One row per candidate answer to a single question, unified across
standalone and pipeline assessments and discriminated by source.
This is the grain to aggregate for question-difficulty and
success-rate analysis: filter by question_id and reduce over
score.

Each row includes the candidate's raw answer and the grader's
evaluation_details, both unbounded JSON. per_page is therefore
capped at 100 here rather than the usual 200.

Deep pagination works as on GET /assessment-results, but with a
lower scan budget: 2 000 rows per source there, 500 here, because
these rows are far heavier. At the default per_page that is 5 pages
before 400 pagination_limit_exceeded; the same escape applies —
pin source= (or pipeline_id) and the limit does not apply at all.