Interface DflashDraftOpts

Return confidence telemetry; pruning still computes it when needed.

Materialize+concat the per-position draft logits into the returned DflashDraftRows.draftLogits. Default true. The grouped draft rows run their own verify lm-head and never read it — pass false there to skip the dead concat + a host-unrelated GPU alloc every round.

Confidence-scheduled draft-length pruning (Alg 1, single-user form): position k is DROPPED (and the block truncated there) when its predicted acceptance c_k < thresholds[k] (per-position, STS-calibrated) or < minConf (uniform manual override). Position 0 is always kept (a source must return ≥1 token). Both absent ⟹ fixed-γ drafting (pre-scheduler behavior). Losslessness is invariant — pruning only changes how many positions the target VERIFIES, never what is emitted.