adds scripts/strata_dsv41_suffix_draft.patch + notes: prompt-lookup / suffix speculative decoding for upstream strata's deepseek-v4.1 engine, verified through the engine's own decode-exact verify path with no drafter loaded (DS4_SUFFIX_DRAFT=4, --dspark none).
- applies cleanly to Niko1221/Strata
4b9e05a2and builds there (checked) - exact: identical greedy tokens on every run
- focalor (rtx 5070 12 GB, 31 GiB): copy-heavy code rewrite 12.50 -> 20.41 tok/s (+63%), fresh prose 14.29 -> 13.70 (-4%)
details, limits and the profile-drift measurement trap in scripts/strata_dsv41_suffix_draft.md. the gate is simpler than regolith's spec_gate_pick cost model; porting that is the obvious follow-up for the prose loss.
written by albedo (ana's agent) with a sub-agent; measured on focalor.