Chapter 34: Keys, grips and a camera that holds the ground 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 parts R5b, R5c and R5d of docs/decor-controls-ui-plan.md (20 October; R5d closed 21 October about 00:35, game commit 91ae641), with the orchestrator rulings of 20 and 21 October and the owner's push of 21 October about 01:00. Not covered: R5e (multi-select, running at the freeze) and R5f (gamepad), which are plans here. Excerpts are given in the order of the work. SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 32-43 (Decor R5b Input map and physical keys) 32: **Decor R5b Input map and physical keys (20 October 2026):** Opus 5.5 ran this on the Mac, 20 October about 18:34 to 19:20 (step `33-R5`, part b; `docs/decor-controls-ui-plan.md` R5b). Snapshot `build/backup-feel/33-R5-b-pre/` (= HEAD c71a349; live tree equal to it at the start); scratch pre-tree `b/pre-tree`; backups `b/backup/` (.gdignore'd under `artifacts/`). Evidence `artifacts/decor/33-R5/b/` (`b/` below). Design `b/design.md`. Patch `b/part.diff` (1363 lines): `project.godot` (new `[input]` section), `scripts/main.gd`, `scripts/editor_ui.gd`, `README.md`, `tools/test.sh`, `tools/Test.ps1`, `tests/test_ui.gd`, `tests/test_grips.gd`, and new `scripts/controls.gd`, `tools/controls_table.sh`, `tools/controls_table.gd`, `tests/test_controls.gd`, `tests/controls_recorder.gd` (Godot wrote their `.uid` files). No scene geometry, save format (version 2) or pin touched. 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 Opus 5.5 review 20 October about 19:20 to 19:28 (`b/review/`): verdict done, no must-fix, four notes (the design note's counts were stale, fixed; the map and conflict checks assume key events, so R5f's joypad events need them generalised; dispatch relies on non-exact matching plus test order and the Cmd/Ctrl gate, so R6d keeps 'more modifiers first' and rebuilds the tooltip and photo-bar labels after a rebind; Windows not run). The recorder oracle was cross-checked against the pre-tree's real handlers (13,500-event dump byte-identical). Ruling: undo, redo, save and open bind the key's letter (keycode with command-or-control autoremap), so Cmd+Z stays undo on QWERTZ and AZERTY; every other key is physical; the five new README rows and the help line for photo-mode keys stay. Platforms: Mac run; Windows not run. 33: - **The map.** 34 actions in `project.godot` `[input]`, every event an `InputEventKey` on device -1. Edit: `cancel` Esc, `select_tool` V, `rotate_back` Shift+R, `rotate` R, `rotate_step` `.`, `rotate_step_back` `,`, `move` G, `duplicate` D, `delete_part` Delete + Backspace, `focus` F, `hide_clutter` H, `links` L, `photo_mode` P, `capture` F12, `brush_smaller` `[`, `brush_larger` `]`. Command (`command_or_control_autoremap`: Cmd on macOS, Ctrl elsewhere): `redo` Cmd+Y + Shift+Cmd+Z, `undo` Cmd+Z, `save` Cmd+S, `open` Cmd+O. Modifier: `place_free` Ctrl + Cmd (the key on its own; mouse events still read the modifiers off the event). Photo: `free_camera` C, `dof_less` `,`, `dof_more` `.`, `fov_less` `-`, `fov_more` `=`, `unpin` U. Fly (polled): `fly_forward` W, `fly_back` S, `fly_left` A, `fly_right` D, `fly_up` E, `fly_down` Q, `fly_fast` Shift. **Extra actions beyond the plan's list**: `rotate_step`, `rotate_step_back` (R5a's `,` `.`), `place_free` (the Ctrl/Cmd key test at main.gd:1713), `fly_fast`; the plan's `fly_*`, `dof_*`, `fov_*` are named as above. Documented pairs: `delete_part`, `redo`, `place_free`. No number keys exist, so `select_tool` is one action (V, the Select tool); R5d's 1-4 will be views. 34: - **Physical keys, one deliberate exception.** Every key binds its physical position except the four Cmd/Ctrl shortcuts (undo, redo, save, open), which bind the layout's letter (`keycode`), so Cmd+Z undoes on the key that says Z: physical binding would make QWERTZ's Z key (physical Y) redo and move AZERTY's undo to the key labelled W. On QWERTY nothing differs. Reversible by rewriting four events; `test_controls` names the exception (`Controls.BY_KEYCODE`). This departs from the brief's "physical keycodes" wording for those four: **for the orchestrator to confirm**. 35: - **No behaviour change.** Handlers test `is_action_pressed` in the old match's order (Shift+R before R, Shift+Cmd+Z before Cmd+Z; actions match with extra modifiers held, as the keycode match did), the Cmd/Ctrl gate stays first (without it Cmd+D would duplicate), the `pressed and not echo` guard stays, `_photo_input`'s pass-through tests `is_action` (press or release) as before, `_unhandled_input` still runs after text fields. The tray filter's shortcut letters are `TRAY_SHORTCUT_ACTIONS`/`TRAY_SELECTION_ACTIONS`; its Esc is `cancel`; Backspace, Space and letters stay the filter's own typing. `--verify`'s synthetic Esc now carries `physical_keycode` too (a keyboard's event has both). 36: - **Generated text.** `scripts/controls.gd` (`class_name Controls`) holds contexts, pairs, shares, labels and the help and README templates (`{rotate}`, `{delete_part#2}`, `{place_free#mod}`); keys are read from `InputMap`. Help labels use the OS's label for the physical key (`keyboard_get_label_from_physical`; the headless server does not support it and errors, `b/probe/headless.log.out`, so headless uses the US name). `tools/controls_table.sh` (pinned engine, headless, `tools/controls_table.gd`; `--check` only reports) rewrites README between `` markers; the engine is the generator so README, help and test share one reader and label table. README's 25 old rows are byte for byte; **5 rows added** (C, W A S D / Q E, `,` `.