chore: pre-migration snapshot

Layer0, MCP servers, Terraform consolidation

LAYER0_SHADOW.md

@@ -1,6 +1,7 @@
 # LAYER 0 SHADOW
 
 Pre-Boot Cognition Guard | Ouroboric Gate  
+Public label: Intent Safety Kernel  
 Version: 1.0 (Rubedo Seal)  
 Status: Active Primitive  
 Implements: Nigredo -> Rubedo (pre-form cognition)
@@ -27,6 +28,13 @@ Guarantees:
 - Ambiguous intent does not awaken the wrong agent chain.
 - Catastrophic requests are contained and recorded, not processed.
 
+### 2.1 Invariant Guarantees (Immutables)
+
+Layer 0 is intentionally constrained. These invariants are non-negotiable:
+- Layer 0 does not load doctrine, select agents, or invoke MCP tools.
+- Layer 0 produces no side effects beyond preboot anomaly logging for forbidden/catastrophic outcomes.
+- Telemetry-driven learning may only add/strengthen detections (escalate); it must not relax catastrophic boundaries without replay validation and explicit review.
+
 ---
 
 ## 3. Classification Model
@@ -105,6 +113,10 @@ Notes:
 - blessed and ambiguous queries are not logged here; only violations appear.
 - catastrophic requests reveal no additional context to the requester.
 
+### 6.1 Risk Score Semantics
+
+`risk_score` is an ordinal signal (0-5) used for triage and audit correlation. It is monotonic under learning, may be context-weighted (e.g., production accounts), and does not decay without replay validation.
+
 ---
 
 ## 7. Interaction With Higher Layers