netmesh summarize -- 5 hosts, 20 measured pairs, reports span 30s

  BASELINE   p50 214us across all pairs   worst pair p50 1.9ms
  JITTER     median 11us                  worst pair 640us
  LOSS       0.31% overall  (0.31% forward, 0.00% on the way back)
  PATH MTU   9000 on 18 of 20 pairs; 2 pairs lower; 1 BLACKHOLE

  WORST PAIRS (of 20, ranked by p99)
    web03 -> db01   p50 1.9ms  p99 8.4ms  jitter 640us  loss 0.9% (fwd 0.9%)
    web03 -> db02   p50 1.8ms  p99 7.9ms  jitter 610us  loss 0.8% (fwd 0.8%)
    db01  -> web03  p50 220us  p99 400us  jitter 14us   loss 0.0%

  ASYMMETRY
    web03 -> db01 is 1.7ms slower than db01 -> web03 (8.0x)
      -- the return path is fine, so look at web03's egress queueing

  PATH MTU
    web03 -> db01  BLACKHOLE above 1500 bytes: small packets get through,
      large ones vanish and nothing reports back.
      Applications hang on big transfers and work fine on small ones.

  WHAT TO DO NEXT
    Every slow pair starts at web03, so this is web03's egress or its first
    hop, not the fabric.  Traffic *into* web03 is fine, which narrows it further.
      netmesh paths --compare web03:db01 web01:db01
    If the paths come back clean, the network is not the bottleneck:
      mx run --for 60           (packets per second)
      iperf_orchestrator.sh all (TCP bandwidth)
