Vision option: larger text and high contrast (tasks C.1–C.3)

What changed

• Preference (C.1). tools/theme_vision.py stores the choice as vision.json beside the theme state (text_scale is default or large; contrast is standard or high). It is written atomically and refuses unknown values and symlinks. The task named service_data.rs. The choice is stored with the theme state instead, because that is the existing theme-choice persistence mechanism the task points to, and it reaches both renderers through the generation.
• Token pipeline (C.2). theme_activate.prepare folds the choice into every generation:
  • A non-default choice changes the generation identity. The default choice keeps the existing identity.
  • High contrast rewrites the palette (black background, white foreground, saturated accents) before the templates run, so every templated consumer picks up the same colours.
  • appearance.json gains sections.accessibility with text-scale (1.0 or 1.25) and high-contrast.
  • Rust render.rs multiplies both context and themed text by the step.
  • nix/card-shell/appearance.c multiplies its font scale by the same token.
• Settings control (C.3). “Vision ›”, beside “Themes ›” in Settings, opens a Vision page with two toggles, Larger text and High contrast. A toggle sends k230-theme vision --text-scale … --contrast … through the theme helper. The helper saves the choice and re-activates the active theme with it. If that fails, the previous choice is restored. With no active theme, the choice is saved and applies with the next activation.

Host evidence (2026-10-16)

Evidence class: host renders and unit tests only. The board check is task C.4 and is recorded separately.

• cargo test --offline --manifest-path nix/rust-shell-client/Cargo.toml --no-fail-fast: 562 lib tests passed (2 ignored), and all integration targets passed.
• PYTHONPATH=tests:tools python3 -m unittest tests.test_theme_vision … (the theme suites listed in the change): 140 tests OK.
• nix build .#card-shell --max-jobs 1 --cores 4 --no-link --print-out-paths: built k230-card-shell (exit 0) from this tree.

Renders

The three generations were prepared on the host from nix/handheld-theme-default/catppuccin with theme_activate.prepare, once per choice. Each prepare was preceded by theme_vision.save(state, choice) and used state_root=<state>, user_themes=<state>/none, builtins=None and tools=HOST_TOOLS. Generation ids:

They were rendered with:

cargo run --offline -q --manifest-path nix/rust-shell-client/Cargo.toml \
  --example render_accessibility_scale_evidence -- <out> <state>/generations \
  default=e1a358a6af0533bf8f1ef70b large=e60e8ad8b173237d6a98737d \
  high=956c733aca3ac6093f1351b0

The example writes 24 frames. Twelve of them are committed in host/: for each choice, Settings and the Vision page at 1280x800, and the Shade and the drawer at 568x1232. The grey area under the portrait Shade is the example’s backdrop, not the shell.

Limits

• These are host renders of the Rust shell. The card-shell headers are covered by tests/test_card_shell_appearance.py (metrics with the token), not by a render.
• With high contrast on, the secondary text keeps the theme’s muted foreground. It is light grey on black, not pure white.
• Persistence across a reboot, and legibility on the real panel (C.4), are not shown here.

Board evidence by injected input (2026-10-16, task C.4 progress)

Evidence class: native grim captures from the board’s compositor, with taps injected through uinput, plus a soft reboot. These are not a camera recording or a real-finger interaction. The output was HDMI at 1280x800.

• Image: kernelMainlineDrmShellLandscapeHdmiBootFiles built from branch vision-scale commit bc14b524. It was installed live with switch-to-configuration test and set as the system profile, so the reboot booted the same system (z42i9i3wq012i8jdglwxbxy88d3im4km-nixos-system-…).
• k230-theme list reported the default choice (text_scale: default, contrast: standard).

Limits

• C.4 also asks for legibility on the real panel, with the operator’s capture-feature.py … --provenance real-touch run. This run used the HDMI output and injected taps. It shows that the choice persists across a reboot and redraws the shell. It is not a glass or camera observation, so C.4 stays open.
• With larger text, the two-line hint in Settings’ Keyboard card sits on the card’s bottom edge (board/05-both-settings.png).

Panel captures (2026-10-16, DSI-1 1232x568 landscape)

Evidence class: native grim captures of the panel output, after the operator unplugged HDMI and turned on Larger text on the device (helperd vision … outcome=ok at 12:19:58). These are compositor frames, not camera photographs of the glass.

The operator’s own look at the glass, and their judgement of legibility, are still needed for C.4.

Redesign: four steps and reflow (2026-10-16, tasks C.5–C.7)

The operator found the drawer “not particularly legible” at the old Large step (125%) and judged a two-way normal/large choice too coarse. The redesign offers four steps: 100, 130, 160 and 200%. An existing large choice maps to 130%. Each surface reflows at a larger step instead of clipping.

Measured against the panel (568x1232 at 4.1”, about 2.07 px per dp), the 100% body text of 17–18 px is about 8.5 sp. At 200% it reaches about 17 sp. Tap targets stay at or above 99 px (48 dp).

