setup: start genie-ai-runtime before memory-heavy services

crates/genie-core/tests/tool_dispatch_test.rs

@@ -264,6 +264,24 @@ fn start_all_uses_configured_llm_backend() {
         contents.contains("is_warmup_unit") && contents.contains("start --no-block"),
         "start_all should queue warmup units without blocking the lifecycle script"
     );
+    let units = contents
+        .split("UNITS=(")
+        .nth(1)
+        .and_then(|s| s.split(")").next())
+        .expect("start_all should declare ordered units");
+    let llm_pos = units
+        .find("\"$configured_llm_unit\"")
+        .expect("start_all should include the configured LLM unit");
+    let homeassistant_pos = units
+        .find("homeassistant.service")
+        .expect("start_all should include Home Assistant");
+    let whisper_pos = units
+        .find("genie-whisper.service")
+        .expect("start_all should include Whisper");
+    assert!(
+        llm_pos < homeassistant_pos && llm_pos < whisper_pos,
+        "start_all should start the configured LLM before memory-heavy services"
+    );
 }
 
 /// Verify genie-ai-runtime service preserves warm GGUF pages across restarts.
@@ -285,8 +303,14 @@ fn genie_ai_runtime_service_preserves_model_page_cache() {
         "genie-ai-runtime.service should use INT8 KV to fit enough context under memory pressure"
     );
     assert!(
-        contents.contains("GENIEPOD_AI_RUNTIME_CONTEXT=2048"),
-        "genie-ai-runtime.service should request the GenieClaw web-chat context size"
+        contents.contains("GENIEPOD_AI_RUNTIME_CONTEXT=8192"),
+        "genie-ai-runtime.service should request the Jetson-tested 8k context size"
+    );
+    assert!(
+        contents.contains(
+            "Before=genie-whisper.service genie-whisper-warmup.service homeassistant.service genie-core.service"
+        ),
+        "genie-ai-runtime.service should reserve KV cache before memory-heavy services"
     );
 }
 

deploy/scripts/start_all.sh

@@ -118,12 +118,12 @@ configured_llm_unit="$(normalize_unit "$raw_llm_unit")"
 configured_warmup_unit="$(warmup_unit_for "$configured_llm_unit")"
 
 UNITS=(
-    homeassistant.service
     genie-audio.service
-    genie-whisper.service
-    genie-whisper-warmup.service
     "$configured_llm_unit"
     "$configured_warmup_unit"
+    homeassistant.service
+    genie-whisper.service
+    genie-whisper-warmup.service
     genie-core.service
     genie-governor.service
     genie-health.service

deploy/systemd/genie-ai-runtime.service

@@ -2,6 +2,11 @@
 Description=GeniePod AI Runtime (Jetson-tuned LLM, OpenAI-compatible)
 Documentation=https://github.com/GeniePod/genie-ai-runtime
 After=network.target
+# Claim the LLM KV cache before memory-heavy voice/container services start.
+# Jetson testing for issue #75 showed the same `-c 4096` request fitting only
+# ~1.7k ctx after the full stack was resident, but fitting 4k/6k/8k ctx when
+# genie-ai-runtime loaded first.
+Before=genie-whisper.service genie-whisper-warmup.service homeassistant.service genie-core.service
 ConditionPathExists=/opt/geniepod/bin/jetson-llm-server
 # Conflicts with genie-llm.service: both bind :8080. systemd will refuse
 # to start the second one while the first is running, so a misconfigured
@@ -12,17 +17,16 @@ Conflicts=genie-llm.service
 Type=simple
 # Keep the GGUF in page cache across restarts when the kernel can. Clearing
 # VM caches here made every runtime restart cold-load Qwen3 again (issue #69).
-# Use INT8 KV so the Jetson service reliably gets enough context for
-# GenieClaw's web prompt even under memory pressure. `-c` is still clamped
-# by runtime memory budget, but INT8 KV roughly doubles the fitted context
-# versus the server's FP16 default.
+# Use INT8 KV so the Jetson service can reserve an 8k context on Orin Nano
+# when systemd starts it before memory-heavy services. `-c` is still clamped
+# by runtime memory budget, so boot/start ordering matters.
 ExecStart=/opt/geniepod/bin/jetson-llm-server \
     -m ${GENIEPOD_LLM_MODEL} \
     -p 8080 \
     -c ${GENIEPOD_AI_RUNTIME_CONTEXT} \
     --int8-kv
 Environment=GENIEPOD_LLM_MODEL=/opt/geniepod/models/Qwen3-4B-Q4_K_M.gguf
-Environment=GENIEPOD_AI_RUNTIME_CONTEXT=2048
+Environment=GENIEPOD_AI_RUNTIME_CONTEXT=8192
 Restart=on-failure
 RestartSec=5
 TimeoutStartSec=120