[dongfengy]

Description

Conversation affinity can balance the number of conversations while leaving their cached prompt-token load uneven. Add opt-in placement by each rank's live pinned prompt tokens, with round-robin tie breaking and the existing request-count limits and sticky bindings preserved.

Enable with TLLM_ADP_TOKEN_AWARE_PLACEMENT=1 alongside conversation affinity. The load estimate replaces a conversation's previous prompt length and expires by write volume or LRU eviction. Explicitly ranked first turns are included in the same accounting. The default routing behavior is unchanged.

Test Coverage

Carries the source change's routing tests for opt-in gating, placement, expiry, LRU bookkeeping, stickiness and rank determinism. Static AST parsing and git diff --check passed. No builds or tests were run, as requested for draft preparation.

PR Checklist

• Reviewed the applicable description, coding-guideline, API, dependency, ownership and documentation requirements.
• This is an opt-in routing policy and introduces no package dependency or LLM-args field.
• Runtime validation through CI.