gd`, `main.gd`, `world_view.gd`, `architecture.gd`, `editor_ui.gd`, `adapt.gd`, `composition.gd`, `creatures.gd`, `data/catalog.tres`, `tools/test.sh`, `tests/test_decor.gd`, `tests/test_bikes.gd`, `tests/test_grounds.gd` (card counts 75 to 77), the new `shaders/fountain_water.gdshader` and the new `tests/test_arcade.gd` (Godot also wrote their `.uid` files). Save format unchanged: version 2, with new optional keys only. Every edit is one undo step. Review: fresh Opus 5.5 review 20 October about 11:10 to 11:20: must-fix two (a ground stroke over 0.5 m left an arcade, fountain or glasshouse foot behind, buried or hanging; each new fountain look cost 2.2 to 3.2 ms to draw through `_lathe`, over the +2 ms release budget) and two smaller (fountain water missing from the shader warm-up; `arcade_seat` read a default walk's jetty edge). Fix pass 1 fixed all four, byte-for-byte geometry (121 looks hash identical), Harbour fountain placement +0.04 ms over a pot. Fresh recheck 12:11 to 12:27: verdict done, two notes (a Length or Arch row change does not re-seat a placed wall walk under a jetty; a cold Size-row undo costs +1.87 ms, warm +0.91 ms). Suite 60 / 32,546 / 0, every pin diff 0, verify clean with no stall. Owner look calls carried to the R8 close: the wall form's limits against freehand walls (back wall 1 to 1.6 m above) and jettied houses (0.25 m stand-off), the back wall covering a house's ground-floor doors and windows, a house dropped into a cloister lawn committing, the flat bright lawn slab, the hidden dusk lantern, a wall walk not re-seating on a Length or Arch change. Rulings in `artifacts/decor/owner-decisions.md`. Platforms: Mac run; Windows not run. 151: - **The kinds** (generic names, K8; tray 75 to 77): 152: - `arcade` (Architecture, Walls, after the stone circle). Rows Form (`arcade_form` "free" or "cloister"; absent: a walk against a wall), Length (`arcade_len` 4 to 20 m, step 0.5; absent 8), Arches (`arcade_arch` "pointed", R7c's two-centred arch; absent round) and Stone (the castle's STONES by `variant`). 2 to 8 arches a face (round(len / 2.5)); square 0.5 m piers with base and impost; voussoir rings laid round each soffit; the face 0.8 m over the crown, then a lean-to of the boathouse's slate (24 degrees, 0.4 m eave); a flagged floor on a plinth sunk 0.55 m. The wall form has a solid back; pointed at a house's side or a freehand wall it seats its back 0.02 m off that face's real outer plane (a house's `Openings.side_frame` plus the furthest jetty under the lean-to's high edge; a wall's nearest segment and thickness), facing out, on the ground (`DioramaState.arcade_seat`, at the hand only: once placed it is a ground part, so moving the house strands nothing). Free: arcaded both sides. Cloister: one part, four walks round a lawn of side `arcade_len`, back walls a pinwheel with a blind arcade and cornice outside, hipped lean-tos. One lantern (two in a cloister) of kind `arcade_lantern`, lamplighter's round, in `WorldView.BELOW_LAMPS`. No walkers, no nap site; solid to tower birds. 153: - `fountain` (Garden, Furniture, after the tree seat). Rows Tiers (`fountain_tiers` 1 to 3; absent 2), Size (`fountain_d` 2 to 5 m; absent 3), Finial (`fountain_finial` "spike"; absent orb) and Stone. Lathed basin and bowls with still water in each; the running water (curtains from each lip, a jet at the collar, ripples) is one node for every fountain (`WorldView.fountain_water`, one draw call, a memoised piece a fountain built from the sync's merged list) under `shaders/fountain_water.gdshader` on `light_time * motion_scale` (Ambient motion Off stills it; automated runs hold it; `--light-pin` moves it in captures). No light. On land or a crag's top, or a flat terrace; never over the canal. 154: - `gate_topper` on the R2a iron gate (row Topper on its card: None, Orb, Pinecone, Stag, Heron, Cat; absent none). Stone (TRIM) toppers on the caps of the posts the gate draws and of the posts a railing run stands for it (the run's cut carries `topper` only when set); a freehand wall's piers keep their coping (said in the row's help). Scaled with the post (1.0 low to 1.35 tall). A topped gate's caps are no nap site. 155: - Refusals with their own hints (`Adapt.HINTS`): arcade water, slope, canvas, ground; fountain water, slope; soft for an edit that did not touch the part, refused for the part the edit changed (`R8D_SOFT`). Settled arcades and fountains draw in R8c's shared chunk meshes (`MERGED_KINDS`). 156: - **Tests.** `tests/test_arcade.gd` (after test_grounds): tests first 22 checks, 21 failing (the gate pins passing, `d/testfirst/`); now 302/0 (`d/dev/test_arcade-4.log.out`). Three fixture fixes in my own new suite before code ran: the cloister and back-wall probes use segment hits, not vertex counts (vertex counts passed vacuously); the stress scene's sixth arcade stood over the water; the seat check reads the jetty under the walk's high edge. 157: - **Unchanged without the new parts** (`d/pins/summary.txt`): R7a 37, R7c 66 and bridge 84 pins diff 0; g0 ident 15/15 diff 0; R8b's 54 boat pins and 959 creature-plan pins identical; the R2a gate pins (`d/probe/gate_pins.gd`: 720 gate looks by width, height, top, ajar, paint and posts, the run, end, wall and fence takeovers and the nap sites; 743 lines) identical pre and live, 216 of them also held in test_arcade. Starters' md5 pins pass in the suites. 158: - **Captures and sheets** (`d/cap/pre-1`, `after`, `pre-2`, `after-2`; `d/metrics/`): `d/sheet/r8d-arcades.png`, `r8d-cloisters.png`, `r8d-fountains.png` (the water two frames apart at 12 s and 12.4 s), `r8d-toppers.png`, `r8d-crag3.png` (all three in R8a's crag scene, day and dusk) and `r8d-existing.png`. Existing views (Castle starter, Harbour, quarter, Amsterdam, R4c courtyard, R2a gates; day and night): integer counts identical pre and live; changes within the noise pairs (gates-r2a-night pre-1 vs after 10.2% against noise 8.6% and 5.6%, pre-2 vs after-2 1.9%). A look pass before the sitting (`d/dev/`): the first arches read flat under the eave, so the face was raised to 0.8 m over the crown and the rings laid round the soffit; the cloister's outer faces gained the blind arcade; the lawn darkened ("plain" renders bright). 159: - **Cost** (AC power throughout): 160: - Release (`d/cost/release-summary.txt`, 3+3 interleaved, medians of 7): Harbour pot placement 26.60 pre, 27.17 live (+0.57 ms); Harbour adapt_all 7.7 vs 7.8 ms. Stress (6 runs, 2 cloisters, 6 fountains, 8 topped gates; the pre-tree reads its 8 plain gates only, so not like for like): adapt_all 2.9 ms against 0.27 (about 0.2 ms a part); arcade placement 13.3, fountain 16.1, pot 11.7, topper row 15.9 ms. Traced: the fountain's +4.4 ms over a pot is the decorations phase (6.2 against 1.9 ms), building its look's mesh once when it starts growing in. 161: - Frame (`d/bench/bench.txt`, dusk 20:45, vsync off, 3 runs each): stress 529 draw calls, p50 4.52 ms, p95 5.37 ms against the 8 plain gates' 415, 3.70 ms, 4.52 ms; 10 arcade lanterns are lit there (plain: none). 162: - Harbour drag (`d/drag/summary.txt`, A0b script, 6+6): handle step p95 59.29 live, 60.11 pre (baseline 53.7 ms; both trees about 5 ms slower today); handle release 13.46 live, 13.61 pre. The slider release is two-moded on both trees (about 40 or 59 ms): 59.04 live, 59.23 pre. 163: - **Full suite** (`d/suite/suite_all.out`): a, b and v at 2880x1800: 60 suites, 32464 checks, 0 failures; --verify passed on try 1 (no stall in the roof checks this time); test_creature_reactions rerun 6871/0. 164: - **Older builds** refuse a scene with an arcade or a fountain ("Unknown or damaged building part."); a gate's `gate_topper` loads there and is ignored. 165: - **Fix pass 1** (11:20 to 12:15, after the fresh review in `d/review/`; evidence `d/fix1/`, starting tree snapshotted in `d/fix1/backup/` (.gdignore'd), its diff `d/fix1/fix.diff` (465 lines) over `scripts/main.gd`, `architecture.gd`, `mesh_batch.gd`, `world_view.gd`, `shader_warmup.gd`, `scene_state.gd`, `tests/test_arcade.gd`, `tests/test_warmup.gd`). What the review found and what changed: 166: - *Must-fix 1, a stroke left the foot behind* (`d/review/probe-follow.log.out`): a raise or lower of more than 0.5 m over an arcade's, a fountain's or R8c's glasshouse's whole site left its foot buried 1 m or its plinth hanging, with the slope line kept for good; a house in the same stroke followed. Now `Main.adapt_all`'s R8C/R8D soft branch takes solve's ground candidate when one holds (glasshouse, arcade, fountain only); where none holds the part stays soft as before. New test_arcade block `follow`: at +1.0, -1.0, +0.4 and -0.4 m the arcade (wall, free pointed, cloister), the fountain and the glasshouse stand with their feet on the ground and nothing else changed, the stroke is one step and one undo (with a later edit undone first) gives back the scene byte for byte, the save round trip is exact, a later edit commits, and a stroke beside them leaves every part byte for byte. The stone circle and the pumpkin already followed before the fix (start tree, `d/fix1/testfirst/test_arcade.log.out`: feet on the ground at every delta) and are not changed; a house keeps its own half metre (stays at ±0.4). 167: - *Must-fix 2, the fountain's lathe cost*: `Architecture._lathe` (and `_water_disc`) now build each solid's faces once into a shape keyed by its profile, sides and centre (`_cached_shape`, cleared whole at 512 entries; a keyed batch still draws face by face) and lay them with the new `MeshBatch.append_faces` (packed appends; the turned normals kept per shape by basis, at most 8). Output byte for byte: 111 looks (every tiers, size, finial and a stone of the fountain; the gate with each topper, low and tall; a plain gate), each drawn under the identity and under a merged piece's turned frame, hash the vertices, normals, colours, UV and UV2 identically before and after (`d/fix1/hash/before.log.out`, `after.log.out`, `diff.txt` empty; all `22816b79…`). Draw of a fountain look: 2.2 to 3.3 ms before; 0.58 to 0.91 ms cold (the first look of a session), 0.05 to 0.08 ms once its shapes are held. Release, Harbour (`d/fix1/release.gd`, `release.sh`; 3 rounds interleaved pre-tree, start tree (the live tree before this pass, `d/fix1/start-tree`), live cold (shapes emptied before every change) and live warm; 9 rounds of pot, fountain placement, Tiers row 2 to 3 and Size row 3 to 4.5 m interleaved inside each run, medians of the difference from the same round's pot; `d/fix1/cost/release-summary.txt`): fountain placement +1.80 ms start, +0.04 cold, -0.77 warm; Tiers row +1.68, -0.82, -1.38 and its undo +3.58, +0.73, +0.63; Size row +1.85, -0.50, -0.62 and its undo +3.88, +1.39, +0.60. The sync's decorations phase on a fountain placement 11.5 to 11.7 ms start, 9.3 to 9.4 cold, 8.4 warm, against a pot's 8.3 to 8.9. All within the +2 ms budget. This probe keeps the scene's fountain (the rows need it), so its pot is not the first sitting's pot: pot placement 27.48 ms pre (harbour-pre.json, no R8d parts) against 29.98 start and 29.6 to 30.1 live (the same on the start tree, so it is the scene's extra parts, not this pass). The first probe design (`d/fix1/cost/v1/`) put the fountain back with a sync before every round and so timed the caches that sync drops; it was replaced. Look-cache keeping (the review's other remedy) was not needed and not done. 168: - *Finding 3*: the fountain water material is in the shader warm-up (`WorldView.fountain_material_of`, made on first use by the first fountain or the warm-up, as P10m's beam); test_warmup checks it is warmed on a mesh before any fountain exists and is the one shared `fountain_water` material (55/0; the start tree fails on the missing function, `d/fix1/testfirst/test_warmup.log.out`). 