kv-router: route vLLM "CPU" KV events to the HostPinned tier (#10368)

[Thachnh]

Cherry-picks upstream ai-dynamo/dynamo ai-dynamo#10368 (d8a94ef28): map vLLM's medium="CPU" KV events to the HostPinned tier.

Why

vLLM's SimpleCPUOffloadConnector tags its KV events medium="CPU". Our fork branched from upstream on 2026-06-02, before ai-dynamo#10368, and only maps "CPU_PINNED"/"CPU_TIER1". So "CPU" fell through from_kv_medium_or_default to the Device tier:

• a CPU-tier eviction deleted a block the GPU still held, so routing thought the pod had lost the prefix;
• a block evicted from GPU but still in the CPU tier looked GPU-resident, so deepapi couldn't rank it below a real GPU hit.

Confirmed on a zai-org/GLM-5.3 prod pod (60 s of its ZMQ stream): GPU stored 7,884 / removed 7,871; CPU stored 8,594 / removed 8,594. The events are balanced and correctly tagged, so the engine is fine and the bug is in the indexer. Every deployed kvtest-* image has it. It affects every model that runs CPU offload with the indexer (GLM-5.3, DeepSeek-V4-Pro, Kimi). On GLM-5.3, DeepApi:kv_tokens reality_best (80.2%) sits below actual (82.4%) over 6 h.

deepapi already consumes tiers (kv_shards_from_indexer_response: gpu vs longest_matched → PRIMARY/SECONDARY), so no backend change is needed.

Changes

• protocols.rs: "CPU" | "CPU_PINNED" | "CPU_TIER1" => HostPinned (upstream)
• lower-tier indexers share the metrics handle, so HostPinned traffic is counted (upstream)
• zmq_wire/tests.rs: conflict resolved by keeping all our tests and adding test_convert_event_cpu_medium_lands_on_host_tier. Upstream's zero-block-size CPU test is dropped because our branch already skips those as partial-prefix entries (90409c7f8).

Validation

• cargo test -p dynamo-kv-router: 550 passed, 0 failed
• image localhost:30500/dynamo-indexer:kvtest-7d461346ca built and pushed
• the same commit on the h24 line applied cleanly (companion PR)

Not done

• Shadow run against live GLM-5.3 pods (pending approval to create the deployment)
• No live indexer swapped. Redeploying kv-indexer:reality wipes a model's routing tree for ~15 min, so the swap is scheduled separately.
• Other upstream indexer fixes since the fork point (fix(kv-router): fail closed on unrecognized KV event media ai-dynamo/dynamo#12456, fix(kv-router): prevent lookup repair from restoring removed entries ai-dynamo/dynamo#11891, fix(kv-router): release unowned radix branches after eviction ai-dynamo/dynamo#14878, fix(kv-router): make standalone indexer cleanup safe ai-dynamo/dynamo#14652, fix(kv-router): restore server-selected buffer-first recovery ai-dynamo/dynamo#14627) are not included. The full port onto v1.5.0 is tracked in DEE-722.

Refs DEE-720, DEE-722

https://claude.ai/code/session_01575gLgRp2Xk4WLR9LsiMu6

[Thachnh]

Update: added b3a2feceac (standalone-indexer TierBridge: promote host-only ancestors when a device store needs them) → image kvtest-11cd973221, live on zai-org/GLM-5.3 reality + routing since 2026-09-24 00:56 UTC (routing on prompt-cache-key throughout).

The bridge does not fix the problem. ParentBlockNotFound is still 150–300/min after warm-up. In the /dump, 458 of 568 rejected stores (81%) have an ancestor in neither tier, so no indexer-side repair can rebuild them. The likely cause is engine-side (the TreeDump / local indexer in our vLLM kv-events patch). That's tracked in DEE-725.

Recommendation for this PR: keep 7d461346ca (ai-dynamo#10368 CPU→HostPinned), since it's the correct tier mapping. Consider reverting b3a2feceac unless DEE-725 shows it helps for the remaining ~19%.

metrics::tests::register_and_encode fails with --features indexer-runtime on the base branch too (pre-existing).