Summary

Genie.poll_for_result was surfacing only the trailing follow-up/clarifying question instead of the full answer for some Genie responses. _parse_attachments (since #419) picked the first text attachment following the last query, but internally-created messages emit a follow-up question text attachment before the final summary. The follow-up has an attachment_id; the final summary text attachment has no attachment_id.

This change makes _parse_attachments:

• Drop text emitted strictly between the first and last query (superseded self-correction attempts), keeping leading and trailing text.
• Among the kept text attachments, return the final summary — the text attachment without an attachment_id — falling back to the first kept text attachment when no id-less attachment is present.

Identifying the summary by the absence of attachment_id matches the agreed short-term approach from the incident discussion (the FinalSummaryAttachment is the only public-API attachment without an id).

Tests

• Added a parametrized _parse_attachments case: follow-up question (with attachment_id) emitted before the id-less summary → the summary wins.
• Full tests/databricks_ai_bridge/test_genie.py suite passes (75 passed).

Context: ES-1953657. Supersedes the concat-all approach in #432.

This pull request and its description were written by Isaac.