[lmeyerov]

Summary

• Preserve scalar carries beside collect(...) across the bounded-reentry collect/unwind rewrite, including plain, aggregate, and property-name-collision projections.
• Preserve native empty-frame schemas for empty prefixes and global empty aggregates on Polars, including avg and max identity rows.
• Keep post-reentry UNWIND behavior and add pandas, Polars, and cuDF regression coverage.

Scope

• Base: master @ 4d627b237d13cbe8326ce1b3bf624f356126db57 (includes fix(gfql): preserve scalar reentry across continuation matches #2253 and GFQL: preserve nullable 64-bit IDs in pandas fallback #2255).
• Head: 42d7667902c0fdf74f547f25f97b9f44274f80b6.
• The non-CHANGELOG product blobs are byte-identical to accepted head be7135fa2c5efd519db30ef3e2080512b553160d; the amendment adds cuDF parameters and avg/max cases.

Evidence

• Focused pandas/Polars matrix: 42 passed, 21 cuDF skipped locally for unavailable CUDA.
• Ruff and git diff --check: pass.
• Exact-head CI/RTD: polling at this SHA.
• Focused cuDF GPU gate: prepared for 50 tests (25 new and 25 existing reentry/carry cuDF IDs), with skip=FAIL; it will run once after exact-head CI/RTD is terminal.
• Earlier full-cuDF gate at be7135fa showed 6 inherited failures, confirmed against master 6a4ee1f0 and recorded in GFQL Cypher cuDF: 6 test_lowering cuDF cases fail on master (list/dynamic expressions raise host-conversion TypeError; cugraph write flips edge orientation) #2269.

Interaction

• Rebased after GFQL: preserve nullable 64-bit IDs in pandas fallback #2255 landed; only the CHANGELOG conflict was resolved, preserving both entries.

Fixes #2267