Adopt a two-drawer UI grammar for global controls and contextual actions

[timohueser]

Origin: the UX redesign that followed #1510. The original squeeze-chord proposal solved discoverability one gesture at a time, but the resulting grammar was still fragmented. This issue replaces it with the two-layer model we arrived at after prototyping the interactions on the real 240×320 RGB222 rendering path.

Prototype branch: prototype/two-drawer-ui-grammar
Prototype commit: 4d36ff8

The branch is a behavioral reference, not a visual specification. Its layout, icon weight, spacing, typography, and final animation curves are intentionally rough. It exists so the interaction can be exercised in the simulator before production architecture and visual polish are committed.

Goal

Give the four-button UI one small, predictable grammar that keeps frequent actions immediate while making secondary actions and settings easy to find:

• ordinary taps operate the current screen;
• Select holds remain local, explicit object/direct-manipulation actions;
• Back hold always escapes to the main menu;
• the upper squeeze opens device-wide controls from the top;
• the lower squeeze opens screen-specific actions and settings from the bottom.

A rider should not need to remember a collection of unrelated hidden gestures, and a screen-specific setting should have one obvious home rather than also appearing deep inside a central settings tree.

Interaction grammar

Always true

Input	Meaning
Select tap	Primary action / confirm on the current screen
Back tap	Back / cancel on ordinary screens
Back hold	Go directly to the main menu from anywhere
Select hold	Local action on the focused object or direct manipulation mode; never the generic way to open a menu
Up + Select	Open the universal quick drawer from the top
Down + Back	Open the current screen's contextual drawer from the bottom, when that screen declares one
Up + Down	Reserved
Select + Back	Reserved; shutdown no longer gets a hidden chord

Object-specific Select holds stay direct. “Forget phone” on the paired phone is the canonical example: burying it in a generic menu would make it less discoverable, not more. The same principle may justify keeping map zoom/pan direct if hardware testing shows that the extra drawer step is too costly.

Riding exception

The two most frequent ride interactions remain one-tap actions with no holds or menus:

• Select tap pauses/resumes the ride.
• Back tap switches ride views.

Secondary ride actions such as Up ahead and Detour move into the contextual drawer. The existing ride compass menu is removed rather than retained as a third navigation grammar.

Universal quick drawer

Up+Select opens a compact sheet that animates down from the top. It is available globally except where a genuinely blocking safety/modal state must suppress all unrelated input.

The root contains exactly four icon-only buttons in the current prototype order:

1. lightbulb — brightness;
2. Bluetooth rune — BLE on/off;
3. gear — central settings;
4. universal power symbol — shutdown.

Do not add labels below these four icons. These symbols are established enough to be clearer and calmer without them. Selection/focus and BLE state still need unambiguous palette-safe visual treatment.

Behavior:

• Brightness: opens a nested value editor with a small number of discrete levels. Up/Down previews, Select commits, Back cancels. Production wiring needs a real backlight platform port and persistence policy.
• BLE: toggles Settings::ble_enabled immediately, persists through the existing settings path, and reaches the board's existing radio-enable seam. The icon and state indicator update immediately.
• Settings: replaces the sheet with the existing central settings entry point. Central settings retain truly device-wide configuration, not screen-specific modifiers.
• Power: opens a clearly distinct confirmation page. A tap must never shut down. Hold Select fills the established guarded-action progress bar; Back or a tap cancels. Completion invokes a platform power-off port (system-off on hardware; a deliberate simulator behavior).

Nested pages transition horizontally while the containing sheet adapts its height. Back returns to the icon row without closing the underlying screen.

Contextual drawer

Down+Back opens an adaptive-height sheet from the bottom only on screens that actually have secondary actions or modifiers. Unsupported screens do nothing; they must not show an empty drawer.

Examples already explored in the prototype:

• ride views: Up ahead, Detour, POIs, Routes, Map display;
• Up ahead: POI filter and source selection;
• route planning from the POI flow: bike type and route options;
• weather screens: refresh and weather-specific settings;
• map: contextual display modifiers such as clock and scale-bar visibility.

