NCCL Events Summary (collective + P2P + RAS)

Aggregate NCCL Inspector events over the last N minutes.

Returns: • Window-level percentiles — mean/p50/p05 bandwidth, p95/p99 duration • by_type block — collective (NCCL ≤ 2.29) + P2P sendrecv (NCCL 2.30+) with separate low-bandwidth thresholds (50 GB/s vs 5 GB/s) • ras block — NCCL RAS subsystem (NCCL 2.24+) totals: critical / warning / max_hang_seconds. RAS events are ingested via POST /api/nccl/ras and automatically reflected here. • ras.diag block — NCCL 2.31.2 ncclras -D diagnostics (24 h lookback, latest result per check × node): checks / nodes / ok / warning / failed / last_run + failed_checks[] with the FAIL detail line.