[Claude Code - Stage 4] Normalize Messages session identity for cache-aware routing

Part of #113 (Claude Code integration). Related to #18 and vllm-project/vllm#48049.

Problem

The Messages path currently forwards Claude Code's x-claude-code-session-id, but agentic-api does not normalize or act on it. Anthropic Messages remains stateless with respect to transcript persistence, but clients can still provide a stable session coordinate for routing, retention, and cache lifecycle policy.

Without one internal contract, agentic-api, llm-d, and vLLM must each understand client-specific headers or lose the session signal.

Proposed contract

• Claude Code: x-claude-code-session-id → session_id
• Generic Messages clients: X-Session-ID → session_id
• Neither header present: leave session_id unset
• Both headers present with different values: return a clear 4xx error
• Scope the normalized value to the authenticated tenant/workspace
• Preserve the original client headers for protocol compatibility and observability
• Do not derive session identity from cache_control, message content, prompt hashes, or request IDs

Ownership

• agentic-api normalizes protocol-specific identity.
• llm-d can use the normalized coordinate for soft affinity, routing, retention, and lifecycle policy.
• vLLM can receive it as typed request metadata once the upstream contract lands.
• Exact KV reuse is still determined by rendered token IDs, canonical block keys, and event-confirmed residency. session_id is not KV block identity and does not prove residency.

Messages remains stateless in the transcript sense: this issue does not add message persistence, rehydration, or a Responses-style stored-response object.

Acceptance criteria

• /v1/messages recognizes x-claude-code-session-id.
• /v1/messages recognizes provider-neutral X-Session-ID.
• Matching values normalize to one internal session coordinate.
• Conflicting values return a deterministic 4xx error.
• Missing values leave the internal session coordinate unset.
• The normalized value is tenant/workspace scoped.
• Original Messages headers and body remain protocol-faithful.
• The canonical session coordinate can be forwarded to llm-d and bridged to vLLM's typed session_id.
• Tests cover Claude Code, generic Messages, missing, matching, conflicting, and cross-tenant cases.

Out of scope

• Persisting Anthropic Messages transcripts
• Hard replica pinning
• Treating session identity as exact KV/cache-block identity
• Implementing llm-d or vLLM lifecycle policy in this repository

Start at the /v1/messages entry point and trace how request metadata can reach the llm-d and vLLM bridge. Add tests covering Claude Code, generic, missing, matching, conflicting, and cross-tenant headers; done means the accepted values normalize consistently while original headers and the stateless Messages body remain intact.