The generation split (thinking / answer / tool args) already tiles its
window off the stream's block-start markers; the same markers now count
into additive-optional per-bucket tallies (reasoningBlocks / textBlocks /
toolArgBlocks), so the card qualifies each decode slice with the number
of blocks actually decoded - the honest per-slice counterpart of TTFT's
call count and tools' run count (one model call emits zero to many
blocks, so a call count there would be a false tally). Overhead stays
unqualified: it is the residue, with nothing behind it to count.
Both log generations carry the markers: the V0 chunk flood counts them
into the step slot alongside the spans, and the V2+ embedded stream
yields them from the same single pass that prices the split
(decodeSpansOfStream -> decodeTallyOfStream). Counts ride the split's
own first-token gate, and rows cached before the fields existed render
the duration alone.
bowenliang123
f723cddb30