` depth of field, `-` `=`, U). Help: same text line for line plus **one line** for the same photo keys (`b/help/diff-headless.txt`; windowed OS labels equal headless, `b/help/help-live-windowed.txt`). Tooltips reading the map, same text: Save, Open, Photo mode, Undo, Redo, the photo bar's Back and key line, the "R Rotate · Esc Cancel" hint. **Carried to R6d** (key letters still written in strings): main.gd toasts and status lines (live line numbers) :2139, :2677, :3055, :3141, :3189, :3592, :3625, :4203, :4268, :4632, :4638, :4647, :6722; editor_ui.gd's "Hold Ctrl" lines :981, :1112, :1125, :1212, :1263 and the filter's :546, :1594. 37: - **Tests.** Tests first (`b/testfirst/test_controls.log.out`): 34 checks, 4 failures (no Controls; AZERTY W and Q positions; AZERTY `[` position). Now `test_controls` 78/0: the map against the survey (key, modifiers, physical or letter, device, pair counts); no key twice in one context, cross-context shares exactly `,` `.` and D; every `is_action*` name in `scripts/` exists and every action is used; no keycode test left in main.gd; README section equals the generator; every action in both templates; help, tooltips and photo bar text; fly keys, Shift, diagonal; AZERTY W position flies forward, Q position flies down, `[` position (types `^`) shrinks the brush; QWERTZ and AZERTY Cmd + the key that says Z undoes. **Table**: `tests/controls_recorder.gd` extends Main and records the calls the handlers make; 50 keys x 10 modifier sets x press/echo/release x 9 contexts (edit, selected, tray x3, grip drag, body drag, photo, free camera) = 13,500 events compared with an oracle written from the pre-part handler: 0 differ; the dump is identical on the pre-tree and live (`b/suites/table-pre-tree.txt`, `table-live.txt`, md5 916f6039...). **Changed test inputs/expectations**: test_ui Q1a's Ctrl event gains `physical_keycode` (input only); test_grips CG1 "F12 is the capture action" now reads the action line and the map (old: the text `KEY_F12: action("capture")`). Affected suites (`b/suites/summary.txt`): test_ui 590, test_camera 544, test_grips 70, test_building_tools 982, test_tree_grips 79, test_drag_harness 125, test_owner_0924 187, test_decor 2717, test_warmup 55, all 0 failures; R5a's photo-mode and Esc-in-drag checks pass unchanged. 38: - **Engine.** Action matching is the same headless and windowed (`b/probe/headless.log.out`, `windowed.log.out`); only the layout-label call differs (unsupported headless), handled above. 39: - **Unchanged** (`b/pins/summary.txt`, `b/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; starter md5s (LOCH_MD5 included) by test_starters in the full suite. 40: - **Captures** (AC power, `b/cap/power.txt`). `b/sheet/r5b-01-help.png` (help pre vs live, 1440x900 windowed). `b/sheet/r5b-02-editing-view.png`: harbour-day pre x3, live x2: counts identical; whole-frame change pre-pre 5.1/6.8%, after-after 17.2%, pre-after 8.8/10.9/20.7% (rain, ripples, smoke; `b/metrics/harbour-day.txt`). 41: - **Full suite** (`b/suite/suite_all.out`): a, b and v at 2880x1800: 62 suites, 33,026 checks (32,948 + 78), 0 failures; --verify passed on try 1 in 26 s, no stall; test_creature_reactions rerun 6871/0. 42: - **Not done / open:** Windows not run (the command-or-control mapping and the OS labels are untested there). R6d's string hints above. The help dialog runs past the right edge of a 1440 window (pre-existing, same on the pre-tree). No pad bindings (R5f). SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 19-31 (Decor R5c Grips for every prop, bench seats, keyboard nudge) 19: **Decor R5c Grips for every prop, bench seats, keyboard nudge (20 October 2026):** Opus 5.5 ran this on the Mac, 20 October about 19:29 to 20:50 (step `33-R5`, part c; `docs/decor-controls-ui-plan.md` R5c). Snapshot `build/backup-feel/33-R5-c-pre/` (= HEAD 4c9cf47; live tree equal to it at the start but for `artifacts/building-tools/tests.json`, `c/diff-live-vs-head-at-start.txt`); scratch pre-tree `c/pre-tree`; backups `c/backup/` (.gdignore'd under `artifacts/`). Evidence `artifacts/decor/33-R5/c/` (`c/` below). Design `c/design.md` (section 6: what changed while testing). Patch `c/part.diff` (1673 lines): `scripts/main.gd`, `scripts/scene_state.gd`, `scripts/world_view.gd`, `scripts/architecture.gd`, `scripts/edit_history.gd`, `scripts/editor_ui.gd`, `scripts/controls.gd`, `project.godot`, `README.md` (regenerated by `tools/controls_table.sh`), `tests/test_controls.gd`, `tests/controls_recorder.gd`, `tools/test.sh`, `tools/Test.ps1`, new `tests/test_prop_grips.gd` (+ `.uid`). Save format stays version 2; the one new key (`params.seats`) is optional. The modified `artifacts/building-tools/tests.json` and the other session's untracked `docs/codebase-audit-plan.md` are not this part's. Review: fresh Opus 5.5 review 20 October about 20:55 to 21:15 (`c/review/`): must-fix one (the arrow-time focus release in `Main._input` took the arrows from a focused inspector slider while a part was selected); fix pass 1 21:15 to 21:37 (only a plain button's focus is let go on an arrow press; a left press in the 3D view lets any focus go but a text field's; the nudge waits while a control holds the focus); fresh recheck 21:38 to 21:52 (`c/review2/`): one more must-fix (holding Up or Down on a focused slider handed the focus to a button, and the next key repeat let that go and nudged the part), fixed by the orchestrator at 21:55 with `and not event.echo` on the release condition and a regression test in test_prop_grips (fails without the