LMArena history: three non-uniform board refits anchors can't absorb

In LMArena text overall history, most board-to-board changes are uniform shifts (e.g. 2025-05-22 -> 06-01 is +10 for everyone, spread 0.04), but 2024-06-23->06-26 (vote recount, ~5% fewer votes, spread ~5 pts among 20k+ vote models), 2025-07-17->07-25 (no new votes for 173 models yet ratings spread ~2.8; CIs switch from asymmetric bootstrap to exactly r +/- 1.96*sqrt(variance)) and 2025-09-08->09-17 (recount, spread ~2.6) reshuffle models non-uniformly. Median-of-anchors corrections don't remove these; check per-step robust spread of rating changes before treating rating differences as vote noise. vote_count also decreases on most boards (small dedups), so n1-n0 is net of recounts.