fix: retry model deletion on Windows when files are locked

[Bortlesboat]

Summary

• Await the unload response instead of fire-and-forget, so the Python process actually releases file handles before deletion begins
• Add a 1.5s delay after successful unload to give Windows time to release lingering file locks
• Retry rm up to 3 times with incremental backoff (1s, 2s, 3s) when the error is EBUSY or EPERM
• Return a descriptive error message if all retries are exhausted

Details

On Windows, when a model is loaded into memory by the Python backend, the OS holds file locks on the model directory. The existing model:delete handler fires the unload request without awaiting it, then immediately attempts rm — which fails with EBUSY or EPERM because handles haven't been released yet.

This fix makes the unload blocking (with a 10s timeout) and adds retry logic so transient lock failures are handled gracefully.

Only the model:delete IPC handler in electron/main/ipc-handlers.ts is modified.

Fixes #12

Test plan

• Load a model, then delete it — should succeed on first try (unload completes, delay lets locks clear)
• Attempt to delete a model whose files are open in another program — should retry 3 times then return a descriptive error
• Delete a model that is not currently loaded — should succeed immediately (unload fails gracefully, rm succeeds)
• Verify on macOS/Linux that behavior is unchanged (no EBUSY/EPERM on those platforms)

🤖 Generated with Claude Code