169: - *Finding 4*: `DioramaState.arcade_seat` reads the jetty at the high edge of the walk being placed (its own length and arch; `Main` passes the candidate), not a default walk's. New test_arcade block `seat_walk`: three jettied medieval houses, every side, four walks (pointed, 4 m, 20 m pointed, default): every back 0.02 m off the furthest the face juts under its own high edge, and the check asserts that some cases differ from the default walk's reading. On the start tree the 4 m walk (its high edge under the jetty the default walk reaches) stood 0.27 m off, 0.25 m too far out, on all 12 house sides (`d/fix1/testfirst/test_arcade.log.out`). 170: - Tests first on the start tree: test_arcade 382 checks, 11 failures (the follow and seat_walk blocks). Now 18 affected suites 0 failures (`d/fix1/suites/summary.txt`: test_arcade 382, test_grounds 201, test_warmup 55, test_adapt 955, test_decor 2717, test_scene 1736, test_building_tools 982, test_starters 397, test_ui 590, test_composition 160, test_creatures 250, test_creature_reactions 6871, test_dusk 461, test_roofs 262, test_crag 200, test_boathouse 200, test_castle_hardening 80, test_build_anim 539). 171: - Pins re-run (`d/fix1/pins/summary.txt`, the review's kit): R7a 37, R7c 66, bridge 84, ident 15, boats 54, creature 959, gates 743, all diff 0 pre against live. 172: - Full suite (`d/fix1/suite/suite_all.out`): a, b and v at 2880x1800: 60 suites, 32546 checks (32464 plus 80 in test_arcade and 2 in test_warmup), 0 failures; --verify passed on try 1 (no stall); test_creature_reactions rerun 6871/0. AC power for every sitting (`d/fix1/cost/power.txt`, `d/fix1/pins/power.txt`). 173: - No save-format change; no existing part's look changed (the hashes and pins above); no captures re-shot (the geometry is byte for byte). 174: - **Owner look calls** (orchestrator rulings, carried to the R8 close; not fixed): the wall form against a freehand wall (its own back wall stands 1 to 1.6 m above the host), against a jettied house (0.25 m stand-off under the jetty), and its solid back wall covering a house's ground-floor doors and windows; a house dropped into a cloister's lawn commits, as the glasshouse allows; the flat bright lawn slab; the dusk lantern hidden under the lean-to. 175: - **Not done / open:** Windows not run. The topper close views are framed loosely (the capture camera's height offset was found by trial). The look calls above. SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 120-149 (Castle R8e Banners, torches, braziers) 120: **Castle R8e Banners, torches, braziers (20 October 2026):** Opus 5.5 ran this on the Mac, 20 October about 12:41 to 14:20 (step `36-R8`, part e; `docs/castle-elements-plan.md` R8e). Snapshot `build/backup-feel/36-R8-e-pre/` (= HEAD f8a26e2); scratch pre-tree `e/pre-tree`; backups `e/backup/pre/` (.gdignore'd). Evidence `artifacts/decor/36-R8/e/` (`e/` below). Design `e/design.md`. Patch `e/part.diff` (2052 lines) covers `scripts/scene_state.gd`, `main.gd`, `world_view.gd`, `architecture.gd`, `fire.gd`, `openings.gd`, `editor_ui.gd`, `adapt.gd`, `composition.gd`, `creatures.gd`, `ivy.gd`, `build_audio.gd`, `data/catalog.tres`, `tools/test.sh`, `tests/test_decor.gd`, `test_bikes.gd`, `test_grounds.gd`, `test_arcade.gd`, `test_ui.gd` (card counts) and the new `tests/test_banners.gd` (Godot wrote its `.uid`). Save format unchanged: version 2, new optional keys only. Every edit is one undo step. The modified `artifacts/building-tools/tests.json` (the suites rewrite it) and the other session's untracked `docs/codebase-audit-plan.md` are not this part's. Review: fresh review 20 October about 14:20 to 14:45 (`e/review/`: `walk/`, `edge/`, probes): must-fix two (inside a placement pass `hang_error` walked every object once for each part on a tower through `tower_wraps`; bartizans were in no clash list with banners or torches) and a wrong figure in this entry (finding 3); fix pass 1 below, 14:47 to 15:30; fresh recheck 15:28 to 15:42 (`e/review2/`): verdict done, two notes (some host reshapes of a scene already holding a bartizan and a banner or torch are refused with the bartizan's line until the part is slid away, matching R7b's crowd rule; the bartizan's square clash box is conservative, refusing banners up to about 1.0 m from the corner). Suite 61 / 32,881 / 0, every pin diff 0, verify clean with no stall. Rulings in `artifacts/decor/owner-decisions.md`: banner and torch release cost (the host's cluster rebuild, 78 ms against 27.6 for a pot on the Harbour, 265 to 272 ms on the Castle starter) accepted as the clock face's class and carried to the R8 close as an owner item; a brazier under a lifted or moved house, the 1 m banner reading as a tile, the torch lean subtle at default zoom and the conservative bartizan box carried as limits and look calls. Platforms: Mac run; Windows not run. 121: - **Fix pass 1** (20 October about 14:47 to 15:30; evidence `e/fix1/`; starting tree snapshot `e/fix1/backup/` (.gdignore'd); patch `e/fix1/fix1.diff`, 501 lines, `scripts/scene_state.gd`, `adapt.gd`, `main.gd`, `tests/test_banners.gd`): 122: - *Must-fix 1 (the per-part walk).* `_pass_walk` (the one walk a pass that builds the kids index) now also lists every bridge and stair under each of its ends' parents (`_stair_index`, valid only while the kids index is), and inside a pass `tower_wraps` calls `stair_wrap` on that tower's own connections only: the same helices in the same order, the same test. Outside a pass it walks as before, counted in `tower_wrap_walks`. The clock face's stair test (`clock_clash` through `Openings.tower_stair_band`) and a tower's drawing in a cap pass share the call, so they take it too, with the same answers (clock behaviour and pins unchanged). Measured with the review's kit plus a walk counter (`e/fix1/walk.sh`, `e/fix1/probe/walkcost.gd`, `e/fix1/walk/summary-before.txt`, `summary-after.txt`, same sitting): the parts' share of one pass on the Harbour went from 0.63 ms (154 objects) and 3.55 ms (1,604) to 0.30 and 0.30 ms: flat; `tower_wraps` full walks in a pass 0 on every scene. On the stress scene grown to 1,482 objects the share is 0.46 to 0.95 ms; traced by kind with the pass's index built first (`e/fix1/probe/perkind.gd`, `e/fix1/walk/perkind-after.log.out`): banners and torches flat (123 and 55 µs on houses, 35 and 24 on towers, against 116, 56, 35, 26 at 32 objects); the growth is the pass's one index walk (17 to 515 µs, which the first part to ask was charged with) and the six braziers (237 to 889 µs), which grow as six pots asked the same do (69 to 572 µs): the generic ground-part check every prop pays, not this part's code. Carried as a note. 123: - *Must-fix 2 (bartizans).* `hang_clash` takes a bartizan (`HANG_BOX_KINDS`) as a box test in the host's frame: the banner's or torch's own box turned onto its face, the bartizan's square box round its axis from its corbel's foot to its finial (round, so any turn fits inside it), `HANG_CLEAR` apart. The bartizan's own check asks it after `crowded()`, so whichever is placed second is refused, with its own line `HANG_BARTIZAN_ERROR` ("A bartizan and a banner or a torch keep clear of each other.") and hint (in `Adapt.HINTS`); Adapt's reseat of a bartizan onto a banner is refused with that line, not the seat's (`hung`, as R7b's `crowded`). A scene already holding both loads (`all_supported` empty) and takes later edits: the banner or torch draws nothing (soft, in `HANG_SOFT`/`R8E_SOFT`), the bartizan stands as it is (soft for it, never listed as levitating, so no shimmer); `Main.adapt_all` refuses only an edit that placed either, changed its rows or moved it by hand (`_hang_own`). A reshape of their house that reseats the bartizan's corner onto the banner is refused with the bartizan's line (w 8 to 9 or 7, d 6 to 7; height 10 to 11 holds: `e/fix1/edge/reshape.log.out`), R7b's rule for an edit that brings two together; a move of the house is not. The review's edge probe re-run (`e/fix1/edge/edge.log.out`): a banner at a bartizan's corner is now refused and the bartizan asked names it; the torch at 3.2 m under a top bartizan stands (their boxes do not meet), as before; the brazier, pot and fountain lines unchanged. 124: - *Tests first* (`e/fix1/testfirst/test_banners.log.out`): 37 new checks (258 to 295) in `tests/test_banners.gd` (`r8e_fix1_wrap_checks`: a wound stair round a tower with banners, torches and a clock, the pass's helices and answers equal to each asked alone, no walk in a pass, one counted outside; the stress pass counts no `tower_wraps` walk; `r8e_fix1_bartizan_checks`: both ways on a house and a tower, torches under a storey-down bartizan, the hint, the hand's refusal through `Adapt.solve`, the soft state with an edit elsewhere, a house move, a reshape, the banner's own row, a new bartizan at commit); 18 failed before the code, 295/0 after. Affected suites (`e/fix1/suites/summary.txt`): 20 suites, 0 failures, 0 script errors. 125: - *Pins* (`e/fix1/pins/summary.txt`, `e/fix1/hash/summary.txt`): R7a 37, R7c 66, bridge 84, ident 15, boats 54, creature 959, R2a gates 743, fire 23, R8d's 111 looks: diff 0 pre against live; starters' md5 in test_starters (pass). 126: - *Finding 3 (no code; the figure corrected below).* Like-for-like control (`e/fix1/control.sh`, `e/fix1/bench/control.out`, 3 rounds, order turned each round, AC power): a pot on the stress scene with the 26 parts 15.36 / 15.49 / 15.47 ms; the same pot on the same scene with the 26 parts removed (live tree) 10.97 / 11.05 / 11.11; the pre-part tree 10.89 / 10.91 / 10.97. So the live code costs a pot nothing without the parts (within 0.1 ms of pre), and the 26 parts' presence costs it +4.4 ms: about +1.8 ms of it in the sync's phases (decorations 2.45 to 3.75, clusters 0.55 to 0.76), the rest in the commit's whole-scene pass over 26 more parts. Not reduced. 