perf(calcite): fold GROUP BY on a dictionary string column per code

[dfa1]

Two commits.

1. GROUP BY on a dictionary string column folds in the scan

Calcite ran GROUP BY by building a boxed row per record, copying it through the interpreter and hashing the key into a HashMap per row (25% of the query in the profile; the reader was under 5%).

A GROUP BY on one string column now folds per dictionary code:

• Reader: Compute.groupedAggregate(chunk, groupColumn, filter, aggColumn) accumulates count, non-null count, sum, min and max per code in primitive arrays (sliced shared dictionaries included) and turns each code into its string once per chunk. Empty when the shapes do not qualify.
• Calcite: VortexTable.groupedFold merges the chunks; VortexAggregatePushDownRule replaces the subtree with one Values row per group, with the same WHERE support as the whole-table rewrite.

Falls back to the scan when a chunk is not dictionary-encoded, an aggregate is not COUNT/SUM/MIN/MAX of a plain numeric column (u64, f16 excluded), there are more than 10,000 groups, or a group value is outside Latin-1.

Calcite fires the rule 6 to 9 times per query, so folds are memoized per planner run (a weak map keyed by the planner): one fold per query, nothing cached across queries.

DictStringBenchmark.sqlGroupBy, single fork, ms/op:

	Before	After
4M synthetic rows, 8 symbols	136.8	4.5 (30×)
sf-salaries (Rust-written, 1.18M rows)	24.4	2.5 (9.6×)

2. Filtered aggregates fold once per query

The same memo now covers VortexTable.filteredFold: a filtered count/min/max over sf-salaries decoded its boundary zones 6 times while planning one query; now once.

Tests

• Unit: Compute.groupedAggregate (null group, long/double lanes with nulls, Rust float order, filter, sliced dictionary, duplicate dictionary strings, malformed and truncated codes, declined shapes); rule rewrites GROUP BY on a dictionary column and leaves numeric, multi-key and AVG groupings alone.
• vortex-jni cross-check: Rust writes a nullable dictionary string column; GROUP BY with COUNT(*), COUNT(col), SUM, MIN, MAX (long and double) under three WHERE clauses equals the aggregates computed from the rows written, and the plan is a Values; a non-Latin-1 key falls back to the scan with the same answer.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WtuVGQ5h5Lk6Zn4JnkxF1K