fix(virality): classify alert levels on percentile ranks of both indicators

[PabloPerezBenavente]

The alert levels compared composite_virality and acceleration_rate against absolute cuts, as if the two ran 0..1 on comparable footing. They do not.

composite_virality is a blend of PERCENT_RANKs: bounded and bell-shaped. In production it reaches only 0.904, with a 99th percentile of 0.805 — so the composite > 0.85 cut selected the top 0.14% of scored narratives, not the "top ~15%" its comment claimed. acceleration_rate is a raw ratio capped at 5.0, with a median of 0; acceleration > 1.2 selected the top 0.064%, and 13 of the 14 narratives that reached it had change_views pinned at exactly the cap. Of 21,913 narratives, three were classified — all via the plateau fallback.

Both indicators are now ranked by PERCENT_RANK within the run's cohort, so each threshold means a knowable fraction of that cohort, and a saturating raw value cannot buy a top rank when the whole tail saturates together.

The levels become a partition of the percentile plane:

VIRAL        composite >= 0.95                          (any acceleration)
EARLY_SURGE  composite <= 0.50  and acceleration >= 0.95
ALERT        0.50 < composite < 0.95 and acceleration >= 0.95
WATCH        0.70 <= composite < 0.95 and 0.70 <= acceleration < 0.95
NONE         otherwise

Every pair of regions differs on at least one axis, so the branches are disjoint as written and their order cannot decide the answer — a test asserts this over the whole grid rather than trusting the chain. VIRAL now reads the composite axis alone: a narrative that large is viral whether or not it is still climbing, which is what the old "plateaued-but-popular" branch was reaching for when it emitted WATCH for exactly that condition, badging the corpus's biggest narratives with its weakest label.

A narrative is classified only when it has both indicators. A missing composite used to be read as 0.0, which satisfies composite < 0.65 and made an unscored narrative eligible for EARLY_SURGE on acceleration alone — absence of data masquerading as a signal, for the 90% of narratives that have no composite. Narratives that cannot be scored now have their alert_level cleared rather than keeping a stale badge.

This changes only how the indicators are compared, not how they are computed.