fix(gfql): multi-hop alias/column collisions are served on pandas and cuDF (#2049)

[lmeyerov]

Stacked on #2046 (retargets to master when it lands). Closes #2049.

Defect. The eager half of the #2039 class: for a non-simple step (hops=2, to_fixed_point=True) whose edge alias names the column its own edge_match filters on, the chain's backward pass re-executed the reversed hop on the forward step's edges, where the alias marker had already replaced that column, so filter_edges_by_dict saw booleans and raised incompatible-column-type on pandas and cuDF. Single-hop steps never hit it (the fast backward branch filters endpoints without re-applying edge_match), which is why #2046 was polars-only.

Fix. _step_with_source_edge_columns (chain.py): when the step's alias collides with an edge column, the backward re-execution runs on the graph's own edge rows for the step's edge ids. Cost-neutral otherwise (only engages on a collision). The rows now match the same chain with a non-colliding alias and match polars.

Not changed, on purpose. Polars serves these shapes already; its colliding column keeps the user values with the marker beside it, which test_alias_scoping_semantics.py pins as the polars alias-scoping contract. #2049 is rescoped to the eager engines accordingly; the collision matrix's multi-hop pins compare rows, not column contents.

Pins. graphistry/tests/compute/test_chain_alias_column_collision.py: the two multi-hop shapes × pandas/cuDF/polars flip from strict expected failure to green (parity with the control alias). Existing alias-scoping, chain, hop, residual-polars and native-seed suites unchanged.

Receipts (local RTX 3080 Ti, cudf 25.10, polars 1.42): 606 passed / 9 xfailed across the collision matrix, chain, hop, alias-scoping, residual-polars, hop-scaling and native-seed suites with the cuDF lane on; comment-density and type-hygiene guards OK; mypy on chain.py identical to master (5 pre-existing); ruff clean. The dgx GPU lane runs after the step-7 measurement window closes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA

[lmeyerov]

READY (3312f00, on master b8f3e99) — CI 69/69 non-skipped check-runs success; the 8 skipped lanes are the gfql path filter (chain.py is outside it; #2057 proposes the fix). tck-gfql run locally from its current main (66fee10, the ref CI would use) against this head: exit 0 — 3658 scenarios, 2964 supported, 1977 translated-supported, 689 xfail (baseline), direct-Cypher non-validation debt 0, no outcome drift. Local suites: 606 passed / 9 xfailed across collision matrix, chain, hop, alias-scoping, residual-polars, hop-scaling and native-seed with the cuDF lane; guards OK; mypy = master; ruff clean.