fix(gfql): treat binding seed ids as identities - #2001

Summary

• treat node ids, not duplicate physical node-table rows, as identities when generic binding traversal starts
• preserve first-encounter order in pandas/cuDF and Polars seed state
• keep relationship rows as the source of MATCH bag multiplicity, including genuine parallel edges

Correctness oracle

The fixture duplicates storage rows for node ids 1 and 2 but adds no graph identities or relationships. Its hand-derived grouped counts are LA=6, NY=2, SF=1; adding one parallel 1 -> 5 relationship changes only LA to 7. Before this fix the generic route returned 10/3/1 and 12/3/1 by multiplying paths from duplicate seed rows.

Semantics boundary

GQL-conformant Cypher preserves binding-row duplicates for plain RETURN / RETURN ALL; only DISTINCT deduplicates (Neo4j Cypher RETURN). This change follows that boundary:

• duplicate physical node-table rows with the same node id collapse only at the initial identity wavefront
• relationship match rows are never collapsed by this change
• a genuine parallel edge therefore remains an additional match row and changes LA from 6 to 7

Validation

• exact reviewed head: a888c33f2d42f626bb0754fb1186b2d1c1bf6561
• topology: four single-parent commits directly on master@d509e428b0b63a7dd6d92ee5271e75afd15ca2bf; zero merge commits
• incremental diff: four intended files, +94/-6; whitespace clean
• prohibited added-line audit: zero Any, object, cast, getattr, setattr, hasattr, or unqualified List/Dict; finite engine values use Literal
• focused safe CPU regression: 4 passed / 4 GPU cells deliberately deselected
• ./bin/lint.sh: syntax, Ruff, type-hygiene, comment-density, and relative-import guards clean
• ./bin/typecheck.sh: mypy 2.3.1, 336 source files, zero issues
• no Docker, local GPU, or broad local suite; the current GitHub Checks panel is authoritative for the full matrix

Review resolutions

• Stacking: the stale predecessor merges were removed. The branch is now four linear commits directly on landed fix(gfql): preserve bag multiplicity for whole-entity projections #2000 master.
• Positive/negative boundary: duplicate physical seed rows are the positive case that must collapse; a real parallel relationship row is the negative control that must remain multiplicative. Both cases run across the engine matrix.
• Anti-vacuity without dynamic patching: the regression uses semantically neutral SKIP 0 to select the generic route and assert_fast_path(..., served=False) to prove the grouped fast path was consulted and declined.

Closes #1996

Reviewer handoff

Please review the incremental four-file diff for:

1. seed-state deduplication is by node id and preserves first encounter order
2. relationship rows, including parallel edges, still determine MATCH bag multiplicity
3. the regression proves the generic route without monkeypatch.setattr or another dynamic-attribute escape

Stack / landing

• predecessor fix(gfql): preserve bag multiplicity for whole-entity projections #2000 is landed at d509e428b0b63a7dd6d92ee5271e75afd15ca2bf
• this PR is based directly on master
• successor fix(gfql): preserve boolean sum zero on polars gpu #2002 must be rebased onto the resulting master after this PR lands
• land only the exact green head using GitHub Create a merge commit; do not delete the branch while successors are being repaired