Percentiles per consumer, still the most useful label

One label added in 2025, still the thing that surfaces problems no aggregate shows.

  all consumers   p95 210ms

  integrator-A    p95 190ms
  integrator-B    p95 205ms
  integrator-C    p95 880ms   ← new, since March
  integrator-D    p95 240ms

C's regression is 4% of traffic and moves the aggregate
by 6ms. the alert that caught it compares each consumer
against the aggregate rather than against a threshold.

A consumer at four times the aggregate is a signal that no absolute threshold catches, because the right threshold differs per endpoint. The cause was an include parameter they had started sending, which was a conversation rather than an index — the second time in a year that has been the answer.