Released-engine integration

Read AGENTS.md, docs/adapters.md and S5 in docs/completion-plan.md from the Git root. The completion plan is the only work queue. Use its current support status; do not treat an upgrade target or upstream feature as qualified behavior.

Select the actual contract

For an upgrade, check the latest official non-prerelease tag and record its commit. Inspect that source and the corresponding released LMCache integration. Use main and related issues/PRs to find fixes; distinguish unreleased fixes explicitly. Freeze the selected release during review and update pins, lockfile and submodule together after the consumed adapter passes. Do not add old/new API fallbacks.

Keep one owner per boundary

• Engines own request scheduling, GPU allocation, native prefix trees, rank collectives, model execution and final request readiness.
• Python translates engine layout, state boundaries, callbacks and CUDA readiness. Existing Rust owners manage shared cache policy, transfers, admission and drain.
• vLLM uses the KV Connector contract; SGLang uses UnifiedRadixCache and the external linker. Preserve valid native HBM hits without external synchronous lookup. Keep unselected backend imports free of native initialization.
• vLLM 0.31.0 workers return KVConnectorTransferResults directly. P/D failed receives must also be finished receives in the same poll; do not restore a custom worker-metadata failure queue. Ordinary cache delivery is best effort, while a P/D producer retains the native reliable-delivery requirement.
• SGLang 0.5.21 cancellation already reaches the external linker through BasePrefixCache.finish(ABORT) and UnifiedRadixCache.release_aborted_request. Keep cancellation/drain in that consumed lifecycle; do not restore a duplicate private Scheduler abort Hook. Queue preparation is opt-in; fork P/D observation Hooks are removed.
• Dense TP shard queries use Manager-owned native fan-out, not Python foreign UDS loops. Official worker handshake exports opaque registered targets; local authenticated configuration validates complete rank/node coverage. Require --enable-query-control; one-node consumption explicitly opts into orbitkv.query_control=true. Keep trusted literal loopback scope and equal TP-only/one-group V1 guards. Speculative/deferred-finalize metadata ordering cannot report native completion safely, so it is rejected in this profile. Worker metadata reports unique global ranks only after wait_restore; release retained source interests only after all declared ranks complete. Run the two-real-Manager installed byte/SSD gate separately from the official one-node handshake serving gate; neither qualifies physical multi-GPU/cross-host TP.
• P/D uses official vLLM NIXL/MultiConnector and SGLang native disaggregation; Manager shared-cache traffic still uses TENT. Do not reintroduce fork factories, callbacks, custom connectors, handshake/proxy or partial-tail cache modes.
• The candidate profile reads/writes cache on P and only saves on D. Native P/D owns incoming destinations; cache retains completed state until its own drain. Preserve reliable P/D delivery and release unselected cache query leases.
• Official vLLM 0.31.0 requires V1 (VLLM_USE_V2_MODEL_RUNNER=0) and one attention cache group. Reject V2/recurrent profiles before connection; the runner and native-prefix monkey patches are removed. Reopen only after consumed released ordering and atomic state hand-off pass GPU/model gates.
• SGLang native P/D candidate enables released deferred KV release on both P and D, but 0.5.21 still releases on timeout without a full drain ACK. Keep transfer cancellation, peer-loss and delayed-ACK page reuse unqualified; ordinary linker cancellation is a separate cache-owned contract.
• vLLM's public reset_cache() now invalidates future external reuse through a fresh prefix-key generation only after all tracked request/query/load/save owners end. HBM-only reset preserves the current generation. Test idle reset_external=true followed by a cold miss and reuse of newly saved state; do not advertise active-request reset, persistent cluster-wide invalidation or live weight changes that bypass the callback. SGLang ordinary flush keeps same-weight external reuse and is not a universal weight-update notification.
• Fixed local PEFT adapters require startup content fingerprints and effective LoRA settings in the deployment namespace. Keep native base/adapter hashes intact; reject resumable/session paths and caller extra-key aliases that bypass safe matching. This immutable profile does not intercept engine SDK/RPC add/remove/reload or live-weight updates; change artifacts only after stopping the engine.
• SGLang 0.5.21 preserves raw ServerArgs and publishes resolved configuration separately. Consume runtime_context.get_lora() for effective enable, startup LoRARef UIDs, backend and target modules, as its native loader/registry do. --lora-paths does not make the raw enable_lora field true.
• Explicit FULL decode Graph cache recovery is independently qualified only for the recorded H20 dense TP1/PP1 serial batch-1 eight-token direct/kernel io_uring profile on both official releases. Use --release-cuda-graph in the installed gate and require native runtime observations; capture-only logs do not qualify. Preset sizes 2/4, long decode/save boundaries, adapters with Graph, concurrency, preemption/cancellation, prefill capture and physical topology remain separate.
• Run native model composition and lifetime gates separately. Output/restart does not prove cancellation, partial-submit, delayed-ACK or page-reuse safety. Earlier fork passes remain upstream contribution evidence, not release support.

Replace internal coupling safely

Audit every remaining Hook against docs/engine-release-audit.md. SGLang 0.5.21 captures graphs before the tree-cache factory; preserve pre-capture events until a safe released callback exists. Its linker lookup lacks pending tickets, so pending-query admission remains a guarded internal Hook; queue preparation is opt-in. Do not describe these as public lifecycle integration. Keep native abort finish, cancellation/drain and page ownership until a replacement is consumed. Do not add a facade merely to hide internal coupling.

For a stuck or incorrect request, read the boundary tracing guide. Enqueue, layer readiness, final native drain and source retirement are distinct. Timeout, process disappearance and metadata expiry cannot prove physical drain.

Verify the claimed profile

Run the relevant gates from AGENTS.md: source-only tests for local contracts; actual engine/GPU tests for changed callbacks, preemption, page reuse and graph ordering. Include cold/partial/full hits, native HBM, restart and failure controls. Preserve adapter/salt/model identity and effective committed lengths. Record TP, PP, attention DP, EP, P/D topology and transport separately; a one-GPU process test does not qualify their distributed combination. Keep evidence outside checkout.

For SGLang copy-backend changes, use benches.sglang_kv_oracle to check the actual released pool, full payload/sentinel coverage, partial destinations, event generations and deliberate corruption through the installed native path. Its deterministic fixture bytes do not dump model-computed KV. Evaluate the consumed model separately with benches.kernel_pairs: freeze fresh independent process pairs, exact input/output and installed artifacts, real shutdown receipts and the original benefit/regression guards. Keep profiler clocks separate. The accepted serial profile and its lifetime limits are recorded in S5.5; neither fixture success nor pair medians select a general backend default.

For upstream delivery, separate baseline registration/configuration/tests/docs, generic lifecycle fixes and TENT transport integration into reviewable PRs. Use a clean engine checkout plus installed wheel as the acceptance environment. Report local qualification, upstream merge and released support separately.