Fix: Publish session.error event for invalid model selection

Problem

When sending a message via the API with an invalid provider or model ID, the error was silently swallowed. API consumers had no way to know something went wrong.

What happened:

1. POST to /session/{id}/message with { model: { providerID: "invalid", ... } }
2. HTTP response returns 200 immediately (before validation)
3. User message is created successfully
4. Session goes busy → idle with no error indication
5. No session.error event is published to /event

Root cause: The Provider.getModel() call happens inside loop() after the HTTP response has already started streaming. When it throws ModelNotFoundError, the error propagates up but is never converted to a session.error event.

Solution

Added error handling around Provider.getModel() in loop() to publish a session.error event before re-throwing. This matches the existing pattern used in the command() function.

After this fix:

data: {"type":"session.error","properties":{"sessionID":"ses_...","error":{"name":"UnknownError","data":{"message":"Model not found: invalid/some-model. Did you mean: anthropic, openai, ...?"}}}}

Changes

• packages/opencode/src/session/prompt.ts: Wrap Provider.getModel() in try-catch, publish session.error on ModelNotFoundError

Fixes #8460