External-engine cache capability inventory

Backlog row (see DEFAULT-ENABLEMENT-NEXT-50.md, line 138):

External engines — Build a cache capability inventory for SGLang, vLLM, llama.cpp, Ollama, and LM Studio. Default/evidence target: each adapter row says passive observe, active warm, exact evict, prefix clone, paged KV, or unknown.

Tracked as issue #1549 (parent epic #1490). This is the first row of the “External engines” lane and the foundation the rest of the lane reads from (the engine.CacheCapability contract #1550, the observation adapters #1551–#1553).

The machine-read row-set lives beside this note at external-engine-cache-capability-inventory.jsonl; the focused witness test in internal/enginecache/capability_inventory_test.go loads that JSONL and fails if any engine is missing a row or any verdict falls outside the closed vocabulary.

Closed capability vocabulary

Every verdict is exactly one of: passive observe, active warm, exact evict, prefix clone, paged KV, unknown. No free-text verdict is allowed (the test enforces membership). A verdict states what fak can prove it does with that engine’s cache today — not the engine’s internal mechanism, and not what a future adapter might add.

Inventory

Engine	Verdict	Why (fak-proven capability)	Evidence anchor
SGLang	`passive observe`	fak observes RadixAttention prefix reuse; its only cache-control primitive is a whole-radix reset, not exact-span eviction.	`internal/engine/sglang.go` (`meta_info.cached_tokens` + radix-residency poll → `PrefixResidencyIndex`); `internal/enginecache/enginecache.go` (`flush_cache` whole-radix reset, `SupportsExactSpan(EngineSGLang)=false`).
vLLM	`passive observe`	fak observes paged-block residency via the KV-event stream; its only cache-control primitive is a whole-prefix reset, not exact-span eviction.	`internal/engine/vllm.go` (KV-event subscription: `BlockStored`/`BlockRemoved`/`AllBlocksCleared`); `internal/enginecache/enginecache.go` (`reset_prefix_cache` whole-prefix reset, `SupportsExactSpan(EngineVLLM)=false`).
llama.cpp / llama-server	`unknown`	No in-tree cache observation adapter. It appears only as a pluggable on-device runtime and as a prefill baseline, neither of which observes or controls its cache.	`internal/engine/on_device.go` (named as an `OnDeviceRuntime`, no cache surface); `internal/benchscore` (prefill throughput baseline only). Observation adapter is item 35 / #1553.
Ollama	`unknown`	No in-tree cache observation adapter. Named only as a pluggable local daemon and as an “Ollama-style” model-pull alias.	`internal/engine/on_device.go` (local daemon runtime, no cache surface); `internal/devindex/verbs.go` (`fak model pull` alias only).
LM Studio	`unknown`	No in-tree adapter, observation lane, or cache-control surface at all.	Named only by `DEFAULT-ENABLEMENT-NEXT-50` row 31. Unknown by honest absence of evidence, not a guessed capability.
LMCache	`unknown`	LMCache persists and reuses KV across CPU RAM, local storage, and remote backends, but fak has no adapter, observation feed, or control surface.	Upstream `LMCache/LMCache@4521c3f9f1b8` README documents the external tier and vLLM integration; in-tree absence keeps the verdict unknown rather than inferring support.

Provenance separation

These are all engine-kernel (external-engine KV) facts. They are not provider dollar-cache rebates, not context-window mechanisms, and not forecast planes — the JSONL carries "provenance":"engine-kernel" on every row so a report never blends an observed engine-KV signal into the provider or forecast lane. vLLM’s internal PagedAttention (paged KV) is recorded as an engine mechanism, deliberately kept separate from fak’s proven verdict (passive observe): fak fronting an engine that uses paged KV is not the same as fak controlling paged KV.

Cold-path correctness

This inventory arms no active cache behavior. Every verdict is passive observe or unknown; no active warm, exact evict, or prefix clone is claimed for any engine. The whole-prefix/whole-radix reset that SGLang and vLLM do expose is the safe over-invalidation superset (enginecache degrades to it and records SupportsExactSpan=false), so a cache miss always re-sends full required context. Cold-path correctness is therefore trivially preserved: there is no hit this inventory lets fak depend on.

Generation classification

Horizon: gen/second-next (Generation G2 - Second Next Gen), per the committed program map GENERATION-CACHE-CONTEXT-PROGRAM-MAP-2026-06-30.md which lists #1549–#1558 (“Cross-engine and architecture options”) in the gen/second-next band and names “capability inventory rows” as the exact promotion evidence for that band. This is a citation of the repo’s own classification, not a guess.

Promotion evidence (gen/second-next → gen/next): this inventory is the “cache capability row” the map requires as the minimum evidence to split the External-engines lane into runnable gen/next adapter work; paired with a per-engine cold-path correctness witness (item 38 / #1538), the lane can promote.
Demotion / retirement evidence: an unknown row with no cheap re-witness path (LM Studio has zero in-tree evidence) stays unknown; if a downstream adapter (#1551–#1553) proves a capability and this inventory is not updated in lockstep, the stale row is demotion evidence against trusting the inventory.
Invalidating assumption: the verdict is fak’s proven-today capability, not the engine’s internal mechanism. If the #1551–#1553 observation adapters land, SGLang/vLLM may move beyond passive observe and llama.cpp may leave unknown — this row-set must be re-witnessed then, or it will misreport capability as the lane advances. The vocabulary is also assumed closed at exactly the six listed terms (#1550’s engine.CacheCapability contract must keep the same terms so the two do not drift).

Witness

internal/enginecache/capability_inventory_test.go (TestExternalEngineCacheCapabilityInventory) loads the JSONL and asserts (a) a row exists for each of SGLang, vLLM, llama.cpp, Ollama, LM Studio, and LMCache, (b) every verdict is a member of the closed vocabulary, and (c) the SGLang/vLLM rows do not claim exact evict, tied to the live enginecache.SupportsExactSpan(...) == false fact so the inventory cannot drift from the code it summarizes. Captured go test output is the repo witness.

External-engine cache capability inventory (SGLang, vLLM, llama.cpp, Ollama, LM Studio)

Completion witness for DEFAULT-ENABLEMENT-NEXT-50 item 31 — one row per external engine, each verdict drawn from the closed cache-capability vocabulary, grounded in what the tree already proves (or marked unknown by honest absence of evidence).