The drawer is the only home for contextual settings. Migrate those settings out of the central settings tree rather than duplicating them. Truly global settings such as language and units remain central. Detailed Bluetooth pairing/device management can remain a settings screen even though the quick drawer owns the radio toggle.

Rows follow the normal Up/Down + Select + Back grammar. Values that need editing open a horizontal nested page:

• Up/Down changes the staged value;
• Select commits and returns to the drawer root;
• Back discards the staged value and returns;
• the committed value remains visibly marked while browsing alternatives.

The screen should provide declarative context content/state; the generic drawer owns navigation, transitions, commit/cancel behavior, and rendering. Avoid screen-specific drawer implementations.

Layering and rendering

Both drawers visually separate themselves from the current screen using the manually tuned device-64 dim LUT proven in the prototype. No alpha blending or expanded color format is needed.

Production behavior should freeze the background frame when a drawer opens:

1. render/capture the current base once through the dim LUT;
2. stop map, weather, and other base-screen visual updates while the drawer is active;
3. animate and update only the drawer/hold overlay regions;
4. invalidate the base once when the drawer closes.

This is especially important for map performance. Use the existing overlay/present damage machinery and verify the actual device transfer cost; the prototype currently favors iteration speed over an optimized frozen-base implementation.

Sheets should normally remain visibly attached to their edge and use only the height their content requires. If content would consume almost the whole display, prefer a bounded sheet with coherent scrolling before automatically turning every drawer into a full-screen page. Final geometry and visual styling require a separate tuning pass on hardware.

Production architecture

The prototype uses simulator/debug entry points. Production work must:

• recognize both chords in the global input plane before screen dispatch;
• swallow all constituent steps, taps, holds, and releases so opening a drawer never also changes the underlying screen or activates its first row;
• decide whether a short chord window/step deferral is needed after hardware testing;
• route Back hold globally to the main menu even from drawer subpages;
• add a small declarative per-screen context model/capability;
• keep the two drawers mutually exclusive and well-defined over existing overlays/modals;
• add real brightness and power platform ports;
• preserve settings dirty/persistence semantics for drawer edits;
• remove migrated contextual rows and the old ride compass menu once their replacements exist;
• localize all drawer row labels, editor text, and confirmation copy.

Do not carry the prototype's debug names, hard-coded English rows, or temporary brightness state into the final architecture.

Prototype coverage

The linked branch already demonstrates:

• real 240×320 RGB222 rendering;
• adaptive top and bottom sheet animations;
• the device-64 dim LUT;
• mutually exclusive simulator chords;
• nested POI-filter, bike-type, and brightness editors with commit/cancel behavior;
• Up-ahead filter state applied back to the underlying screen;
• a real persisted BLE toggle;
• guarded power confirmation and in-progress hold rendering;
• deterministic headless screenshot tokens and captured PNG/GIF design states under design/context-drawer/ and design/quick-drawer/.

Run it with:

cargo run -p obc-sim --release -- apps/obc-sim/assets/grimsel-demo.obcm

Use the physical controls or keyboard equivalents:

• Left Arrow + Enter: top quick drawer;
• Right Arrow + Backspace: bottom contextual drawer;
• Left/Right: previous/next;
• Enter: Select;
• Backspace: Back.

Acceptance criteria

• The grammar above is consistent across the complete screen inventory, with intentional exceptions documented.
• Back hold always reaches the main menu.
• Ride pause/resume and ride-view switching remain one tap.
• Object-specific Select holds such as Forget phone remain direct.
• Quick controls work from every eligible screen and never leak constituent gestures.
• The contextual drawer exists only where content is declared.
• Every contextual setting has one home; migrated settings disappear from central settings.
• The ride compass menu is removed.
• BLE toggles the real radio setting; brightness and shutdown reach real platform ports.
• Power always requires explicit guarded confirmation and cancels cleanly.
• Drawer entry freezes/dims the base, and animation/redraw cost is measured on hardware—especially over the map.
• All UI fits the 240×320 display and RGB222 palette in every supported language.
• The final visual design is reviewed on simulator captures and physical hardware; the prototype styling is not accepted by default.

