What — The prompt-safety linear-time regression guard intermittently fails on loaded runners (three hits in 24h across unrelated PRs).
Why — It asserts a single-sample timing ratio (e.g. expected >= 96.35ms, got 94.76ms — a 1.6% miss); at that granularity CI measures the scheduler, not the algorithm.
Scope
modules/ai/tests/ai.prompt-safety.helper.unit.tests.js REGRESSION GUARD "tag matching stays linear...": replace the single-run threshold with a robust form — repeat runs and compare medians, assert the linear/quadratic growth RATIO with a generous margin, or grow the input until the sample is >>1ms. The guard's intent (no quadratic blowup) stays; the flake goes.- Evidence: failed on unrelated diffs (a downstream absorption and a config-only change) with deltas within scheduler noise.
Scope: validated 2026-07-25
Created via /dev:issue
What — The prompt-safety linear-time regression guard intermittently fails on loaded runners (three hits in 24h across unrelated PRs).
Why — It asserts a single-sample timing ratio (e.g. expected >= 96.35ms, got 94.76ms — a 1.6% miss); at that granularity CI measures the scheduler, not the algorithm.
Scope
modules/ai/tests/ai.prompt-safety.helper.unit.tests.jsREGRESSION GUARD "tag matching stays linear...": replace the single-run threshold with a robust form — repeat runs and compare medians, assert the linear/quadratic growth RATIO with a generous margin, or grow the input until the sample is >>1ms. The guard's intent (no quadratic blowup) stays; the flake goes.Scope: validated 2026-07-25
Created via /dev:issue