test(gfql): run the #1788/#1790 suites on all four engines, in a lane that exists

[lmeyerov]

CB3 (a) and (b): the two test files the owner flagged, both engine-parametrized properly and both now running in a lane that exists.

What was wrong

file	before
test_rows_table_named_middle.py (#1788)	pandas + polars only; module-level pytest.importorskip("polars"); not in bin/test-polars.sh → the test-gfql-core lane installs no polars, so the whole file skipped there and ran in no CI lane at all
test_rewrite_param_discard.py (#1790)	named cuDF but skipped polars-gpu outright; also absent from bin/test-polars.sh, so every polars assertion in it was dead in CI

Both files pin which table a query returns — engine-agnostic semantics, not a polars question.

What this changes

Fixed engine list (pandas, cudf, polars, polars-gpu) — deliberately not available_nonpandas_engines(), which builds its list by importability so a missing engine vanishes from the report instead of showing as SKIPPED. Both files added to bin/test-polars.sh; the module-level importorskip is gone, which also makes 5 pandas cases live in test-gfql-core.

Two probe traps, both hit by trying rather than reasoned about

1. A probe that runs the shape under test disarms its own file. My first version smoke-tested with rows(table='edges'). Reverting the #1788 table guard then made the probe raise, and all 20 parameters reported SKIPPED instead of failing — a regression disguised as a missing environment. The smoke query is now a plain traversal with no rows() in it.

2. A probe that swallows every exception is worse than no probe. A transient MemoryError: std::bad_alloc: out_of_memory: CUDA error (failed to allocate 32 bytes) … cudaErrorMemoryAllocation in a fresh GPU container silently dropped cuDF from a run that otherwise passed. A skipped GPU parameter reads as evidence of passing. engine_skip_reason (new, in polars_test_utils.py) now classifies: a missing module skips, a recognisable GPU-stack error skips with its text quoted in the skip reason, and any other failure propagates.

GPU receipts

dgx GB10 · graphistry/test-rapids-official:26.02-gfql-polars · docker run --gpus all · cuDF 26.02.01 / polars 1.35.2:

45 passed, 3 xfailed, 0 skipped        (repeated — all four engines running)

Mutation-checked on the same box, reverting the #1788 and #1790 table guards:

24 failed, 21 passed, 3 xfailed, 0 skipped
   6 [pandas]   6 [cudf]   6 [polars]   6 [polars-gpu]

Evenly spread across all four engines, so the added parameters are not decorative.

Two real gaps the new parameters surfaced — pinned, not papered over

• GFQL: indexed bindings bypass never engages on engine='polars-gpu' (silent loss of the index path) #1803 — the indexed bindings bypass gate (index/bindings.py:307) admits only (PANDAS, CUDF, POLARS), while the native polars chain hands it Engine.POLARS_GPU whenever the GPU target is active. polars-gpu therefore always reports served: False, reason: 'unsupported_engine' and silently runs the canonical scan. Values are unaffected; the optimization is lost with no signal. Strict xfail.
• GFQL: native polars bindings builder ignores rows(alias_prefilters=...) — 5 rows on pandas/cuDF vs 12 on polars #1804 — the native polars bindings builder never receives alias_prefilters: 5 rows on pandas/cuDF vs 12 on polars for the same query. It was already xfailed before this change, but with no issue number, so there was nothing to close and nothing to find. Now filed and referenced from the marker.

Strict xfails in both cases, so whoever fixes either is told to delete the marker rather than leaving a stale "known gap" behind.

Runtime delta: zero

No production line changed — the diff is two test files, one shared test helper, bin/test-polars.sh and the CHANGELOG. Per CB5 no pyg-bench lane run is required, and that claim is checkable from the diff rather than merely asserted.

Note for the reviewer

bin/test-polars.sh is also touched by #1801 and by the CA4 allowlist PR; the edits are to different lines of the same array and are independently mergeable.

🤖 Generated with Claude Code

https://claude.ai/code/session_015YsqAZQLbqjSDrYSFz2GoB

[lmeyerov]

Overlap note for whoever merges this. #1805 (fix/gfql-polars-lane-invisible-tests, opened concurrently) also adds test_rows_table_named_middle.py and test_rewrite_param_discard.py to bin/test-polars.sh, and restructures that array — so the two will conflict textually even though they agree on the outcome.

Left in here deliberately rather than removed: this PR's whole point is that these two files' non-pandas parameters actually execute, and dropping the lane lines would make that contingent on a PR that has not landed. If #1805 goes first, resolve by keeping #1805's array and dropping the two lines from this diff — nothing else here touches that file.

The substance of this PR (the four-engine parametrization, the classified availability check, the GPU receipts, and the two strict xfails for #1803/#1804) is independent of #1805 and does not overlap it.

Also worth flagging against #1797 (the polars coverage cell sitting at its timeout ceiling): this PR's additions measured ~15s per lane run locally, and CI came back green on all 77 checks including every test-polars matrix cell — so it did not tip that cell over. That is an observation from one run, not a guarantee.

[lmeyerov]

static typing

[lmeyerov]

Merged master (a10173575) in as f6fa7fb7e. Both conflicts resolved, and the conflict exposed two defects in this branch that I fixed rather than carried forward.

Defect 1 — bin/test-polars.sh would have run two modules twice. This branch adds test_rows_table_named_middle.py and test_rewrite_param_discard.py to POLARS_TEST_FILES. But #1805 — which is this branch's merge-base (49db91cc9) — already added both, at lines 43 and 50. So the addition was redundant at the moment it was written, and taking the branch side would have listed each path twice and run each module twice. That lane's py3.12 cell is the coverage cell and has been up against its budget: 442s on master's latest green run, cap 600s. Resolution is master's file verbatim, which also keeps #1814's xdist plumbing that the branch side predates. bin/test-polars.sh no longer appears in this PR's diff at all.

Defect 2 — a false claim in the CHANGELOG. The entry said "neither file was in bin/test-polars.sh, so test_rows_table_named_middle.py … ran in no CI lane at all". Same root cause: both files were listed at the base, so the polars params did run. Rewritten to the residual that is actually true — the suites were parametrized over pandas + polars only, so cuDF and polars-gpu never exercised the table guards these suites exist to pin; and the module-level pytest.importorskip("polars") skipped the whole file in test-gfql-core, so its pandas cases ran only in the polars lane. Every measured claim is untouched (45 passed / 3 xfailed / 0 skipped on GB10; 24 mutation failures, 6 per engine).

Typing on engine_skip_reason, per review. The engine → required-modules map was a dict literal rebuilt inline on every call inside the for; it is now a module-level _ENGINE_REQUIRED_MODULES: Mapping[str, Tuple[str, ...]]. Signature is (engine: str, smoke: Callable[[], object]) -> Optional[str]. gpu_environment_reason's "str | None" string annotation is normalized to Optional[str] to match — this module is imported under the py3.8 lanes, so the deferred-annotation form was load-bearing, not stylistic. flake8 and mypy 1.18.2 clean under the repo config.

Net diff vs master is now CHANGELOG + the three files under graphistry/tests/compute/gfql/ — no production line and no CI-manifest line changed.