127: - *Full suite* (`e/fix1/suite/suite_all.out`): a, b and v at 2880x1800: 61 suites, 32,881 checks, 0 failures; --verify passed on try 1 (no stall); test_creature_reactions rerun 6871/0. AC power throughout (`e/fix1/*/power.txt`, `e/fix1/suite-power.txt`). 128: - *Rulings (orchestrator, 20 October about 14:55, reversible by the owner; no code):* the banner and torch release cost (the host's cluster rebuild: 78 ms against 27.6 for a pot on the Harbour; 265 to 272 ms on the Castle starter's 13-member cluster, a clock face at the same spot 260 to 270 ms) is accepted, the same class as R7d's clock faces; the size of the hitch goes to the owner at the R8 close (a possible later part: draw banners and torches apart from the host's mesh). A brazier standing under a house (a house lifted 3 m, or moved onto it) is a limit carried to the R8 close, like the pot and the fountain. The 1 m banner reads as a tile and the torch lean is subtle at default zoom: look calls for the owner. 129: - **The kinds** (generic names, K8; no crest or device; tray 77 to 80, Details 44 to 47; cards, not new tray groups): 130: - `banner` (Details, Signs, wall). Rows Length (`banner_len` 1 to 6 m, step 0.5; absent 3), Field and Trim (`banner_field`, `banner_trim`: wine, navy, forest, ochre, slate, cream, black; absent wine and cream; any pair, equal allowed). An iron rod (R2's SIGN_IRON) 0.1 m off the wall on two brackets with ball ends; the cloth `clamp(0.6 + 0.12 L, 0.7, 1.3)` wide with a border and a band under the rod in the trim; **swallowtail hem** (chosen over straight). Cloth through `MeshBatch.cloth_box`/`cloth_face` in CLOTH_HANG, hoisted at the rod: the flags' and the stall valance's wind, no spring lines, still under Motion Off, half rate under Reduced. 131: - `torch` (Details, Lighting, wall). R8b's `Fire.torch` exactly; row Lean (`torch_lean` left or right, 18 degrees about the wall's normal; absent upright). One height: its plate at 3.2 m over the wall's foot. 132: - `brazier` (Details, Lighting, surface). Row Size (`brazier_size` small or large; absent medium): bowl 0.5 / 0.7 / 1.0 m across, rim 0.75 / 0.95 / 1.15 m, three splayed legs with rails, coals, in R2's iron. On land (a quay and a crag's top count) with the ground under its ring within 0.4 m of its foot, or on a flat terrace; never over water. 133: - Hosts: a house standing alone (any side, on the real outer plane of the storey its rod or plate is in: `Openings.side_frame` plus `Architecture.face_jetty`) and a tower: round (where pointed), octagonal (a face's middle, on its plane) or D-plan (the round front only, the whole cloth clear of the flanks). Not hosts: a joined house, a gable or roof, a freehand wall or curtain connection (a path stroke opens gaps in them, which would strand a part), the boathouse. Seat `DioramaState.hang_seat`: the cloth's tails stay 2.9 m over the foot (over every door and path arch), the rod 0.25 m under the wall head, 0.12 m in from the corners. Refusals with their own hints (`Adapt.HINTS`, seven lines). 134: - Clashes asked both ways (`hang_clash`, from the pass's kids index): placed windows, roses, clocks, gargoyles, the wall details and other banners and torches on the same face (a tower: within a quarter turn), 0.15 m apart, and a tower's helix stair; since fix pass 1 also a bartizan on the same house or tower, box to box (above). Derived pieces give way: a house front's composed windows and trim under them (`wall_roses`), a tower's rings, strips and generated windows (`Openings.tower_clock_band`), derived pipes and window boxes (`dressing_mounts`), ivy (`Ivy.BLOCKERS`), buttress piers (`BUTTRESS_FACE_KINDS`); roof clutter keeps off a terrace brazier (it is a placed part on the deck); ground clutter and the walkers keep off a brazier's ring (`Composition`, `Creatures._blocked`). No walker or nap site on any of them; the cat is not drawn to a brazier (its cells are blocked, as a fountain's). 135: - Soft states: a banner or torch whose host is moved or turned rides with it (a wall part's `transform_of`); resized, restyled or joined, Adapt's new "hang" move reseats it on its own face; where no seat holds it stays in the save and draws nothing (`DioramaState.hang_soft_ids`), carves nothing, and is refused only for an edit that placed it, moved it by hand or changed its rows (`Main._hang_own`). A brazier's water and slope lines are soft for an edit elsewhere and follow the ground candidate (R8d's branch). A deleted host takes its parts. 136: - Fire: one flame. `Fire.flame` (two warm cones, crossed pool cards, one light) is factored out of `Fire.torch`; the torch calls it with its own sizes and the brazier scaled to its bowl. R7e's viaduct fire bowls are not re-routed. Both lit in the lamplighter's round (fixture order 1, each part its own turn by its frame); light kinds `torch` and `brazier`, already in `WorldView.BELOW_LAMPS`. Flicker is P8b's night-life breathe of the pool cards and light (seeded per part, still under Off); **no embers**. All materials (warm, pool, metal, painted_metal, wood, cloth, plain) were already in the warm-up. 137: - Drawing: settled parts join R8c's chunk meshes (`MERGED_KINDS`); looks by length and colours, lean, size. A host's cluster signature carries its banners and torches (kind, place, turn, length; not colours or lean), so a placement or a Length row rebuilds the host once and a Colour or Lean row does not. 138: - **Tests.** `tests/test_banners.gd` (after test_arcade): tests first 20 checks, 20 failing (`e/testfirst/`); now 258/0 (`e/suite/suites/test_banners.log.out`). Fixture fixes in my own new suite: windows and the clock were made at world points (make_object reads a child's point as world), the 6 m case moved off the 9 m house (it has no room there, as another check says), the quay-edge brazier moved 0.65 m out. A house side turn is stored to 1e-4 (as a tower's) so a short-precision JSON save reads back exactly. Affected suites (`e/affected/summary.txt`, 29 suites, 0 failures after two fixes: a branch I had inserted in the cluster signature had swallowed R7d1's buttress-run line, and test_ui's Details count). 139: - **Unchanged without the new parts** (`e/pins/summary.txt`, final code): R7a 37, R7c 66, bridge 84 diff 0; ident 15/15; boats 54, creature 959 and R2a gates 743 identical pre and live; starters' md5 pins pass in test_starters. Fire hash (`e/probe/firehash.gd`, `e/hash/summary.txt`): R8b's torch alone (3 frames x 2 transforms) and every boathouse look (8 x 2), groups and emitters, byte for byte; R8d's 111 looks (`d/fix1/meshhash.gd`) identical. 140: - **Captures and sheets** (`e/cap/pre-1`, `after`, `pre-2`, `after-2`; `e/metrics/`; AC): `e/sheet/r8e-banners.png` (1, 3, 6 m on a 12 m house front and a 14 m round tower, day and dusk, each in a gust and a lull), `r8e-torches.png` (door torches day and dusk, the boathouse's own torches, crag house and tower), `r8e-braziers.png` (three sizes on a terrace and the ground, dusk close; on the crag top), `r8e-crag.png` (all three in R8a's crag scene with R8b's boathouse, day and dusk), `r8e-existing.png`. Existing views (Castle starter, Harbour, quarter, Amsterdam, courtyard; day and night): integer counts identical; change within the noise pairs. The cloth moves: `e/metrics/cloth-pair.txt`, `e/sheet/r8e-cloth-house.png`, `r8e-cloth-tower.png` (the cloth rect changes 7.8 to 18.9% between a flap step and a lull, a plain-wall rect beside it 0%); each banner flaps on its own steps, so the tower's 6 m banner is shown at 18.25 s (it rests at 20.5 s; `e/dev/flap-tower/sweep.txt`). The banner views in `cap/after` were re-shot at 20.5 s (19 s caught the house's flap at rest; `cap/after/REPLACED-BANNER-VIEWS.txt`). 141: - **Cost** (AC power throughout, Low Power Mode 0; `e/bench/power.txt`, `e/cap/power.txt`, `e/drag/power.txt`): 142: - Release (`e/release.sh`, `e/probe/release.gd`, 3+3 interleaved, medians of 9; `e/bench/release.out`, `release-summary.txt`). Harbour with 10 torches and 6 braziers (pre-tree reads it without them): pot 27.58 pre, 29.79 live; brazier placement 28.35 (**met**, -1.0 to -2.2 against the round's pot); banner **78.30**, torch **76.60** (over the +2 ms budget against a pot). Traced: the extra is the host's cluster rebuild (clusters phase 53 to 54 ms of it), the same a clock face pays placed at the same tower spot today: 77.55 pre, 80.34 live (split control in both trees). The +2.2 ms on the live pot is the 16 parts in the scene: on the plain Harbour live and pre are equal (`e/dev/split2/summary.txt`: 27.45/26.08 against 27.45/27.98). Stress (6 hosts, 10 banners, 10 torches, 6 braziers; pre reads the hosts alone): pot 11.10 pre, 15.64 live; brazier 14.06; banner 45.78, torch 47.16 (their host is a 10 m house, rebuilt: clusters 33 to 35 ms); clock on a tower 33.92 pre, 34.78 live. Traced (`e/dev/split/trace*.log.out`): before a keyed memo the stress scene's parts cost +2.8 ms in adapt_all and +2.6 ms in the sync per commit; `hang_error` now answers from a memo keyed by the part, its host, the host's parts, its join and the tower's helices (`_hang_memo`, bounded 2048), and the soft gate's tower stairs come from one walk a pass (`_wraps_walk`): now +0.9 and +1.1 ms (soft gate 1.76 to 0.25 ms). 143: - Lights (`e/probe/lights.gd`, `e/bench/lights.log.out`): Harbour with 10 torches and 6 braziers at five camera targets: 16 of 16 shared lights given in both scenes, 0 to fires, **0 town lamps lose their light**; the fires' 16 sources are ranked below every lamp. Seven tower window sources give way to the torches (122 to 115). 144: - Frame at dusk (`e/bench/bench.txt`, 20:45, vsync off, creatures live, 3 runs each): stress 748 draw calls, p50 4.96 ms, p95 5.47; banners only (hosts plus 10 banners) 645, 4.69, 4.91; hosts alone 621, 4.68, 5.09. Banner cloth cost: about 0 ms a frame (24 draw calls). 16 lights lit in all three, none starved. 145: - Harbour drag (`e/drag/summary.txt`, `handle-steps.txt`, A0b script, cold, 6+6): handle kind p95 61.11 live against 55.84 pre (baseline 53.7; R8d measured about 59 to 60 on both). Not a step cost: handler p95 and step medians match on both trees, and the handle-drag-height step's own p95 is 62.5 live against 61.8 pre (medians of runs); the kind's p95 moves with its after-frame tail, which is two-moded (15 or 41 to 52 ms) and landed low in 2 of 6 pre runs and 0 of 6 live runs. Releases within ±0.5 ms. A watch item. 146: - **Full suite** (`e/suite/suite_all.out`): a, b and v at 2880x1800: 61 suites, 32841 checks, 0 failures; --verify passed on try 1 (no stall); test_creature_reactions rerun 6871/0. 147: - **Older builds** refuse a scene with a banner, a torch or a brazier ("Unknown or damaged building part.", the load check of a kind it does not know; not run on an older build). 