GFQL: indexed bindings bypass never engages on engine='polars-gpu' (silent loss of the index path) #1803

What

graphistry/compute/gfql/lazy/engine/polars/chain.py::_try_indexed_middle_polars deliberately passes Engine.POLARS_GPU when the GPU execution target is active:

engine = Engine.POLARS_GPU if active_target() == ExecutionTarget.GPU else Engine.POLARS
return try_indexed_connected_bindings_state(g, middle, engine=engine), True

but graphistry/compute/gfql/index/bindings.py:307 — the first gate in try_indexed_connected_bindings_state — admits only three engines:

if (
    engine not in (Engine.PANDAS, Engine.CUDF, Engine.POLARS)
    ...
):
    return None

So on engine='polars-gpu' the indexed fixed-hop bindings bypass always declines, and the query silently runs the canonical scan instead.

Evidence

Measured on the dgx GB10 box, graphistry/test-rapids-official:26.02-gfql-polars, docker run --gpus all, cuDF 26.02 / cudf_polars present. index_trace() for [n({id:1}, name='a'), e_forward(name='r'), n(name='b'), rows()] with index_policy='force' on a gfql_index_all()-ed graph:

{'operation': 'indexed_traversal', 'seam': 'connected_bindings', 'engine': 'polars-gpu',
 'served': False, 'reason': 'unsupported_engine', 'path': 'scan',
 'decision_reason': 'unsupported_engine'}

The same query on engine='polars' is served by the bypass.

Impact

Not a wrong answer — every value/parity case in graphistry/tests/compute/gfql/test_rewrite_param_discard.py passes on polars-gpu (45 passed / 2 xfailed on the GPU run). What is lost is the optimization, with no signal at all: a user who indexes a polars graph and runs it on the GPU target gets the scan path and no indication that the index was skipped. That matters because polars-gpu is one of the two engines the perf work targets.

Why it looks like an oversight rather than a decision

The polars-specific branches inside bindings.py (lines 147, 163, 172, 196, 233, 403, 453) are all spelled engine == Engine.POLARS. A polars-gpu graph carries polars frames, so those branches are the ones it would need — i.e. widening the gate to POLARS_GPU also means widening those == Engine.POLARS tests to in (Engine.POLARS, Engine.POLARS_GPU). That is a small, mechanical change, but it is a change (and needs its own GPU parity run), so it is filed rather than folded into the test PR that found it.

Pinned

test_indexed_bypass_still_serves_a_bare_rows[polars-gpu] carries pytest.mark.xfail(strict=True) referencing this issue, so whoever widens the gate is told to delete the marker rather than leaving a stale "known gap".

Found by

Adding the polars-gpu parameter to test_rewrite_param_discard.py (CB3(b) — the file previously skipped polars-gpu). First GPU run, first finding.