Rank operations by latency, volume, or inconsistency.

Groups spans by (project, service, span name) over a bounded window and returns count, error rate, total/min/max/avg/stddev duration, p50/p95/p99, and two variability ratios per operation. Sorting is what makes it useful:

• sort_by=total_time (default) — where the wall-clock actually goes.
• sort_by=p95 / p99 — what users actually feel.
• sort_by=variability or tail_ratio — operations whose spread is the problem: the ones that take 40ms most of the time and 4s the rest.
• span_name=payments.charge — the worst this one operation ever got, in max_duration_ms.

Pair the variability sorts with min_count — a ratio computed from three samples is noise, and without a floor it outranks every real signal.

Two bounds are enforced rather than clamped, so a result never claims to cover more than it does: at most 50 projects, and a window no wider than 31 days. Both return 400. Unlike the trace list this report has no early exit — it aggregates every span in the window before it can rank anything.