The pipeline has three stages:
  1. Planner — breaks the query into 3-5 targeted sub-questions (GPT-4o)
  2. Searcher — runs parallel Tavily web searches with credibility scoring
  3. Synthesizer — produces a structured markdown report (GPT-4o)

Results are cached for 24 hours. Identical (query + context + depth) combinations
return instantly on subsequent calls.

Args:
    params (ResearchInput): Validated input containing:
        - query (str): The research question (3-2000 chars)
        - search_depth (SearchDepth): 'basic' for speed or 'advanced' for depth
        - conversation_context (str): Prior research for multi-turn follow-ups

Returns:
    str: Markdown research report with executive summary, key findings per
         sub-question, knowledge gaps, and numbered source citations with
         credibility tags (high/medium/unverified).