Expose workload-aware model selection + normalized cache metrics for gateway routing

[stackbilt-admin]

Summary

llm-gateway now does route-class-aware provider/model selection and adds durable response/prefix caching with TTL. We need llm-providers to expose stronger provider-native model/caching capabilities so gateway policy can stay thin and avoid hardcoded provider quirks.

Why this matters

Current gateway integration has to hardcode Groq/Cerebras behavior to get good results (model picks by route class, tool stripping on cheap routes). That logic belongs primarily in llm-providers so all clients benefit and policy remains centralized.

Requested enhancements in llm-providers

1. Route/workload-aware model defaults

• Add first-class API for model selection by workload class (e.g. summary, planning, code_draft, long_context, tool_loop) instead of a single generic default.
• Keep provider-specific best models configurable without per-consumer hardcoding.

2. Provider-native prompt caching normalization

• Ensure supported providers expose prompt-cache controls and consistently report cache usage.
• Normalize usage fields across providers to include (when available):
  • cachedInputTokens
  • cacheReadInputTokens
  • cacheWriteInputTokens

3. Model capability metadata surface

• Expose capability metadata per model (tools, streaming, context window, reasoning profile, structured-output reliability) so router layers can make policy decisions without static registries.

4. Stable override contract

• Provide a documented override path (env/config/hook) for consumers to set provider+workload model preferences without forking selection logic.

Gateway-side context (already implemented)

• Per-route provider model overrides for Groq/Cerebras.
• Response cache with TTL and max-entries eviction for safe routes.
• Cache-hit/usage telemetry integration.

Acceptance criteria

• A gateway consumer can ask for a provider default model using workload intent and get stable, provider-tuned outputs.
• Cache usage fields are normalized enough that gateway telemetry does not need provider-specific parsing.
• Consumer no longer needs hardcoded per-provider model maps for basic routing quality.