The stroked icon set

Change: openspec/changes/the-shell-has-one-visual-language, task 4.1.

What changed

• glyph.rs is the shell’s own icon set: eleven monochrome glyphs drawn as cairo paths on a 24-unit grid.
  • Glyphs: Wi‑Fi, sun, small sun, speaker, muted speaker, keyboard, motion, key backlight, reboot, power and search.
  • Strokes have round caps and joins. The stroke is 3 px at 48 px (glyph::ROW), 4 px at 64 px (glyph::DRAWER), and never under 2 px.
• Where they replace older drawing:
  • The slider-end suns and the volume speaker (in Settings and the Shade) were a filled disc and a filled cone (paint_sun, paint_speaker).
  • The volume HUD’s speaker.
  • The drawer search glyph, which was an ad hoc circle and line.
• New leading glyphs on the Settings rows (Wi‑Fi, Brightness, Volume, Keyboard, Motion, Key backlight), the Reboot and Power off cards, and the Shade’s Keyboard toggle.
• Colour. Every glyph is in the foreground colour; the small sun and the muted speaker are muted. The bright-end sun and the unmuted speaker used to be in the accent; they are now in the foreground colour, in line with the one-accent rule.
• Size: a deviation from design §11. Design §11 gives a fixed 48 px row icon. A fixed 48 px column with the text beside it cut the Keyboard value (“On-screen keyb…”) at 100% in the side-by-side layout, and stayed small beside 200% text.
  • Instead, a row’s glyph sits on the label line, sized to 1.3 times that line’s text and centred on it (paint_line_glyph).
  • Only the label line is indented; the value and caption keep the full width.
  • This makes the glyph about 34 px at 100% and 68 px at 200%. The 48/64 px sizes are what the test checks.
• App icons still come from the theme and are unchanged. That includes the drawer’s 64 px app icons.

Host evidence (2026-10-17)

• cargo test --offline --manifest-path nix/rust-shell-client/Cargo.toml --tests:
  • All targets passed (lib: 574 passed, 2 ignored).
  • glyph::tests::every_glyph_inks_its_box_at_the_row_and_drawer_sizes_and_stays_inside renders each glyph at 48 and 64 px. It requires more than 4% and less than 50% of the box to be inked, and no ink outside the box.
  • the_stroke_is_three_pixels_at_the_row_size_and_never_under_two checks the stroke widths.
  • the_glyphs_are_distinct checks that no two glyphs render the same.
  • The accent-group test from 3.3 still passes.
• Pixel-identity baselines. The docs/evidence/shell-responsive/drawer-* frames were regenerated deliberately for the search glyph. The Settings and Home baselines did not change: their fixture has no services, so it shows “Loading settings”.
• Renders. render_glyph_evidence drew the glyph sheet, and render_accessibility_scale_evidence drew the surfaces, using the catppuccin and catppuccin-latte generations from docs/evidence/card-jobs/README.md. Each surface sheet is at 50%.

Limits

• The Volume row’s “No audio device found” was cut off at 200% in portrait. Fixed afterwards (2026-10-17): the line now uses themed_role_fit like the other row values. It shrinks to fit the column, never below the Body role, and a host render of largest-settings-568x1232 shows it whole.
• The board captures are below. The operator’s look at the icons (5.2) is separate and still open.

Board evidence (2026-10-17)

• Build. kernelMainlineDrmShellLandscapeHdmiBootFiles was built at 5f244c0c and deployed live with switch-to-configuration test. The active system was /nix/store/a4241ifl36r41l3ccx24prrsq5p2q1n0-nixos-system-nixos-26.11.20260919.20b1ddd.
• Capture. DSI-1 was the only output, at 1232x568 and 100%, with the retro-82 theme.
• Navigation. Surfaces were opened with k230-shell-rust --surface. Settings was scrolled with an injected uinput swipe (/tmp/bp.py). The frames are native grim captures.

• Observation: audio arrives late after a restart. The deploy restarted PipeWire and the shell together (shell-ui started at 01:36:40). The shell’s journal shows its first default-sink snapshot at 69.6 s: the volume HUD opened at 01:37:56. The two Shade captures without a volume slider were mapped at 12.2 s and 43.6 s, before that snapshot. The capture with the slider was mapped at 126.8 s. The missing slider was PipeWire still starting, not a rendering fault.