Verification

• Input-plane unit tests for both chord orders/timings, constituent-event swallowing, release suppression, and unchanged ordinary taps/holds/steps.
• App tests for global Back hold, drawer mutual exclusion, supported/unsupported contexts, commit/cancel, settings persistence, and guarded power.
• Overlay/present tests proving the frozen base and bounded drawer damage.
• Headless renders over map, ride, Up ahead, route planning, weather, settings, and representative modal states in every language.
• Device profiling of open/close animation transfer time and redraw regions.
• obc test -p obc-app, obc test -p obc-sim, strict clippy, and affected board checks once platform ports are introduced.

[timohueser]

Production slice map (post-S4)

Prepared 2026-08-26. The prototype branch stays a behavioral reference and does not merge. Every slice is written against the world after S4 #1447: Caps carries a RenderKeyKind per screen row, App::pass compares exact stack-local keys, and manual map_dirty sites are gone.

Architectural commitments (differ from the prototype):

1. Frozen base is a render-key property, not a code path. A drawer row's key kind shadows the keys of the screens below it: while a drawer is on the stack, the key builder emits only the drawer facts (identity, page, selection, staged value, committed value) and never reads camera/fix/weather facts. The base cannot dirty while a drawer is open; close changes the key variant and produces exactly one base invalidation. No capture buffer, no second framebuffer.
2. The dim LUT is a colour policy, not a layer. Two-Canvas split in App::draw_frame: base through dim_color ∘ color_fn, sheet through color_fn. Four u8 in .rodata, zero RAM. ContextBackdrop::{None, Stipple} never ships.
3. Drawer state costs zero resident bytes when it fits the existing Screen slot (size_of::<Screen>() == 104, pinned by a test). Both drawer states are ~24–32 B. App::quick_brightness does not survive; brightness is a Settings row.
4. The resident risk is D1, not the drawers: UiRuntime owns the InputPlane, so bytes added to Gestures land inside App, against ~72 B of headroom.

D1 — Four-button edges: the platform seam moves into the shared recognizer

Button gains Up/Down; all four controls forward debounced edges; auto-repeat moves from obc-platform into the shared recognizer so hosts and hardware share one timing model. InputEvent::Step survives as the injected variant (wheel, scripts, debug link) — this keeps the slice snapshot-neutral. No chord recognition and no step deferral here: the first step still fires on press, exactly as today.

• Files: obc-ports/src/lib.rs, obc-platform/src/button_input.rs, obc-platform/src/debug_link.rs, obc-app/src/input.rs, obc-sim input/gui, web-demo + desktop button mappers.
• Deletes: Timing::{auto_repeat, repeat_delay_ms, repeat_interval_ms}, ButtonInput::step, up_repeat/down_repeat; the four board-side repeat tests are ported into input.rs, not dropped. Grep guard for auto_repeat.
• Start gate: S4 merged (preparable during S4 review; rebase after).
• Tests: recognizer units (wrap-tolerant repeat, catch-up-not-burst, release silence, Select-held-while-Down independence); snapshots byte-identical.
• Resource: two direction states in Gestures; the naive shape is ~48 B inside App — require a packed form targeting ≤ 24 B total and report the measured size_of::<App>() delta.

D2 — The chord plane, the overlay drawer host, and the universal quick drawer

Chord recognition (Chord::{Quick, Context}) lands in Gestures, above screen dispatch, with full constituent swallowing; the latch clears only when both buttons are up. The two-plane firmware carries the chord on its own channel and signals the input wake. Caps::overlay() and the first Overlay rows appear. App gains one drawer owner enforcing mutual exclusion plus a suppression set for genuinely blocking modals. The quick drawer ships complete: four unlabelled icons, nested brightness editor, guarded power confirmation. Brightness and power reach real obc-ports traits (Backlight, PowerOff) with board and sim implementations in the same PR (a port with no caller is forbidden speculative capability).

