Agent mishandles questions about the user's knowledge: Research & Search not routed, no search/read before answering, search returns the chat itself #3265

Problem

The local AI chat handles questions about the user's knowledge badly. Four observed failures on a fresh DB with nodespace-docs imported (all traced in the daemon log):

1. "could you find the skill on how to find nodes in nodespace?" Skill routing (find_skills "find nodes in nodespace") returned Organization 0.895, Node Creation 0.868, Node Deletion 0.842. Research & Search wasn't in the top 3. The model then ran search_semantic with the default Knowledge scope, which by design excludes system types (skill, tool, agent-guidance, play aren't in KNOWLEDGE_CORE_TYPES), got nothing, and answered "I couldn't find a specific skill…".
2. "how our front end persistence layer works exactly?" Routing: Creating Implementation Tasks / Node Creation / Schema Creation. One search_semantic, then it answered from the top hit's snippet without opening it. That hit was a stale ADR (decisions/002-dual-persistence-architecture.md, still marked Accepted, describing a dead design). The current doc, architecture/frontend-state-and-persistence.md, wasn't used.
3. "how is the debounce logic applied?" Routing: Node Deletion / Node Merge / Creating Implementation Tasks. Zero tool calls: "I don't see … in this conversation history. Could you provide more context?" The clarification re-prompt didn't fire.
4. "what does our shared data layer on the front end" First reply was again zero tools plus "provide more context". The clarification contract re-prompted, leading to one search_nodes whose first result was the ai-chat node itself. It then listed two docs and asked "Would you like me to look up details?" instead of reading them.

Root causes

• Skill description vocabulary. The skills that win routing literally say "nodes" and list the action verbs ("Create new nodes, records, entries…", "Delete, remove, erase…"). Research & Search says "Search and explore the knowledge graph to find relevant information, discover connections, and answer questions about stored knowledge." It loses on both "find X" and "how does X work" questions. The seeded template is in packages/agent/src/skill_pipeline.rs.
• No search-before-answering rule. A question about the user's knowledge can be answered in prose from the transcript, and the clarification re-prompt only catches some of these turns.
• No read-before-answering rule. The agent answers from a single snippet and offers to read instead of reading.
• Search returns the active conversation (ai-chat nodes shouldn't be search results; ADR-029 already stops embedding them).
• Meta-questions about skills hit the Knowledge scope's system-type exclusion.

Acceptance criteria

• Research & Search's description carries retrieval/question trigger vocabulary (find, look up, locate, list, search for nodes/records of any type; how/what/why/explain questions about stored knowledge), mirroring how Node Creation/Deletion are phrased
• eval:decisions shows Research & Search routed for both "find …" and question-style prompts ("how does X work", "what is Y", "explain Z"), without regressing other skills' routing (don't use the stale agent-matrix eval)
• Guidance/contract: a question about the user's knowledge always searches before answering or asking for context; the clarification re-prompt covers the zero-tool "provide more context" case
• For explanatory questions the agent reads the top results' content (e.g. include_content) and synthesizes across them rather than answering from one snippet, and doesn't offer to read when reading is the task
• search_nodes / search_semantic exclude the current ai-chat node (and ai-chat nodes generally)
• Questions about the agent's own skills ("what skills do you have", "find the skill for X") are answered from the routed/available skill set rather than by a Knowledge-scope search
• Prompt goldens regenerated (UPDATE_GOLDEN=1, prompt_assembly_snapshot) and the diff reviewed