feat(scripts): separate real usage shifts from PyPI measurement artefacts

Two habits make download data interpretable, and both are enforced here rather
than left to the reader. Normalize against a control package: the 2026-08-25
step moved dashscope and modelscope identically across all three operating
systems with no release from either, so it was a PyPI log change, not a usage
change -- without the control it reads as a 46% collapse in demand. And compare
weekday medians inside one segment, never absolutes across a break date, since
a level shift rescales everything after it.

Breakpoint detection is corrected for the bias a centred median introduces: it
flips as soon as the new level owns the majority of the window, landing raw
detection up to window/2 days early, so each run of flagged days collapses onto
the split point minimizing within-side deviation of a two-level fit. Four
independent series now land exactly on 2026-08-25. Clusters are sign-checked
before being called platform-wide, because a log-export change rescales in one
direction and opposite-signed shifts sharing a few days are coincidence.

A burst that ignores weekends is batch traffic, not people, so spans report
weekend-only and weekday-only outright instead of discarding the strongest
signal as a missing ratio.

Author: zhansheng.lzs