• Deletes every prototype leftover: debug_toggle_quick_drawer / debug_toggle_context_drawer, App::quick_brightness, stipple backdrop, the c/S/C/q script tokens, all hard-coded English. Grep guards.
• Start gate: D1.
• Tests: input-plane units (both press orders, both release orders, late-partner and early-release failure modes, a chord never emits any constituent gesture, latch re-arm); app tests (mutual exclusion, suppression over Passkey/DfuInstalling/MapTransfer, drawer Back keeps the base, brightness staged/committed/cancelled, BLE toggle flips Settings::ble_enabled through the existing dirty path, power needs a completed hold); render-key tests (with a drawer on top the key contains no base facts; a camera move produces no dirt; close produces exactly one). New snapshot frames in all four languages.
• Resource: Settings +1 B; dim LUT 4 B .rodata; flash estimate 3–5 KB (glyphs + ~14 catalog keys × 4 languages); re-pin measured_flash.

D3 — The declarative context model, the ride context, and the end of the compass menu

Screens declare context data, not behaviour: Screen::context() returns Option<&'static ContextMenu> (static row table + typed value-binding enum); the generic drawer owns navigation, nested pages, commit/cancel, rendering. Unsupported screens return None and the chord does nothing. Same push: Gesture::BackHold becomes a global escape to the main menu resolved in App above screen dispatch, and screen/ride_menu.rs is deleted whole (check which Msg::RideMenu* catalog keys detour.rs/up_ahead.rs still read before removing rows).

• Start gate: D2.
• Tests: Back-hold from every screen family including drawer subpages; ride context opens over Map/Statistics/Climb/RideControl and nowhere else. Snapshots re-pin with justification — compass-menu frames out, ride-context frames in; update the coverage note in ui-snapshots.sh.

D4 — The remaining contexts; settings leave their second home (4 parallel PRs)

Each PR moves an editor's home and deletes the duplicated central rows in the same push — one home per setting. Persistence semantics untouched (the existing before/after == dirty check covers drawer edits).

PR	Context	Deletes from central settings
D4a	Up ahead — POI filter, source	RideUpAhead row
D4b	Weather — refresh now, interval	settings/weather.rs + root row
D4c	Map — clock, scale bar (+ contours, see decisions)	three toggles in settings/display.rs
D4d	Route plan — bike type, route options	settings/bike_type.rs + RideBikeType row, if ruled contextual

• Start gate: D3; D4a–D4d mutually parallel.
• Tests: commit writes the field and flags a save; Back discards; committed value stays marked. A one-home guard test: no Msg key drawn by both a settings screen and a drawer row.

D5 — Frozen-base hardening, device profiling, and the visual tuning pass

Bound drawer damage to the sheet rectangle via ScreenTick::region (open = full-dirty, steady/growing = sheet-only, close = full-dirty). Measure open/close transfer time and repaint regions on hardware, especially over the map. Then the geometry/typography pass (prototype styling is not accepted by default) and the four-language overflow sweep at 240×320.

• Start gate: D4a–D4d all merged.
• Tests: overlay/present tests proving the base is not re-rendered while a drawer is open and damage is sheet-bounded; full headless sweep × 4 languages; device profiling recorded here. Snapshots re-pin — the slice where pixels are allowed to move.

Parallelism

S4 #1447 ─┬─ D1 ── D2 ── D3 ──┬── D4a ──┐
          │  (D1 preparable   ├── D4b ──┤
          │   during S4)      ├── D4c ──┼── D5
                              └── D4d ──┘

