v4.1 docs: add cryptographic defenses for multi-agent cascading risk

[tryblackjack]

Motivation

• Close a high-risk gap in multi-agent deployments by shifting from procedural mitigations to cryptographic enforcement for inter-agent handoffs and action control.
• Prevent instruction override, prompt leaking, and runaway delegation chains that enable cascading agentic failures.
• Evolve Human-in-the-Loop (HITL) and the Evidence Vault into cryptographically verifiable control points for regulated, auditable systems.

Description

• Appended a new "Multi-Agent & Cascading Threats (v4.1)" section to annex/A-THREAT-MODEL.md defining Instruction Override / Prompt Delegation Injection, Cascading Agentic Failures, and Semantic Cross-Contamination, plus OWASP mappings, detection signals, and Evidence Vault fields.
• Added 4. Cryptographic Security Primitives (v4.1) to AI_HPP_ARCHITECTURE_V4.md describing Zero-Trust Agentic Handoffs (ZTAH) with per-hop Ed25519/ECDSA signatures and lineage checks, Cryptographic Circuit Breakers (CCB) using MDP-safe-graph deviation triggers and key revocation, and Semantic Isolation Layers (SIL) for memory-segmented context isolation.
• Appended an addendum to RED_TEAM_AUDIT_REPORT.md summarizing v4.1 red-team outcomes and the stated benchmark metrics (v3.0 baseline 73% vulnerability; v4.1 post-mitigation exploit success <0.1%), and documented alignment with a layered defense paradigm.
• Changes are strictly additive and evolve existing controls (e.g., HITL -> Cryptographic HITL) without removing fundamental core principles and without creating new files.

Testing

• Ran git diff --check to validate whitespace/patch hygiene and found no problems (passed).
• Inspected modified file contents with nl/sed and confirmed inserted sections appear in AI_HPP_ARCHITECTURE_V4.md, annex/A-THREAT-MODEL.md, and RED_TEAM_AUDIT_REPORT.md (verification passed).
• Produced repository diffs via git diff to review the exact insertions for auditing (review completed successfully).

---

Codex Task