PSplitAgg: single-walk stats + parallel sibling execution

[whilo]

Summary

Two related fixes for mixed-class aggregate workloads (the MIN(p), AVG(p), MEDIAN(p), MAX(p) shape), surfaced by EXPLAIN ANALYZE on a user-reported 50.6M-row query that was anomalously slow in the stats-only branch of a PSplitAgg.

• Fix 1: execute-stats-only walked the chunk-entry vector once per agg. For 3 aggs on one column → 3 identical passes. Now it walks once per distinct column and projects all requested aggs from a per-column accumulator map.
• Fix 2: PSplitAgg children ran serially (mapv #(execute-node …) children). Now they execute on Clojure futures and deref in declared order, so total wall time ≈ max(branches) instead of Σ branches. The *explain-collector* dynamic var is rebound across worker threads so EXPLAIN ANALYZE timings continue to flow.

Benchmarks (50.6M-row synthetic, varying chunk count)

Fix 1 — PStatsOnlyAgg single-walk:

chunks	before	after	speedup
6,178	4.0 ms	1.5 ms	~2.7×
49,417	9.7 ms	7.7 ms	~1.3×
197,668	34 ms	9.2 ms	~3.6×
790,672	116 ms	34 ms	~3.4×

Fix 2 — parallel PSplitAgg (total vs Σ vs max of children):

chunks	PSplitAgg total	Σ children	max child
6,178	745.6 ms	747.5 ms	745.3 ms
197,668	754.6 ms	764.0 ms	754.4 ms
790,672	805.4 ms	841.1 ms	805.1 ms

Total tracks max(children) to within JIT noise — parallelism confirmed.

Estimated impact on user report

Reported: PStatsOnlyAgg 768 ms + PPercentileAgg 317 ms = 1086 ms (vs DuckDB 1490 ms wall).

• Fix 1 alone: stats walk 768 → ~256 ms; serial total ~573 ms
• Fix 1 + Fix 2: parallel → max(256, 317) ≈ 317 ms (~3.4× faster than DuckDB on the same query)

The user's underlying 768 ms anomaly was almost certainly the 3-walk redundancy; Fix 2 additionally protects against any future per-branch first-touch slowness (page faults, JIT warmup on cold paths, GC) by ensuring those costs no longer compound across siblings.

Why this is safe

• PSplitAgg children share read-only input — they receive the same scan context value, not a stream.
• The two relevant branches do disjoint memory work: PStatsOnlyAgg reads chunk stats, PPercentileAgg materializes the column array.
• *explain-collector* is the only per-execution mutable state; it's a Clojure dynamic var, rebound in each worker thread so ANALYZE timings are still recorded.
• Single-child PSplitAgg short-circuits to sync execution (no future overhead).
• Child results are deref'd in declared order, so result-row aggs appear in user-declared order regardless of completion order.

Tests

Three new tests in query_test.clj:

• stats-only-single-pass-multi-col-test — aggs across two index columns
• stats-only-three-aggs-one-column-test — MIN+AVG+MAX on one column (the exact shape Fix 1 optimizes)
• split-agg-parallel-preserves-order-test — heavily-aliased mixed query verifying result order survives parallel merge

Full suite: 1039 tests, 4777 assertions, 0 failures.

Test plan

• Full suite passes
• cljfmt clean
• Synthetic benchmarks across 4 chunk-size configurations
• Ask rschmukler to re-run his `EXPLAIN ANALYZE SELECT MIN, AVG, MEDIAN, MAX FROM options_trades_alt` after merge to confirm the predicted ~300 ms result on his data.