Open decisions for the owner

1. Chord window / step deferral. The prototype defers the first Step by 100 ms unless released sooner. On Map, Step is zoom — deferral makes zoom feel laggy; no deferral leaks one zoom step when a chord opens. Options: (a) fire on press and accept the leak, (b) defer only where a drawer chord is actually armed, (c) defer everywhere. D1 ships (a); D2 implements the verdict after hardware testing.
2. Back hold in Map pan mode is bound to toggle_pan_family; "Back hold always reaches the main menu" cannot also do that. Documented exception, or re-bind pan?
3. Map zoom/pan stays direct? If pan stays a Select-hold, the map context drawer carries only display modifiers.
4. Brightness: level count (prototype: 5), whether level 0 exists, persisted vs session-only, and whether Up/Down previews live on the panel (the port then needs preview + revert-on-cancel).
5. Bike type's home: contextual to route planning, or device-global? Decides whether D4d exists.
6. Contours toggle (provisional pending Elevation EL10d — on-glass review: keep or drop contours #1097): migrate into the map drawer, or delete in D4c?
7. Simulator shutdown: quit the process, or hold a deliberate "powered off" frame? Affects the snapshot recipe.
8. Quick drawer → central settings: Replace the sheet (prototype) or Push? Where does Back from settings return?
9. Reserved chords Up+Down / Select+Back: recognize-and-swallow now so they never leak, or leave unrecognized until they mean something?

🤖 Generated with Claude Code

[timohueser]

Owner dispositions for the open decisions (2026-08-26)

Decisions 1, 2, 3, 5, and 6 are owner rulings from today's session. Decisions 4, 7, 8, and 9 are orchestrator defaults; the owner can veto them before the affected slice starts.

1. Chord window / step deferral — ruled. D2 ships the symmetric 100 ms deferral, both press orders accepted, full constituent swallowing. The window width is the explicit on-glass tuning item. If hardware testing shows the deferral is perceptible (Map zoom is the sensitive case), the fallback to evaluate is the strict action-first order — not a narrower window. D1 ships today's fire-on-press unchanged.
2. Back hold in Map pan mode — ruled. No exception to the global-escape law: the pan-family toggle moves off Back-hold to a different binding. The exact rebinding is design work for the D3 child issue.
3. Map zoom/pan — ruled. Stays direct. The map's contextual drawer carries only display modifiers (clock, scale bar, contours).
4. Brightness — default. Five levels, no level 0, persisted through the settings path, live preview on Up/Down. The Backlight port therefore needs preview and revert-on-cancel, not only a setter.
5. Bike type — ruled. Contextual to route planning. D4d exists; the central bike-type screen is deleted in it.
6. Contours toggle — ruled. Migrates into the map drawer in D4c. The Elevation EL10d — on-glass review: keep or drop contours #1097 verdict later decides whether it survives at all.
7. Simulator shutdown — default. The sim renders the powering-off frame, then the process exits.
8. Quick drawer → central settings — default. Keep the issue's own wording: the entry replaces the sheet; Back from central settings returns to the base screen.
9. Reserved chords — default. Up+Down and Select+Back are recognized and swallowed from D2 on, so they can never leak two independent gestures; they perform nothing until a meaning is assigned.

🤖 Generated with Claude Code

[timohueser]

D1–D3 are merged — the planned pause for the on-glass session starts here

PRs #1537 (D1, four-button edges), #1547 (D2, chord plane + quick drawer + Backlight/PowerOff ports), and #1555 (D3, declarative context model + ride context + global escape; the compass menu is deleted) are all on develop. D4a–D4d and D5 wait until the on-glass round below is recorded.

Setup

1. The board runs the debug-UART soak build. Flash the current develop release image first: pkill probe-rs, two plain probe-rs download --verify passes, settle, standalone verify, THEN attach separately (the access-protection incident rule).
2. Card state: v14 regbez map plus the litter routes from the soak era.