• Steps (C.5). tools/theme_vision.py TEXT_SCALES holds default, large, larger and largest (1.0, 1.3, 1.6 and 2.0). The helper’s --text-scale choices come from that table. The Rust and appearance.c bounds are raised to 2.0. The Vision page shows the four steps as one segmented control, plus a High contrast card.
• Drawer (C.6). At a larger step:
  • The minimum tile width is 112 px times the step, so the drawer uses fewer columns: 2 in portrait and 4 in landscape at 200%.
  • Icons grow toward 99 px.
  • Names wrap onto two lines before they ellipsize.
  • The search pill and Help grow to fit their text.
  • The sheet is opaque whenever a vision option is set.
  • At 100% the drawer is pixel-identical: tests/responsive_pixel_identity.rs passes.
• Shade, Settings and Vision (C.7).
  • Shade header. Lines stack by their drawn height. The Settings and “Dismiss all” bands widen with the text, up to half the column. Notification rows grow to hold three lines.
  • Settings rows. Each row gets the height its lines need, plus a full slider band on slider rows.
  • Settings header. The Vision and Themes links get their own row, with tap bands sized for the text.
  • Settings scroll. At 200% the rows are taller than the screen, so the Settings body scrolls under a fixed header. A drag on a slider row still belongs to the slider.
  • Chooser and Vision headers. Back and Close widen to fit the text.
  • Fitting secondary lines. Values and secondary lines shrink to fit before they ellipsize, but never below their 100% size.

Host evidence

Evidence class: host renders and unit tests. The board and glass checks are task C.8.

• cargo test --offline --manifest-path nix/rust-shell-client/Cargo.toml --no-fail-fast: 568 lib tests passed (2 ignored), and all integration targets passed, including responsive_pixel_identity.
  • navigation::the_largest_text_step_reflows_the_drawer_without_overlap_or_tiny_targets checks that, at 200% in both orientations, tiles do not overlap, icons are at least 99 px and the search pill clears Help.
  • icon::drawer_names_wrap_without_ellipsizing_at_the_largest_text_step checks that no fixture name is ellipsized.
  • service_ui::the_largest_text_step_scrolls_settings_and_keeps_every_row_reachable checks each row’s text height against its card. It also checks that a tap on Power off after scrolling reaches it.
  • service_ui::the_largest_text_step_grows_notification_rows_and_header_links checks the Shade.
  • theme_ui::the_vision_layout_never_clips_and_keeps_the_tap_floor_at_every_step checks the Vision page.
• PYTHONPATH=tests:tools python3 -m unittest over the tests/test_theme*.py suites and tests.test_card_shell_appearance: OK.

The five generations were prepared as described above, from nix/handheld-theme-default/catppuccin, with theme_vision.save before each prepare:

render_accessibility_scale_evidence was run with those five names. It drew 40 frames: each name for the drawer, Shade, Settings and Vision, at 568x1232 and 1232x568. The fixture app names include two-word names (“System Monitor”, “Image Viewer”).

Each file in steps/ puts the five frames of one surface and orientation side by side (portrait) or stacked (landscape), in the table’s order, at half size:

Limits

• At 200%, a notification’s summary and body each stay on one line, so long ones still ellipsize, as the body already did at 100%.
• In landscape at 200%, Settings needs a scroll to reach the lower rows of each column.
• The Home deck, the card shell and app windows are outside this change. Only the battery ring was fitted, in 91a60bfb.
• None of this has been seen on the panel yet. That is C.8.

Panel captures of the redesign (2026-10-16, task C.8 progress)

The board ran /nix/store/sbn9x5qi1a3qlrd6sh0npjnxwiqyd70y-nixos-system-…, built from integration branch integ-cma-popup at 0c438090. That branch contains 8799961c (Home widgets) and the fit-margin commit on this branch. It was deployed live with switch-to-configuration test.

Method: the step was chosen by injected taps on the Vision page (tools/board-pointer.py). Each surface was opened with the shell socket route command. Each frame is a native grim capture of DSI-1 at 1232x568. After each change the theme helper logged outcome=ok, and k230-theme list read larger/standard or largest/standard.

Evidence class: injected input and native captures. These are not real-finger interaction and not a judgement on glass.

After the capture, the board was set back to largest/standard.

Limits

• Home widgets do not grow with the step. They keep the text sizes they have at 100%, scaled only by their cell.
• At 200% the Shade’s title and its “Swipe up…” caption shrink to fit beside the Settings link and the battery label. In 16-panel-200-shade.png they are smaller than the rest of the Shade.
• The clock widget has text too small to read on the panel (the operator’s note below). It needs its own redesign.

Operator judgement on glass (2026-10-16)

Evidence class: the operator looked at the landscape panel. No camera recording was made. Verbatim: “it’s all good however the clock widget could use a redesign it’s got some text too small to see really but yeah it’s good”. This accepts C.8 and closes C.4.