Summary

This PR does one thing: optimize the common undirected single-hop GFQL traversal path in hop().

Scope:

• add a narrow fast path in graphistry/compute/hop.py for direction='undirected', single-hop, no-predicate/no-label traversal
• move the small typed dataframe access shims used by that path into graphistry/compute/dataframe_utils.py
• add targeted regression coverage in graphistry/tests/compute/test_hop.py
• add a matching CHANGELOG.md note

The broader RAPIDS work was split back out of this PR:

• #977: pure cuDF/cuGraph regression repro
• #978: cuGraph projection caching / reuse
• #988: DGX RAPIDS GPU matrix helper workflow

Why

The accepted optimization is real, local, and reviewable on its own. The broader benchmark, to_cugraph(), and RAPIDS compatibility work did not belong in the same PR.

Validation

Local targeted regression suite:

PYTHONPATH=. pytest -q \
  graphistry/tests/compute/test_hop.py \
  graphistry/tests/test_compute_hops.py \
  graphistry/tests/compute/test_chain.py \
  graphistry/tests/test_compute_chain.py

Result:

• 167 passed, 6 skipped

DGX GPU validation of the kept hop() change:

• gplus, RAPIDS 25.02: 2.6495s -> 1.5985s (-39.67%)
• gplus, RAPIDS 26.02: 3.6406s -> 2.2109s (-39.27%)
• twitter, RAPIDS 26.02: 0.2666s -> 0.2370s (-11.10%)