148: - **Not done / open:** Windows not run. Banner and torch placement misses the +2 ms release budget by the host rebuild every fitting pays (above); not reduced (a rebuild only when something actually gives way would need the derived layouts); accepted by the orchestrator's ruling (fix pass 1, above). A pot placed on the stress scene costs **+4.5 ms** with the 26 parts present (15.64 against 11.10 ms on the pre-tree in the first bench, decorations phase +1.2 ms; corrected by fix pass 1 from "+2 ms", which was wrong); the like-for-like control (fix pass 1) gives +4.4 ms against the same scene without them on the live tree, which equals the pre-tree. A whole-scene pass on a scene grown to 1,482 objects pays the braziers' generic ground check as it pays a pot's (fix pass 1). The Harbour handle kind p95 (above). Embers not drawn. The wall-torch Lean is subtle at the default zoom. A ground brazier, as a fountain, is not refused for standing under a house's walls when placed programmatically (the hand cannot point there). SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 108-119 (Castle R8f Loch castle starter) 108: **Castle R8f Loch castle starter (20 October 2026):** Opus 5.5 ran this on the Mac, 20 October about 15:43 to 16:40 (step `36-R8`, part f, the starter half of `docs/castle-elements-plan.md` R8f; the stage close and the M-R4 exports are a later job and were not done). Snapshot `build/backup-feel/36-R8-f-pre/` (= HEAD 1e9a79f); scratch pre-tree `f/pre-tree`; backups `f/backup/` (.gdignore'd). Evidence `artifacts/decor/36-R8/f/` (`f/` below). Design `f/design.md`; dev prototype and look passes `f/dev/` (`proto.gd`, `look1` to `look4`). Patch `f/part.diff` (514 lines): the new `scripts/loch_scene.gd` (Godot wrote its `.uid`), `scripts/main.gd`, `scripts/editor_ui.gd`, `tests/test_starters.gd`, `tests/test_ui.gd`. No new kind, key, rule or game behaviour; save format unchanged (version 2). The modified `artifacts/building-tools/tests.json` (the suites rewrite it) and the untracked `docs/codebase-audit-plan.md` are not this part's. Review: fresh Opus 5.5 review 20 October 16:40 to 17:05 (`f/review/`): must-fix one (the `--drag` fixture `loch` was missing from `DragHarness.FIXTURES`, so the parser refused every script naming it; fixed by the orchestrator at the close with a test, test_drag_harness 125/0) and a wording fix (`--verify`, not `--smoke`, runs the integration check). Cues judged by the reviewer on its own crop: 1, 2, 4, 5, 6 read; 3 (the hall reads, the buttresses only close up, the roof not steep) and 7 (warm lights read, the sky is pink-violet) in part. Existing views unchanged (the 19.5% and 12.2% night figures were a noisy capture; pre-4 against after-4 2.9% and 7.3%). Tower release +6 to +8 ms on this scene accepted as the 54-part scene's own cost (`adapt_all` flat per part; the Loch's own edits 21-56 ms against 72 on the Harbour and 212-282 on Four turrets), carried to the owner. Owner items for the R8 close: the crag's Height row is refused up and down on this starter (both bridges run from crag-top terraces to off-crag terraces) and a viaduct or covered bridge cannot land on a crag top; the cottage chimney reads as a tall black pipe in front of the owl tower; the covered bridge's trestles read as iron; the pink-violet dusk sky; buttress readability; LOCH_MD5 is stable on three builds but not tested on Windows. Suite 61 / 32,947 / 0, every pin diff 0. Platforms: Mac run; Windows not run. 109: - **The starter.** "Loch castle" (generic name, K8; no crest, no house colours; banners navy and forest), in the scene menu right after "Castle" (`EditorUI.MENU_ACTIONS` gains `loch`; the separator before Help moves from index 8 to 9). `scripts/loch_scene.gd` (`extends RefCounted`, `static func build()`, preloaded as `LochScene` in `main.gd`, the Amsterdam and quarter idiom; `sample_scene.gd` untouched). Main opens it as the other starters (one undo step, "open loch castle", `scene_name` "Loch castle"); `--drag` fixture `loch` (added to `DragHarness.FIXTURES` at the close; accepted by the parser in test_drag_harness, 125 checks; a windowed `--drag` run on it was not made); `--verify` (its integration checks; not `--smoke`) opens it after the Castle and checks it is supported, whole and named. It opens at dusk (Night, 20:45, dry) with the starters' fog and smoke. Every part is seated by the game's own seat functions at build time (`buttress_seat`, `rose_seat`, `belfry_seat`, `hang_seat`, `roof_seat_point`, `Magnets.boathouse`); build 49 ms. 110: - **What it holds** (54 parts, pinned per kind in `LOCH_COUNTS`): one 8 m crag (16-point outline, about 29 x 22 m); the great tower (flat, 13 m, machicolations) with a 3 m spire tower on its terrace: the walk and parapet round the crown are the viewing ring, its vane at 30.5 m, 4.8 m over the next top; a clump of four spire towers (11, 9.5, 8.5, 6.5 m); the hall (canal, 12 x 7 x 8 m, eight stained double lights, round in the middle of the south front, a buttress run each side, a rose on the east gable, a belfry); an octagonal ogee tower, a ruined tower, two merloned freehand walls, a fountain between two braziers, two banners; a viaduct (4 arches by the shipped count, 23 m, deck 10.6 m over the water, fire bowls, sweep 0.08) from a one-storey gatehouse's terrace to a mainland gate tower; a covered bridge (14 m, deck 13.2 m over the water) from a north-east bridge tower to the islet's owl tower (`perches`); on the islet a stone circle (7 stones, sundial), the round cottage (3.2 m medieval tower, 2.6 m, steep `spire` cone, a chimney where the hand seats it: the cone's apex) and four pumpkins; a stone boathouse at the crag's south foot where its magnet puts it, a cliff stair from the waterline shelf, four lantern rowing boats; a glasshouse and five conifers on the mainland. Not used: arcade/cloister (no room on the top without crowding), gate toppers (no gate), torches (the gatehouse is one storey; the boathouse brings its own). 111: - **One finding, no code.** A viaduct or a covered bridge cannot land on a crag's top: its deck meets the rock, and the shipped code refuses it (test_crag `r8a_ride_style_checks` already says so). Both long bridges therefore land on terraces at each end (gatehouse and mainland gate tower level at 10.75; bridge tower and owl tower level at 13.35), which the shipped code excuses. Not a stop: the shipped kinds do what the scene needs this way. 112: - **Tests.** Tests first (`f/testfirst/`): test_starters 412 checks, 3 failing (no starter script, the menu twice); test_ui 590, 1 failing (menu order). Now test_starters 463/0: new block r8f (`LOCH_MD5` f10d62aad985d2910866afe9bd5b4114, `LOCH_COUNTS`, the twelve other starter documents' md5 `STARTER_MD5S` read on the pre-tree, `f/pins/starter-md5-pre.txt`; validates as version 2; JSON round trip byte for byte; builds the same twice; every part's `placement_error` empty (not even soft), `Adapt.solve` adapts and refuses nothing, `all_supported`, `levitating()`, `hang_soft_ids()` empty; the pieces and cue 1 measured; creatures: no walker or cat cell and no nap site on the crag, the owl tower's birds come, fly 30 s clear of every tower, land on its ledges, nothing walks, sits or naps on the crag; in the editor: the menu, one undo step, its camera and dusk, supported, a delete of one part of each of its 20 kinds one step and undone byte for byte, Size row on a pumpkin, Lantern row on a boat and Height row on the great tower one step each and undone byte for byte, undo back to the previous scene); the atmosphere checks now include it. **Changed expectation:** test_ui's menu order gains `loch` after `castle`. Affected suites (`f/suites/summary.txt`): 23 suites, 0 failures, 0 script errors. 113: - **Unchanged** (`f/pins/summary.txt`, `f/hash/summary.txt`, pre-tree against live): R7a 37, R7c 66, bridge 84, ident 15, boats 54, creature 959, gates 743, fire 23, R8d's 111 looks: diff 0; every existing starter's md5 (`f/pins/starter-md5-diff.txt`): the only difference is the added `loch` line. 114: - **Captures and sheets** (`f/caps.sh`; `f/cap/pre-1`, `after`, `pre-2`, `after-2`; `f/metrics.sh`; AC power, `f/cap/power.txt`). `f/sheet/r8f-01-cue-sheet.png` (dusk 20:45 from three cameras, cues marked; laid out as R7's `close/cues`), `r8f-02-cues-day-verdicts.png`, `r8f-03-loch-day.png`, `r8f-04-loch-close.png` (hall front, the ring, viaduct, boathouse and stair, islet day and dusk), `r8f-05-existing.png`. The low and high cue views were re-shot once for framing (`f/cap/after/RESHOT-LOW-VIEWS.txt`). **Cue by cue, dusk, low across the water** (`f/cap/after/cue-low-dusk`): 1 reads (the crown and vane clear over every roof, the ring visible); 2 reads (four spires in a clump); 3 reads as a hall among towers (stained lights, the rose lit, the belfry), its buttress piers only close up; 4 reads (the crag from the loch, lights mirrored); 5 reads for both (the viaduct's four tall arches, bowls lit; the covered bridge long and thin on high trestles); 6 reads (ruined tower, ogee octagon, flat and spired towers, round lights among pointed); 7 reads (windows, stained glass, bowls, braziers, glasshouse, boathouse lamps, lantern boats), but the 20:45 sky is pink-violet rather than cool blue, as at the R7 close. From the high camera cue 1 reads less (the height foreshortened). Existing views (Four turrets, Harbour, quarter, Amsterdam, R4c courtyard): integer counts identical; changes within the noise pairs except castle-starter-night pre-2 against after-2 19.5% (noise 1.5% and 15.0%; pre-1 against after 10.2%) and amsterdam-night pre-2 against after-2 12.2% (noise 9.4%, 8.9%; pre-1 against after 5.0%): rain and ripple noise, no drawing code changed. 115: - **Cost** (AC power throughout). Frame at dusk, vsync off, creatures live, each starter at its own camera, 3 runs rotating (`f/bench/bench.txt`): Loch castle 680 draw calls, p50 5.22 ms, p95 5.62; Four turrets 1029, 6.70, 7.47; Harbour 971, 6.65, 6.83; 16 lights of 16, none starved. Placement release (`f/cost/release-summary.txt`, 9 rounds x 3 runs a scene): pot 26.41 ms on the Loch against 26.42 on the Harbour (24.05 Four turrets); rowing boat 27.22 against 27.13 (met); a medium brazier on the crag top 28.19 (+1.8 against the Harbour's pot); **a small flat tower 38.70 against 30.70 on the Harbour (+8.0 ms, over the +2 ms budget)** and 31.91 on Four turrets. Traced (`f/cost/split/`, `f/cost/trace.log.out`): on the Loch `Main.adapt_all` costs about +5 to +6 ms more than on the Harbour for any placement (the solve pass alone +2.1 ms: 12 towers 3.9 ms, the crag 0.7, the two bridges 0.6; the rest of the extra not split further), which a pot or a boat gets back in the sync (decorations 3.1 against 7.5 ms) and a building does not (its sync +2.3 ms, terrain phase 7.6 against 3.0). No game code is added by R8f: these are the shipped kinds' costs on this scene. 