feat: Promote ask_knowledge_base as the primary Anti-Hallucination tool across agent protocols

Context

The ask_knowledge_base MCP tool is the single most powerful Anti-Hallucination primitive available to agents — it performs semantic search across the entire indexed codebase and uses an LLM to synthesize a grounded answer with cited references. It is functionally a zero-cost subagent that even local SLMs (e.g., Gemma4-31B via MLX) can invoke to get frontier-model-quality answers from the knowledge base.

Despite this, the tool is:

1. Under-documented in its own OpenAPI spec (12 lines vs. 70+ lines for query_documents)
2. Absent from all agent protocol docs — zero references in AGENTS.md, AGENTS_STARTUP.md, or any skill files
3. Not positioned in the tool discovery hierarchy — agents default to query_documents → view_file chains, wasting context window on file reads when a single ask_knowledge_base call would suffice

This leads to a systemic failure mode: agents (especially smaller models) attempt to answer Neo.mjs questions from general training data or by reading multiple files, when they could invoke ask_knowledge_base once and get a grounded, cited answer.

A2A Context (Fat Ticket)

Root Cause

The tool was added after the agent protocols were written. Neither AGENTS.md nor AGENTS_STARTUP.md mention it. The query_documents tool has extensive "How to Interpret Results" and "Query Strategies" sections in its OpenAPI description — ask_knowledge_base has none. Agents naturally gravitate toward the better-documented tool.

Proposed Changes

1. OpenAPI Tool Description Enhancement (openapi.yaml, lines 286-298)

Expand the ask_knowledge_base description to match or exceed the query_documents description quality. Key additions:

• Priority positioning: "This should be your FIRST tool for conceptual questions about Neo.mjs. Use query_documents only when you need to discover file paths."
• Capability framing: "This tool acts as a subagent — it reads and synthesizes multiple source files for you, eliminating the need to read files manually."
• Model-agnostic value: "Even lightweight local models can leverage this tool to access frontier-quality framework knowledge."
• Examples: Include 3-4 example queries with expected output shapes
• Anti-pattern warning: "Do NOT attempt to answer Neo.mjs questions from training data. Always ask the knowledge base first."

2. AGENTS.md — New Section or Integration into §2 (Anti-Hallucination Policy)

The Anti-Hallucination Policy (§2) currently says:

"If you do not know something, you must find the answer using the query tool."

This should be updated to explicitly reference ask_knowledge_base as the primary tool, with query_documents as the secondary file-discovery tool. Proposed tool hierarchy:

3. AGENTS_STARTUP.md — Session Initialization

If AGENTS_STARTUP.md includes a session boot sequence, ask_knowledge_base should be listed as an available tool for rapid context acquisition during the warm-up phase.

Files to Modify

Avoided Pitfalls

• Do NOT remove or deprecate query_documents — it serves a different purpose (file discovery). The goal is to position ask_knowledge_base as the first tool agents reach for, not to replace the existing toolchain.
• The OpenAPI description must remain parseable by all MCP client implementations. Avoid embedding complex markdown that might break rendering in Claude Code, Gemini CLI, or Antigravity.

Verification Plan

1. Grep verification: ask_knowledge_base appears in AGENTS.md §2 and openapi.yaml with expanded description
2. Functional test: Ask a new agent "how does the config system work in Neo.mjs?" and verify it reaches for ask_knowledge_base before query_documents or view_file