Token usage no longer flashes zero

After a chat round finished, the token chip on a thread could briefly show ↑0 ↓0 · $0 before snapping to the real numbers. That was not a display delay — the usage record was genuinely written twice: once empty, and again a moment later with the round's counts.

The empty write is gone. A round's token counts are now saved in a single step, so the chip shows the true figures as soon as the record exists. This also closes a rarer and worse case: if that second write was lost — under heavy load it could be abandoned after 15 seconds, silently — the zeros became the permanent record of a round that really had consumed tokens. There is no longer a second write to lose on a thread's first round with a model, and when one is abandoned it now says so in the log instead of failing quietly.