116: - **Full suite** (`f/suite/suite_all.out`): a, b and v at 2880x1800: 61 suites, 32,947 checks, 0 failures; --verify passed on try 1 (no stall); test_creature_reactions rerun 6871/0. 117: - **Older builds** do not list the starter; its saved scene (`f/scenes/loch.json`) opens on the pre-part tree 1e9a79f, validated and all supported (`f/probe/pre_load.log.out`). 118: - **Not done / open:** the R8 stage close and the M-R4 exports (separate job). Windows not run. The tower placement release on the Loch (+8.0 ms against the Harbour, traced above, not reduced: it is the shipped crag, tower and bridge machinery). The dusk sky is pink-violet, not cool blue (cue 7, a look call). The buttress piers read only close up (cue 3). The cottage's chimney stands at its cone's apex, where the hand seats a chimney on a round tower. SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 62-107 (Castle R8 Castle grounds stage close and milestone M-R4) 62: **Castle R8 Castle grounds stage close and milestone M-R4 (20 October 2026):** the stage-end overhead of stage R8 "Castle grounds" (step `36-R8`, docs/castle-elements-plan.md part "R8f ... Then the stage close") and milestone M-R4 (both exports and the Windows checklist; plan section 4: "A milestone M-R4 after R8f rebuilds both exports, with the Loch castle starter as the thing to show"). Opus 5.5 ran it on this Mac, 16:58-17:51, on the final R8 code: `main` 1024118 (`close/head.txt`; `close/status.txt`: no code change since, working tree as at the start with the tests.json churn and the other session's untracked `docs/codebase-audit-plan.md`). No game code changed; the scenes built for the captures (`close/scenes/`) are evidence data. Parts: R8a 0440347 (crag and cliff stairs), R8b 03ed81f (boathouse and lantern boats), R8c a2d33b8 (glasshouse, owl-tower perches, stone circle, pumpkins), R8d f8a26e2 (arcade and cloister, fountain, gate toppers), R8e 1e9a79f (banners, wall torches, braziers), R8f 1024118 (the Loch castle starter); bb15b93 (a handoff note) in the window. Evidence: `artifacts/decor/36-R8/close/` (below, `close/`). Platforms: Mac run; Windows exported, not run. 63: - **Before side:** `close/pre-tree` from `tools/feel_snapshot.sh tree 36-R8` (`build/backup-feel/36-R8` = f90e005, the R7x close). It differs from live in 25 files (scripts, `data/catalog.tres`, tests, `tools/test.sh`) plus the new `crag.gd`, `fire.gd`, `loch_scene.gd`, `fountain_water.gdshader` and five new suites (`close/diff-pre-vs-live.txt`). Kit: the R7 close's scripts copied into `close/` with paths adapted (`setup.sh`, `suite.sh`, `caps.sh`, `journey.sh`, `journey-note.js`, `metrics.sh`, `frame-table.js`, `prot.js`, `bench.sh`, `bench-summary.js`, `load.sh` / `load.gd` / `load-table.js`, `drag/drag.sh` / `dragall.sh` / `summ.js`, `crop.js`, `diffmap.js`, `rectmean.js`, `compose.js` (now wraps long panel labels), `gd.sh`, `chain.sh`, `progress.sh`), plus `close/scenes/build.gd` (the stress grounds), `close/sheet/specs.js` and, for M-R4, the R3 close's export kit (`close/export/m-r4_export.sh`, `START-HERE.txt`, `smoke/smoke.sh`) and `close/export/loch_md5.gd`. 64: - **Power:** every run on AC, 100% charged, Low Power Mode 0 (`close/suite/suite_all.out`, `close/cap/power.txt`, `close/bench/power.txt`, `close/load/power.txt`, `close/drag/power.txt`, `close/export/build-info.txt`, `close/export/smoke/power.txt`; 40 "AC Power" lines, no battery line). No run on battery. 65: - **Stage-end full suite** (`close/suite.sh` a, b, then v = `tools/suite_all.sh`, `VERIFY_RES=2880x1800` for v; `close/suite/suite_all.out`, 17:00-17:16, every suite on try 1): **61 suites, 32,948 checks, 0 failures** (R8f's 32,947 plus the orchestrator's `loch` fixture parser check). `--verify` 2880x1800 exit 0 on try 1 (26 s), `PIXELPUNK_INTEGRATION []`: **no verify stall** (R8c's watch item did not recur in R8d, R8e, R8f or here). test_creature_reactions 6,871/0, and 6,871/0 on the rerun. 66: - **The stress grounds** (`close/scenes/stress.json` from `close/scenes/build.gd`, headless on live with the parts' own builders: seat functions, the boathouse magnet, `Adapt.solve` as `Main.adapt_all`; `close/scenes/build.log.out`): a 10 m crag in a north-east water corner with a cliff stair from its shelf and a boathouse at its foot (the magnet's third release point), 6 glasshouses, 6 perched towers, 3 host houses and 3 host towers carrying 10 banners and 10 torches, 4 stone circles, 6 braziers, 2 cloisters, 3 fountains, 8 topped gates, 20 pumpkins: 84 objects, nothing removed, `all_supported` and `validate_document` empty. It cannot load on the pre-tree. 67: - **Captures** (`close/caps.sh`, 1600x1000, one sitting 17:16-17:31 by `close/chain.sh`, every view on try 1; `close/cap-sitting.txt`). g1 = the R7 close's 23 views (pre-1, after, pre-2). cues: the Loch castle (`36-R8/f/scenes/loch.json` = `LochScene.build()`, md5 = `LOCH_MD5`) from R8f's three cameras, dusk 20:45 and day 14:00, rain 0, twice (cues, cues-2: integer counts identical, frame >8 luma 0.2-1.4%). parts: 29 views on the parts' own scenes, the Loch close-ups and the stress grounds. 68: - **Metrics** (`close/metrics.sh` -> `close/metrics/`): **integer counts identical pre-1 -> after in all 23 views** (`metrics/counts.txt`); 21 of 23 views have both cross legs above the noise leg (`metrics/frame-table.txt`), the diff maps (`close/look/courtyard-night-diffmap.png`, `castle-night-diffmap.png`, each with its `-noise-` twin; looked at) show rain rings, streaks and clouds over the whole frame, as at the R4 and R7 closes; harbour-night's 15.2% sits at its own noise pair's 15.0%. Protected looks (`metrics/protected.md`, 59 rows): p50 within 0.5 and hex equal or one unit off in 58; the 59th, the Harbour night zenith, 54.5 -> 53.9 against pre-1, is 53.9 -> 53.9 against pre-2, its noise pair differing 23.6% (clouds). No stop. No R8 part changes an existing scene's drawing (every part's pins held, and so do the starters' md5 pins in test_starters). 69: - **Journey** (`artifacts/feel-journey/36-R8/`, the six views via `close/journey.sh`, 17:16): against 35-R7 (`36-R8/counts-vs-35-R7.txt`, `frame-vs-35-R7.txt`) **integer counts identical in all six views**; frame >8 luma harbour-day 0.6%, harbour-lane 2.3%, harbour-night 1.9%, quarter-day 9.6%, quarter-night 12.1%, styles-day 16.0% (rain and ripples; `close/look/journey-styles-day-diffmap.png`, looked at). By part: R8a to R8e add kinds and keys no starter uses (each held the starters' md5 pins); R8f adds a starter and changes none. **harbour-lane was re-shot**: its first capture (17:16:23) looked out over open water from the wrong place (draw calls 98 against 796), a one-off: the close's own after view of the same camera in the same sitting matched the pre-stage tree, and the re-shot view (17:43) is identical to 35-R7 in counts; the flaked capture is kept in `close/look/journey-lane-first/`. **New:** the Loch castle at dusk, default and low across the water, under `36-R8/pkg/`. `notes.json` has a `36-R8` entry (`close/journey-note.js`, backup `notes.json.bak-before-36-R8`); `index.html` (43 steps) and `journey-standalone.html` (114.92 MB) rebuilt with `node tools/feel_journey_index.js --digest`. 70: - **Whole-stage A/B bench** (`close/bench.sh` = the R7 close's recipe; `close/bench-t1.out`, `close/bench/summary.txt`; 17:31-17:35, every run on try 1). pre and live interleaved per scene, round 0 discarded, rounds 1-3; `--disable-vsync`, 1920x1080, smoke, `--churn=100000`, rain 0.45, `--motion-live`, 01:00; 491 samples each. Medians pre -> live (paired median; rounds live slower): 72: | Scene | 01:00 p50 | 01:00 p95 | draw calls | 73: |---|---|---|---| 74: | Dense | 15.442 -> 15.507 (+0.004, 2/3) | 17.172 -> 17.196 (−0.061, 1/3) | 2587 -> 2587 | 75: | Harbour | 8.372 -> 8.390 (+0.006, 2/3) | 9.747 -> 9.758 (+0.011, 2/3) | 983 -> 983 | 76: | Stress grounds (live only, dusk 20:45) | 7.678 (7.678/7.611/7.681) | 8.725 (8.399/8.725/8.941) | 1043 | 78: The stop line (dense p95 over about +0.5 ms) is **not hit**. Both trees ran about 1 ms slower than at the R7 close (machine state; the paired legs are equal). The stress grounds cannot run on the pre-tree (`STRESS=1 close/bench.sh 20.75`, `close/bench-stress.out`). 79: - **First sync** (`close/load.sh` with `close/load.gd`; headless, pre and live 3+3 interleaved with the first side alternating, the Loch castle and the stress grounds live only; 17:36-17:39, every run on try 1; `close/load.out`, `close/load-table.txt`), medians pre -> live: **Harbour 244.4 -> 248.6 (+4.2), dense 3315.8 -> 3378.9 (+63.1), Amsterdam 612.5 -> 627.9 (+15.4), courtyard 301.7 -> 309.5 (+7.8), Castle 703.1 -> 708.1 (+5.0) ms; the Loch castle 607.1 ms (610/607/605, `LochScene.build()`; it cannot load on the pre-tree, which has no `loch_scene.gd`); the stress grounds 686.5 ms.** The dense rise is in the terrain (+13) and clusters (+28) phases (PHASES lines), about 1.9% with no R8 part in the scene; not traced (carried). 80: - **Drags** (`close/drag/dragall.sh`; cold, the A0b Harbour script `data/drag/baseline.json`, 1600x1000; pre and live interleaved per run, first side alternating; runs 1-3 at 17:39-17:41 and, because the slider release split, runs 4-6 at 17:41-17:43; all try 1, 0 errors; `close/drag/summary.txt`, `summary-6runs.txt`). Six-run medians: **handle step p95 pre 55.62 / live 55.92 ms against the 53.7 ms baseline** (R8's reviews saw 55-60 on both trees); handle release 12.75 / 11.37. **Slider release (the R7 close's watch item):** the script's Height slider releases in 20.5 ms on the pre-tree in 6 of 6 runs, and on live in 21.1 ms (3 runs) or **49.4 ms (3 runs)**; the Width slider 28.0 pre, 29.05 live, with one live run at 64.7 ms. The by-kind median is 24.28 pre against 39.25 live. A two-moded release on the final tree only, over the +2 ms budget in half the runs; not traced (carried, recommended below). Other kinds within +0.2 to +0.8 ms. Digests equal on both trees. 81: - **Silhouette cues** (plan section 1; R8f's acceptance "all seven cues visible in a dusk capture from across the water"; `close/cap/cues/`, sheet pages 1 and 2): 82: 1. **Height variety: reads.** The great tower (13 m with a 3 m spire tower, vane 30.5 m) stands clear over every roof from the low and default cameras; from the high camera it reads less. 83: 2. **Clustered steep spires: reads.** Four spires of 11, 9.5, 8.5 and 6.5 m in a clump. 84: 3. **A cathedral in a fortress: reads in part.** The hall reads as a church (lit stained double lights, the rose, the belfry); its buttress piers show only close up or on the long side from high, and its roof is not steep. 85: 4. **Cliff siting over water: reads.