Commit Beelink mini defaults: CPU inference, 32k context

Stop forcing Vulkan NGL in compose override (that overrode .env).
Raise mini profile context to 32768 for ~28–32GB RAM hosts.

docker-compose.override.yml

@@ -1,22 +1,11 @@
-# GPU access for Vulkan (bonsai) + VA-API (jellyfin).
+# Default override for mini / Beelink-class hosts.
+# - bonsai: CPU only (do NOT force NGL/devices here — that overrides .env)
+# - jellyfin: /dev/dri for VA-API when available
 #
-# Do NOT use group names like "video"/"render" here — Docker resolves them
+# mid/gaming: re-add bonsai devices + set BONSAI_NGL in .env via apply-profile.sh
-# inside the container image, which often has no "render" group and fails with:
+# Do NOT use group names (video/render) — Docker resolves them inside the image.
-#   unable to find group render: no matching entries in group file
-#
-# Options:
-#   A) Rely on host /dev/dri being 0666 (Ansible users role udev rule) — no group_add.
-#   B) Use numeric host GIDs:  getent group video render
-#
-# Ansible regenerates this file with host GIDs when you run the stack role.
 
 services:
-  bonsai:
-    devices:
-      - /dev/dri:/dev/dri
-    environment:
-      NGL: "99"
-
   jellyfin:
     devices:
       - /dev/dri:/dev/dri

profiles/mini.env

@@ -5,7 +5,8 @@ STACK_PROFILE=mini
 # Ternary-Bonsai 27B fits in system RAM; keep CPU
 BONSAI_BACKEND=cpu
 BONSAI_NGL=0
-BONSAI_CTX=16384
+# 32k uses free system RAM well on ~32GB mini PCs (weights ~7GB)
+BONSAI_CTX=32768
 BONSAI_KV4=0
 BONSAI_THREADS=12
 BONSAI_TEMP=0.7