test: cover the Polars SQL surface added since py-1.44.0 (30 sqlp tests)

[jqnatividad]

Context

50c2cf18f bumped Polars to rev 9d5804d. 20 commits touched crates/polars-sql between upstream py-1.44.0 (9ae4e57f5) and that pin, and none of them had test coverage here — the previous bump (6d33fb41f, to d84c1d4) added zero new test functions to tests/test_sqlp.rs, only adjusting existing ones for join-order fallout. So the whole py-1.44.0..9d5804d window was uncovered, not just the last few days of it.

This adds 30 tests and 7 shared fixtures (sqlp suite 106 → 136), plus one Cargo feature and the CHANGELOG entries for the surface that arrives with the bump.

Every expectation was captured by running the freshly rebuilt binary and pinning its actual output. Upstream's own tests drive SQLContext directly, so their SQL syntax transfers but their expected values do not — qsv sqlp layers CSV type inference, null rendering and float formatting on top.

New syntax and functions

	upstream
GROUP BY GROUPING SETS / ROLLUP / CUBE, GROUPING(), GROUPING_ID()	#29278
bare date-part fns YEAR QUARTER MONTH WEEK DAY/DAYOFMONTH DAYOFWEEK DAYOFYEAR HOUR MINUTE SECOND	#29269
APPROX_QUANTILE, 2–4 args	#29288
typed DATE '…' / TIMESTAMP '…' literals	#29007
date ± integer arithmetic; Decimal non-equi joins	#29156
parenthesized JOIN … ON, and aliases inside those parens	#28967, #29158
OVER on multi-argument aggregates	#29160
NULLS FIRST/LAST inside a window's own ORDER BY	#29159
EXISTS/NOT EXISTS, correlated scalar subqueries, CTE shadowing, case-insensitive relation names, ORDER BY over an unselected aggregate	#29006, #29010

Behavior changes now pinned

• CAST(<string> AS DATE/TIME/TIMESTAMP) parses instead of casting, and is strict where TRY_CAST is not. Two distinct failures: some rows parse → conversion from `str` to `date` failed; none parse → could not find an appropriate format to parse dates (#28062, #28986).
• Postgres scope strictness — an aliased relation's original name is out of scope, so SELECT t1.a FROM t1 AS f now errors (#28937).
• A scalar subquery's aggregate binds to its own relation (#28939), and unaliased constants with GROUP BY are named literal, literal:1, literal:2 rather than colliding (#29367).

The one non-test change

Cargo.toml enables polars' approx_quantile feature. Upstream cfg-gates the function and ships the feature only inside its docs-selection/full umbrellas, neither of which qsv uses, so sqlp rejected APPROX_QUANTILE outright before this. Verified empirically, not inferred: unsupported function 'approx_quantile'. qsvmcp and qsvdp share the polars dep entry and both still build; qsvlite has no polars.

A full sweep of polars-sql's function registry at the pin confirms approx_quantile is the only gate qsv was missing — the complete set is approx_quantile, bitwise, list_eval, nightly, rank; qsv has list_eval and rank, and bitwise is deliberately off.

Found an upstream bug while hardening this

Addressing a review finding about tied peer rows in the #29159 tests turned up a real defect, filed as pola-rs/polars#29390: a window ORDER BY honors NULLS FIRST/LAST only with a single key. Add a second key and the clause is silently ignored — null placement then tracks only the sort direction.

Positions of the two null rows under ROW_NUMBER(), options spelled identically on every key:

window ORDER BY	1 key	2 keys
a NULLS LAST	5, 6	1, 2	❌
a NULLS FIRST	1, 2	1, 2	✅ coincidence
a DESC NULLS LAST	5, 6	5, 6	✅ coincidence
a DESC NULLS FIRST	1, 2	5, 6	❌

The same two-key ORDER BY at the top level is correct, so it is window-specific. Root cause: Expr::over_with_options (polars-plan/src/dsl/mod.rs:853-861) collapses several keys into one as_struct(...), and a struct holding a null field is not itself null, so SortOptions::nulls_last has nothing to act on. descending survives the same path, which is why only null placement breaks.

sqlp_window_order_by_multiple_keys_limitation pins all of it, so the assertions fail the moment upstream fixes it. The CHANGELOG bullet is scoped to a single sort key accordingly.

Two traps worth knowing, both of which produced a test that asserted nothing until fixed

• A projection of only a nulled column writes the row as an empty line, which the test CSV reader then drops — so a TRY_CAST assertion could not tell a nulled row from a missing one. The cast fixtures carry an id column for exactly this reason.
• sqlp_join_non_equi_decimal originally joined on whole numbers, where an integer comparison gives the same answer and the test proved nothing about Decimal. It now uses fractional values where 1.05 < 1.10 is true but any integer truncation makes it false.

Out of scope, recorded so nobody re-investigates: ARRAY_INNER_PRODUCT/ARRAY_DOT_PRODUCT (#29033) needs the Array dtype, which qsv does not enable, and there is no path from CSV input to an Array-typed column through sqlp.

Order determinism

sqlp has no --maintain-order and its join output order has been nondeterministic since 6d33fb41f, so every join test orders by a unique key rather than relying on an incidental order, and each was confirmed byte-identical over 12 runs. The window tests that have peer rows project only determinate columns instead of adding a tiebreak key — see the upstream bug above for why a tiebreak is not available. APPROX_QUANTILE is sketch-based, so its values are pinned exactly only after confirming byte-identical output over 20 runs; a tolerance band against QUANTILE_CONT would have hidden numeric drift.

Verification

• Full suite 3,960 passed / 0 failed, 72 ignored
• cargo t sqlp 5× consecutively: 136/136 each run
• 13 mutation tests on the least-obvious assertions, all 13 caught
• scripts/double-run-check.py --check clean
• No new clippy warnings; cargo +nightly fmt clean
• cargo check on qsvmcp and qsvdp both pass with the new feature

No docs regeneration needed — no docopt USAGE change, so no --generate-help-md and no --update-mcp-skills.

🤖 Generated with Claude Code

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}