perf(bench): add a repeatable benchmark suite - #252

The benchmark suite's process attribution showed a server holding ~132 MB
before serving a request, with ~72 MB of it attributable to importing
server/app.ts — mostly shiki, @pierre/diffs, markdown-it and
@mermaid-js/parser, loaded at boot whether or not the workspace ever used
them.

Two dynamic imports, both at the point of use:

- app.ts loads richRender.ts when it renders a markdown/code/diff/terminal
  surface. It sits after the html and mermaid branches, so an html-only
  workspace never loads any of it.
- postSurfaces.ts loads @pierre/diffs and @mermaid-js/parser when validating a
  published diff or mermaid surface.

Both halves were needed: leaving either static keeps the whole graph resident
and neither saves anything. Idle RSS 132 MB -> 102 MB, boot 486 ms -> 275 ms.

The mermaid parser and the diff parsers now load outside the try/catch that
reports parse failures, so a module-load error can't be reported to the user as
invalid mermaid (a 400 where it should be a 500).

A dynamic import is the kind of change that works on Node and fails only once
deployed, so test/workerIntegration now renders all four rich surface kinds on
real workerd and asserts on markup only the real renderers emit — the existing
html render deliberately never reaches that path.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MGMuudU8anVrxB7sjY9hpt

CodeQL's high-severity alert on this PR was regex injection at bench/run.ts:104
— a command-line argument going straight into the RegExp constructor.

--filter now takes comma-separated literal substrings, OR'd, case-insensitive.
That removes the taint flow, and it is the better CLI anyway: metric names are
full of regex metacharacters ("GET /s/:id code (cache hit)"), so the obvious
move — pasting a name off the results table — used to match nothing. Now it
works, and no hand-written pattern can backtrack a benchmark run into a hang.

Lowercasing happens inside the matcher rather than in the arg parser, so a
caller constructing a context directly still gets case-insensitive matching;
a test covers that seam along with the pasted-name and OR cases.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MGMuudU8anVrxB7sjY9hpt

The /s/:id benches asked for `?part=0` with no `&theme=&mode=`. The viewer
always sends both (Card.tsx builds every surface iframe src that way), and the
server renders differently when the mode is pinned — so the suite was measuring
a URL shape no client ever requests.

That is the failure mode a benchmark can't afford: profiling a shiki change
against these numbers showed no improvement at all, while the same change
through the real viewer URL was 47% faster and 30% smaller. A bench that misses
the win is worse than no bench, because it argues against the fix.

Baseline re-recorded for the corrected URLs.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MGMuudU8anVrxB7sjY9hpt