The visual work and the reason the visual work kept regressing, in one pass.
Picker and palette
- Rows drop their card borders: a row is transparent until hover/selection.
Five ringed white rectangles in a column read as a form, not a list, and
the panel-to-row lightness difference already separates them.
- Option paths start at one shared column instead of being right-aligned.
Right alignment floated every row's path START with its own length
(cmd.exe began ~300px right of Nushell's), so scanning meant hunting
left and right per line.
- Footer hints drop their chips. A chip background means "clickable" here;
the arrow/Enter/Esc legends aren't buttons, so key and description
separate by ink weight instead.
Keycaps go back to one chip per combination
Per-key chips are a macOS form, where the modifiers are single-character
glyphs that tile evenly. Ctrl/Shift/Alt are words, so splitting produced a
row of ragged blocks; and Windows users already read Ctrl+Shift+P as one
string in VS Code and Office.
Carets that look alive
The blink phase hung off the wall clock, so a field taking focus had a 50%
chance of showing its caret up to half a second late — which reads as "not
focused". Phase now starts at the last edit: lit immediately, steady while
typing, breathing only after the hand stops, at the system's
GetCaretBlinkTime (including the accessibility setting that disables
blinking). The palette was also missing from the fast-redraw list, so its
caret froze once the open animation ended. The palette caret is now a
1.5px beam rather than a full-cell glyph, which lets placeholder and real
input share one x — the first keystroke no longer shifts the line.
Overlays share one recipe
Confirm dialogs and the SSH editor gain a real drop shadow and the common
radius; both previously drew hairline + fill only, so a dialog demanding an
answer floated lower than the palette you dismiss with Esc. The SSH
editor's dim was a hard-coded 67% black ignoring the theme. Dropdowns cast
a shadow instead of a glow (a glow spreads brightness without establishing
height) and are opaque, since a menu sits directly on the content it hides.
Why this kept regressing
ui/mod.rs has said "layer contract: no magic numbers" for a while, and
widgets.rs violated it — a constraint expressed only in a comment is not a
constraint. ui/guardrails.rs turns three of them into failing tests, each
carrying a budget equal to today's real count: new violations fail the
build, and clearing old ones fails it too until the budget is lowered, so
the number only goes down. This pass: radius 121 -> 94, strokes 69 -> 56,
glow-as-shadow 12 -> 9.
Probe script now runs a COPY of the debug binary. Running the build output
directly makes the next cargo build die with an opaque mingw link error,
and it invites cleaning up by image name — which would kill the user's own
window, possibly the one hosting this session.
kuddev
b795ebbc85