Prefix-cache budget sweep alternatives — 2026-08-10

Verdict

INCOMPLETE. This packet executes fak’s prefix-cache sweep and a no-cache baseline. libCacheSim, Caffeine, and Redis/Valkey have zero measurements until their real simulators/servers consume the same trace; #6146 tracks those witnesses.

Same-workload contract

Every arm receives five timestamped accesses containing exact repeated prefixes and divergent children, the same four token budgets, an unbounded ceiling pass, and the same write-delay and token-cost assumptions. Full witnesses report hit correctness, reuse ratio/tokens, evictions, ROI-knee error, simulation/runtime latency and throughput, CPU/RSS/storage/network, and total cost.

Arm	Class	Local availability	Honest result
fak native prefix-cache sweep	native	yes	valid four-point curve, positive ceiling reuse, finite knee
no prefix cache	tuned baseline	yes	zero saved tokens; no ROI knee
libCacheSim	external	no	zero measurements
Caffeine simulator	external	no	zero measurements
Redis or Valkey maxmemory policies	external	no	zero measurements

Repository inspection found no equivalent first-class fak integration distinct from the native radix-backed sweep. Adapters and in-memory mocks are not external cache witnesses.

Local native witness

go test ./internal/cachesweep -bench BenchmarkSweepPrefixCacheBudgets -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 22,746; 24,891; 28,420; 30,736; 29,412 ns/op. Median: 28,420 ns/five-pass sweep, 82,136 B/op, 120 allocs/op. This measures local simulation only, not external cache latency or infrastructure cost.

Reproduce

go test ./internal/cachesweep -run TestCompareLocalKeepsCacheSimulatorAlternativesExplicit
go test ./internal/cachesweep -bench BenchmarkSweepPrefixCacheBudgets -benchmem -run '^$' -count=5
go test ./internal/nativebench