Observed
Two PR CI runs failed on 2026-09-02 (runs 33627119623 / 33628354448, PRs #263 / #266) on the same single test — ParseCostProbeSpec "should stay far below the 1 ms budget on every realistic body" (core/src/test/scala/app/softnetwork/elastic/client/ParseCostProbeSpec.scala:102):
Cause
The ceiling was calibrated as "~15× the worst measured median" — but that median (0.067 ms at 32 KB) was measured on Apple silicon (the scaladoc records it). A loaded shared GitHub runner measures >20× that baseline; the same failing job shows a 116 ms max on the 3 KB body, which is pure runner noise. The probe measures the raw json4s-jackson parse cost directly, so nothing in a PR's diff can affect it — a red run says nothing about the PR. Introduced by #238 (78d91b7d).
Fix
Keep the probe's stated intent (an order-of-magnitude guard that "catches a change that makes body inspection categorically expensive, never a slow machine") and make the ceiling honest for the machines that actually run it: raise MedianCeilingNanos to 10 ms (~7× the worst CI-observed median, still ~150× the Apple-silicon median) and record the CI measurements in the scaladoc so the next calibration has both data points.
Observed
Two PR CI runs failed on 2026-09-02 (runs 33627119623 / 33628354448, PRs #263 / #266) on the same single test —
ParseCostProbeSpec"should stay far below the 1 ms budget on every realistic body" (core/src/test/scala/app/softnetwork/elastic/client/ParseCostProbeSpec.scala:102):MedianCeilingNanos = 1_000_000.origin/mainbaseline the same day and passed — the test straddles the budget depending on which runner the job lands on.Cause
The ceiling was calibrated as "~15× the worst measured median" — but that median (0.067 ms at 32 KB) was measured on Apple silicon (the scaladoc records it). A loaded shared GitHub runner measures >20× that baseline; the same failing job shows a 116 ms
maxon the 3 KB body, which is pure runner noise. The probe measures the raw json4s-jacksonparsecost directly, so nothing in a PR's diff can affect it — a red run says nothing about the PR. Introduced by #238 (78d91b7d).Fix
Keep the probe's stated intent (an order-of-magnitude guard that "catches a change that makes body inspection categorically expensive, never a slow machine") and make the ceiling honest for the machines that actually run it: raise
MedianCeilingNanosto 10 ms (~7× the worst CI-observed median, still ~150× the Apple-silicon median) and record the CI measurements in the scaladoc so the next calibration has both data points.