** The crag stands in open water with a rocky shore and the lights lie in the water at dusk; at 8 m it is modest beside the towers. 86: 5. **Long horizontal links: reads.** The viaduct (four tall arches, fire bowls lit) and the covered bridge, long and thin on high trestles (which read as iron). 87: 6. **An irregular plan that grew: reads.** Spired, flat, ogee-octagon and ruined towers, a gatehouse, round and pointed lights, the islet's cottage and circle. 88: 7. **Warm lights at dusk: reads in part.** Windows, stained glass, bowls, braziers, glasshouse, boathouse lamps and boat lanterns read; the 20:45 sky is pink-violet, not the plan's cool blue (as at the R7 close; a lighting call for every starter). 89: The same verdicts as R8f's implementer and reviewer. 90: - **Owner sheet** `close/sheet/` (`close/sheet.sh`: `sheet/specs.js` -> `sheet/spec/*.json` -> `close/compose.js`; existing scenes through `tools/art_metrics.js sheet`; `sheet.md` in plain words; `carried-notes.md`). Every page was looked at: **`r8-01-cue-sheet.png` first** (three cameras at dusk, cues marked, a closer crop), `r8-02-cues-day-verdicts.png`, `r8-03-crag-boathouse.png`, `r8-04-grounds.png`, `r8-05-arcade-fountain-toppers.png`, `r8-06-banners-torches-braziers.png` (with the stress grounds), `r8-07-loch-closeups.png`, **`r8-08-owner-items.png` and `r8-09-owner-items-2.png`** (A to M, one sentence each: the banner and torch release hitch, 78 ms against 27.6 for a pot on the Harbour and 265-272 ms on the Castle starter; the Loch's crag Height row refused and no bridge landing on a crag top; the cottage chimney as a black pipe; the trestles reading as iron; the pink-violet dusk; the arcade wall form's limits; a house in a cloister lawn; the flat lawn slab; the hidden dusk lantern; a brazier under a lifted or moved house; the 1 m banner as a tile; the torch lean; the conservative bartizan box and host reshapes refused with its line), `r8-10-existing-castle-courtyard.png`, `r8-11-existing-harbour-amsterdam-quarter.png` (before and after; the titles run off the right edge, as before). 91: - **M-R4 exports** (`close/export/m-r4_export.sh` = the M-R2 script with M-R4 names; `m-r4_export.out`, `build-info.txt`, `build-macos.out`, `build-windows.out`, `export-windows.log`, `codesign.txt`; 17:44, exit 0, no ERROR line in either export log). 92: - The previous `build/macos/PixelPunk.app` and `build/windows/PixelPunk.exe` (M-R2's, 13 October) are moved to `close/export/previous-macos/`, `previous-windows/`. 93: - On the Desktop: `PixelPunk-M-R4-macOS.zip` 67,142,035 bytes, sha256 `72b4272fb4b4b0e3dd391bfa9f14a043e9529b4a606edcbc32507e0ab8c6567a`; `PixelPunk-M-R4-Windows.zip` 44,028,313 bytes, sha256 `4ab6f4603d476ed936d262fd8100fd5fb12de88283b7def8add82e5e531a3e85` (PixelPunk.exe 113,063,936 bytes, sha256 `fa9dcc47…555f`, START HERE.txt, README.md, THIRD_PARTY_NOTICES.txt); `PixelPunk-M-R4.app` (182 MB; ad-hoc signed, universal x86_64 and arm64, `codesign.txt`). Recorded in `close/export/build-sha256.txt` and `desktop-files.txt`. Every other Desktop item untouched (`desktop-listing-before.txt` | `-after.txt`: only the three M-R4 items added; the older M-R1, M-R2, M6, R1a and R7 builds sit in the owner's "PP Dev builds" folder, untouched). 94: - `START HERE.txt` (source `close/export/START-HERE.txt`): the Loch castle as the thing to show; what is new since M-R2: R4's lift bridge and waterside props, BK1's bike styles, R7's towers, openings, buttresses, clock and belfry, viaduct, covered bridge, finishes, and R8's crag, boathouse and lanterns, glasshouse, perches, circle, pumpkin, arcade and cloister, fountain, toppers, banners, torches, braziers; one undo step each. 95: - **Mac smoke** (`close/export/smoke/smoke.sh`, 17:45, the Desktop `PixelPunk-M-R4.app` binary, 1600x1000; `summary.txt`): `default` (Harbour at night, `--smoke`) exit 0, 969 draw calls, 138 parts, capture_failures []; **`loch-open`** (`--drag` with `smoke/open-loch.json`: the harness opens the `loch` fixture, `LochScene.build()` in the exported binary, settles 30 frames) exit 0, 54 parts, fixture loch, drag errors []; **`castle-open`** (the `castle` fixture, Four turrets) exit 0, 102 parts; `loch-dusk` (`--smoke` on the Loch document at 20:45) exit 0, 670 draw calls, capture_failures [] (looked at: `smoke/loch-dusk/courtyard.png`, the castle as in the cue captures). No PixelPunk process left. Each log holds the known `terrain_flowers.svg` loader line (as at M-R2) and `loch-dusk` the "2 resources still in use at exit" line. 96: - **LOCH_MD5** (`close/export/loch_md5.gd`, headless on live; `export/loch_md5.log.out`): `LochScene.build()` compact md5 f10d62aad985d2910866afe9bd5b4114 = `LOCH_MD5` (MATCH); opened through the editor's own `action("loch")` the same (MATCH, scene name "Loch castle"); saved as a player's Save writes it, `export/loch-save/Loch check.json` (26,433 bytes) has md5 `d2933e919e08f7894c969ea04afe5d1e`, the figure the Windows checklist compares with `certutil`. 97: - **Windows: exported, not run.** 98: - **Windows checklist** `artifacts/decor/m-r4/windows-checklist.md` (owner E14; every entry "Windows exported, not run"): the zip's sha256 and the PC's details; no R5b rows (R5b not built); 22 rows: unzip and launch; the Loch castle; the body font (Pixelify Sans, Segoe UI as the fallback) at 1100x700, 1600x1000 and full screen; Castle then Loch again; **LOCH_MD5 by Save and `certutil -hashfile ... MD5` against `d2933e91…5d1e`**; a crag, a cliff stair, a boathouse, a glasshouse, an arcade, a fountain, a topper row, a banner, a torch and a brazier; a grip drag; undo and redo of each; save and reopen; photo mode; quit. 99: - **Prune** (owner, 12 October): `node tools/prune_evidence_png.js artifacts/decor/36-R8/close` after the sheet moved 188 PNGs (136.9 MB) to the macOS Trash [local Trash path omitted]; the 72 PNGs the sheet and this entry cite (`cap/cues`, `cap/parts`, `export/smoke/loch-dusk`) were moved back, so 116 stay in the Trash: pre-2 (46), the cue noise twin `cues-2` (12), the other smoke frames (6) and the `--verify` integration shots (52) (`close/pruned-png.txt`, with the restoration line). pre-1, after, the cue and part captures, `look/` and the sheet stay. 100: - **Carried notes** (`close/sheet/carried-notes.md`, one line each with its part and class, from every R8 Review line, "Not done / open" block and the rulings of 15, 16 and 20 October, plus the R7 items still open). Player-facing among them: R8a a house left straddling a crag edge, a later house on a stair deck turning or covering the stair; R8d a wall walk not re-seating under a jetty; R8e a brazier (as a pot or fountain) under a lifted or moved house; and this close's Harbour Height-slider release mode. The R8c verify stall watch is closed (not seen since). R7's still-carried player items (flying footings, clock hands' grow-in, arch rings in the solids, the one-edge ruin line, the victorian Finish row) by reference. 101: - **Recommendation (the orchestrator decides):** a small follow-up part "R8x" before R5b with (1) a trace and fix of the Harbour Height-slider release mode (+28 ms in half the live runs, never on the pre-stage tree; the budget scene has no castle part); (2) a refusal with a hint for a house dropped into a cloister lawn and for a brazier, pot or fountain left under a house lifted or moved onto it (one ground-under-a-building check); (3) the Loch cottage's chimney off the apex or dropped (starter data, re-pin `LOCH_MD5`) if the owner agrees. The banner and torch hitch is the owner's call (a later draw-apart part); the rest are looks, limits or hardening that can wait. 102: - **This close (new):** the slider release mode; the first-sync rises (dense +63.1 ms, Amsterdam +15.4; not traced); the flaked journey capture (tooling, re-shot). 103: - **Not done / open:** 104: - The owner's answers on the sheet's items A to M; the orchestrator's decision on R8x. 105: - Windows run (checklist written). The `LOCH_MD5` check on Windows is the checklist's row 8. 106: - Bench at 01:00 only (the stress grounds at dusk); drags cold and the Harbour script only; the Loch castle and the stress grounds measured on the final tree only; the slider release mode and the first-sync rises not traced; no arrows on any page (automated captures hide grips). SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 44-61 (Castle R8x Harbour slider-release trace) 44: **Castle R8x Harbour slider-release trace (20 October 2026):** an orchestrator-added follow-up after the R8 close (a2f639b): trace the Harbour Height-slider release that the close measured at 20.5 ms on the pre-stage tree in 6 of 6 runs, and at 21.1 or 49.4 ms on the final code. Opus 5.5 ran it on this Mac, 17:52-18:35. Evidence: `artifacts/decor/36-R8/x/` (`x/` below). **No game code changed.** Snapshot `build/backup-feel/36-R8-x-pre/` (= a2f639b). Review: orchestrator, 20 October about 18:40 (a trace, no game code changed, so no fresh review): the 49 ms mode is A2c's paced live rebuild (`LIVE_FULL_GAP_USEC`, 80 ms): a release is slow exactly when the drag's last step is still waiting in that gap, and the same 28 ms build lands before the release in the fast mode; it also occurs on the pre-stage tree (Height 5/40 pooled, 1/10 in the clean repro) and no R8 part introduced it. Real R8 cost in the fast mode is +0.7 ms, inside the +2 ms budget. Ruled: not a regression; future benches report the release with the catch-up build or the waiting flag. The dense first-sync rise stays untraced and is not this cause. Platforms: Mac run; Windows not run. 45: - **Recipe:** the close's drag recipe (`data/drag/baseline.json`, cold, 1600x1000), run by `x/drag.sh` and `x/dragab.sh` with the trees interleaved and the first side rotating. Every run was on AC power (`x/*/power.txt`: 104 "AC Power" lines, no battery). Every run passed on try 1 with drag errors []. The first bisect call passed all seven trees as one path because of zsh word splitting, so nothing ran; it is kept in `x/bisect-failed-wordsplit/`. 