bench: an M3 Ultra joins the table, and takes the top of the picker

32 cores, cross-origin isolated, Chrome 153. 30 rows, every column present,
zero WRONG, zero n/a, zero unverified checksums. WebGPU lands where a machine
this size should: matmul 320x, ode-rk4 1071x, path-trace 795x over hand-written
JavaScript, and several rows measured at 1-3 ms.

FILED AS 2.23.0, THOUGH IT WAS MEASURED ON 2.24.0. The entire diff between the
two releases is a build-date banner and one try/catch:

    - const getContext = require('gl');
    + let getContext = null;
    + try { getContext = require('gl'); } catch (e) {}

That is the headless-gl backend declining to throw when its optional native
build is absent. In the browser bundle `require('gl')` is already
`require_empty_module()`, so the change wraps a no-op — nothing that executes in
a browser differs between the two versions. Labelling it 2.23.0 keeps all three
2.23.0 tables directly comparable rather than splitting them across a version
boundary that has no runtime behind it. The version it was recorded under is
here in the history, which is the honest place for it.

It sorts to the top by date, so the page now opens on it and the signature list
derives from it — hence the reordering there. The poster is unaffected: it is
generated from the 5090 explicitly, not from whatever happens to be newest.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RJujXBsBhXBbQ72tS7E7GV

Author: fuzzie360