perf(llm): cascade routing — avoid Vec clone per request, add cost_tiers config #1724
Description
Context
Follow-up from cascade routing implementation (PR #1721, closes #1339).
Issues
PERF-CASCADE-01: RouterProvider::clone() on every chat()/chat_stream() call causes 2 Vec heap allocations before any LLM work. providers: Vec<AnyProvider> is cloned in both ordered_providers() and self.clone(). Fix: use Arc<[AnyProvider]> (immutable slice, clone = ref-count bump).
cost_tiers deferred: Providers are ordered by chain order, not explicit cost ranking. Add cost_tiers config field to CascadeConfig for explicit cheapest-first ordering independent of chain order.
Priority
Low — overhead is dominated by LLM I/O by orders of magnitude.