A — Drawer grammar (the new material)

1. Chord hit-rate: Up+Select and Down+Back, 20× each, both press orders, with gloves. A failure is a leaked step/gesture or no drawer. The window is 100 ms (DEFAULT_CHORD_MS) — the tunable.
2. Deferral over Map zoom: hold Up/Down on the map. If the ~100 ms first-step delay is perceptible, the ruled fallback is the strict action-first order — a decision, not a retune.
3. Thumb-roll mid-squeeze: roll a finger off and back onto either constituent during a slow squeeze — nothing may fire (the round-3 leak fix).
4. Quick drawer: three glyphs on the board (no backlight hardware exists — the bulb hides; owner veto point: is the three-icon root acceptable vs the spec's four?). Glyph legibility at arm's length in daylight. BLE toggle against the real radio (confirm with the phone). Settings entry replaces the sheet; Back lands on the base.
5. Power: a tap never commits; the hold fills the ring; Back or tap cancels; completion reaches real system-off — record the wake behavior. A squeeze on POWERING OFF does nothing (three guarded doors + the remote-DFU fourth).
6. Dim LUT: map legibility under the sheet; the DIM_LEVEL verdict.
7. Sheet-open animation: the map repaints ~11–14× per open. Known and expected — D5 bounds it. Judge feel only; do not file it.
8. Context drawer: opens from all four riding views mid-ride; the inert Detour row on a browse Map reads as unavailable (judge against the amber cursor and a live row); each row lands on its screen; one Back returns to the ride view.
9. Global escape: BackHold from deep settings, a paused ride, a drawer subpage, and the power confirmation (not POWERING OFF); the recovery card refuses it; repeated laps do not grow the stack; the Menu keeps your last station.
10. Pan ring: Select tap cycles Route Move → Free Move → Zoom (Free ↔ Zoom with no route). Gloves feel; the displaced Route+Zoom combination costs one extra tap — verdict. The reserved-latch mis-squeeze (Up+Down then reaching for Back) — measure whether it happens with real thumbs.
11. Card over sheet: land a card (e.g. start a map transfer) with a drawer open — the drawer closes; ending the transfer does not restore a stale editor.
12. Languages: the drawer labels in de/fr/es on the panel — BLUETOOTH INACTIVO is the width constraint.

B — Carried debt from the refactor merges

13. S4 S4 — Replace manual dirty mirrors with declared render keys #1447: wake rate and frame latency on a parked device — no repaint churn on an idle map.
14. feat(catalog): the trip cascade is the domain's order, and the residual loses a command #1535: held-to-delete a large trip folder — the cascade is one commit per pass; time it.
15. Visible weather policy is decided by WeatherDomain, not assembled by each host's render call #1549 (four checks, issue stays open for them): cold dashboard open → exactly one urgent fetch; the UPDATING cue rises and falls with a rider-ordered fetch AND a cadence fetch nobody ordered; a resample repaints the open dashboard card; Home does not repaint during a background resample.
16. The ride close is a Recorder verdict, not a legacy command and a catalog re-feed #1554: save a ride — it appears in the catalog once (one read per save); discard a ride; power-cut mid-ride → the recovery offer at next boot; a discarded recovery does not leak totals into the next ride.

Decisions the session feeds

The chord window number (or the strict-order fallback), the three-icon board root, DIM_LEVEL, the pan-ring feel, and the D5 visual-pass priorities.

🤖 Generated with Claude Code

[timohueser]

On-glass round 1 partial results (2026-08-27) — session paused, resumes after the region slice

Recorded so far: the up-ahead context drawer is wanted back (→ D4a, in implementation); the brightness station must exist on the board behind the Backlight HAL seam (→ #1558, in implementation); drawer motion does not read on the panel — verdict: pull D5's frozen-base/region-bounded rendering forward into #1559, then retune open/close animation on top.

The owner defers the remaining checklist items (chord hit-rate, DIM_LEVEL, pan ring, languages, carried-debt items) until #1559 is on the board, so motion and feel are judged once, on the real render path. Checklist above stays the reference.

🤖 Generated with Claude Code

[timohueser]

Curated on-glass list (2026-08-28) — the session list

This replaces the checklists that were spread over a dozen epics. Those checklists are retired.
Everything that still matters is below. The earlier checklist comment stays as the reference for
the detail behind each drawer line.

Setup

1. The board already runs cde75a53, flashed 2026-08-28. Do not reflash unless a check fails
   or you need a different feature set.
2. Card state: the v14 regbez map plus the litter routes.
3. If you must flash: pkill probe-rs, two plain probe-rs download --verify passes, let it
   settle, verify standalone, and only then attach.

