Аудит адаптивного контура W-TinyLFU в Caffeine

testing

audit-adaptivity is a Claude Code skill that audits the W-TinyLFU adaptive control loop in Caffeine for implementation defects — not algorithm quality — covering the hill-climber, probe machine, and admission-window / main-region resize logic. The scope includes `climb`, `determineAdjustment`, `densityClimb`, `armProbe`, `walkStep`, `probeEnding`, `increaseWindow`, `decreaseWindow`, and `setMaximumSize` inside the recently rebuilt `WindowClimber` class, which now adds a density-signal path and a probe machine for large caches (>4096 capacity). Findings must be concrete defects: wrong arithmetic, sign errors, violated structural invariants (e.g., `1 <= starvation.rung <= 64`, undo returning exactly to `walk.baseWindow`), races, NaN, or overflow. Convergence rate, oscillation, and tuning-constant choices are explicitly out of scope and should not be reported.