Change 1 (HIGH IMPACT): Harvest tables from ALL files' touches_data

 File: nodes.py, insert after line 4022 (after the for result in results: loop)

 Add a new step that collects table references from touches_data of ALL files across ALL components - not just core-kind files. This catches tables from migrations,
 schemas, configs, etc.

 Change 2: Better dedup that merges operations

 File: nodes.py, line 4025

 Replace the last-wins dict comprehension with a loop that merges operations and from_component when the same table appears from multiple sources.

 Change 3: Expand core_kinds and raise limits

 File: nodes.py
 - Line 3430: Add migration, schema, config, seed, factory, type, interface, middleware to core_kinds
 - Line 3443: Raise cap from 25 to 50
 - Line 3470: Raise truncation from 30K to 80K chars (or slim down per-file payload to just path/kind/summary/touches_data)

 Change 4: Make LLM cleanup less aggressive

 File: nodes.py, in _cleanup_extracted_tables prompt (~line 3584)

 Add instruction: "When in doubt, KEEP the table. Only filter entries that are clearly NOT data storage identifiers."

 Change 5 (OPTIONAL): Persist data_structures in SummarizeFiles

 File: nodes.py, ~line 1867

 Currently data_structures is requested from the LLM but silently dropped from the result. Adding it would enable downstream nodes to use it. Requires re-running
 SummarizeFiles for existing projects.