Page score set downloads without walking every skipped row #901

Summary

Scripts page through large score sets with start and limit, and each page makes the database read and discard every row before start. On a 648,000-variant score set a page near the end reads more than half a million full rows to return a few thousand. This issue finds each page's first row through the index alone and then reads only the page's own rows, with no change to the API.

Problem

fetch_variant_csv_data (lib/csv/fetch.py) applies start as SQL OFFSET. With the variant-number index the rows come out in order, but OFFSET still fetches each skipped row from the table, including its data column, before dropping it.

Scope

• Select the page's variant ids in a subquery that reads only the index: SELECT id FROM variants WHERE scoreset_id = :id ORDER BY <variant number>, id OFFSET :start LIMIT :limit. The index on (scoreset_id, variant number, id) covers it.
• Fetch the full rows, and their joined mappings, for those ids only.
• Apply it wherever start or limit is set; unpaged requests keep today's query.

start and limit keep their meaning, and the output doesn't change.

Acceptance criteria

• EXPLAIN of the id subquery shows an index-only scan.
• In staging, EXPLAIN ANALYZE of the queries for start=600000&limit=50000 on urn:mavedb:00000053-a-1 totals under 1 second.
• Paged output is byte-for-byte identical to the current output for the same start and limit.