46: - **Reproduction** (`x/repro/`, `node x/summ.js x/repro`; 10 live runs against 10 on `x/tree-36-R8` = f90e005, without instrumentation): Height slow (over 35 ms) in **5/10 live and 1/10 pre** (pre-3: 48.3 ms). Width slow in 3/10 live (about 65 ms) and 0/10 pre. **The handle release has the same two modes** (13.5 or 47 ms): 2/10 live and 4/10 pre. So the slow mode is not on the final code only. 47: - **Phase** (`x/instrument.pl`, `instrument2.pl` and `instrument3.pl` patch only the scratch trees `x/tree-*`; `x/phase*/`, `x/phase*-summary.txt`). The extra 28 ms is in **`clusters`**: the release builds the main house's cluster `7,8,13` inline (27.4-27.9 ms on Height, 32.7-33.2 ms on the handle) after cluster 14. Every other phase stays put: adapt_all 7.6, terrain 0.5, set_tool 3.7. The cluster is rebuilt **exactly when the drag's last step was still waiting in WorldView's pacing gap at the release**. That gap is `LIVE_FULL_GAP_USEC`, 80 ms (A2c, owner 24 September: a drag rebuilds at full detail at most once per gap, and the release builds a step that is still waiting). Over 150 instrumented releases, 42 were slow and every one was waiting at the release; 108 were fast and none was waiting (`node x/settle.js x/bisect x/phase3`, `x/settle.txt`). In the fast mode the same full build of the last step happened earlier: inside the step's own frame ("built") or in `_process` once the gap ended ("catch"; 31-39 ms, `node x/timeline.js x/bisect`). The total work is the same in both modes; the release metric counts it in one mode and not in the other. Which mode a run lands in depends on a race about 230 ms into the scrub: either a step arrives just after the gap ends, or the `_process` catch-up builds first. The rest of the timeline follows from that (`x/timeline.js`). The rejected hypotheses: R8d's and R8e's caches (the mode is on the R8c tree and the pre-stage tree), a time-driven tick or GC (the cost is a named cluster build), and run order (both modes on both sides). 48: - **Bisect** (`x/bisect.sh`, 7 trees x 6 runs, 18:16-18:30; `x/pooled.txt` pools every sitting including the close's own 6+6 runs): 49: | Tree | Height slow | Handle slow | Width slow | Fast Height release (median) | 50: |---|---|---|---|---| 51: | pre f90e005 (pooled 40 runs) | 2/6 (5/40) | 0/6 (11/40) | 0/6 (0/40) | 20.54 | 52: | b (after R8a) | 3/6 | 4/6 | 0/6 | 20.99 | 53: | c (after R8b) | 2/6 | 1/6 | 1/6 | 20.99 | 54: | d (after R8c) | 3/6 | 0/6 | 0/6 | 20.97 | 55: | e (after R8d) | 3/6 | 3/6 | 2/6 | 21.12 | 56: | f (after R8e) | 2/6 | 5/6 | 0/6 | 21.22 | 57: | head a2f639b (pooled 40 runs) | 3/6 (19/40) | 3/6 (19/40) | 0/6 (4/40) | 21.25 | 58: **No part introduces the 49 ms mode**, because it already exists on the pre-stage tree. From R8a on, the Height slow odds rise from about 13% to about half, and they then wander from part to part (handle 0/6 to 5/6) with no trend. Per-step costs barely move: the Height step handler goes 5.63 -> 5.75 ms after R8a and 5.84 at head, and the live build goes 31.1 -> 31.8 ms (`node x/ticks.js x/bisect `). A shift that small tilts the race; it does not cause the mode. The fast-mode release itself rises **+0.7 ms** pre -> head (`x/fast-phases.txt`): adapt_all 7.18 -> 7.69 (R8a +0.25, then +0.05 to +0.13 a part) and clusters 5.80 -> 6.00 (signatures 2.48 -> 2.58). That is inside the +2 ms budget. 59: - **Cause and ruling:** the cause is A2c's paced live rebuild, which predates R8 and is not R8 code. A one-frame release after a 24-step scrub sometimes lands inside the 80 ms gap, and the release then pays the full house build that would otherwise land a frame or two earlier. It is a property of the measurement and the pacing design, and the same on every tree, so **no change was made**. All R8 did was tilt the odds through its sub-millisecond costs. So no test, no pin check and no suite run were needed (no code changed; every pin stays as at the close). Suggestion for the orchestrator: in future benches, report the release together with the catch-up build (as `x/settle.js` does) or the waiting flag, or let the script hold for one gap before a release, so that a bench does not read this race as a regression. The first-sync rises (dense terrain +13 ms, clusters +28 ms) cannot have this cause, because a first sync is a forced whole build with no pacing. They were not traced (no code was changed). 60: - **Not done / open:** the first-sync rises were not traced; no change to the pacing design or the harness (both are outside R8); the scratch trees `x/tree-*` (about 30 MB each, instrumented) are kept for re-runs and can be pruned. SOURCE: artifacts/decor/owner-decisions.md (frozen 21 October 2026), lines 183-221 (rulings of 15, 16 and 20 October for R8a to R8x, and the owner approval of 15 October) 183: Orchestrator rulings, 15 October 2026 about 16:03 (reversible by the owner), from the R8a review (`artifacts/decor/36-R8/a/review/`, verdict must-fix open): (1) Terrain chunks: the bank's drawing must not depend on crags its fingerprint does not cover (lessons 5 and 8). The terrain chunk is drawn from the heightfield alone (a land query that ignores crags; the crag draws its own sides and top); if any chunk piece truly needs the crag, the chunk fingerprint includes every crag overlapping it (outline, height, seed). A test: draw, move away and undo a crag over a bank, and the chunk's drawing equals the never-touched chunk byte for byte. (2) The cat never naps where no route reaches: nap sites on anything standing on a crag top (and the crag top itself) are excluded while crag tops are unwalked, with a test. (3) The crag index is kept across `crowd_dirty` when the replaced object is neither a crag nor a cliff stair, so `adapt_all` walks once (walk-count test). (4) A crag must not bury a ground landing: a crag stroke, move or resize whose outline would cover a bridge's, walkway's or jetty's ground end is refused with a hint naming the landing ("Keep the crag clear of the bridge landing."), checked both ways (a landing set inside a crag's outline seats on its top or is refused), never stranding later edits. (5) The entry says that an older build refuses a scene holding a crag ("Unknown or damaged building part."), and lists as limits carried to the R8 close: creatures do not walk crag tops and cliff stairs carry no walkers (as tower wraps today); a house straddling the edge after a crag move keeps standing with its own Height row refused until moved; a later house on the stair's deck turns the stair the other way, and with both ways blocked the deck runs under it; a second crag beside a stair is not refused; the faceting reads as a regular sawtooth close up (to show the owner at the R8 close). Fix pass 1 by the implementer. 185: Orchestrator ruling, 15 October 2026 about 16:57 (reversible by the owner), from the R8a recheck (`artifacts/decor/36-R8/a/review2/`, verdict must-fix open; the chunk, nap and index fixes hold): what stands on a crag top rides with it. When an edit changes a crag's height or outline, every connection end with a ground parent (bridges, walkways, jettys and the cliff stair's head) that sat on its old top and is still inside its outline moves to the new top before the pass, so `crag_buries` no longer sees it; each bridge style's own rules then decide (equal landings, the viaduct's single height, the covered bridge's level) and refuse with that style's own line and hint (lesson 3), and a cliff stair's head always follows its crag (its plan is never left empty after a Height row). Lowering a crag under a bridge end on its top moves the end down with it, so no bridge is left levitating. The Height row's preview shows the refusal's hint when the release will be refused. Tests for each bridge style (stone, timber, japanese, modern, rope, jetty, bascule; viaduct and covered where the crowd check allows two crags), lowering, and a cliff stair after a 1 m and a 4 m Height change. The entry's fix-pass-1 views line is corrected (castle-crag-low, crag8-close, crags-high, shore-moss and stair-bank also beyond noise; the castle-crag-high change is stray path edging now gone; suite from 16:30). Fix pass 2 (the last agent round). 187: Owner, 15 October 2026 about 17:38: "All good and push". Taken as approving the R7 stage close's owner items (sheet `artifacts/decor/35-R7/close/sheet/r7-07-owner-items.png`: gargoyles and pinnacles dark beside rubble piers kept; the D-plan tower's rubble half-gable kept; slate piers beside a grey gargoyle on a timber hall kept; the round spire back halves, fixed in R7x) and the work through R8a. Pushed: origin/main 05454c5..0440347 (11 commits, R7c to R8a). 189: Orchestrator rulings, 15 October 2026 about 19:34 (reversible by the owner), from the R8b review (`artifacts/decor/36-R8/b/review/`, verdict must-fix open): (1) A berthed boat rides with its boathouse: when `crag_carry` (or any reseat that moves a boathouse: a quay edit, a turn) moves a boathouse, the boats berthed in its channel move with it by the same transform before the pass, so a crag raise, move or reshape is never refused for a boat the player did not touch; tests for raising to 12 and 14 m, moving and reshaping with a berthed boat, and undo. (2) A boat too long for the channel gets the real reason: a boat longer than the berth limit released in the channel is refused with a hint that names the limit (for example "This boat is too long to berth here: boats up to 3.7 m fit the channel."), or seats stern-in poking out of the door as a placed one may; either is fine, the hint must not tell the player to do what they just did (lesson 3). (3) The rock check runs before the water check, so a boathouse pushed into a crag gets the rock line. (4) The entry's older-build line quotes "Unknown or damaged building part." (5) The entry gives the stress boat placement's second sitting (+3.0 ms against pre; +2.95 against the same scene live) beside the first, and accepts it: `boathouse_error` is flat with scene size (0.52 to 0.59 ms for 6 houses, one index walk a pass), the Harbour has no boathouse; a memo keyed on terrain and position is carried as later hardening. Fix pass 1 by the implementer. 191: Orchestrator rulings, 15 October 2026 about 20:11 (reversible by the owner), from the R8b recheck (`artifacts/decor/36-R8/b/review2/`, verdict must-fix open; the crag carries, turns, deletes, lantern boats and long-boat hints hold): (1) A boathouse moved by hand takes its berthed boats in the move check too (`validate_candidate` carries the berths on its check copy), so a short hand move is never refused for its own boat; the boat follows the drag's ghost if cheap, otherwise at release (say which). (2) The boathouse's checks run water first with crags off, then rock, then crag tops, so a house on land beside a land crag gets the water line and a house pushed into a crag gets the rock line, each with a hint the player can act on. (3) A long boat moored stern-in but a little skewed or off the axis gets the long-boat line with a hint naming "straight along the channel" (no bound that sends it back to the berth hint). (4) `carry_berths` reads the boathouse index's boats, not every object (lesson 1), with a walk-count test. (5) Carried as a limit to the R8 close, not changed: a boat waiting just outside the water door is not berthed, so a crag raise or reshape that moves the house onto it is refused with the boat line (soft; the player moves the boat). Fix pass 2 (the last agent round). 