=== A2: 2 x 0.50 CPU / 256M ===
  run 1: rps=1020.576 err=0 p50=11.476 p95=44.082 p99=73.405
  run 2: rps=946.385 err=0 p50=12.05 p95=48.707 p99=78.884
  run 3: rps=1053.114 err=0 p50=11.35 p95=41.956 p99=71.286
=== B4: 4 x 0.50 CPU / 256M ===
  run 1: rps=1221.047 err=0 p50=10.126 p95=34.09 p99=54.566
  run 2: rps=1205.189 err=0 p50=10.131 p95=34.418 p99=54.313
  run 3: rps=1215.869 err=0 p50=10.404 p95=33.065 p99=51.443
A2  2x0.50 CPU  rps=1020.6 [946.4..1053.1]  p50=11.5 p95=44.1 p99=73.4  errors=0
B4  4x0.50 CPU  rps=1215.9 [1205.2..1221.0]  p50=10.1 p95=34.1 p99=54.3  errors=0
wrote evidence/scaling/keepalive/summary.csv evidence/scaling/keepalive/rps.svg

=== rolling update ===
14:22:50 recreate api-1 (old container df4ba056eb90)
14:22:56 api-1 healthy (new container 53a60bfc9073)
14:23:06 recreate api-2 (old container 6b53175a3693)
14:23:12 api-2 healthy (new container 047f8e956a97)
{
  "lab": 3,
  "workload": "closed-loop POST messages; no automatic retries",
  "requested_duration_seconds": 60.0,
  "elapsed_seconds": 60.012,
  "concurrency": 4,
  "successful_requests": 53067,
  "failed_requests": 0,
  "error_ratio": 0.0,
  "successful_rps": 884.267,
  "successful_latency_ms": {
    "p50": 3.936,
    "p95": 6.112,
    "p99": 18.691
  },
  "note": "Latency percentiles cover successes only. Timeouts/errors are counted separately; closed-loop has coordinated omission and client overhead."
}
