Implement install stages S3–S5: package, policy, and skills sync.

Attach/onboard sandbox from agents/hermes, additive OpenShell policy overlays, and nemohermes skill install for scaffold skills. No doctor/connect and no push.

agents/hermes/config/inference.yaml

@@ -0,0 +1,22 @@
+# agents/hermes/config/inference.yaml
+# Inference config fragment for nemohermes onboard.
+#
+# This fragment is merged into the sandbox's managed config during onboard.
+# Values are resolved from .env at install time.
+#
+# Platform-first: the actual config write is done by nemohermes inference set,
+# not by hand-editing in-sandbox files.
+
+# Main model endpoint (OpenAI-compatible)
+# Resolved from LUMINA_INFERENCE_BASE_URL and LUMINA_INFERENCE_MODEL
+inference:
+  provider: compatible-endpoint
+  # model: set by nemohermes inference set from .env
+  # base_url: set by nemohermes inference set from .env
+
+# Auxiliary vision model
+# Resolved from LUMINA_VISION_MODEL
+auxiliary:
+  vision:
+    # model: set from LUMINA_VISION_MODEL (often same as main if multimodal)
+    pass