max / quasi

0.70.0: a change is complete before it is a write `8032fe61` and `de2376bd`, one answer. Only half of `8032fe61`, deliberately: that task carries a standing instruction not to answer it as written, and the half it means is re-filed rather than settled. Two shipped renderers disagreed about what a member means. quasi-webview hx-trigger="change" blur, Enter, a range released quasi-immediate buffer != last frame every keystroke, every drag frame quasi-tui every keystroke So a search box was one request per letter on two hosts and one per search on the third, and audiofiles' bounded `row_height` posted a row height of 3 on the way to 30, outside the bounds the field's own hint states. One description behaving two ways is the drift this stack exists to end, and the webview was again the host that was already right. `Field::changes` means the change is complete. A value the reader builds up, typed into or dragged across, writes when they are finished with the control; a value chosen in one go was complete when it changed and still writes then, which is what a browser's `change` does for those too. Leaving a control nobody altered writes nothing, which is the other half of that promise. Each `View` grew a `written` map beside `edits`, because a completed value needs a baseline: `edits` is what is in the box and `written` is what the route has been told. Reading an untouched box as empty wrote every prefilled field walked past, which the new terminal test caught. `de2376bd` closes with no member. It asked for `Field::settles` beside `changes` for three sliders writing once per drag frame; under this a drag is one write, so option (a) is unnecessary rather than declined. The shipped app's own `widgets::settled` helper, the workaround that prompted it, is what `changes` now means. Live search costs nothing under this, which is what makes it a rule rather than a restriction: a box asking a route about what is being typed carries a `Consult` with its own `after`, a wait the description states. Writes complete; questions debounce. Both renderers still fire consults per keystroke. Two quasi-tui tests asserted the old behaviour and were rewritten rather than deleted; one was really about ticks travelling with a picker's value, which is unchanged. The new egui test types into a focused box and asserts nothing fires until the caret leaves, which is the shape of the `row_height` defect, and it fails against the old code. `push_deadlines` split out of `field_node` while that was over clippy's line bound, and it is a whole idea: a keystroke pushes every question the box raises further out, which is what makes it a debounce. Consumers forward-fixed in the same pass: MNW server, goingson and audiofiles move to `^0.70` and each tests green.