Expose external provider models through GET /v1/models

[urucoder]

Summary

Controllers with configured external OpenAI-compatible providers only surfaced those providers' models via GET /studio/provider-models. Since models are discovered through GET /v1/models, external provider models were invisible in the model picker even when the provider was configured and reachable.

Changes

• Shared discovery helper: extracted discoverProviderModels and enabledProvidersWithApiKey into controller/src/services/provider-routing.ts, replacing the inline providerModels function previously duplicated in provider-routes.ts.
• /studio/provider-models: now consumes the shared helper; behavior unchanged.
• GET /v1/models: now also queries enabled, API-key-configured providers and appends their models as qualified IDs (<provider-id>/<upstream-model-id>), alongside existing recipe/active-inference entries. Per-provider discovery failures are swallowed (Effect.option) so one unreachable provider never fails the whole endpoint, matching /studio/provider-models semantics.
• No changes needed in the proxy (openai-routes.ts) — it already parses the qualified ID and rewrites the upstream request model correctly.

Example: with provider my-provider (base_url http://<host>:8000) enabled and reporting example-model-1, GET /v1/models now includes:

{
  "id": "my-provider/example-model-1",
  "object": "model",
  "owned_by": "my-provider",
  "active": false,
  "metadata": { "vision": false }
}

Validation

cd controller && bun run typecheck
cd controller && bun run lint
cd controller && bun run check
node scripts/project.mjs validate-contracts
node scripts/project.mjs validate-structure
node scripts/project.mjs audit-layout

Manually verified end-to-end against a live controller and a mock upstream provider server: GET /v1/models returned the qualified model ID, and a chat completion request using that ID was routed upstream with the model field rewritten back to the original provider model ID.

UI changes

None — controller-only change. Effect is a discoverability fix for the existing model picker.

Risks / rollout notes

• Purely additive to /v1/models response; existing consumers unaffected.
• Provider discovery adds one upstream HTTP call per enabled/configured provider per /v1/models request (10s timeout, best-effort, run concurrently) — negligible latency impact for typical provider counts.

Closes #410

[urucoder]

Closes #410