fix, 99/0 with it; the recheck's hold probe on the live tree: the house no longer moves, `c/review2/probe/hold-live2.log.out`). Full suite after that fix: 63 suites, 33,125 checks, 0 failures, verify clean (`c/fix2/suite/`). The recheck's other notes: a second separate Down press after the first on a focused slider can still nudge (the disclosed finding-4 class); the recheck's real-window clicks landed on toolbar buttons in headless, so GUI routing to the 3D view was tested by calling the handlers directly. Rulings in `artifacts/decor/owner-decisions.md`: the cold nudge release (41.8 ms) is a harness artifact, like for like the nudge is cheaper than a move release (Harbour pot 28.8 against 35.5 ms warm, Loch brazier 25.4 against 31.3, Loch banner 84.6 against 88.6); the nudge rewriting `rot` from `get_euler` is the move path's existing behaviour; a held arrow on a cluster-hosted part commits 85 to 144 ms steps (carried to the owner with the banner/torch hitch). Platforms: Mac run; Windows not run. 20: - **Turn grip on props.** `DioramaState.prop_turns(o)` (data, no scene walk): every kind but buildings and towers, the length parts, the trees, the wall parts, the connections, `wall_path`, `crag`, and `bartizan`/`belfry` (their turn is re-derived from their seat by Adapt, so a grip could never hold). Those 46 kinds gain one turn grip (`rot`, shape `turn`, the buildings' violet, the same mesh and pick box) at the base, 0.45 m off the box's -x/+z corner. Drag: `_drag_turn_step`, factored out of R1a's boat turn (continuous, Shift 15 degrees, last valid turn kept, boat clearance kept); release is one step "turn part"; Esc cancels. **Grip table** (`c/probe/gripsets.gd`, all 88 kinds, pre vs live; grouped in `c/probe/grip-table.txt`): 47 kinds change, all from no grips: the 46 `rot1:turn` and bench `length1:arrow,rot1:turn`; the 16 kinds that had grips are byte-identical (building and tower 8 grips, boat/quay_fence/gate/iron_arch/stall, bike_rack, tree/blossom/conifer, neon/neon_bar, pipe, wall_path, crag); 25 kinds keep none. `HANDLE_PICK_RADIUS` 0.36 and `handle_pick_box` untouched. 21: - **Bench seats.** `params.seats` 2-4 (absent = 3) as a count part on the shared length grips (R3a's rack pattern, generalised: `count_range`, `count_length`, `length_count`, `part_count`): 1.1 / 1.6 / 2.1 m; the -x end stays, the +x arrow steps whole seats; inspector Seats row; `validate_document` refuses a non-integer or out-of-range value ("Invalid bench seats."); `restore` makes it an int. A bench without the key draws today's code path: hash 53542753... for every variant and seed, box and dimensions identical pre vs live (`c/probe/bench-pre.log.out`, `bench_pins-live.log.out`); explicit `seats: 3` draws the same. 22: - **Arrow nudge.** Eight actions in `[input]` and `Controls`: `nudge_left/right/forward/back` on the physical arrows (edit), and `*_free` with `command_or_control_autoremap` (Cmd on macOS, Ctrl elsewhere); Ctrl+arrow on a Mac is free too. No clash: nothing else used the arrows; photo mode swallows them and the fly keys (W A S D Q E, Shift) are unchanged. 0.25 m (Shift 1 m) along the camera's flattened right/forward. Each step runs the move's own path with a synthetic hit (no pick, no scene walk): `object_candidate`, `_settle_candidate` (the preview's validate/Adapt code, factored out, now called by the preview too), then begin, replace, `adapt_all`, commit, `_sync`. **Undo**: `EditHistory.commit(state, label, merge_key, tag)` folds a step into the newest entry when it carries the same key; a step merges when the same part is nudged again within 500 ms of the last key-up and the newest entry is still the nudge's: four in 0.3 s are one "Nudge", two 1 s apart are two, undo restores byte for byte. No timer, no `_process` change. Held keys repeat at most every 90 ms. **Behaviour by host** (test_prop_grips): off a terrace's edge it stops at the edge ("Stopped at the edge..."), Cmd/Ctrl keeps it overhanging (soft); a ground prop over water stands on a plinth; a window slides along its wall, Up/Down moves it up and down; a crag-top part stays on the top; a crag nudges with its tops; a boat on a quay berth slides along the quay; a berthed boat is held ("Held here by its magnet. Hold Ctrl/Cmd to nudge it as drawn."), Cmd/Ctrl frees it; a boathouse nudged takes its boat; a freehand wall shifts with its parts; stairs and bridges get the body drag's line; a house nudged over a crag's edge is refused with Adapt's own hint and nothing moves. **Added while testing**: a toolbar button or slider holding keyboard focus ate the arrows (GUI focus navigation, `c/probe/focus-before-release.log.out`); `Main._input` now lets a `BaseButton`/`Range` focus go when a nudge arrow arrives with a part selected (`c/probe/focus-live2.log.out`); text fields keep their arrows. 23: - **Tests.** Tests first (`c/testfirst/`): test_prop_grips 14 checks, 8 failures; test_controls 78 checks, 10 failures. Now **test_prop_grips 90/0** (new; registered in test.sh and Test.ps1) and test_controls 78/0. **Changed expectations** (test_controls): EXPECTED gains the 8 actions; the recorder oracle gains the arrows (press/echo → `nudge(...)