fix(comparison): make decision tolerance adaptive to observed page jitter

[ebrasha]

What changed

The final decision in _comparison() used a hard-coded DIFF_TOLERANCE = 0.05.
On dynamic targets, normal page noise (rotating banners, timestamps, ad slots, slight markup drift) easily crosses that 5% line and gets read as a real True/False signal.

To fix this, I made the tolerance adaptive:

• when the median-based kb.matchRatio gets locked, I also compute jitter using MAD (Median Absolute Deviation) over the same calibration buffer
• jitter goes into kb.matchRatioJitter
• the final decision now uses tolerance = max(DIFF_TOLERANCE, JITTER_TOLERANCE_MULTIPLIER * jitter)
• on context change, both buffer and jitter get reset so old noise doesn't bleed into a new baseline

The static DIFF_TOLERANCE still acts as a hard floor — pages that are actually stable behave exactly like before.

Why this is better

5% was a reasonable default 10 years ago, but on modern dynamic pages it's just too tight. We were paying the cost in false positives.

• using MAD keeps the noise estimate robust against single bad samples (same reason we use median for the baseline itself)
• multiplier of 3.0 with MAD gives roughly ~2σ-equivalent for normally distributed noise, which is a sane signal-detection threshold
• stable targets see no change because the floor still kicks in
• noisy targets get a wider band that actually matches their real noise level, so we stop labeling random page drift as a successful boolean payload

Scope

lib/request/comparison.py