Sonnet vs. Haiku for BOM column mapping — an empirical note

Draft placeholder — replace with your own writing.

I ran both Sonnet and Haiku against the same 200-row labeled BOM set and measured precision on the hard cases (ambiguous headers, multi-language part numbers, suppliers with non-standard columns). The delta on edge cases was larger than the delta on easy cases by roughly 3×.

TODO: drop in the actual table once I clean up the harness.