perf(renderer): fetch the viewport once per frame, add glyph atlas and scroll blit

[brianegan]

We embed ghostty-web in a VS Code extension as the terminal for coding CLIs.
Streaming output felt slow enough to be distracting, so we profiled it instead
of guessing. This PR is the renderer half of what came out.

Fetching the viewport once per frame

getLine(y) on the WASM terminal calls getViewport(), which walks every cell
of every row and then throws away all but one. The renderer called it once per
visible row, so a single frame did around forty full viewport walks. That is
O(rows squared) crossings, and it degrades quadratically as the terminal gets
taller:

rows 24    15.6 ms
rows 40    56.3 ms
rows 50   129.6 ms

getViewportLines() walks once and slices, and the renderer caches it per
frame. On a 120x40 streaming frame that single change is about 11 of the 12x in
this PR. The atlas and the blit are the remaining 11%.

Also worth reporting: dirtyRows averaged 0.19 per frame while 36.58 rows were
being painted. Ghostty reports needsFullRedraw() on every scroll, because
relocating every row's content cannot be expressed as a row set, so the per-row
dirty flags were doing nothing during streaming.

Glyph atlas

Rasterises each distinct glyph once, keyed by text, colour, bold and italic.
Slots are sized from actualBoundingBox* ink extents rather than the cell box,
so italics, combining marks and full-height powerline glyphs do not clip.
Engines without those metrics fall back to a padded cell box. Shelf-packed, and
after four repacks it disables itself and reverts to fillText rather than
thrashing.

One thing worth knowing before you review it. A first version built a
${style}|${color}|${text} cache key per cell, allocated thousands of strings a
frame, and was a net loss at 0.96x. Two-level buckets with the outer one
memoised across a run turned it into a win. Chromium's Skia already caches glyph
masks internally, so a userland atlas only pays off if the lookup is cheaper
than fillText, which is a narrower margin than it sounds.

Scroll blit

The shift is derived rather than reported. Row y always shows absolute line
scrollbackLength - viewportY + y, so holding a line fixed across two frames
gives shift = (viewportY - lastViewportY) - (scrollbackLength - lastLength).
Scrolling into history and streaming at the bottom both fall out of that one
formula.

It is treated as a hypothesis, not a fact. Every retained row is checked against
two independent 32-bit hashes of what is actually on the canvas before its
pixels are reused, and mismatches get repainted. Alternate-screen programs and
scrolling regions move content without growing the scrollback, and the hash
catches them. If more than half the viewport fails verification it falls back to
a full repaint.

Two mistakes fixed after measuring, in case they read as odd:

The blit was gated on !forceAll, which meant needsFullRedraw() disabled it
on exactly the frames it existed for. A caller-requested repaint still blocks
it, but a full-dirty buffer does not, because the hashes are better informed
than a viewport-wide flag.

Every blit was staged through a scratch canvas, copying the full bitmap twice
per frame, which made it a net loss at 5.78 ms against 5.30. An opaque
background now moves pixels in one self-copy. The scratch is only used for
translucent themes, where clearing first would destroy the overlapping source.

Options

glyphAtlas and scrollBlit land on ITerminalOptions, defaulting on, with
live toggling. Mainly so the two can be measured independently, which is how the
numbers above were separated.

Verification

tests/e2e/10-renderer-optimisations.spec.ts renders identical content with each
optimisation on and off and diffs the bitmaps.

blit vs full repaint      0 differing pixels
atlas vs fillText         max delta 2/255 on 5.4% of pixels
rows painted, 40 frames   800 -> 80

The atlas delta is antialiasing rounding from rasterising at device scale in an
unscaled context instead of CSS scale in a scaled one. It is spread evenly
across rows. A positioning bug would show max delta 255 in clusters.

438 unit tests, the e2e suite, tsc, biome and prettier all clean.

Stacked on #28.

[sonarqubecloud]

Quality Gate passed

Issues
13 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud