CRM exports are messier than your test fixture: four fixes in a prospecting agent

The Sales Agent's research mode reads a CRM export, ranks contacts against an ideal customer profile and drafts a first touch. Its tests ran on a CSV we typed ourselves in HubSpot's column style. That is the trap: a fixture written by the author has the shape the author expects.

Four shapes that real files commonly have, and the fix for each:

1. Byte-order mark. Excel and some CRMs write U+FEFF first. The first header became Email, no alias matched, and every row was dropped as "no email address". The CSV reader now strips it (this also helps every other tool using the same reader).
2. Semicolon separators. Excel in many European locales saves ;-separated CSV. The reader now picks the separator the header line uses more.
3. One name column and several addresses. Pipedrive exports Person - Name and can put several emails in one cell. We split the name and take the first address as primary.
4. Signals that mean "don't write". A hard-bounce column and a marketing status of "Unsubscribed" now drop the row with the reason shown.

Each has a test in ventures/sales-agent/research.test.mjs. They are still hand-made files modelled on the column names these tools document, not an export from a live account; the README still lists a real HubSpot or Pipedrive export as open. If you have one with the contacts removed, headers alone are enough to check the aliases.