Record no-spec ubatch benchmark

docs/LLM/dealign-ornith-crack.md

@@ -75,8 +75,12 @@ On Dankserver the available buffer types are `CPU` and `CUDA0`.
 
 MoE model with a bundled MTP head -> native `draft-mtp`. The head is
 auto-detected from the target (b10121 creates the MTP draft context against
-the model file). ngram under MTP hurts on this family; MTP alone is the
-proven best. `--spec-draft-n-max 4` is the knee.
+the model file). ngram under MTP hurts on this family; MTP alone was
+previously selected for the production configuration. A later 8K API
+benchmark found that this MTP path produced lower generation throughput than
+`--spec-type none` for that workload, so speculative decoding should be
+re-evaluated before treating it as optimal. `--spec-draft-n-max 4` remains the
+configured value.
 
 ```
 --spec-type draft-mtp