### What happened?

On a proxy configured with RDS IAM auth (proactive token refresh), the writer Prisma query-engine is periodically recreated via `recreate_prisma_client` (kill-then-construct). This is a *planned* recreate and is coordinated correctly since the engine-death-watcher race fix (#29176).

However, during the brief (~0.5–1s) recreate window, a concurrent `/health/readiness` probe calls `PrismaClient.health_check()` → `query_raw("SELECT 1")`, which fails with:

``` httpx.ConnectError: All connection attempts failed ```

This transient, expected failure is reported to `proxy_logging_obj.failure_handler(...)`, which drives the `db_exceptions` alerting. Because engine recreates happen every IAM-refresh cycle (~every 10–15 min per process), this produces recurring **false-positive DB-exception alerts** even though the reconnect is benign — no pod restart, no request-failure impact.

There are two distinct problems:

**1. Mislabeled error string.** In `litellm/proxy/utils.py`, the `health_check()` exception handler builds:

```python error_msg = f"LiteLLM Prisma Client Exception disconnect(): {str(e)}" ```

but this is inside `health_check()` (the `failure_handler` call eve…