Bring merged #17 and #18 into main

.env.example

@@ -105,27 +105,54 @@ BONSAI_API_KEY=bonsai-change-me
 #QCM_QWEN9_CTX=8192
 #QCM_QWEN4_CTX=8192
 #QCM_GRANITE_CTX=8192
+#QCM_QWEN36_CTX=8192
+#QCM_GEMMA4_CTX=8192
 #QCM_AGENT_UBATCH=256
 #QCM_AGENT_CACHE_RAM=1024
+# Existing Qwen3.5/Granite agent presets keep their prior 256-token reuse
+# granularity. New Qwen3.6/Gemma start with Prism's default 0; compare both
+# with the dedicated prompt-prefix probe before changing production settings.
+#QCM_AGENT_CACHE_REUSE=256
 # GPU layer limits can be reduced for partial offload when a model does not fit.
 #BONSAI_NGL=99
 #QCM_QWEN9_NGL=99
 #QCM_QWEN4_NGL=99
 #QCM_GRANITE_NGL=99
+# For Qwen3.6, leaving NGL unset lets Prism --fit choose a hybrid GPU/CPU
+# tensor split. Setting it overrides that fit choice: sweep measured values on
+# the RTX 3070 Ti rather than copying the dense-model 99-layer setting.
+# With explicit NGL, FIT_TARGET cannot guarantee the reserved VRAM margin.
+#QCM_QWEN36_NGL=12  # example sweep value only; not a tested default
+#QCM_QWEN36_FIT_TARGET=1536
+#QCM_GEMMA4_NGL=99
 # Flash attention: on/off/auto. Bonsai retains its prior on setting.
 #BONSAI_FA=on
 #QCM_QWEN_FA=auto
 #QCM_GRANITE_FA=auto
+#QCM_QWEN36_FA=auto
+#QCM_GEMMA4_FA=auto
 # Begin with f16; q8_0 is the first conservative quantised-cache comparison.
 # Quantised V requires flash attention. Verify tools and long turns after changes.
 #QCM_QWEN_CTK=f16
 #QCM_QWEN_CTV=f16
 #QCM_GRANITE_CTK=f16
 #QCM_GRANITE_CTV=f16
+#QCM_QWEN36_CTK=f16
+#QCM_QWEN36_CTV=f16
+#QCM_GEMMA4_CTK=f16
+#QCM_GEMMA4_CTV=f16
 # Qwen agent default: no thinking. Compare on and a bounded budget separately.
 # Use the runtime reasoning option; enable_thinking CLI kwargs are deprecated.
 #QCM_QWEN_REASONING=off
 #QCM_QWEN_REASONING_BUDGET=-1
+# Qwen3.6 and Gemma are separate text/tool-only presets; no mmproj is loaded.
+# Select one pinned quant per alias and download that exact file with
+# ./fetch-models.sh --model <id> [--quant <quant>]. Do not switch quant or
+# model within each browser action; measure a whole workflow.
+#QCM_QWEN36_QUANT=Q4_K_XL
+#QCM_GEMMA4_QUANT=QAT_Q4_0
+#QCM_QWEN36_REASONING=off
+#QCM_QWEN36_REASONING_BUDGET=-1
 
 # --- ./update.sh (Linux/Docker only) -----------------------------------------
 # These matter because update.sh rolls the stack back when it decides a model