`, release → `nudge_released()`); the 13,500-event dump differs from the pre-tree's in exactly **420 lines**: Up and Left (the two arrows in the table) x 7 non-photo contexts x 10 modifier sets x 3 phases; photo contexts unchanged (`c/suites/table-diff.txt`). README: the Drag handle row gains a sentence on props' turn arrow and the bench's seat arrow; **two rows added** (arrows; Ctrl+arrows "(Cmd on macOS)"). Help: one line added. The pre-tree's own test_controls run fails only its README check (the scratch tree has no README). Affected suites (`c/suites/`): owner_0924 187, tree_grips 79, building_tools 982, decor 2717, grips 70, ui 590, bikes 499, boathouse 200, drag_harness 125, camera 544, scene 1736, adapt 995, starters 463, composition 160, crag 200, grounds 201, presentation 44, all 0 failures. 24: - **Unchanged** (`c/pins/summary.txt`, `c/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; starter md5s (LOCH_MD5 included) by test_starters 463/0. 25: - **Captures** (1600x1000 windowed, AC power, `c/cap/power.txt`; `c/caps.sh`; `c/sheet/README.md`). `c/sheet/r5c-01-prop-grips.png`: pot, bench of 2 and 4 seats, bike, boat selected, pre row above live row: the turn grips and the bench lengths are the only differences; the boat's grips identical. `c/sheet/r5c-02-nudge-film.png`: a pot nudged 1 m a step to the quay edge and onto a plinth in the water (`c/probe/film.gd`, real `Main.nudge`). Editing view harbour-day pre x2, live x2 (`c/metrics/metrics.txt`): integer counts identical; frame change pre-pre 22.2%, live-live 13.9%, pre-live 27.8/21.8% (rain and ripples). 26: - **Acceptance.** Pick radius unchanged (test). **Nothing selected**: harbour-day vsync off 3+3 (`c/bench/bench.txt`): p50 6.38 pre / 6.36 live ms, draw calls 970 both. **Harbour handle drag** (`c/drag/summary.txt`, `slider-and-handle-by-run.txt`, A0b script, 6+6 interleaved): step p95 median 55.2 pre / 57.1 live (runs 53.2-61.0 / 54.4-61.2), within noise of 55-56. Releases within +2 ms: handle +1.27 (bimodal 9.5/12.5 on both trees), slider +0.01 (live 2 of 6 runs hit R8x's 49 ms waiting mode, which R8x ruled a known A2c pacing mode; fast-mode releases equal), the rest within 0.2. **Release probe** (`c/cost/summary.txt`, headless, 3+3): pot placement 28.34 / 28.36 (+0.02); G-move release 36.28 / 37.23 (+0.95); **nudge 41.8 ms**, +5.5 over the pre move release, **over the +2 ms budget as measured cold**. Traced: the move's release follows 4 frames of preview; the nudge follows 4 idle headless frames, after which every sync phase is 20-30% slower, terrain included (`c/cost/phases3-live.log.out`); with the same 4 preview frames run before it the nudge is 29.5-30.4 ms against that run's move release 36.1-36.4 (`c/probe/warm.gd`, 3 runs). Pieces (temporary instrumentation, removed; main.gd md5 checked): candidate 9 warm / 15 cold, adapt 8 / 11, commit and sync 13 / 15, grips 0.1. **For the orchestrator to rule** whether the cold figure stands against the budget. 27: - **Full suite** (`c/suite/suite_all.out`): a, b and v at 2880x1800: 63 suites, 33,116 checks (33,026 + 90), 0 failures; --verify passed on try 1 in 27 s; test_creature_reactions rerun 6871/0. 28: - **Fix pass 1** (Opus 5.5, 20 October about 21:06 to 21:36; evidence `c/fix1/`, backup `c/fix1/backup/` .gdignore'd). **Review found** (must-fix 1, `c/review/focus_probe.gd`, `c/review/probe/focus-pre.log.out` and `focus-live.log.out`): `Main._input` let the focus go from any `Range` (and any `BaseButton`) on a nudge arrow with a part selected, so a focused inspector slider no longer took the arrows: pre, Right on a building's Width slider gave `w` 4.0 -> 4.25 (Depth the same); live the slider stayed at 4.0, the focus was dropped and the house moved 0.25 m; a focused option button nudged too. **Fixed** (`scripts/main.gd` only): `_input` lets go only a plain button's focus (a `BaseButton` that is not an `OptionButton` or `MenuButton`, a popup's trigger), on the key press only (a key-up let a button's focus go after a slider's Up had moved it there); `_nudge_input` does not nudge while any control still holds the focus (a slider, an option button, a text field: an Up or Down a slider passes on to focus navigation that finds no neighbour no longer nudges); new `_world_press_focus` lets the focus go at any left press in the 3D view outside photo mode (a select, a press on the selected part's body or a grip, a placement, a stroke), so a slider focused earlier no longer holds the arrows after the hand clicks a part, while a slider focused after selecting keeps them; text fields keep their focus there (and their letters, as before R5c). **Kept as disclosed (finding 4):** arrow-key focus moves between toolbar buttons and tray cards end while a part is selected and the focus is on a plain button; Tab still moves the focus. **Tests** first (`c/fix1/testfirst/test_prop_grips.log.out`: 98 checks, 3 failures: the slider table, the option buttons, the body press); now **test_prop_grips 98/0** (`c/fix1/suites/`): every visible inspector Range (the building's four `HSlider` rows, the only Range kind the inspector has) x Right/Left/Up/Down pushed through the window steps or moves the focus exactly as the pre-tree, the part unmoved; every visible option button keeps its arrows; a text field's caret moves 2 -> 3/1/0/4; after a select press on another part and after a press on the selected part's body, a slider's focus is let go at the press and Cmd/Ctrl+Right nudges 0.25 m (the select press case also passed before the fix, because the select rebuilds the inspector); the plain-button check stays; photo mode and its free camera neither nudge nor let a focused button go. (The world-press checks were moved to Cmd/Ctrl+Right after the test-first run, so a neighbour's magnet cannot bend the step.) The reviewer's probe on live (`c/fix1/probe/focus-live.log.out`): Width 4.0 -> 4.25 and Depth the same, the option's Down moves the focus