fix(gfql): polars chain re-executes steps on the graph's original edge columns (#2039)

[lmeyerov]

[n({"id": 30}, name="m"), e_forward({"type": "HAS_CREATOR"}, name="type"), n(name="p")] raised GFQLSchemaError [incompatible-column-type] on polars while pandas served it: the forward step stamps the alias marker type onto its edge frame, and the backward pass (g_step_full) and the pruned re-execution (g_sub) fed that stamped frame back into the hop, whose edge_match then compared the boolean marker to 'HAS_CREATOR'. Both sites now take the step's edge rows joined back to the graph's original edge columns (semi-join on the chain's edge id), so re-execution sees the user values; the combine's marker placement is unchanged.

Pins: test_polars_alias_column_collision_2039.py (edge-alias collision, node+edge collision, node-only collision: same rows on pandas and polars under policy off and use; a 2-hop with the colliding alias is served), and the #1911 alias-scoping suite unchanged except test_rows_route_edge_alias_colliding_with_its_own_type_filter, which keeps pinning polars' rows-route decline on that shape (now a typed GFQLValidationError from the rows route rather than the chain) — that residual stays open on #2039. Local: polars chain/hop/narrow-combine/conformance/alias-scoping suites 2031 passed. Real-GPU (polars-gpu) run to follow on dgx.

Fixes the native op-list half of #2039.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA

[lmeyerov]

Real-GPU receipt (dgx-spark, RAPIDS 26.02 image, cudf 26.02.01 / polars 1.35.2): test_polars_alias_column_collision_2039.py + test_engine_polars_gpu.py + polars chain/hop + alias-scoping suites at 662e557 = 1508 passed, 0 failed, 0 skipped.

[lmeyerov]

READY (662e557): 77/77; polars re-executes steps on the graph's original edge columns (#2039); real-GPU 1508 passed / 0 skipped; the rows-route residual stays pinned and open on #2039.

[lmeyerov]

Review follow-up (owner: "other specializations of the same bug" + mirror-path tests), head 0de93ce.

Probe. 22 alias/column collision shapes × pandas/cuDF/polars × index policy off/use against the pandas full path (edge alias = filtered / unfiltered / src / dst / edge-id column; node alias = filtered column on the seed or the destination, node-id binding, an edge column name; hops=2; to_fixed_point; unseeded; step 2 of 2; all three colliding). Three more specializations surfaced:

1. polars node side — fixed here. A destination alias that names its own filtered column (n({"type": "p"}, name="type")) raised the same incompatible-column-type, and the final alias tagging joined the marker in as type_right with nulls where pandas replaces the column. The node step now filters the graph's own values for that column and the marker replaces a colliding column (marker-authoritative, as pandas' combine).
2. multi-hop edge alias (every engine) — gfql: an edge alias that collides with the column its own filter uses breaks multi-hop steps (pandas/cuDF raise, polars mis-names) #2049. hops=2 / to_fixed_point with an edge alias = the filtered column: pandas and cuDF raise, polars emits type_right. Separate PR.
3. alias = source / destination / edge-id binding column — gfql: an alias named like the source, destination or edge-id binding column is not rejected (silent clobber on pandas/cuDF, raw polars SchemaError) #2050. pandas/cuDF silently clobber the binding (the seed's edges vanish, or the edge id becomes True); polars raises a raw SchemaError. Only the node-id binding is validated today. Separate PR (validator).

Pins. Tests moved to the mirror path graphistry/tests/compute/gfql/lazy/engine/polars/test_chain_alias_column_collision_2039.py; new cross-engine matrix graphistry/tests/compute/test_chain_alias_column_collision.py (12 served shapes × 3 engines × scan/indexed = every single-hop collision matches the pandas full path; strict xfails for #2049 and #2050 that flip when they land; node-binding rejection pinned as the positive control). Local: 696 passed / 15 xfailed across the polars suites, cuDF 25.10 included; mypy unchanged; ruff clean. Still stacked on #2045; CI running.

[lmeyerov]

READY (04ff8ed) — rebased onto master after #2045 landed; CI 77/77 check-runs success (the earlier red was the pre-move #2020 pin path the union rebase re-added to the polars lane list; dropped). Content unchanged since the review follow-up: polars node-side collision fix + cross-engine collision matrix + mirror-path tests.

[lmeyerov]

note: #2049 goes into the eager case, and we'll separately handle alt specializations across recent fixes