Resolve deployment-owned vLLM serving plans for Slurm

Priority Level

High

Task Summary

Implement the package-owned vLLM resolver that converts one typed deployment declaration plus planner-supplied placement into process, readiness, backend-endpoint, and logical-endpoint records.

Technical Details & Implementation Plan

• Dispatch on the deployment server discriminator with vLLM as the only implemented backend.
• Keep model alias, model source, and optional served model name distinct.
• Resolve tensor parallelism within a node and pipeline parallelism across nodes in a replica group.
• Derive replica groups, ranks, lanes, rendezvous inputs, and lane-head backend endpoints from planner-owned placement.
• Carry per-deployment startup and distributed-initialization deadlines, launch standoff/stagger, queue backpressure, readiness configuration, tokenized extra arguments, and typed environment bindings.
• Reject extra arguments that override compiler/runtime-owned model identity, ports, topology, rendezvous, launch timing, or middleware behavior.
• Emit typed logical-endpoint inputs so the runtime can aggregate readiness and load-balance healthy backends while preserving overload responses.
• Preserve the inspected vLLM runtime version as digest-bound provenance without maintaining a package-version compatibility matrix.

Acceptance criteria

• A single-node deployment resolves deterministically without importing scheduler or shell modules.
• A multi-node deployment derives the expected replica groups, ranks, pipeline parallelism, and lane-head endpoints.
• Two deployments may select different serving images without endpoint or resource identity collisions.
• Invalid node, GPU, tensor-parallel, replica-group, expert-parallel, image-kind, and resolved-behavior combinations fail before submission.
• Default and overridden launch-timing and queue-backpressure values serialize in golden records.
• A rank failure is represented as failure of the coordinated deployment; no per-replica restart contract is introduced.

Out of scope

• Slurm submission or node allocation.
• Shell rendering, process supervision, and cleanup.
• Image building or registry mutation.
• Dynamo, SGLang, serving plugins, and per-replica recovery.

Investigation / Context

This is a feature lane under #850. Portable deployment intent remains separate from backend-specific process resolution so configuration stays declarative.

Agent Plan / Findings

Build against the reviewed shared configuration, image-inspection, placement, and runtime record contracts. Keep the resolver pure with focused single-node and multi-node golden tests.

Dependencies

Blocked by the shared-contract and fake-infrastructure foundation work tracked by #850.