Upgrade Beta 1 to GSQ-RCO IQ3_S

compose.yaml

@@ -235,9 +235,10 @@ services:
       - --spec-draft-p-min
       - "0.05"
 
-  # Beta 1 keeps the complete GSQ-RCO text model and KV cache on the RTX 5080.
-  # Only the multimodal projector runs on the RTX 3060. The measured hard
-  # boundary is 196608 tokens; 192K deliberately retains runtime headroom.
+  # Beta 1 keeps the complete GSQ-RCO IQ3_S text model and KV cache on the RTX
+  # 5080. Only the multimodal projector runs on the RTX 3060. The larger IQ3_S
+  # build starts conservatively at 112K until its hard context boundary has
+  # been measured; the former IQ3_XXS result does not transfer to this file.
   llama-beta1:
     <<: *llama-common
     container_name: mike-ai-llama-beta1
@@ -258,7 +259,7 @@ services:
       - --alias
       - qwen-beta-1
       - --ctx-size
-      - "${BETA1_CONTEXT:-192000}"
+      - "${BETA1_CONTEXT:-112000}"
       - --flash-attn
       - "on"
       - --cache-type-k