fix(providers): discard stale model discovery results

[Quine-rq]

The model picker combines the current provider's saved models with discovery results. Switching provider, endpoint, credentials or profile keeps the previous request's models available while the replacement request loads. A response arriving after the picker closes can also restore results into an idle discovery state.

Clear models, free-model metadata and the cache flag for each new request, and cancel response handling in the effect cleanup. Keep the existing debounce and allow discovery to restart normally when reopened.

Eight hook regressions cover request identity changes, disabled discovery, late responses, closing/reopening the same provider and unmounting. The initial provider-switch and disabled-response cases both fail before the fix.

Validation: 2,316 tests across 222 files, type checks, uncached ESLint, production build and lat checks passed. Tests exercise the real hook with controlled discovery IPC; packaged desktop/live-provider validation remains pending.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no actionable correctness, security, or quality issues were identified.

Summary

This PR prevents stale model-discovery responses from updating renderer state after the request identity changes, discovery is disabled, or the hook unmounts.

• Clears models, free-model metadata, and cache state whenever a new discovery lifecycle begins.
• Uses effect-scoped cancellation to discard superseded or late asynchronous responses.
• Adds regression coverage for provider, endpoint, credential, profile, disablement, reopen, and unmount scenarios.
• Documents the discovery request lifecycle in the provider setup guide.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Discovery inputs change] --> B[Clean up previous effect]
  B --> C[Mark previous request cancelled]
  C --> D[Clear prior discovery state]
  D --> E{Enabled and provider present?}
  E -- No --> F[Set status to idle]
  E -- Yes --> G[Set status to loading]
  G --> H[Wait for debounce]
  H --> I[Request provider models]
  I --> J{Effect cancelled?}
  J -- Yes --> K[Discard response]
  J -- No --> L[Publish models, metadata, and status]

Loading

_{Reviews (1) · Last reviewed commit: "fix(providers): discard stale model disc..."}