Three tools with reasonable naming (verb_noun pattern) and descriptions, but critical gaps in schema documentation and parameter descriptions. All tools lack explicit input parameter type definitions and descriptions in the visible code. Descriptions are adequate (100-200 chars) but parameters are underdocumented. No output schemas documented. Error handling and recovery guidance absent. Tools appear to be READ_ONLY (good for safety) but lack formal annotations.

• Add explicit JSON Schema type definitions for all parameters. For 'query' (string), add minLength/maxLength. For 'top_k' (integer), add minimum (1) and maximum (100). For 'ranker', declare as enum: ['importance'].
• Document output schemas for each tool. E.g., search_pairs should return {results: [{pair_id: int, score: float, preview: string}], total: int}. compact_session should return {session_id: string, pairs: [...], dropped_count: int}.
• Add error handling guidance to descriptions. E.g., 'If substrate files are missing, returns error code 404 with message "Substrate not initialized. Call substrate_info() first to verify."'
• Declare tool annotations in the MCP server registration. Mark all three tools with readOnlyHint=true and idempotentHint=true to signal to agents that they are safe to retry.
• Expand parameter descriptions with context and constraints. E.g., 'k_drop (0.0 - 1.0): fraction of lowest-ranked pairs to exclude from output. Default 0.5 removes bottom 50% by importance score.' Explain 'honor_manifesto': 'If true, filters results to respect labeling constraints defined in manifesto.yaml.'