Add bedrock (bedrock/us.anthropic.claude-haiku-4-5) to the negative-spend
sweep, the surface BerriAI/litellm#25846's negative streaming-cost bug
actually lived on, and register it in the suite's driver_models with AWS
creds rather than a single api_key. Rework the sweep constants into a small
frozen _SweepModel(call, row_marker) so a friendly bedrock alias whose spend
row resolves to bedrock/us.anthropic... still matches; the two claude markers
are prefix-qualified so a bedrock-anthropic row can't stand in for
anthropic-direct or vice versa.
Add a live spend-tracking e2e that sweeps every chat provider the proxy
exposes (non-streaming and streaming) plus an embedding on one key, then
asserts every logged spend row is non-negative. A negative cost is a real
billing regression; cache-token accounting that lets a derived token count
fall below zero (BerriAI/litellm#25846) surfaces here as spend < 0, and the
streaming leg is exercised because that is the path the regression rode in on.
The per-provider positive-row check keeps the guard non-vacuous, so a pipeline
that silently logs 0 or drops the row fails instead of sliding past a bare
non-negative check.
Register the covering cell quota_management.spend_tracking.non_negative.
never_negative and note the test in the coverage matrix.