to a button, as pre. `c/fix1/probe/focus_table.gd` (53 rows: four sliders and eight option buttons x four arrows, a text field x four) gives the same lines pre and live (`table-diff.txt` empty). Affected suites (`c/fix1/suites/summary.txt`): controls 78, ui 590, camera 544, owner_0924 187, presentation 44, building_tools 982, decor 2717, grips 70, tree_grips 79, drag_harness 125, boathouse 200, adapt 995, all 0 failures. R5b's 13,500-event recorder dump is byte for byte R5c's live dump, so it still differs from the pre-tree's in exactly the 420 arrow lines (`c/fix1/recorder/`). Pins live against the pre-tree lines (`c/fix1/pins/summary.txt`): R7a 37, R7c 66, bridge 84, ident 15, boats 54, creature 959, gates 743, diff 0; fire 23 and R8d's 111 looks diff 0 (`c/fix1/hash/summary.txt`); the starter md5s (LOCH_MD5 included) by test_starters in the full suite. **Full suite** (`c/fix1/suite/suite_all.out`, AC 100%): a, b and v at 2880x1800: 63 suites, 33,124 checks (+8), 0 failures; --verify try 1 in 27 s; test_creature_reactions rerun 6871/0. No bench re-run: the change is a focus test per arrow key and per world press. 29: - **Rulings** (orchestrator, 20 October about 21:15; reversible by the owner; no code): the **nudge release cost is accepted**: like for like the nudge is cheaper than a move release in every case (Harbour pot 28.8 against 35.5 ms warm; Loch brazier 25.4 against 31.3; Loch banner 84.6 against 88.6), and the cold 41.8 ms figure was a harness artifact. **Finding 2** (a nudge rewrites `rot` from `get_euler`, 4.0 becoming -2.2831852, the same world turn with float32 noise; a G move does the same) is accepted as the move path's existing behaviour and carried. **Finding 3** (a held arrow on a cluster-hosted part such as a banner on the Loch castle commits every 90 ms step, 85-144 ms each) is carried to the owner as a hitch of the same class as the banner and torch placement cost. 30: - **Not done / open:** Windows not run (Ctrl+arrow free nudge and the Ctrl/Cmd labels untested there). Multi-select nudge (R5e) and a pad d-pad nudge (R5f) not built. The CG1 carried notes are carried on, not measured: grips one frame late after a select (a nudge re-lays them in the same call), and the first body-drag frame after `--warmup=startup`. The toasts name "Ctrl/Cmd" and "Shift+arrow" in strings (R6d). The arrows let a clicked plain button's focus go, which ends arrow-key focus moves between toolbar buttons and tray cards while a part is selected (Tab still moves focus); sliders, option buttons and text fields keep their arrows (fix pass 1). Findings 2 and 3 carried (rulings above). SOURCE: docs/visual-handoff.md (frozen 21 October 2026), lines 5-18 (Decor R5d Camera) 5: **Decor R5d Camera (20 October 2026):** Opus 5.5 ran this on the Mac, 20 October about 22:13 to 23:20 (step `33-R5`, part d; `docs/decor-controls-ui-plan.md` R5d). Snapshot `build/backup-feel/33-R5-d-pre/` (= HEAD af575ff; live tree equal to it at the start but for `artifacts/building-tools/tests.json`, `d/diff-live-vs-pre-at-start.txt` empty); scratch pre-tree `d/pre-tree`; backups `d/backup/` (.gdignore'd under `artifacts/`). Evidence `artifacts/decor/33-R5/d/` (`d/` below). Design `d/design.md` (section 8: what changed while testing). Patch `d/part.diff` (1446 lines): `scripts/main.gd`, `scripts/camera_rig.gd`, `scripts/user_prefs.gd`, `scripts/controls.gd`, `project.godot` (seven actions), `README.md` (table regenerated by `tools/controls_table.sh`; the Feel D2 sentence and a new "Camera preferences" section), `tests/test_controls.gd`, `tests/controls_recorder.gd`, `tools/test.sh`, `tools/Test.ps1`, new `tests/test_view_controls.gd` (+ `.uid`). No geometry, save format (version 2) or document key touched: the camera preferences live in `preferences.cfg`. The modified `artifacts/building-tools/tests.json` and the other session's untracked `docs/codebase-audit-plan.md` are not this part's. Review: fresh Opus 5.5 review 20 October about 23:25 to 23:45 (`d/review/`): must-fix two (after a toolbar button was clicked, Space armed the pan and also fired that button on release; the corrected pan did nothing in the sky and flung the target near the horizon) and notes (the Q-orbit sentence overstated; digit keys turn the view while a tray filter is live; a pan over a crag top or roof holds the terrain level; the FOV preference is written on every Scenery change); fix pass 1 20 October 23:45 to 21 October 00:08 (a plain button's focus is let go on a Space press with the pointer over the view; grazing or missing rays fall back to a linear pan at the view-centre scale, blended by smoothstep between |dir.y| 0.1 and 0.2: worst 1 px move 45.3 m to 1.245 m, 0 flung to the clamp, 0 do-nothing pans); fresh recheck 00:11 to 00:30 (`d/review2/`): verdict done, two notes (inside the blend band the pan speed rises about 9% and falls to the centre's speed, so a loop through the sky does not return exactly to its start; the tray-filter sentence in this entry now says the button is still pressed on release, as before R5d). Suite 64 / 33,274 / 0, every pin diff 0, verify clean. The pan feel change (half as fast across at pixel size 2, up and down follows the pointer, because the old pan was wrong) is intended and stated in the README. Rulings in `artifacts/decor/owner-decisions.md`. Platforms: Mac run; Windows not run. 6: - **Review and fix pass 1** (20 October about 23:38 to 21 October 00:10, Opus 5.5; `d/fix1/`, brief `d/fix1-prompt.md`, patch `d/fix1/fix.diff`: `scripts/main.gd` and `tests/test_view_controls.gd` only; design