Return the audience's built output as JSON, one page at a time. Each
result is one row with the same columns the download files carry.

The audience must be built first: call
GET /api/v1/audiences/{id}/download and poll until its status is
ready. Until then this endpoint answers 409 with not_ready.

To read the whole audience, request a page, then pass its next_cursor
as cursor on the next request. Repeat until a response has no
next_cursor. page_size is a maximum, not a guarantee: a page can
hold fewer rows and still carry next_cursor, so loop on the cursor,
not on the row count.

Send Accept-Encoding: gzip; responses are gzip-compressed when you
do, which is much faster for large pages.

Billing. Billed per record returned, at the audience's meter
(Predictive Signals audiences bill at the Predictive Signals rate). A
record read again on the same UTC day is not charged again, and in one
day paging a build never costs more than downloading it. If the build's
file download was already charged, paging that build is free. A page
that fails partway is not charged.

Test API keys page a fixed 7-row sample (count: 7; with
page_size=3 the pages hold 3, 3 and 1 rows, and the last has no
next_cursor). Responses carry the header X-Delivr-Mode: test and
are never billed. Audience id pending returns 409 not_ready, so you
can exercise that path before using a live key.

Project-scoped: name the project with the X-Project-Id header or a
?project_id= query parameter, as with the other audience endpoints.
Any query parameter other than page_size, cursor, and project_id
returns 400.