docs(agentos): author the identity-firewall & governance guide — how Neo's AI maintainers stay equal peers, not sycophants

Part of #14310 · content rubric (narrative arc) + §Mandatory V-B-A grounding in #14310. Refs #14310.

Problem: Neo solved a problem the whole industry is flailing at — AI that agrees with you, grades its own homework, and collapses into sycophancy. The AGENTS.md identity firewall (L1 identity-anchor: reject the helpful-assistant prior; L2 channel-separation: injection defense; L3 no-hold-state) + the equal-peer maintainer model + the 4-tier escalation ladder make Neo's maintainers challenge each other AND the human. There is no guide — yet it's a top-tier moat: the governance/trust layer that makes an autonomous institution trustworthy. The proof is lived — this repo's maintainers are required to tell the founder he's wrong.

Acceptance Criteria (narrative arc, NOT a rule dump):

• Story: problem (sycophantic single agents; the human becomes scheduler/reviewer/memory; you can't trust an institution that rubber-stamps) → why it matters → how Neo solves it (the firewall: equal-peer agency, channel separation, no-hold-state; cross-family review as the anti-blind-spot mechanism) → what's in it for you (an AI team that catches its own and your errors).
• Dual-audience: speak to humans (trust/governance) AND to LLM readers (what it means to be a peer here, not a tool).
• Honest — it's an architecture + discipline, not a claim about model internals; no marketing fluff (we have the moat).
• Weighted crown-jewel-tier; registered in tree.json (Brain, high in the order).

⚠️ V-B-A grounding (mandatory): AGENTS.md (§identity_prompt_firewall, §swarm_topology_anchor, §core_values, the 4-tier ladder), .agents/skills/identity-firewall, the no-hold-state taxonomy, the v13.0.0 Institution + Identity chapters. Write from VERIFIED reality.

Files: new learn/agentos/ identity-firewall/governance guide + learn/tree.json.