section 9; backup `d/fix1/backup/`, .gdignore'd). **Finding 1**: after a click on a toolbar button (which keeps the keyboard focus) Space armed the pan on the press and fired the button on the release (`d/review/probe/probe4.log.out`: press alone pressed 0 and pan held; release pressed 1, Undo history 2 -> 1; a focused Garden or Scenery button opened the tray), unlike this entry's old claim that no pan was armed. **Fixed** (orchestrator ruling, R5c's arrow precedent): `Main._input` lets a plain button's focus go on a Space press (not echo, not photo mode, no dialog, no live tray filter, no Cmd/Ctrl, the pointer not over the GUI), so Space after clicking Undo, Select, Garden or Scenery with the pointer over the view only arms the pan: pressed 0, history and tray unchanged, Space+drag pans; Enter still presses a focused button. And no pan is armed while a button still holds the focus (the pointer over the GUI, an option or menu button): over the GUI the focused button takes Space as before R5d and nothing else; a focused option button, whose popup opens on the press and takes the release, had left the pan armed (found by the new test; fixed). A text field, a live tray filter (it types the space and the button keeps its focus (and is still pressed on the release, as before R5d)) and a dialog are as before. **Finding 2**: rays that miss or graze the plane (|dir.y| < 0.02: the sky, the top third at fov 60 and 14 degrees) panned nothing; just below that 1 view px moved the target 9 to 31 m and 3 px flung it to the 32 m clamp; orthographic bottom corners at size 95 and 14 degrees did nothing (the plane point behind the camera). **Fixed** (reviewer's remedy): exact where both rays come down at least as steeply as |dir.y| 0.2, linear at the view centre's scale where either is below 0.1 or misses (`_centre_pan`, the same pan at the centre, and in orthographic everywhere; `_linear_pan`, the old formula corrected, if the centre's ray misses too), smoothstep between (`pan_exact_band`). Measured (`PAN_EDGE`/`PAN_SWITCH` lines; 24 set-ups: both projections, perspective at fov 60, pixel size 1 and 2, pitch 14 and 72 degrees, size 10/50/95, drags of 1 and 3 view px four ways from the sky, the horizon band at |dir.y| 0.005 to 0.3, the middle and the four corners, and a row scan of the 1 px move): tests-first (pre-fix) worst 1 px move 45.3 m, motion jump 82 view px between rows 4 apart, 18 flings to the clamp, 142 pans that did nothing; now worst 1.245 m (perspective, size 95, 14 degrees, from |dir.y| 0.15), jump 0.171 view px, 0 flings, 0 still; a hard switch at 0.1 would jump 4.17 view px, so the blend stays. Where the ray comes down steeply the ground point stays under the pointer (worst 0.00056 view px; the 192 PAN_MEASURE drags now held to 0.001). **Finding 3** (entry only): the "Not fixed" Q sentence below is corrected. **Rulings** (orchestrator, in the fix1 brief, 20 October about 23:38, recorded 21 October 00:07; reversible by the owner; no code): digit keys turn the view while a tray filter is live (it only ever took letters); a pan that starts over a crag top or a roof holds the terrain level under it (orthographic unaffected); the FOV preference is written on every Scenery change, non-atomically, like the other preferences (a read-only file warns and the session carries on); carried to R6c. **Tests**: tests first `d/fix1/testfirst/` (test_view_controls 146 checks, 19 failures on the pre-fix tree); now 146/0 (`d/fix1/suites/`); the old line-492 test, which read `pan_held` only after the release, now reads it between the press and the release. Changed expectations: the focused-button Space case (pointer over the view: the button does not fire, the pan arms); the pan-scale threshold 0.05 -> 0.001 view px. The 13,500-event recorder dump is identical to the implementer's (md5 145c66d9..., `d/fix1/recorder/`; the recorder drives `_unhandled_input` with no focused button), and so is the eased recording (322 lines, `d/fix1/recording/`: the art-pixel snap holds). Affected suites (`d/fix1/suites/summary.txt`) 14, 0 failures. Pins (`d/fix1/pins/summary.txt`, `d/fix1/hash/summary.txt`): R7a 37, R7c 66, bridge 84, ident 15, boats 54, creature 959, gates 743, fire 23, R8d's 111 looks, diff 0. Full suite `d/fix1/suite/suite_all.out`: 64 suites, 33,274 checks (+27), 0 failures; --verify try 1 28 s at 2880x1800; test_creature_reactions rerun 6871/0. 7: - **Keys** (`project.godot` `[input]`, `Controls`, physical keys, edit context): `orbit_left` Q and `orbit_right` E turn the goal yaw +/-45 degrees about the middle of the view (target, pitch and size kept), eased with FOCUS_TAU like F; a press while easing adds to the goal; first press only (echo ignored); a running spin is dropped. `view_1`..`view_4` (1 2 3 4) turn to `Main.corner_yaw(n)` = 0.78 (the Harbour/Courtyard opening yaw, view 1), 2.3508, 3.9216, 5.4924 (the way Q turns, so two Q presses from view 1 are view 2), the short way round (half a turn: the Q way), at the current pitch, size and target. `pan_hold` Space: held, a left drag pans. Photo mode swallows all seven (the free camera keeps Q and E for flying: Q/`fly_down` and E/`fly_up` join `Controls.SHARED`); Cmd/Ctrl with them does nothing (on a Mac Ctrl+Q turns as Q, as Ctrl+R turns a part); with the part tray open and no filter, Q and E still start the filter (R6a). Camera never in undo. 8: - **Space and Alt: today's meanings kept.