### What?
Fixes an issue where only the fields from the first batch of documents
were used to generate CSV column headers during streaming exports.
### Why?
Previously, columns were determined during the first streaming batch. If
a field appeared only in later documents, it was omitted from the CSV
entirely — leading to incomplete exports when fields were sparsely
populated across the dataset.
### How?
- Adds a **pre-scan step** before streaming begins to collect all column
keys across all pages
- Uses this superset of keys to define the final CSV header
- Ensures every row is padded to match the full column set
This matches the behavior of non-streamed exports and guarantees that
the streamed CSV output includes all relevant fields, regardless of when
they appear in pagination.
34c3a5193b