### Check for existing issues

- [x] I searched open and closed issues before filing - [x] I found no existing issue for last-write-wins spend resets in `RouterBudgetLimiting`

### What happened?

When a provider, deployment, or tag budget window is considered expired, `_handle_new_budget_window` resets the spend key with a plain `SET` to the current response's cost

Concurrent responses crossing the same window boundary can all enter the reset path. Each writes only its own cost, so the final counter contains whichever write wins instead of the sum of all response costs

The start-time and spend keys are also written separately, so another worker can observe a mixed window state. In multi-replica deployments this directly undermines the shared budget counter at the point where traffic crosses a budget boundary

A separate reliability concern in the same accounting path is that queued Redis increments are dispatched with `asyncio.create_task` and the queue is cleared immediately, so the sync method does not wait for the remote write to complete

The reproduction uses synthetic cache keys, timestamps, and costs and makes no provider or network call

### Steps to Reproduce

1. Check …