[krishung5]

Summary

Add encoder-only execution and engine-owned encoder-output handoffs to OpenEngine. A client receives an opaque reference from an encoder and forwards it to a compatible generation engine; tensor contents stay outside the inference RPC payload.

• Add typed EncoderOptions, EncoderReady, and EncoderCacheRef messages.
• Negotiate encoder/consumer roles and transfer backends through ModelInfo.encoder.
• Add Control.TransferEncoderCache; success means transfer completed, and an uncertain single-use handoff must not be replayed.
• Document capability negotiation, invalid states, terminal events, transfer completion, cancellation, and ownership in the API reference.

Why

Refs #17.

This is a smaller first step toward the encoder-disaggregation use case in #17. It reuses Generate and keeps backend-specific artifact layouts opaque rather than introducing the separate media service and portable tensor descriptors proposed there. That design choice is open for review; this PR does not complete the broader issue or add a new EngineRole enum value.

The reference implementation covers image E+PD and E+P+D at TP/PP/CP=1. E+P+D consumes the encoder reference at P, then uses the existing native KV-session handoff for D; it needs no additional encoder-protocol fields. Video/audio encoder handoff, reusable encoder caches, explicit artifact-release APIs, and portable tensor descriptors remain outside this change.

Validation

• buf build
• buf lint
• Buf formatting and breaking-change checks against main
• Markdown lint passed for the updated API reference; existing links were preserved
• Commits include a DCO sign-off

The additions retain existing field numbers and RPCs. Matching TRT-LLM and Dynamo implementations have CPU regression coverage and prior cross-host native gRPC and Dynamo HTTP validation for image E+PD and E+P+D. The dependent PRs record their exact validation scope and runtime prerequisites.

Protocol checklist

• The canonical schema and API documentation are synchronized
• Presence, defaults, and invalid states are defined
• Acceptance, streaming, terminal, and error behavior are defined
• Discovery and capability impacts are documented
• Intentional schema rewrites are called out clearly: no existing fields or RPCs are removed or renumbered

Additional context

Dependent implementations: TRT-LLM transport, TRT-LLM serving, and Dynamo integration. They pin this proposal's immutable source revision. This PR can be reviewed independently of the known CUDA IPC shutdown cleanup issue in the TRT-LLM reference implementation.