[sgl-router] Serve chat completions through the policies_reorg resolver

[sherlockwu]

Motivation

Sixth PR of the engine-selection redesign, on top of #40277. The server now routes every chat completion through the new resolver. The legacy policies module stays in the tree (its Policy trait, factory and helpers still compile) so this PR is only the switch; the removal and the _reorg rename follow.

Modifications

• server/app_context.rs: AppContext::with_engine_state(config, tokenizers, proxy, registry, active_load, kv_index, prefix_index) builds Dependencies (affinity store with its sweeper, prefix source from the remote indexer or the local KV index, metrics) and BucketResolver from build_pools. AppContext::new is the four-argument convenience without engine state. The legacy policies, bucket_selector, prefix_index and radix_tree_prefix_provider fields are gone.
• server/routes/chat.rs: classifies the deployment through PdPoolResolver to choose the first stage, builds one PickRequest with a shared LoadView and PrefixMemo, picks prefill then decode through ctx.buckets, and maps PickError onto the existing API errors and selection-failure metrics. Unknown models are 404 before tokenization.
• main.rs no longer builds the legacy policy registry.
• buckets_reorg::Pools carries needs_request_tokens and needs_dispatch_timestamps; Bucket keeps its pending-prefill budget so the affinity group applies each engine's own bucket budget (PendingPrefill is a per-engine map).
• Metrics: AffinityPolicy records sgl_router_sticky_total outcomes for sticky; CacheAwarePolicy records cache admission evaluations and rejections and the cache-aware decision. Decision reasons use the legacy labels (assigned, remap, session_primary, sticky_hit, cache_candidate, no_cache_candidate).
• Removed: policies/selection.rs, policies/decode.rs, their component test, tests/proxy/shared_prefill_admission.rs (it drove the legacy proposal and backup contract directly), the two indexer unit tests in chat.rs, and the CacheCandidatesExhausted failure reason.

Behavior change worth a look: with buckets, an engine that the affinity group rejected is evaluated again by its own size bucket, so sgl_router_cache_admission_rejected_total can count one request twice; the pending-prefill proxy test now asserts at least one rejection.

Checklist

• cargo clippy --locked --workspace --all-targets -- -D warnings clean
• cargo test --locked --workspace -- --skip parity_matrix passes (965 tests)

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): 🚫 Run #35411243786
Latest PR Test (Extra): 🚫 Run #35411243644
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[sherlockwu]

Closing: the series stops at the power-of-two PR (#40271) for now.