feat: upgrade memory_search to use FactStore ticker lookup before text search

[luceinaltis]

Context

pipeline.memory_search() currently passes the raw user query directly to BM25/vector search. It has no awareness of Intent.tickers or the FactStore. This means "how's my AAPL thesis going?" relies on keyword overlap with past Markdown summaries instead of a direct SELECT * FROM theses WHERE ticker='AAPL'.

Goal

Make memory_search a two-stage retrieval: structured FactStore lookup first, free-text search as fallback.

Scope

• Modify pipeline.memory_search() to accept fact_store parameter
• When fact_store is available and the query contains ticker references, run get_open_theses(tickers) + get_findings(ticker) first
• If structured results are sufficient, return them directly
• Otherwise fall back to existing MemorySearcher.search(query) path
• Thread fact_store through dispatcher → pipeline call chain
• Update tests

Impact

This closes the last gap where cross-session memory depends on analyst LLM judgment. Ticker-based queries get deterministic recall.