The lazy editor (Exp A) unmounts on exit, so the editor's own deferred
focus-return (useReactiveTextInput) targeted a detached node and focus was
lost to <body>, breaking table keyboard navigation after submit/escape.
Return focus to the parent cell deterministically in a layout effect once the
display element is back in the DOM (rather than a setTimeout that races the
next keypress), mirroring the time/duration cells.
Update the e2e tests to the lazy interaction model (click the cell to open the
on-demand editor, assert it unmounts on exit) — matching the existing timer
cell test. Verified end-to-end: 202-cuesheet (3/3) and 206-url-preset (5/5).
Render a lightweight, focusable TextLikeInput for cue/title/note/custom cells
and only mount the real input/textarea editor (with reactive-input hooks and
autosize) on click/focus, mirroring the existing time/duration cells.
Measured baseline vs Exp A at CPU 4x (median of 3, ~1140 row mounts):
- longFrames 201 -> 127 (-37%), maxFrameMs 517 -> 417 (-19%)
- longtask total 58.9s -> 30.7s (-48%), EventRow.mount avg 179 -> 146ms (-18%)
- cell.autosize and cell.reactiveInputInit eliminated from the scroll path
- per-row re-renders 4648 -> 3048 (-34%): stable display height, no virtuoso churn
(the opposite of the reverted lazy-autosize Exp B, which doubled re-renders)
Editing verified end-to-end: click mounts+focuses the editor, edits persist,
editor unmounts back to the display on commit. TextLikeInput now fills its cell
so the whole cell is clickable to edit.
Claude