The old approach added ±12 noise to scores that span -365..+100 (recency
uses raw Julian days), so the perturbation could never reorder videos that
differed by more than 24 points — which is almost all of them. Every
reshuffle returned the same ranking.
Fixes:
- Per-channel candidate window: rn <= 15 (was rn <= 5) for a much wider pool
- Candidate pool: up to 600 per page (was limit * 4 = 100)
- Non-overlapping page offsets: page N pulls SQL rows N*600 .. (N+1)*600 so
pagination actually moves through new material instead of re-reading the same top-100
- Replaced ±12 perturbation with proper tier-based random sampling:
top 40% → 60% of page, mid 40% → 30%, bottom 20% → 10% wildcards
Each reshuffle picks a genuinely different mix from the score-ranked pool
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Mattias Thall
146a044e69