Decouple inference request state from the HTTP administration shell

[tiammomo]

Summary

The application state is owned by the HTTP routes module and exposes all 17 runtime and administration dependencies to inference handlers and provider adapters, creating a reverse dependency from provider code into transport composition.

Actual behavior

src/routes.rs defines AppState; src/server.rs imports it from routes; OpenAI-compatible, Anthropic, and token-counting provider modules accept routes::AppState. The inference path therefore compiles against administration-only dependencies and the application composition root is owned by its HTTP adapter.

Expected behavior

An application-owned top-level state composes the process dependencies, while Axum FromRef exposes a typed inference substate containing only dependencies required by client inference and provider execution. Provider adapters no longer import the routes module.

Steps to reproduce

Inspect src/routes.rs:207, src/server.rs:25, and imports of routes::AppState under src/providers on the verified #45 branch.

Environment

ModelPort main at 14ea1f5; implementation depends on the domain-router composition in #44/#45; Axum 0.8.9.

Acceptance criteria

• Move application-state ownership out of src/routes.rs into an application-layer module; model an inference substate and derive or implement Axum FromRef from the top-level state; migrate client inference and provider adapter entry points to the inference substate; ensure provider modules do not import routes; keep route composition, public APIs, authentication, routing, quotas, rate limits, streaming, metrics, persistence, configuration, migrations, and dependencies behavior-compatible; add compile-time or architectural regression coverage for the dependency boundary; document the state ownership rule and the remaining administration-state follow-up; pass scripts/check-all.sh.