The first non-search merge on an ingest was made both ways (partials scaled, and
the scale from the fulls alone) and PartialScalingWins picked one. When the prior
is to scale the partials and that merge resolves its error model, the decision is
the prior whatever the other arm says, so the fulls-alone merge was a whole final
merge thrown away - 16% of all scaling time over the battery (238 of 258
decisions), e.g. ~1.5 s of a 22 s GPU run on an in-house cytochrome C set, 8-13 s on
the largest open-arm sets.
That arm is now made only when the partials' merge fails to resolve its error
model (then the old sequence runs unchanged) or when the prior is the fulls
(small molecules; unchanged). The partials' arm starts from g_partial = 1, the
only state the fulls-alone arm left for the arm after it (corr is restored,
g_full reset by the combine, device g/scaled cleared per call, the sweep /
rejection / damage diagnostics reset where they are measured), so the merge is
bit-identical to the one the pair returned.
The log's decision line no longer quotes the fulls-alone ISa when that merge is
not made.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5K8jvPPbmCrbqnWkddTuB
leonarski_fandClaude Opus 5.5
fc358720ed