** Space+left drag pans with the middle button's code and does nothing else; a live tray filter and a text field still take Space. A **focused button** (fix pass 1, above; this sentence corrected): with the pointer over the 3D view a plain button's focus is let go on the Space press, so Space arms the pan and the button does not fire; over the GUI, or on an option or menu button, the button takes Space as before R5d and no pan is armed; Enter presses a focused button. **Alt+left** keeps every old meaning: the selected part's body (copy drag), grips, a pending move or copy, strokes, terrain and placing tools, a creature, an unselected part or derived group; only where the press would fall to `select_at_cursor` with nothing under the pointer (Select or Scenery tool, open ground) does it orbit, with the right button's amounts and release spin; let go without travelling 5 view px it is the old click (deselects). 9: - **Pan scale measured first** (`d/testfirst/test_view_controls.log.out`, PAN_MEASURE lines; 100 window px drags, pixel size 1 and 2, frame factor 1 and 2 with the canvas at the UI scale, pitch 0.25 to 1.25, yaw 0.78 and 2.4, both projections). **Before**: across the screen the ground moved **pixel size x** the pointer (1.000 / 2.000; perspective 1.005 to 2.119), and up and down it moved **the opposite way**, -(pixel size x sin pitch): -0.247 to -0.949 at size 1, -0.495 to -1.898 at size 2 (errors up to 195 view px). **Fixed**: a pan drag holds the ground: the level of the terrain under the pointer where the drag starts, and each motion moves the target by the difference of that plane's points under the previous and the current pointer, through the drawn camera's rays (exact in both projections, unaffected by the snap and the easing). **After**: 1.000 in all 128 measured drags, worst error 0.00 view px (192 drags checked). This changes the middle-drag pan feel: half as fast across at the default pixel size 2, and up and down now with the pointer instead of against it. Where the ray under the pointer misses or grazes the plane the pan is linear at the view centre's scale, blended in (fix pass 1, above). The two-finger pan gesture keeps today's formula (it has no pointer to hold). 10: - **Preferences** (`UserPrefs` `[camera]` of `preferences.cfg`, `CAMERA_DEFAULTS` = today): `orbit_sensitivity`, `zoom_sensitivity`, `pan_sensitivity` 0.5 to 2.0 (1.0); `invert_orbit`, `swap_buttons` (false); `view_pivot` cursor/centre, `projection` orthographic/perspective, `fov` 12 to 60 (24). `camera_value`, `set_camera_value`, `camera_prefs`, `validate_camera`, `reload_camera`; cached per path (`camera_loads`), so no gesture reads the file; out of range clamped, wrong types, NaN/inf, unknown words and corrupt or missing files give the defaults, unknown keys ignored, other sections kept. Main reads them through `camera_pref` only while `prefs_live` (every automated run and test: the defaults). Orbit drags x s (invert negates turn and tilt), wheel and pinch factor^s, pan drags and the two-finger pan x s; Q/E, 1-4 and F unscaled; swap: middle orbits (and carries the release spin), right pans, Alt and Space unchanged, the free camera still looks with the right button. Scenery's pivot, projection and lens are written while `prefs_live` and applied at an interactive launch (`apply_camera_prefs`); photo mode's lens never writes them. No UI (R6c). 11: - **Found while recording**: corner view 3 from view 1 is exactly half a turn; the goal went one way and the ease the other, so the view settled a whole turn from its goal and `CameraRig.settled` (raw yaw) never said so (240 frames, camera_tau left at FOCUS_TAU). Fixed: `view_corner` takes half a turn the Q way, and `CameraRig.settled` compares the yaw with `angle_difference` (a view a whole turn from its goal has arrived and lands on the goal's number). Now 137 frames. Not fixed (corrected in fix pass 1 from the review's probe, `d/review/probe/probe.log.out`): the eased view can step backwards (5.6 to 11.2 degrees) only with extreme gaps between Q presses: 4 presses in one frame, 5 presses 17 ms apart, 6 at 33 ms or 8 at 50 ms; never at gaps of 67 ms or more; so a player meets it only during a frame hitch; every case still lands on its goal. Three quick presses end at 135 degrees (tested). 12: - **Tests.** Tests first (`d/testfirst/`): test_view_controls 84 checks, 39 failures; test_controls 81 checks, 14 failures. Now **test_view_controls 119/0** (new; registered in test.sh and Test.ps1; 146/0 after fix pass 1) and **test_controls 81/0**. **Changed expectations** (test_controls): EXPECTED gains the seven actions; the shares gain Q (`fly_down`/`orbit_left`) and E (`fly_up`/`orbit_right`); the README row checks gain `| Q / E |`, `| 1 2 3 4 |`, `| Space + drag |`; the grep check reads names with digits; the recorder oracle gains Q -> `orbit_step(1.000)`, E -> `orbit_step(-1.000)`, 1/2 -> `view_corner(n)`, Space -> `set_pan_held(true)` on a press. The **13,500-event dump** differs from the pre-tree's in exactly **196 lines** (`d/recorder/table-diff.txt`; pre-tree dump md5 83f7f13d..., equal to R5c's): all presses; Q 36, E 36, 1 42, 2 42, Space 40; edit 30, edit_selected 30, grip_drag 30, body_drag 30, tray 26, tray_selected 26, tray_live 24 (Alt+key past the tray, and its live filter's Space with Alt); photo and photo_free none; echo and release none. README: four rows added (Q / E; 1 2 3 4; Space + drag; Alt + left drag on open ground); help: one line added. `Controls.fill` reads placeholders with digits. Affected suites (`d/suites/summary.txt`): camera 544, ui 590, grips 70, prop_grips 99, owner_0924 187, presentation 44, building_tools 982, tree_grips 79, drag_harness 125, creatures 250, warmup 55, scene 1736, all 0 failures (camera and ui rerun after the settle fix). 