fix-ci-e2e-flake

First classify the failure: run the spec locally with --repeat-each=5, check whether retry recovers it, and inspect git history for recent changes to the spec. Then read the call log failure shape—fluctuating counts mean background noise, stable wrong values mean the tested subject never became ready—and harden the assertion accordingly.

A flake passes locally with --repeat-each and recovers on CI retry, while a regression fails deterministically everywhere and retry cannot save it. Also check whether the same commit fails consistently, and use git stash on a clean baseline to compare behavior.

CI runners are slower, so a poll window equal to the global test timeout gets cut off before the condition is met. Common fixes include adding test.slow(), raising the specific poll timeout, and removing hardcoded timeouts that override CI-scaled config values.

The element is visible but a sibling element overlaps it at the click point, so the hit-target check fails. This often happens with horizontal layout overflow at CI viewport sizes; fix by adjusting layout assumptions or scrolling the target into a clickable position.

No. Robustness means excluding background noise while keeping the tested assertion at full strength. If the only way to pass is loosening the real assertion, treat it as a likely genuine regression rather than a flake.

After every diagnosed flake, append the failure signal, root cause, and fix to the case reference file and update the signal index. This builds a lookup table so future failures with the same signature are resolved in minutes.