expr: don't read unordered stats bounds as an empty range

[def-]

Motivation

Filter pushdown discarded persist parts that held rows matching the query, so the query returned too few rows. '-NaN' is ordinary user input, so no corrupt storage is needed to reach this.

Closes: PER-53

Description

Two independent causes, both fixed here.

ResultSpec::value_between collapsed min > max to nothing(), i.e. "no value can be here". That only holds for bounds a caller derived coherently, and persist part statistics are not such a caller. Arrow orders floats totally, putting -NaN below -Infinity, while OrderedFloat, the Datum order value_between compares in, ranks every NaN above every finite value. A part holding -NaN therefore reports lower = -NaN against a finite upper, the range reads back as empty, and the part is discarded along with every other row in it. Unordered bounds now widen instead of collapsing.

Widening alone is not sufficient. A part holding NaNs of both signs records the total-order bounds (-NaN, +NaN), which decode under OrderedFloat to NaN == NaN, a valid-looking, non-inverted range claiming the part holds nothing but NaN. A filter that a finite row matches but NaN does not, for example value = 0.0, still discarded the part. By the time value_between runs the NaN signs are erased, and (NaN, NaN) is also the honest spec of a genuinely all-NaN part, so the guard has to live in the stats decode, which still sees the sign bits. mz_repr::stats::col_values now refuses to produce float bounds from a negative-NaN lower, unless the upper shows the whole column is NaN.

Verification

Unit regressions for both symptoms in persist_source.rs's filter_pushdown_audit module, negative_nan_does_not_discard_matching_part and mixed_sign_nans_do_not_discard_matching_part, each verified to fail without its half of the fix. Peek, dataflow, and EXPLAIN FILTER PUSHDOWN cases for both shapes in test/sqllogictest/explain/pushdown.slt.

🤖 Generated with Claude Code

[linear-code]

PER-53