feat(fuzz): coverage distributions — score/dimension/latency spread + per-cell cost

[drewstone]

Answers "why is there only one dimension of scale?" — there isn't anymore, and the aggregates stopped lying by averaging.

• Distribution { mean, median, p90, min, max, n } everywhere a mean used to be: per-cell headline score, every judged dimension, evaluation latency.
• Latency: engine-measured wall-clock per evaluation (consumer-supplied latencyMs overrides, e.g. to exclude judge time); per-cell + capsule-wide distributions; median-latency KPI with a p90≫median warning accent.
• Per-cell cost: known dollars + tracked-but-unknown runs counted apart; fields absent entirely when cost tracking is unwired — never a fabricated $0.
• Capsule stats: robustness: Distribution | null over per-cell means (cells weigh equally — variance steering sends more runs to weak cells, so run-weighting would bias low), latencyMs: Distribution | null over all runs. Bare meanRobustness is gone.
• HTML: tiles color by mean, tooltips carry median/min/latency/cost; KPIs add cell spread (min–max).
• Determinism preserved (latency excluded from the seed-determinism contract — it is wall-clock by nature).

Greenfield replacement, no compat aliases. In-repo consumers updated; product fuzz CLIs that read stats.meanRobustness / coverage[].robustness update at their next agent-eval bump.

9 new tests (distribution math incl. empty-sample throw, latency override, per-cell cost reconciliation against capsule totals); full suite 2264 passing; typecheck + build clean. 0.91.0.

[tangletools]

✅ Auto-approved PR — 323b8c19

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-11T00:35:14Z}