Fix exercise inconsistencies across all three tracks

[noravth]

Summary

• Python track: Align appraiser_agent role/goal/backstory with solution; fix LLM model string (sap/gpt-4o → sap/anthropic--claude-4.5-opus); remove forward reference to solve_crime from Exercise 06; fix solve_crime context to include all 3 tasks; fix method ordering list in Exercise 07
• Python-LangGraph track: Fix _lead_detective_prompt in Exercise 07 to keep 4-parameter signature from Exercise 06; fix ainvoke → invoke in Exercise 06; wire up WEB_RESEARCHER_AGENT prompt so it's actually used (LLM synthesizes tool results)
• JavaScript track: Fix intelligenceResearcherNode to call LLM with AGENT_CONFIGS.intelligenceResearcher.systemPrompt to synthesize tool results; update Exercise 06 to match
• Both LangGraph tracks: intelligence_researcher_node now follows the same pattern as other nodes — runs tool calls first, then passes raw results to LLM for synthesis