Establish author baseline fixtures and offline tests, make SemanticProject the sole authority for graph edges, classify named steps, and tighten local reuse so unknown effects are not pure and MLSP binds by SymbolId. Co-authored-by: Cursor <cursoragent@cursor.com>