input, render: keep fast touch typing whole, fire pattern nodes once, and make the software image path cheap

Two ways a tap was lost on a phone. The gesture machine serves one finger: a second finger landing while the first was still down went to the auxiliary slot and never reached it, so with two thumbs every overlapping tap on a keypad vanished. A finger landing on a primary that is still a plain press (GestureState::is_plain_press: no drag, scroll, swipe or slider in flight) now commits that press as a tap at its last position and takes the primary slot; the earlier finger's release reports through on_touch_up, and a second finger during a pattern stroke or a slider drag stays out as before. The touch-up body moves to release_primary so both paths share it. The other loss needed one finger: the 8 px travel that turns a press into a scroll, swipe or drag, and the 6 px stray that cancels a touch long-press, were compared against physical coordinates, so at buffer scale 2 they were 4 and 3 logical px and an ordinary thumb tap became a vertical drag whose release skipped the button. SwipeConfig carries the surface's scale and the machine reads the thresholds through slop() and stray(), which are logical.
Pattern grids fired on_node on every motion event within a node's snap radius, and the Slider move outcome requested a repaint each time, so drawing a pattern rebuilt and re-rastered the surface at the touchscreen's motion rate. The machine remembers the node under the finger (last_pattern_node) and emits only on change; the handlers split pattern_node_at from pattern_msg, with pattern_node_msg kept on top of them; a move outcome without a message no longer requests a redraw.
SoftwareCanvas::draw_image_data premultiplied the whole source in floating point on every call, so a full-surface backdrop paid a pass over every pixel per frame before the blit. Premultiplied copies now live in a 32 MiB least-recently-drawn cache keyed by size, content fingerprint and alpha, the key the GLES texture cache already used; fingerprint_rgba moves to render::helpers and both backends share it. The entry records whether the source was fully opaque, and an opaque image drawn 1:1 with no clip and full alpha is copied row by row into the canvas instead of going through tiny-skia's bilinear pipeline.
docs/architecture.md describes the primary-slot hand-over, docs/widgets.md the once-per-node contract, and the CHANGELOG lists the five fixes.

CHANGELOG.md

@@ -50,6 +50,11 @@ All notable changes to `ltk` are documented here. The format is based on [Keep a
 
 ### Fixed
 
+- **Fast two-thumb typing no longer drops taps.** A finger landing while another was still down went to the auxiliary slot and never reached the gesture machine, so on a touch keypad every overlapping tap was lost. A finger landing on a primary that is still a plain press (`GestureState::is_plain_press`: no drag, scroll, swipe or slider in flight) now commits that press as a tap at its last position and takes the primary slot; the earlier finger's release reports through `App::on_touch_up`. Presses that have become a drag keep the old behaviour, so a second finger during a pattern stroke or a slider drag still stays out of the gesture machine.
+- **`PatternGrid::on_node` fires once per node, not once per motion event.** Every touch motion within a node's snap radius re-emitted the node's message and requested a repaint, so drawing a pattern rebuilt and re-rastered the surface at the touchscreen's motion rate. The gesture machine now remembers the node the finger is over (`last_pattern_node`) and emits only on change; a `Slider` move outcome without a message no longer requests a redraw.
+- **Gesture thresholds are logical pixels.** The 8 px travel that turns a press into a scroll, swipe or drag, and the 6 px stray that cancels a touch long-press, were compared against physical coordinates, so on a scale-2 panel they were 4 and 3 logical px and an ordinary thumb tap became a vertical drag whose release skipped the button. `SwipeConfig` now carries the surface's buffer `scale` and the machine reads the thresholds through `SwipeConfig::slop` / `stray`.
+- **A 1:1 opaque image is a row copy in the software renderer.** With the premultiplied copy cached, drawing it unscaled with no clip and full alpha now copies rows into the canvas instead of running tiny-skia's bilinear pipeline over every pixel, which is what a full-surface backdrop cost on every frame.
+- **The software renderer premultiplies an image once, not on every draw.** `SoftwareCanvas::draw_image_data` converted the whole source to premultiplied alpha in floating point on each call, so a full-surface backdrop cost a pass over every pixel per frame on top of the blit. Premultiplied copies are now kept in a 32 MiB least-recently-drawn cache keyed by size, content fingerprint and alpha, the same key the GLES texture cache uses (`fingerprint_rgba` moves to `render::helpers` to be shared).
 - **Rows no longer spill their children out of the layout.** `Row` distributed leftover width but had no notion of a shortfall: `leftover` floored at zero, so a cluster wider than its rect kept every preferred width and overflowed — symmetrically, since the no-spacer branch centres it, which is why both end labels of a segmented control were clipped at once. The shortfall now comes off the widest children first (water filling), so a long title gives up its width while the icon beside it keeps its size, and equal siblings — a segmented control — share it evenly; pinned spacers keep their width, and `Row::no_shrink()` opts out for strips deliberately wider than their viewport. `Button` elides its label into the rect it is granted, so shrinking degrades to a truncated label rather than clipped glyphs; the truncation rule moves to one crate-internal helper shared with `ListItem`.
 - **`ListItem` labels no longer run under the trailing slot.** The label and subtitle were painted with no width limit, so a title longer than the row overlapped the trailing text or the disclosure icon instead of truncating. The trailing slots are now laid out first and both lines are elided with an ellipsis against the space they leave, matching what `Text` already did.
 - **Text inputs no longer insert mid-string when the value grows after focus.** Focusing a text input pinned the cursor to a snapshot of `value.len()`; if the value kept growing without the widget seeing keystrokes (a field fed over IPC), the first normally-delivered key inserted at the stale position. The focus-time cursor is now an end-of-value sentinel that every consumer clamps to the current value length, collapsing to a concrete position on the first real keystroke or click.