[adsharma]

What was broken

CI (build (ubuntu-latest, 3.12) on PR #20, also reproducible on main) fails in tests/test_polars.py::PolarsTests::test_project:

AssertionError: DataFrames are different (value mismatch for column ":id")
[left]:  [1, 4, 2]   expected
[right]: [1, 4, 3]   actual (varies run to run; also seen [1, 4, 9])

This failure is pre-existing and unrelated to PR #20 — it fails identically on main (a8cdb82).

Root cause

Nondeterministic tie order in the Polars sort, via three compounding facts:

1. Test data has ties by construction. User.get() assigns age=random.choice([16,17,18]). With the test's random.seed(100), ages are {1:16, 2:17, 3:17, 4:16, 5:18, 6:17, 7:18, 8:17, 9:17}, so where(age >= 16) keeps all 9 rows and order_by(age) has a 5-way tie at age 17. take(3) then picks [1, 4, <first 17-year-old in sort output>].
2. The builder requests an unstable sort. visit_order_by applied LazyFrame.sort(field), whose defaults are maintain_order=False, multithreaded=True — equal elements may come out in any order.
3. Proven flaky, not version-specific. Three identical local runs of the same query returned [1, 4, 2], [1, 4, 3], [1, 4, 2]. It passed on main back on Oct 5 purely by scheduling luck; polars is unpinned (>= 0.12.0), so any version/threading change reshuffles the ties.

Change

fquery/polars_builder.py: generalize the deferred-op stack entries from (func, params) to (func, params, kwargs) and pass maintain_order=True for sorts, so ties keep input order — matching the in-memory executor's behavior.

Verification

• tests/test_polars.py: 6/6 consecutive runs pass (previously failed intermittently).
• Full suite: 59 passed (baseline on main: 58 passed + the 1 polars failure).
• black + flake8 clean on the touched file.