feat(site): surface latency and tokens as leaderboard metrics

[jessie1111101]

Summary

Efficiency data was already in Firestore and the dashboard could never show it. Every ResultRow carries latencySec and the token buckets, but derive() didn't project them into setups — and the browser only reads setups. This adds them as two metrics alongside the quality scores.

No producer or ingest changes needed. The data has been landing in the results collection all along; the missing link was one function plus the UI's percentage assumptions.

What's here

• derive() (and its seeder mirror) average latency and total tokens per task and per history point. Efficiency is treated as telemetry, not a score: it's averaged over all rows rather than the scored subset, and it survives a cell where nothing scored — an unscored iteration still burned wall-clock and tokens.
• Metrics now carry presentation rules (unit, lowerIsBetter, percentage). Quality metrics keep their existing behaviour verbatim; latency and tokens format by unit (42.7s, 38.4k), rank ascending, and scale their bar against the largest value on screen — inverted, so the fastest setup still gets the fullest bar and "longer is better" stays true everywhere.
• Direction is respected throughout: leaderboard sort, the detail page's Best Task card (which would otherwise show the slowest task), and the trend chart's y-axis, which previously clamped to [0,100] and would have flattened every latency series onto the top gridline.
• Tokens prefer the producer's own totalTokens when present and otherwise sum the captured buckets, staying null when the harness recorded no usage — so "not measured" never renders as zero.

Notes

• Stacked on feat(site): dark mode (persisted toggle + theme-aware chart) #216 → feat(site): surface v1 composite + correctness + safety + catastrophic (frontend Phase 1) #206; retargets to main as those land.
• Sorting puts a setup with no value for the metric last in both directions, rather than treating null as 0 and making it look like the best latency.
• turns is deliberately not included — it's still unmerged in feat(results): emit 'turns' efficiency axis on the result row #215. It slots into the same machinery as a third key when that lands.
• Cost isn't here either: it needs a price table over the cached/cache-write buckets, which is its own change.

Test plan

• vitest — 107 passed (12 files), including new coverage for the formatter, direction, and bar-scaling
• vite build --mode staging clean
• Seeded the emulator and confirmed latency and tokens now appear on all derived setup docs (132 occurrences each)