Drawers

4. Chord hit-rate: Up+Select and Down+Back, 20 times each, both press orders, with gloves. A
   leaked step or no drawer is a failure. The 100 ms window is the tunable.
5. Deferral over Map zoom: hold Up/Down on the map. If the first-step delay is perceptible,
   rule the strict action-first order — this is a decision, not a retune.
6. Thumb roll mid-squeeze: roll a finger off and back onto either constituent during a slow
   squeeze. Nothing must fire.
7. Quick drawer: read the four glyphs at arm's length in daylight; toggle BLE against the real
   phone; the settings entry replaces the sheet and Back lands on the base screen.
8. Power: a tap never commits, the hold fills the ring, Back or a tap cancels, completion
   reaches real system-off. Record the wake behavior.
9. Sheet open at 440 ms: judge the feel of both sheets and the step counts (12 and 9).
10. DIM_LEVEL: give the verdict for menu and list backdrops only — the map is no longer
    dimmed under a sheet.
11. Undimmed map under a sheet: judge it on a dense scene. The panel is the judge, not the
    still in PR Drawer motion: the sheet opens over a base that is left alone #1562.
12. Context drawers: open from all four riding views mid-ride; judge the up-ahead editor feel
    (D4a, PR D4a: the Up-ahead context, its nested value editor, and one home per control #1561); each row lands on its screen; one Back returns to the ride view.
13. Global escape: BackHold from deep settings, a paused ride, a drawer subpage and the power
    confirmation. The recovery card refuses it and repeated laps do not grow the stack.
14. Pan ring: a Select tap cycles Route Move → Free Move → Zoom. Judge the gloves feel and the
    one extra tap that Route+Zoom now costs.
15. Card over sheet: start a map transfer with a drawer open. The drawer closes, and ending the
    transfer does not restore a stale editor.
16. Languages: read the drawer labels in de, fr and es on the panel. BLUETOOTH INACTIVO is
    the width constraint.

Backlight (PR #1560)

17. THE GATE — duty stepping on the logic analyzer (sigrok-pico rig): 1 kHz period, 4 / 16 /
    36 / 64 / 100 % high time through all five levels. No host test can reach the
    inverted-versus-normal polarity choice. A reversed ladder looks correct everywhere but a scope.
18. Boot level: after a power cycle the panel comes up at the persisted level, not at full.
19. Live preview: stepping the editor moves the PWM immediately, and Back reverts it in one
    frame.
20. LED2 ladder: the steps look evenly spaced — that is the whole claim the square law makes.
21. LED2 after a completed power-off hold: system OFF does not drive the line low, so this says
    whether a fitted lamp would stay lit on a switched-off device.
22. Bring-up flash: judge whether the full-brightness flash at boot is acceptable to a rider set
    to level 0.

Carried debt

23. S4 S4 — Replace manual dirty mirrors with declared render keys #1447: wake rate and frame latency on a parked device — no repaint churn on an idle map.
24. feat(catalog): the trip cascade is the domain's order, and the residual loses a command #1535: hold to delete a large trip folder and time the cascade (one commit per pass).
25. Visible weather policy is decided by WeatherDomain, not assembled by each host's render call #1549, four checks: a cold dashboard open makes exactly one urgent fetch; the UPDATING cue
    rises and falls for a rider-ordered fetch and for a cadence fetch nobody ordered; a resample
    repaints the open dashboard card; Home does not repaint during a background resample.
26. The ride close is a Recorder verdict, not a legacy command and a catalog re-feed #1554: a saved ride appears in the catalog once; a discarded ride goes; a power cut mid-ride
    gives the recovery offer at the next boot; a discarded recovery does not leak totals into
    the next ride.

Two notes, not checks. A page slide over a far-zoom map still costs a base render per frame —
the one place the old lag remains (PR #1562). An upload popup under a drawer can outlive its 30 s
auto-close by up to about 12 s, because the frozen base does not tick; it is self-healing and
invisible under the sheet.

New items from tonight's work get appended here.