{"engine":"sglang","display":"SGLang","verdict":"passive observe","provenance":"engine-kernel","engine_internal_mechanism":"RadixAttention prefix/radix KV cache","fak_control":"whole-radix reset only (flush_cache); exact-span eviction not exposed","cold_path":"no active warm/evict armed by fak; a miss re-sends full context","evidence":"internal/engine/sglang.go — RadixAttention prefix-hit count (meta_info.cached_tokens) + radix-residency poll feed PrefixResidencyIndex (observe only); internal/enginecache/enginecache.go — flush_cache whole-radix reset, SupportsExactSpan(EngineSGLang)=false (exact-evict gap, NOT exact evict)"}
{"engine":"vllm","display":"vLLM","verdict":"passive observe","provenance":"engine-kernel","engine_internal_mechanism":"PagedAttention paged KV + prefix cache (engine-internal; fak does not control it)","fak_control":"whole-prefix reset only (reset_prefix_cache); exact-span eviction not exposed","cold_path":"no active warm/evict armed by fak; a miss re-sends full context","evidence":"internal/engine/vllm.go — KV-event subscription (BlockStored/BlockRemoved/AllBlocksCleared) folds paged-block residency (observe only); internal/enginecache/enginecache.go — reset_prefix_cache whole-prefix reset, SupportsExactSpan(EngineVLLM)=false (exact-evict gap). vLLM's paged KV is an engine mechanism, not a fak-proven capability — kept separate."}
{"engine":"llama.cpp","display":"llama.cpp / llama-server","verdict":"unknown","provenance":"engine-kernel","engine_internal_mechanism":"unknown to fak — no in-tree cache observation adapter","fak_control":"none","cold_path":"n/a — no active cache behavior in-tree","evidence":"internal/engine/on_device.go names llama.cpp (CGo) / llama-server as a pluggable OnDeviceRuntime but exposes NO cache surface; internal/benchscore uses llama.cpp only as a prefill throughput baseline, not a cache adapter. The observation adapter is out of scope here (item 35 / #1553)."}
{"engine":"ollama","display":"Ollama","verdict":"unknown","provenance":"engine-kernel","engine_internal_mechanism":"unknown to fak — no in-tree cache observation adapter","fak_control":"none","cold_path":"n/a — no active cache behavior in-tree","evidence":"internal/engine/on_device.go names Ollama's local daemon as a pluggable OnDeviceRuntime (no cache surface); internal/devindex/verbs.go 'fak model pull' is an Ollama-style download alias only. Neither observes nor controls Ollama's cache."}
{"engine":"lm-studio","display":"LM Studio","verdict":"unknown","provenance":"engine-kernel","engine_internal_mechanism":"unknown to fak — no in-tree evidence","fak_control":"none","cold_path":"n/a — no active cache behavior in-tree","evidence":"No in-tree adapter, observation lane, or cache-control surface for LM Studio; named only by DEFAULT-ENABLEMENT-NEXT-50 row 31. Verdict is unknown by honest absence of evidence, not a guessed capability."}
{"engine":"lmcache","display":"LMCache","verdict":"unknown","provenance":"engine-kernel","engine_internal_mechanism":"persistent tiered KV offload/reuse across CPU RAM, local storage, and remote backends (external to the serving engine)","fak_control":"none","cold_path":"n/a — fak has no LMCache observation or control adapter","evidence":"Upstream LMCache/LMCache README @ 4521c3f9f1b8 documents persistent tiered KV offload/reuse across CPU RAM, local storage, and remote backends, with vLLM integration. In-tree grep finds no LMCache adapter or control surface; the row is unknown, not inferred support."}