Interface ArmReport

Aggregated metrics: mean for numbers, fraction-true (0..1) for booleans.

Per-metric mean / std / se / n, so an A/B gap can be read for significance.

Stub invocations no rule answered, across this arm's trials. Non-empty means the arm measured at least one answer nobody wrote: vigiles eval exits 2 on it and --update refuses to record the lock. Empty when the eval has no stubs.

Cost / latency / token totals + means for this arm.