Cancelled and failed runs now count their tokens

When you stopped a run mid-flight, or a run ended in an error, the tokens the model had already consumed could vanish from usage and cost reporting — and whatever was recorded was attributed to the model you asked for, not the model that actually ran (a delegated sub-run could even land under "(unknown)").

Cancelled and errored rounds now record their consumed tokens the same way completed rounds do — keyed to the model that actually served the round, which is the model the provider reports when it routes your request elsewhere. The message shows that same model. An errored round no longer loses its usage record when writing the error message itself fails.

Where a provider reports usage only at the very end of a stream and the round stops before that report arrives, there is still no count to record — nothing is estimated in its place.