ben-manes / caffeine / #5823

100%

push

github

record the prefix-cache and warm-up findings

The window climber is judged against policies that see the same keys
(LRU, ARC, the best fixed window). A policy that reads what the access
stream omits, such as a disk cache's block addresses or an LLM server's
request kind, turn and prompt depth, can beat it on its own traffic;
that gap is a separate comparison, not a climber defect. The design
rules, hill-climber.md, the adaptive-window page and /audit-regret now
say so.

Replaying a short trace back to back reuses its keys, so the frequency
sketch remembers the previous pass and flatters small windows. The
prefix-cache rows looked like a steady-state loss for that reason; with
fresh keys per pass the shipped climber converges within two to four
passes. On a prefix cache a hit after a request's first missing block
saves nothing, so block hit rates overstate a small window's benefit.
/climber-gate's harness notes and the methodology cover both.

The graveyard gains the LRU-estimate audit trigger (no gain once the
replays are honest, at 2.7-3.7% throughput for 5-10K entries) and the
warm-up accelerations: a geometric jump, momentum, short samples, and a
fill-drop gate that is clean on real traces but fails slowswap and
mixture_d050_long because the jump parks a coarse landing or shifts the
later audits. Seeds vary admission draws but not a trap's timing, so
both skills now say to move a fixed-time regime change before reading
an eight-seed loss as a mechanism.

4823 of 4832 branches covered (99.81%)

9268 of 9269 relevant lines covered (99.99%)

1.0 hits per line