fix(group): bound the tied groups the radix top-N selection keeps

[ser-vasilich]

Bug

The native top-N of the radix grouping path (agg_radix_select_topn, src/ops/agg_engine.c) keeps every group at or beyond the threshold, then heap-sorts the kept set by first row to emit it in first-seen order. Groups strictly beyond the threshold number fewer than N by construction, but the groups at the threshold can be nearly all of them: a count-per-group top-10 over a near-unique key pair has a threshold of 1, every group ties it, and the "kept superset" is the whole grouping — a comparison sort over ~100M pairs for a query whose answer is ten rows.

On a 100M-row table, by: [WatchID ClientIP] c: (count …) desc: c take: 10 went from 3.4 s to 78 s, and the same query under a row selection from 0.45 s to 6.5 s; 53% of the profile is agg_sort_pairs_by_key.

Fix

The emitted prefix is the tied groups' first-seen order, so only the N tied groups with the smallest first row can ever be taken. The selection now collects the strictly-better groups (fewer than N) and, among the groups equal to the threshold, exactly the N with the smallest first row through a bounded max-heap, the same scheme agg_radix_select_first_n uses. The sort then runs over at most 2N entries. When more than N groups turn out strictly better than the threshold (which the threshold contract excludes) the selection declines and the caller takes the full path, as it does for an aggregate without a scalar order.

The rows emitted are the ones the full kept set produced: identical output on every shape checked, including the ascending direction and a row selection. The same query is now 1.8 s and 0.35 s, faster than before the native selection.

Tests

• test/test_agg_contract.c agg_contract/radix_native_topn: the pinned kept count of the tie-heavy case becomes N (it counted the whole tie set before).
• test/rfl/group/emit_filter_v2_route.rfl: a near-unique two-key grouping whose threshold every group ties, in both directions and under a row selection, against the full grouping.