test(v4): stop redos tests failing on CPU contention - #6360

Fixes #6359.

Both tests in redos.test.ts decide pass/fail from wall-clock timing, so CPU contention from parallel test files flips them. On ead9fcb3 the file failed in 2 of 6 full-suite runs and passed every time it ran alone, which also makes the pre-push hook reject unrelated pushes.

no built-in pattern is ReDoS-vulnerable treated any non-safe status as a failure. unknown only means the checker gave up, and recheck's fuzz checker derives vulnerable from how long an attack string runs — so the verdict tracks available CPU rather than the pattern. Squeezing attackTimeout reproduces it on an idle machine, non-monotonically:

fuzz  extendedDuration  10000ms=unknown  1000ms=vulnerable  100ms=vulnerable  10ms=vulnerable
auto  email             10000ms=safe     1000ms=vulnerable  100ms=vulnerable  10ms=safe

The flagged patterns are fine — matched against adversarial input they scale linearly and stay under a millisecond out to 6.4KB. So this now ignores unknown and re-confirms a vulnerable verdict with a generous budget before failing. A genuinely exponential pattern exhausts any budget, so detection is unchanged: ^(a+)+$, ^(a|a)*$, ^([a-zA-Z0-9]+)*@x$ and ^(\w+\s?)*$ are all still caught.

emoji rejects a backtracking payload in linear time asserted a 100ms budget on a single cold call. The fixed pattern matches that payload in ~0.2ms and stays flat out to 200 emoji, so the threshold sat in JIT and scheduler noise while the regression it guards against took ~1.9s. It now warms up, takes the best of three, and allows 500ms — still three orders of magnitude below the exponential case.