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.

policy/openshell/base.yaml

@@ -0,0 +1,33 @@
+# policy/openshell/base.yaml
+# Lumina base policy — reference document.
+#
+# This file documents the deny-by-default intent for the Lumina sandbox.
+# It is NOT applied directly. The live policy is managed by nemohermes
+# policy-add/policy-remove commands.
+#
+# Apply via: nemohermes <name> policy-add --from-file <path>
+#
+# ── Deny-by-default intent ────────────────────────────────────────────────
+#
+# The sandbox starts with NO network egress except what is explicitly allowed
+# by applied policy presets. The base tier (balanced) provides:
+#
+#   npm, pypi, huggingface, brew  — package management (always allowed)
+#
+# Lumina adds overlays for:
+#   - inference endpoints (main model + vision)
+#   - SaaS integrations (Square, QBO, Vagaro) — when connected
+#   - messaging channels (WhatsApp, Telegram, Email) — when connected
+#
+# ── Always deny ────────────────────────────────────────────────────────────
+# - Social publish APIs (no silent publish)
+# - Payment/refund/payout/bill-pay endpoints
+# - Any host not explicitly listed in an overlay
+#
+# ── Policy lifecycle ───────────────────────────────────────────────────────
+# 1. Install (S5): apply base overlays (inference)
+# 2. Connect (S7): apply SaaS/channel overlays as integrations are enabled
+# 3. Upgrade: re-apply all overlays (idempotent)
+# 4. Doctor (S6): verify active policy matches expected state
+#
+# See docs/POLICY.md for full policy documentation.