193: Orchestrator ruling, 15 October 2026 about 22:27 (reversible by the owner), from R8c's hand-back: tower birds (the perched towers' rock pigeons, 6 a tower) are counted in their own cap of 18, outside the 30-body creature cap, as fireflies and moths are; ordinary pigeons keep their own cap. The stress release (+3.3 to +3.9 ms an edit with 27 more parts present) and a perched tower inside a cluster getting no birds are sent to the fresh review before any ruling. 195: Orchestrator rulings, 15 October 2026 about 22:38 (reversible by the owner), from the R8c review (`artifacts/decor/36-R8/c/review/`, verdict must-fix open; the cost is accepted: flat with scene size, within 0.6 to 2 ms of the same slots filled with benches and lamps, no walk per part; memoising `circle_feet` is carried as hardening): (1) A perched tower draws its dovecotes and boards and gets its flock whether it stands alone, is joined by curtain walls, or is merged with a hall (the castle case R8f needs): `_tower_alone` stops excluding wall-joined towers, and `ClusterArchitecture` draws the perched windows on a clustered tower with perch sites derived from what is drawn; the Perches row never does nothing silently. (2) No bird sits on a perch that a later crag, wall part or other solid covers (Motion Off included). (3) Meadow tufts and planting keep out of a glasshouse's floor (its footprint joins the footprint index only when one is present, so other scenes' signatures stay). (4) A stone circle across a crag edge is refused with a hint, soft for strokes, as a house over the edge is. (5) Rule 5 (at most 12 MultiMeshes a kind) holds for any number of circles and glasshouses: stone circles share their MultiMeshes across circles, and glasshouse lengths share meshes (quantised or built into the batch rather than one MultiMesh each); say how. (6) The entry's Harbour perched-tower lines are corrected and the Perches-row cost re-measured on a tower that draws its ledges. Fix pass 1 by the implementer. 197: Orchestrator rulings, 15 October 2026 about 23:27 (reversible by the owner), from the R8c recheck (`artifacts/decor/36-R8/c/review2/`, verdict must-fix open; clustered perches, covered perches, the glasshouse floor, the chunk keying and the crag-edge circles hold; the Perches row's +2.38 ms is own-edit content cost, the same as the Variant row on that tower, accepted): (1) The perch check refuses only the Perches row itself, or a row on the tower that takes away the last perch slot it had; an unrelated row on a perched tower whose slots another edit (a hall raised round it) already took away is never refused for perches, and such a tower is a soft state that draws no ledges and gets no birds until a window opens. (2) Settled glasshouses keep the lamplighter's round: the shared chunk mesh sits at its chunk's centre and each part's glass sweeps with its own light (the same turn before and after settling, with a test). (3) An edit to one part in a shared chunk does not rebuild its neighbours' pieces: each part's built piece is memoised by id, look key and frame, so the chunk rebuild only re-joins (with a release test of 20 in one chunk against one alone). (4) The entry's dusk sentence is corrected, and it notes that two cue towers' ledges covered by a neighbour's box get no perch sites (no bird on an undrawn ledge). Fix pass 2 (the last agent round). 199: Orchestrator ruling, 16 October 2026 about 00:15 (reversible by the owner), closing R8c on fix pass 2 (`artifacts/decor/36-R8/c/fix2/`; full suite 59 / 32,124 / 0, every pin diff 0): (1) The glasshouse's lamplighter turn shifting when it settles (the sweep is per chunk once settled, per part while growing in; glass and light always agree) departs from ruling 2 of 23:27 ("the same turn before and after settling") and is accepted as a note: a one-second change on a placed glasshouse at dusk, carried to the R8 close. (2) The verify stall (twice in the roof checks, then two clean passes on the same code, cause not found) is carried as a watch item: if it recurs in R8d's full suite, that part's fix pass traces it. R8c closed. 201: ## R8d review ruling, 20 October 2026 about 11:20 202: Review (fresh, Opus): must-fix open. 1) A raise or lower stroke over 0.5 m leaves an arcade, fountain or glasshouse foot behind (soft branch of adapt_all). 2) Fountain look draw 2.2-3.2 ms (_lathe per-face GDScript), +4.3 ms over a pot on release, over the +2 ms budget. Ruled: fix both (glasshouse too); also warm-up for the fountain water material and arcade_seat jetty check on the walk being placed. Owner look calls carried to the R8 close: wall form limits against freehand walls and jettied houses, the back wall covering a house's ground-floor doors and windows, a house dropped into a cloister lawn committing, the flat bright lawn slab, the hidden dusk lantern. 204: ## R8d close, 20 October 2026 about 12:30 205: Recheck verdict done (two notes). Accepted: a wall walk does not re-seat on a Length or Arch change under a jetty (same class as the ruled stand-off); the cold Size-row undo at +1.87 ms (warm +0.91). Carried to the R8 close with the earlier look calls. Closed and committed locally. 207: ## R8e review ruling, 20 October 2026 about 14:55 208: Review (fresh, Opus): must-fix open. 1) hang_error walks every object once per tower part via tower_wraps (lesson 1). 2) Bartizans are not in the banner/torch clash list either way. Ruled: fix both; correct the entry's pot-cost line. Accepted (reversible by the owner): banner and torch release cost is the host's cluster rebuild (78 ms against 27.6 for a pot on the Harbour, 265-272 ms on the Castle starter, the same class as a clock face at 260-270 ms; R7d precedent); the size of the hitch is carried to the owner at the R8 close. Limits carried: a brazier under a lifted or moved house (as pot and fountain); the 1 m banner reads as a tile and the torch lean is subtle at default zoom. 210: ## R8e close, 20 October 2026 about 15:45 211: Recheck verdict done (two notes). Carried to the R8 close: some host reshapes in a scene holding a bartizan and a banner or torch are refused with the bartizan's line (R7b's rule); the bartizan's clash box is conservative (up to about 1.0 m from the corner). Closed and committed locally. 213: ## R8f close, 20 October 2026 about 17:00 214: Review (fresh, Opus): must-fix one, the `--drag` fixture `loch` missing from `DragHarness.FIXTURES`; fixed by the orchestrator with a parser test. Accepted: the tower release +6 to +8 ms on the Loch castle as the 54-part scene's own cost (adapt_all flat per part). Owner items carried to the R8 close sheet: the crag Height row refused on this starter and no bridge landing on a crag top; the cottage chimney; the covered bridge's trestles read as iron; the pink-violet dusk sky; buttress readability; LOCH_MD5 untested on Windows. 216: ## R8 stage close and M-R4, 20 October 2026 about 17:55 217: Close done. New finding: the Harbour Height-slider release is bimodal on the final code (21 ms or 49 ms, 3 of 6 runs each) against a steady 20.5 ms on the pre-stage tree. Ruled: trace it (bisect the per-part snapshots) before R5b as part R8x; no behaviour change otherwise. Open for the owner: the Loch cottage chimney, refusing a house dropped into a cloister lawn or a brazier/pot/fountain under a lifted or moved house, the banner and torch placement hitch. 219: ## R8x, 20 October 2026 about 18:40 220: Trace done, no code change: the bimodal Height-slider release is A2c's paced live rebuild (LIVE_FULL_GAP_USEC 80 ms), a timing race present on the pre-stage tree too; no R8 part introduced it; R8's real cost +0.7 ms. Ruled not a regression. Benches report the release with the catch-up build or the waiting flag from now on. SOURCE: artifacts/decor/owner-decisions.md (frozen 21 October 2026), lines 237-256 (owner push of 21 October and the R8 sheet answers A to M) 237: ## Owner, 21 October 2026 about 01:00 238: "gamepad connected, push it". Pushed origin/main 0440347..91ae641 (R8b to R8f, the R8 close, R8x, R5b, R5c, R5d). A gamepad is connected to the Mac for R5f's live check. R5e (running) comes first; later commits stay local until the owner says. 240: ## Owner, R8 close sheet items A to M, 21 October 2026 241: Drafted by the assistant at the owner's request and recorded on the owner's word ("write them into owner-decisions.md"); each is reversible. Sheet pages `artifacts/decor/36-R8/close/sheet/r8-08-owner-items.png` and `r8-09-owner-items-2.png`; context in `carried-notes.md` beside them. 242: - **A. Banner and torch placement hitch: keep.** Accept the host-house rebuild (78 ms Harbour, 265 to 272 ms castle keep), as for the clock face. A later "draw apart" part only if it bothers the owner. 243: - **B. Loch castle crag Height row refused: keep.** Both long bridges land on terraces, so the refusal is right and carries its hint. 244: - **C. Cottage chimney: change.** Drop it or seat it off the cone's apex (starter data only), then re-pin `LOCH_MD5`. 245: - **D. Covered bridge trestles read as iron: change, low priority.** Recolour toward weathered timber when convenient; blocks nothing. 246: - **E. 20:45 pink-violet sky (cue 7): defer.** A separate lighting decision for every starter, not an R8 part. 247: - **F. Arcade wall form: keep the 1 to 1.6 m back-wall height and the 0.25 m jetty offset as form limits; change the door cover.** The back wall refuses or steps off when it would hide a house's ground-floor doors and windows. If that is too large for now, record it as a limit instead. 248: - **G. House dropped into a cloister lawn: change.** Refuse with a hint, sharing one ground-under-a-building check with J. 249: - **H. Flat bright lawn slab: keep.** Cosmetic; revisit with other lawn work. 250: - **I. Hidden dusk lantern: keep for now.** Move it out from under the lean-to in a polish pass. 251: - **J. Brazier under a lifted or moved house: change.** Refuse with a hint, or let the prop lift with the house; same check as G (brazier, pot, fountain). 252: - **K. 1 m banner reads as a tile: keep.** If it bothers the owner, raise the minimum to 1.5 m rather than redraw it. 253: - **L. Torch lean subtle at default zoom: keep.** The lean is correct and shows close up. 254: - **M. Bartizans beside banners or torches (conservative clash box): keep.** Over-refusing is safer; the refusal names the bartizan's line and sliding the banner away clears it. Tighten the box later if it annoys. 256: **Follow-up queued 21 October 01:08 as stage R9 (plan `docs/castle-followup-plan.md`; board stage `R9`):** R9a takes C, F's door cover, G and J, with D as an optional recolour; R9b adds alternative distant views (mountains, forest, hills), with the owner's S1 to S7 answered 21 October (every recommendation taken). Queued after R5f, before R6b. The Harbour slider-release trace named on the sheet as item 1 is already done as R8x (no code change).