14cef649

fix(web): the page never ran; unzoom was unreachable; a paste wrapped every chunk

a73x   2026-08-13 12:21

fix(web): the page never ran; unzoom was unreachable; a paste wrapped every chunk

`WebAssembly.instantiate` has two return shapes and the shell picked the
wrong one: given BYTES it resolves to {module, instance}, but given an
already-compiled Module — which is what compileStreaming hands back — it
resolves to the Instance itself. `const { instance } = ...` was therefore
undefined, every tile threw on `.exports`, and the rejection went
nowhere because nothing was reading it. Adding the .catch on tile.start()
(a MINOR item in the review) is what surfaced it: driving the real page
under headless Chrome printed the TypeError, and every tile had been
'gone' from the first frame. verify.js passes bytes, so it takes the
other shape and was right all along — which is exactly why no test saw
this.

With the page actually running, the rest of the round:

  - Unzoom was unreachable. The zoomed tile was `inset: 0; z-index: 10`
    over a `z-index: 5` shade, so the shade's click handler could never
    fire and the tile covered the viewport. The tile is now inset
    (3vh 4vw) at z-index 20 with the shade at 10: the visible ring of
    shade around it IS the unzoom control. Escape is deliberately NOT
    bound — it belongs to the application, and vim needs it. The zoomed
    grid is measured from the element's own box, so the CSS stays the
    single source of the geometry. Verified over CDP: elementFromPoint at
    the viewport corner hits #shade, and clicking it unzooms.
  - Wall tiles drew at full logical size into a big bitmap and let CSS
    squash it. They now draw THROUGH a scaled context, with the backing
    store at CSS pixels × devicePixelRatio, so glyphs rasterize once at
    the size they are actually seen.
  - Composed IME text was being sent as a bracketed PASTE. It is typing,
    and it now goes out as plain bytes (mux_text_encode) — bracketing it
    would make vim skip paste mode's indentation and a shell's
    bracketed-paste guard refuse to run it.
  - A chunked paste wrapped EVERY 32 KiB chunk, putting a paste-end in
    the middle of the pasted text. mux_paste_begin/mux_paste_end are now
    separate exports and the shell sends begin + N unwrapped chunks +
    end: one wrap around the whole paste, which is what keymap.zig's
    pasteInto has always said it was. mux_paste_encode is retired — one
    way to spell a paste, not two.

    KNOWN LIMITATION, deliberate for v1: the wrap is unconditional. The
    daemon's pty_mode frame carries icanon/echo and not mode 2004, so
    nothing on this side can tell whether the application asked for
    brackets; one that did not sees the literal markers. Teaching
    pty_mode to carry the bit is a daemon protocol change and is
    deferred.

  - The browser leg reconnects on its own, on client.zig's nextBackoffMs
    schedule (0, then 200ms doubling to a 2s cap, no retry cap). Nothing
    was reconnecting this socket, so a restarted hub left every tile
    permanently 'gone' with a reload as the only cure. The badge reads
    'reconnecting' while retrying and 'gone' only once the retries are
    also failing. Verified live: up → kill hub → reconnecting → gone →
    restart hub → up, with no reload.
  - mux_apply_frame < 0 tears down and re-inits the core rather than
    painting on, and a frame over the staging cap re-attaches instead of
    returning silently.
  - gotState follows Replica.state_since_attach: a delta's ARRIVAL alone
    proves the attach was admitted, so exit_status after a bad delta
    reads as the shell exiting rather than "session full".

verify.js gains the new exports and, more usefully, reads mux.js and
asserts every mux_* the shell calls is a real export — the pair that has
to agree, and the one a renamed export breaks with nothing else noticing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>