### What happened?

A model group configured as LiteLLM's native adaptive router returns **HTTP 500 on every request, permanently**, with:

``` gammavariate: alpha and beta must be > 0.0 ```

surfacing from `litellm.router.py::async_function_with_fallbacks()`. Once it starts it never recovers, including across proxy restarts. Deleting one row from `LiteLLM_AdaptiveRouterState` and restarting fixes it until the condition recurs.

The state row responsible:

``` router_name | request_type | model_name | alpha | beta | total_samples sovereign-router | general | gpt-oss-cloud | 1 | 0 | 1 ```

### Root cause

`litellm/router_strategy/adaptive_router/update_queue.py` — the Prisma upsert's `update` branch correctly uses `{"increment": ...}`, but the **`create` branch writes the raw delta as the absolute value**:

```python "alpha": payload["delta_alpha"], "beta": payload["delta_beta"], ```

`adaptive_router.py::_compute_bandit_delta` can legitimately produce `d_alpha = 0.0` or `d_beta = 0.0` (e.g. `d_beta = misalignment + stagnation + disengagement + failure + 0.5*loop`, all zero on a clean satisfied turn). So the **first** flush for a new `(router_name, r…