Bound on the two one-shot counter probes (the flush's baseline read and
reportFlushVerdict's follow-up).
Both are trivial property reads that answer in milliseconds on a live page,
so this is not sized for the WORK — it is sized for the wait to be
SCHEDULED. Deliberately smaller than the 15 s the spec's two detector probes
get, because those probes' answers are what the tests assert on whereas a
missed probe here only downgrades a verdict to a warning. A larger bound
would not rescue the case it might seem to: main-thread starvation was
measured to outlast any bound that fits in the test budget (a trivial
evaluate unanswered for 5 s, twelve times in a row, ~78 s in all, while the
page went on rendering), so past a few seconds the extra latency buys no
extra answers.
Bound on the two one-shot counter probes (the flush's baseline read and reportFlushVerdict's follow-up).
Both are trivial property reads that answer in milliseconds on a live page, so this is not sized for the WORK — it is sized for the wait to be SCHEDULED. Deliberately smaller than the 15 s the spec's two detector probes get, because those probes' answers are what the tests assert on whereas a missed probe here only downgrades a verdict to a warning. A larger bound would not rescue the case it might seem to: main-thread starvation was measured to outlast any bound that fits in the test budget (a trivial evaluate unanswered for 5 s, twelve times in a row, ~78 s in all, while the page went on rendering), so past a few seconds the extra latency buys no extra answers.