DATE_FORMAT: compile the pattern once, render into a reused buffer

Applies to: DATE_FORMAT expression evaluation

chrono's format(pattern) re-parses the strftime pattern inside every Display, and the original loop rendered each row into a fresh String — a per-row parse plus a per-row allocation for a pattern that is a validated literal (the JVM encoder admits only translated literal patterns). A Fluss-rung profile surfaced it: StrftimeItems::next under spec_to_string, once per row.

First round: compile once, reuse the buffer

CompiledFormat parses the pattern to owned Items once per distinct pattern per batch (in practice once), formats with format_with_items, and writes into one reused buffer that the Arrow builder copies from — the same compile-once principle applies to regex.

Measured (Criterion date_format/*, 4096 rows, the Nexmark yyyy-MM-dd pattern): the old formulation (per_row_parse) runs 670 µs/batch (~6.1 Melem/s); the compiled path 378 µs (~10.8 Melem/s) — 1.77× on the hot loop, kept as an A/B pair in the bench suite.

On the matrix, DATE_FORMAT is a ~10% slice of its queries (q10/q14/q15/q16/q17), so the end-to-end effect sits inside combined-run noise at 500K events — the loop win is the stable number, not an end-to-end one.

Second round: skip chrono's rendering machinery entirely

A 2026-07-12 q17 profile still had DelayedFormat::fmt (per-item core::fmt dispatch and padding) as the bulk of the remaining cost. Patterns made only of literals and zero-padded date/time fields — every Nexmark pattern — now lower once to a digit-writing plan that pushes ASCII digits straight into the reused buffer. Anything else (and a year outside 4 digits, which chrono prints unpadded) falls back to format_with_items per row, so output stays byte-identical — a Rust parity sweep pins the two renderings against each other.

Criterion (date_format/digit_plan vs compiled, same 4096-row batch): 362 µs → 78.7 µs — 4.6× on the hot loop, 11.3 → 52 Melem/s.

As with the first round, this is a loop-level win: the end-to-end matrix effect remains inside combined-run noise at 500K events for the queries that use DATE_FORMAT.