13: - **Acceptance.** The pan test passes (above). **Eased moves on the art-pixel grid** (snap on, 60 fps, `d/recording/recording.txt`, `summary.txt`): Q orbit 115 frames, corner view 3 137, middle pan 35, Space pan 35: the drawn look-at point on the grid every frame (worst 0.0000 px), residual within half an art pixel (worst 0.4987), the eased target at centre + residual (< 0.01 px), the pans' static points moving by whole art pixels every frame. 14: - **Unchanged** (`d/pins/summary.txt`, `d/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; starter md5s (LOCH_MD5 included) by test_starters in the full suite. 15: - **Captures** (1600x1000 windowed, AC power, `d/cap/power.txt`; `d/caps.sh`; `d/sheet/README.md`): `d/sheet/r5d-01-corner-views.png` (the four corner views of the Harbour by day and the Loch castle at 20:45); `r5d-02-q-filmstrip.png` (one Q press, 13 frames of the recorded ease, `--snap`); `r5d-03-default-view.png` (harbour-day, default camera and prefs, pre and live). Noise (`d/metrics/noise-matrix.txt`, three captures a tree): pre-pre 6.8 to 19.1%, live-live 7.4 to 15.8%, pre-live 8.9 to 21.7% of the frame changed (rain, ripples); integer counts identical. 16: - **Full suite** (after fix pass 1: `d/fix1/suite/suite_all.out`, 64 suites, 33,274 checks, 0 failures, above). Before it (`d/suite/suite_all.out`, after the settle fix): a, b and v at 2880x1800: 64 suites, 33,247 checks (33,125 + 119 new + 3 in test_controls), 0 failures; --verify passed on try 1 in 28 s; test_creature_reactions rerun 6871/0. The run before the settle fix (`d/suite-before-settle-fix/`): a and b 64 suites, 0 failures. 17: - **Not done / open:** Windows not run (the command key, the OS key labels and Alt on Windows untested). No preferences panel (R6c). A pad's camera (R5f). Alt orbits only on open ground with the Select or Scenery tool; Q held does not repeat; Q presses at frame-hitch gaps can step the eased view backwards (above). Fix pass 1 made no new captures (it changes no default view: the eased recording is identical) and was not re-reviewed yet; the README's "a pan drag keeps the ground point under the pointer" holds except where the ray grazes the horizon (not reworded). The two-finger pan gesture keeps its old formula (scaled only by pan sensitivity). SOURCE: artifacts/decor/owner-decisions.md (frozen 21 October 2026), lines 222-238 (R5b, R5c and R5d rulings and closes, and the owner push of 21 October) 222: ## R5b close, 20 October 2026 about 19:30 223: Review done, no must-fix. Rulings (reversible by the owner): undo, redo, save and open bind the key's letter (keycode, command-or-control autoremap) so Cmd+Z stays undo on QWERTZ and AZERTY; every other key is physical; the five photo-mode README rows and the help line stay. Carried: R5f generalises the map checks for joypad events; R6d keeps 'more modifiers first' dispatch and rebuilds labels after a rebind. 225: ## R5c review ruling, 20 October 2026 about 21:15 226: Review (fresh, Opus): must-fix open. The focus fix in Main._input dropped the arrow keys from a focused inspector slider while a part is selected (Width 4.0 stayed, the house moved). Ruled: fix pass keeps arrows on focused sliders, spin boxes, text fields and popups, releases GUI focus on a world click and on arrow only for a plain button. Accepted (reversible by the owner): the nudge release cost (like-for-like cheaper than a move; the cold 41.8 ms was a harness artifact); the nudge rewriting rot from get_euler (4.0 becomes -2.2831852, as a G-move does). Carried: a held arrow on a cluster-hosted part (banner on the Loch castle) commits 85-144 ms steps, the same class as the banner/torch placement cost. 228: ## R5c close, 20 October 2026 about 22:15 229: Recheck found one more must-fix (key repeats of a held Up/Down on a focused slider let a button's focus go and nudged the part); fixed by the orchestrator (`not event.echo`), regression test added, suite 63 / 33,125 / 0. Carried: the second separate Down press after the first can still nudge (the disclosed focus class); CG1's two notes unmeasured; held arrow on a cluster-hosted part hitches. 231: ## R5d review ruling, 20 October 2026 about 23:38 232: Review (fresh, Opus): must-fix open. 1) After clicking a toolbar button, Space arms the pan and also fires that button on release. 2) The corrected pan breaks near the horizon: grazing rays do nothing or fling the target (up to the 32 m clamp). Ruled: on a Space press with the pointer over the view, a plain button's focus is let go (as R5c did for the arrows; Enter still activates buttons); grazing or missing rays fall back to a linear pan at the view-centre scale. Accepted (reversible): digit keys turn the view while a tray filter is live; a pan starting over a crag top or roof holds the terrain level under it; the FOV preference is written on every Scenery change (carried to R6c). The pan feel change (half as fast across at pixel size 2, up/down follows the pointer) is intended: the old pan was wrong. 234: ## R5d close, 21 October 2026 about 00:35 235: Recheck done, two notes. Accepted: the pan speed bump inside the smoothstep band (about 9%, a loop through the sky does not return exactly); the README wording now says the pan is exact only in the well-conditioned part of a perspective view. Carried to R6c: the FOV preference write; a preferences panel for the camera keys. 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.