lib/ui: labels that measure themselves, grids that don't own your model

[InauguralPhysicist]

Second half of the lib/ui gap series from DeslanStudio's desktop-shell port — the widget-level fixes. Stacked on #615 (merged).

What changed

label carries w/h (#561). It was the only text widget without them, and _layout_toolbar/_layout_box read child.w/child.h unconditionally — so the most natural toolbar child crashed the layout engine outright. Measured from the text at construction and refreshed on every _layout pass: a label's text changes at runtime (a BPM or time display restamps every frame) and a size stamped once goes stale the first time the clock ticks.

A label draws no chrome, so unlike badge there's nothing to pad — its box is exactly its text, and spacing comes from the container's gap. (The issue suggested + pad; this seemed the more truthful bounding box. Easy to change if you'd rather.)

grid.owns_cells (#572). The widget flipped cells[r][c] itself and only then reported, so an app whose model is the source of truth had to overwrite grid.cells after every click and every model-side change (undo, pattern switch, randomize) or watch the two drift. owns_cells = 0 makes input report (row, col) and touch nothing. The keyboard path in ui_focus.eigs flips cells too and honors the same rule — otherwise mouse and keyboard would disagree about who owns the pattern.

Grid row-label gutter (#572). ax - 60 at scale 1 was hardcoded; DeslanStudio positions its grid at x=76 purely to make that constant land inside its panel. Now row_label_w/row_label_scale with the historical values as defaults, and vertical centering uses the real text_height of the scale rather than an assumed 7px.

Piano key release (#570). The note-off was a piano_kb special case in dispatch's mouseup that only ran when the mouseup hit-tested back to the same widget. Press a key and drag off it — or release over a button, whose branch returned early — and the note sounded forever. piano_kb now registers a clear_pressed entry and dispatch runs the registry release walk on every mouseup, retiring the special case.

Hover feedback (#594). checkbox and toggle received .hover from dispatch and their renderers ignored it. Both now draw the translucent shade button already used, factored into a shared _draw_hover_shade — an overlay rather than a color swap, so it composes with a per-widget bg (#566) instead of erasing it.

One scope judgment to flag

#570 asked for "either a vertical: 1 mode or a documented note that the widget is a horizontal trigger strip only." I took the second option. DeslanStudio paints its own key sidebar on a canvas and the upstream Qt port does the same, so a vertical mode would ship with zero consumers exercising it — against the forcing-function rule. The widget's scope is now documented in its header and STDLIB instead. Say the word and I'll build the mode.

The other two #570 claims (pressed_note never clears; on_note never called with 0) no longer reproduced — see the evidence comment on the issue. The real bug was the residual above, which the original report didn't have.

Verification

Red-then-green, each fix stashed independently:

• label w measured from its text — expected 36, got null, plus the issue's exact crash: cannot apply '-' to num and null at _layout_toolbar
• non-owning grid leaves the cell untouched — expected 0, got 1
• key released even though mouseup missed it — expected -1, got 0
• hovered checkbox draws a shade — expected 1, got 0

test_ui 143 → 176 assertions. Render behavior is asserted by capturing what the renderer emits (the shade is the only gfx_rrect carrying an alpha).

• Release suite: 2890/2890
• ASan+UBSan, detect_leaks=1: clean, leak tally still 0

The lone ASan invariant_weak.eigs timeout is #616 — unrelated, and it reproduced identically on #615's branch.

Closes #561
Closes #572
Closes #594