fix(comparison): avoid first-sample lock in matchRatio calibration

[ebrasha]

What changed

kb.matchRatio used to get locked on the very first valid ratio it saw.
That was fragile — if the first response happened to be noisy (CSRF token, ad slot, timestamp widget, A/B fragment), the whole baseline for that parameter ended up biased from the start.

To fix this, I switched the calibration from single-shot to multi-sample:

• valid ratios are now collected into a small buffer instead of locking immediately
• once we hit a minimum sample count, kb.matchRatio gets set from the median of that buffer
• there's a hard cap on the buffer so it can't grow forever
• on context change (different page template / textOnly / titles mode), the buffer resets so old samples don't bleed into a new baseline

Why this is better

The real problem was outliers. One bad first response could throw the whole detection logic off for that parameter. Median makes those single-sample spikes basically harmless.

In practice:

• the baseline takes a few more samples to settle, but it settles cleaner
• false positives that came from early-response noise drop a lot
• nothing else in the comparison flow gets touched — only the calibration step itself was rewritten

Scope

Only lib/request/comparison.py is touched.
The behavioral change is limited to how kb.matchRatio is determined.
Detection paths, ratio bounds, and decision logic stay as-is.