chore: frank-default-small — ngram-mod, drop DFlash draft

- spec-type → ngram-mod (no separate draft model)
- drop --model-draft, --spec-draft-n-max; add --spec-ngram-mod-n-min/max/match
- DFlash batch default fallback SPEC_TYPE → ngram-mod
- env: SPEC_TYPE=ngram-mod, NGRAM_MOD_N_MIN=48, N_MAX=64, N_MATCH=24
- DRAFT_FILE env var emptied (not used under ngram-mod)

scripts/frank-default-small.env

@@ -1,4 +1,4 @@
-# Tunables for frank-default-small (HauhauCS Q6_K_P + DFlash draft, dankserver)
+# Tunables for frank-default-small (HauhauCS Q6_K_P, ngram-mod spec, dankserver)
 # Runtime: llama-server b10108; RTX 3090 24GB; Ryzen 7 7700X; 62GB RAM.
 MODEL_DIR=/home/dank/models
 ALIAS=frank-default-small
@@ -8,8 +8,13 @@ PORT=7770
 # (Reduced from 262144/stream per operator: allows the larger batch to fit VRAM.)
 CTX_SIZE=600000
 PARALLEL=3
-SPEC_TYPE=draft-dflash
-SPEC_N_MAX=7
+SPEC_TYPE=ngram-mod
+# Ngram-mod speculative decoding: look back this many tokens to guess the
+# next tokens. Defaults (48/64/24) are a reasonable starting point; tune if
+# you want more or fewer draft lookahead tokens.
+NGRAM_MOD_N_MIN=48
+NGRAM_MOD_N_MAX=64
+NGRAM_MOD_N_MATCH=24
 CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=f16
 # All 40 layers keep their MoE expert weights on CPU (targeted expert offload).