Full-trace analysis backing figure 2a on the real 2h cluster trace:
- f2a_reuse_topology_analyze.py: infinite-KV-cache (LRU) decomposition of
prefix-cache reuse hits into intra-session vs cross-session, by most-recent
prior holder of each content-addressed block.
- f2a_mixture_sweep.py: sensitivity of the intra/cross split to the
single-turn session fraction (tests whether the 93%-intra sample vs 54.6%
full-trace gap is session-mixture selection bias) -- keep all multi-turn
sessions, downsample single-turn to each target fraction, reclassify.
Includes the result JSONs for both.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
