refactor(api): prune unused/redundant/reconstructable RPCs

[connorcarpenter15]

Summary

Trims the OpenEngine v1 contract to the RPCs and message schemas a client actually needs. The Dynamo sidecar is the single OpenEngine client for the vLLM, SGLang, and TensorRT-LLM servers; scope is defined by what it calls.

Removed — 4 genuinely-unused RPCs + their message trees

• Inference.Embed / Classify / Score — the sidecar only calls Generate. Deleted embedding.proto, classification.proto, scoring.proto, and the shared tasks.proto; dropped ModelInfo.tasks (field 27) and TaskCapabilities from model.proto.
• Control.SubscribeRuntimeEvents — no client or server implementation. Removed SubscribeRuntimeEvents{Request,Response}, RuntimeEvent, RuntimeEventType from observability.proto (GetLoad's LoadInfo/RankLoadInfo stay).

Removed — 1 redundant RPC (data preserved elsewhere)

• Control.GetKvConnectorInfo + GetKvConnectorInfoRequest — returned the exact KvConnectorInfo already on ServerInfo.kv_connector (GetServerInfo). All three engines populate both surfaces from the same value; the sidecar fetched ServerInfo, then re-fetched the identical connector info and discarded the first. Follows the contract's static-vs-dynamic split: static discovery data (parallelism, capacity, connector) rides on ServerInfo; only dynamic state (LoadInfo via GetLoad) keeps a poll RPC. KvConnectorInfo is kept — now ServerInfo.kv_connector's sole source.

Removed — 1 reconstructable RPC

• Control.Drain + DrainRequest/DrainResponse/DrainState — graceful shutdown doesn't need an engine-facing drain RPC. The client stops admitting (discovery unregister), lets in-flight Generate streams complete, polls GetLoad for quiescence, and cancels stragglers via gRPC stream cancellation into the engine's existing Abort path. Dropped lifecycle.proto's now-unused error.proto import. Kept the drain-adjacent state signals that stand on their own — HEALTH_STATE_DRAINING, KvConnectorInfo.supports_drain (validated by the sidecar for disagg connectors), ERROR_CODE_DRAINING — since an engine still drains on its own SIGTERM.

Retained — SubscribeKvEvents

Not unused: the sidecar consumes it on the grpc/protobuf KV-event transport advertised by GetKvEventSources (kv.rs subscribe_loop → Control::subscribe_kv_events). Only TensorRT-LLM serves it, because its KV events are available solely through an in-process API — vLLM/SGLang advertise a native ZMQ/msgpack publisher instead.

Final Control surface

GetServerInfo · GetModelInfo · GetLoad · Health · Abort · LoadLora · UnloadLora · ListLoras · GetKvEventSources · SubscribeKvEvents (+ Inference.Generate).

Validation

• protoc -I proto --experimental_allow_proto3_optional builds the full descriptor set clean.
• markdownlint-cli2 clean on all Markdown.

Consumer follow-up (separate repos, not this PR)

• Embed/Classify/Score/SubscribeRuntimeEvents: engine servers implement these as trait stubs; regenerated stubs drop the methods, so the stubs should be removed.
• GetKvConnectorInfo: the sidecar reads engine.kv_connector from GetServerInfo and deletes its second discovery call (client.rs:144–152); the three servers drop the GetKvConnectorInfo method but keep the kv_connector_info() helper that feeds GetServerInfo.
• Drain: the sidecar's drain() override (engine.rs:736) and the three servers' Drain handlers are removed; the sidecar's LLMEngine::drain reverts to the default no-op and graceful shutdown relies on the is_quiescent/GetLoad + grace-period + stream-cancellation path already in Worker.

Compatibility

Wire-breaking (removes RPCs and message types). No known consumer depends on the removed surface; GetKvConnectorInfo's data is preserved on ServerInfo.kv_connector.