extract_and_group stored the per-error-type percentage already rounded to
2 decimals, and is_important() compares that stored value against the
threshold. Rounding before the comparison can push a rate that is strictly
below the threshold up to it:
8 Captcha errors / 267 sites = 2.99625%
round(2.99625, 2) == 3.0 -> is_important() True -> spurious
'Too many errors of type "Captcha" (3.0%)' warning
This is the residual of the same class fixed in #2788 (scale-then-round):
scaling to a percentage first stopped 2.5% from becoming 3.0%, but rounding
the percentage still feeds the threshold check. The display sites already
round (round(e["perc"], 2)), so storing the raw percentage keeps the shown
values unchanged while making the threshold comparison exact.
Genuine at-or-above-threshold rates still fire (3/100 = 3.0%, DNS 10%).
Adds a regression test for the 2.996% -> 3.0% rounding-up case.
`extract_and_group` computed the per-error-type percentage as
`round(count / len(search_res), 2) * 100` — it rounded the *fraction* to 2
decimals (i.e. to the nearest 1%) and only then scaled by 100, so `perc`
always had whole-percent granularity and was mis-rounded.
A genuine 2.5% error rate (1 error out of 40 sites) became
`round(0.025, 2) * 100 = 3.0%` and tripped the 3% "important" threshold in
`is_important()`, firing a spurious "Too many errors of type ..." warning.
Likewise 9.5% DNS errors rounded to 10.0% and crossed the DNS override.
Scale to a percentage *before* rounding: `round(count / len(search_res) * 100, 2)`.
The existing tests use integer-landing ratios (25/100, 5/100, 3/100) that are
identical under both formulas, so none of them change.
Adds a regression test for a sub-threshold non-integer rate (2.5%) staying silent.
