benchmark: refuse to publish a run that mostly answered 5xx

[MDA2AV]

Follow-up to the hono-bun log in #1251.

There is no error gate anywhere. save_result writes the JSON
unconditionally; status_5xx is recorded and never acted on, so a run can fail
almost entirely and still publish a number.

That is worse than it sounds, because an error is cheaper to serve than a real
response — a server that fails fast scores higher than one that works.

php-fpm publishes 294,864 rps for a baseline-h2 run in which
1,462,506 of 1,474,323 responses — 99.2% — were 5xx, and that number counts
toward its composite score.

Fix

save_result computes the 5xx share of all responses and refuses to write the
result above HTTPARENA_MAX_ERROR_PCT (default 5). The container log is still
written so the run can be diagnosed, and any stale JSON for that profile/conns is
removed.

Blast radius

Seven already-published runs are above the threshold and are dropped here:

framework	profile	err%	was
php-fpm	baseline-h2-1024	99.2%	294,864 rps
symfony-spawn-franken	baseline-h2-256	72.9%	28,243 rps
symfony-spawn-franken	static-h2-1024	56.5%	52,582 rps
symfony-spawn-franken	static-h2-256	47.3%	63,518 rps
symfony-spawn-franken	baseline-h2-1024	36.5%	97,817 rps
rage	api-16-1024	15.2%	9,825 rps
rage	api-4-256	8.8%	13,999 rps

Those three frameworks now score 0 on the dropped profiles, the same as any entry
that does not publish one. php-fpm and symfony-spawn-franken lose most of their
h2 numbers, which is the intended outcome — those runs were not results.

The 24 other runs carrying some 5xx are all under 5% and are left alone; the
worst is hono-bun's fortunes at 3.7%.

site/leaderboard/data.js regenerated; check_badge_parity.js passes —
751 ranks match.

Judgement call

5% is the threshold the codebase already documented in a comment that no longer
matched any code. If you would rather draw the line somewhere else, it is one env
var, and re-running the purge is a one-liner.

🤖 Generated with Claude Code