Desktop: a failed provider deployment leaves an agent whose "Run on" settings can never be corrected

**Describe the bug**

When creating an agent on a remote provider and the deployment fails, the agent is still created and kept, but its "Run on" settings can never be corrected. The edit dialog shows them read-only ("Where an agent runs can't be changed afterwards — create a new agent to run somewhere else"), which is a reasonable rule for an agent that deployed successfully and holds a key on that deployment. For one that never deployed, there is nothing to protect, and the only way out is to delete the agent and create it again.

You can press "Deploy" to retry, but only with the exact settings that just failed. If the failure was caused by those settings, the retry fails too.

**Steps to reproduce**
1. Create an agent, pick a remote provider under "Run on", and enter a configuration that will fail (wrong region, missing credential, unreachable endpoint).
2. Submit. The deployment fails and the agent appears in the list as not deployed.
3. Open the agent and choose "Edit agent". The "Run on" section is read-only.
4. There is no way to fix the setting that caused the failure; "Deploy" only retries the same configuration.

**Expected behavior**

An agent that has never deployed should allow its "Run on" settings to be edited, so the failed deployment can be corrected and retried. Locking them down makes sense once a deployment exists.

**Version and platform**
- Buzz version: 0.5.23
- OS: macOS

**Logs / additional context**

Searched for existing reports on failed deployments and on editing provider settings after creation; none found.

Start from the Edit agent entry point and the Deploy action described in the reproduction steps, then trace how deployment state controls whether the "Run on" section is read-only. Reproduce a failed remote-provider deployment and verify that an agent with no successful deployment can edit those settings and retry, while successfully deployed agents remain locked.

Written by the indexing model from the issue text.