fix(server): serve last settled models from /api/model while plugins reload - #57011

Issue for this PR

N/A

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Serves the last settled model catalog and default model from GET /api/model and GET /api/model/default while plugins/config are reloading instead of blocking on awaitPlugins and failing with 503 ServiceUnavailableError after 5 s.

When global config (opencode.json) is edited on a server with many loaded Locations (e.g. 133 booted Locations in our Studio server trace), PluginSupervisor holds Plugin.awaitActivation across all loaded Locations while debouncing and re-running Plugin.activate inside State.batch. During that reload window:

1. Plugin.awaitActivation can stay pending for >5 s across the loaded Location set (5.1–11.6 s in opencode.log), causing awaitPlugins in ModelHandler to fail GET /api/model with HTTP 503 (Location plugins did not finish activating within 5 seconds).
2. Inside Plugin.activate's State.batch, closing a changed plugin's scope removes its provider.transform / model.transform registrations and invalidates Provider.Service / Model.Service before the replacement finishes loading, so rebuilding on the HTTP fiber mid-batch would observe an incomplete catalog.

Changes in packages/server/src/handlers/model.ts:

• Cache the last settled { available, default } snapshot per Model.Interface in a WeakMap whenever Plugin.awaitActivation is open (Duration.zero check).
• When Plugin.awaitActivation is pending on a Location that has already settled a snapshot, immediately return that last settled snapshot (model.updated notifies clients once the reload batch finishes).
• On a cold Location that has not yet settled its first snapshot, continue waiting on awaitPlugins, and if the 5 s wait times out, fall back to the currently available models rather than returning 503.

How did you verify your code works?

cd packages/server && bun test test/model.test.ts && bun run typecheck

• Preserves the existing cold-start test (waits for plugin initialization before listing models).
• Adds serves the last settled model list and default model while plugins reload, which holds a plugin reload mid-batch after v1 scope teardown, verifies GET /api/model and GET /api/model/default return 200 with chat-v1 immediately, then releases the reload gate and verifies both endpoints return chat-v2.

Screenshots / recordings

N/A

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR