Hardcoded reason: "overloaded" is semantically inaccurate

LiveSessionModelSwitchError signals a session-stored model conflict, not an infrastructure overload. Using "overloaded" here will appear in the attempts array and the final FailoverError summary logged when all candidates are exhausted, making it harder to distinguish a genuine overload from a session state mismatch.

Looking at failover-policy.ts, valid FailoverReason values include "unknown", which is a much more accurate stand-in since the session mismatch may or may not have been caused by overload on the original model:

This is purely an observability concern — the core fix (preventing re-throw and infinite retry) is unaffected by the reason value.

This is a comment left during a code review.
Path: src/agents/model-fallback.ts
Line: 793-797

Comment:
**Hardcoded `reason: "overloaded"` is semantically inaccurate**

`LiveSessionModelSwitchError` signals a session-stored model conflict, not an infrastructure overload. Using `"overloaded"` here will appear in the `attempts` array and the final `FailoverError` summary logged when all candidates are exhausted, making it harder to distinguish a genuine overload from a session state mismatch.

Looking at `failover-policy.ts`, valid `FailoverReason` values include `"unknown"`, which is a much more accurate stand-in since the session mismatch may or may not have been caused by overload on the original model:

```suggestion
        const switchNormalized = new FailoverError(switchMsg, {
          reason: "unknown",
          provider: candidate.provider,
          model: candidate.model,
        });
```

This is purely an observability concern — the core fix (preventing re-throw and infinite retry) is unaffected by the `reason` value.

How can I resolve this? If you propose a fix, please make it concise.

Preserve live-switch error on sole fallback candidate

This branch converts every LiveSessionModelSwitchError into a generic FailoverError and continues, including when the current candidate is the last/only candidate. That removes the signal that outer retry loops depend on (runCronIsolatedAgentTurn and runAgentTurnWithFallback only retry on LiveSessionModelSwitchError), so runs with no remaining fallbacks (for example fallbacksOverride: [] or an empty fallback config) can no longer apply the requested live model/auth-profile switch and now fail immediately instead of retrying with the switched selection. The conversion should be limited to cases where another candidate remains, while preserving the original error on the final attempt.

Useful? React with 👍 / 👎.

Re-throw live-switch error on final candidate

Handling LiveSessionModelSwitchError unconditionally as a FailoverError here means the fallback loop never re-throws the original switch signal on the last/only candidate. In single-candidate runs (for example fallbacksOverride: [] or configs without fallbacks), this now surfaces as a generic failover and bypasses outer retry handlers that only react to LiveSessionModelSwitchError (src/cron/isolated-agent/run.ts:604 and src/auto-reply/reply/agent-runner-execution.ts:624), so the requested provider/model/auth-profile switch is not applied and the run fails immediately. Fresh evidence from this commit is that cron now has its own retry cap (src/cron/isolated-agent/run.ts:597-611), so preserving the original switch error on the final candidate no longer implies an unbounded loop.

Useful? React with 👍 / 👎.