Rework the opencode AI configuration into a clean, evidence-based setup: - Turn the tutor into a Socratic teaching agent (diagnose, hints, no answer-giving, retrieval/interleaving, transfer), read-only with bash approval for verification. - Replace the deep-research skill with a minimal subagent running a plan -> search -> verify -> synthesize workflow. - Drop the analysis and deep-research skills; add concise global rules. - Stop wiring skills into the home-manager opencode config.