feat(site): lead the benchmarks page with a scoreboard and drawn ratios

[Exoridus]

The page opened with a paragraph of generated prose and then repeated every
comparison as a wide table whose mechanism column wrapped over several lines,
so the numbers were the hardest thing on it to find. It ran to about 10500 px
at 1440 px wide, and the practices it is published under sat above the results.

What changed

• A scoreboard leads, one line per arm: a rendering backend against one
  opponent, or one physics arm, with that pair's outcomes as a strip and as
  counts. The split is per arm and never per backend - a backend line pooled two
  different opponents into one strip and showed a mix that belonged to neither.
  Nothing is summed across lines and no line is ranked against another.
• One table per domain, columns are the pairs and rows are archetypes, so a
  single archetype reads across both backends and every opponent on one line.
  ExoJS gets a column of its own per backend instead of being repeated inside
  every cell.
• The ratio is drawn: a bar on a log axis with the published factor beside
  it. The axis is scaled to the widest ratio in that table, up to two decades;
  a bar past that is drawn cut and keeps its exact factor.
• Nothing is behind a toggle. The medians, the p95 of the same window and
  the range the pooled runs observed sit under every bar, on lines that never
  wrap.
• Methodology, reproduction and fairness move below the tables, and the
  structural evidence, the omitted rows, the WebGL1 arms and the caveats become
  collapsed asides under them.

A comparison whose runs disagreed no longer prints a verdict placeholder and a
sentence about what each run concluded. It keeps its numbers and shows the ratio
band those runs spanned, drawn without a side. That band is the one figure the
page derives rather than reads, computed from the published extremes of both
arms, so a cell the measurement could not settle still says what was measured.
It is never a verdict: it has no side, and a band straddling 1.00 is the
statement that the runs could not separate the pair.

Unchanged: every number and verdict still comes from the signed profile, nothing
is aggregated into a score or an overall winner, losses are published like wins,
rows the harness left out are listed with their reasons, and the page still
renders with no profile at all. The harness and the profile JSONs are untouched.

Layout

DocsLayout gains a wide variant: the docs shell caps itself at 1260 px for
prose, which a table page has no use for. The benchmark hues become their own
tokens - the shared semantic palette is tuned for the dark surface and washes
out under a 6 px bar on the light one.

BenchMeasurement.astro is gone; the cell composes both arms now.

Validation

site/pnpm check-ts, pnpm gates typecheck, pnpm gates lint, pnpm gates sync, pnpm gates site, git diff --check, and the browser lane the pre-push
hook selected (site:build + example smoke, 24 passed). Checked in the dev
server at 1600 px and 760 px in both themes: no body overflow, no truncated
cell, page height about 3870 px.

https://claude.ai/code/session_01Y3YY9obdzhWK7nBPby2jK5

[codecov]

Bundle Report

Changes will increase total bundle size by 137 bytes (0.0%) ⬆️. This is within the configured threshold ✅

Detailed changes

Bundle name	Size	Change
site-server-esm	14.3MB	137 bytes (0.0%) ⬆️

Affected Assets, Files, and Routes:

view changes for bundle: site-server-esm

Assets Changed:

Asset Name	Size Change	Total Size	Change (%)
_astro/ExoHeader.B8nrBBs4.js (New)	28.96kB	28.96kB	100.0% 🚀
_astro/ExoHeader.DESjr-0D.js (Deleted)	-28.82kB	0 bytes	-100.0% 🗑️