### Check for existing issues

Related but distinct: #11864, #17093, #9298 all report `NotConnectedError` in different environments (containers/K8s/restricted images). This report is for a bare uv/venv install on Debian, where the same command succeeds as root in the foreground and fails under systemd.

### What happened?

LiteLLM proxy fails to start under systemd with `prisma.engine.errors.NotConnectedError` on every version from **1.87.0 through 1.94.1** inclusive. Version **1.86.0** starts cleanly with the exact same environment, config, database, and systemd unit. The regression window is bounded.

The failure is not "Prisma binaries missing" (#13793, #22760) — the engine binary is present, linked correctly, executable, and runs standalone. The failure is later, inside `health_check() → query_raw() → _build_request()`, where the query engine subprocess should already have an HTTP URL and doesn't.

### Environment

- Linux distribution: Debian 13.5 (bookworm) in Proxmox LXC - Proxmox VE: 9.1 - Python: 3.13.5 - Install method: `uv add 'litellm[proxy]==<version>'` in a venv at `/opt/litellm/.venv` - Runner: systemd unit invoking `/opt/litellm/.venv/bin/litellm --config ...` (orig…