[lmeyerov]

Summary

• Preserve engine-aware row-prefix slicing so pandas does not scalarize whole-node reentry prefixes.
• Correlate hidden WITH scalar carries through surviving Polars edge steps, including multi-edge and forward/reverse/undirected traversal.
• Keep valid hidden reentry property access out of destination-alias masking and handle empty Polars prefixes.
• Cover pandas/Polars positive, partial, empty, multi-edge, convergence, multiplicity, no-reentry, and typed-scalar cases. Add a direct cuDF convergent positive/partial-rejection test.

Validation

• Focused CPU issue_2252 or issue_1413 or reentry selection: 141 passed, 1,591 deselected.
• Ruff passed. Type-hygiene guard: 4,439 grandfathered findings, no growth. Local mypy 1.18.2 reports the same 15 baseline diagnostics on d53 and this head; hosted CI typing lanes pass.
• Exact-head CI run 40901849902 completed SUCCESS: all 90 jobs are terminal, with no failed conclusions. CodeQL run 40901850068 and RTD build 35501473 are SUCCESS.
• Paired F9 TCK on exact e131: all three direct-Cypher cases pass on pandas and Polars in normal mode and with all six routes off. The same three pass under hosted-like pandas 3.1/NumPy 2.5. Exact product master a649 is red on the two nonempty cases as expected; the empty case passes.
• TCK [BUG] memoization fails on df's with lists #212's full hosted CI currently fails earlier on the independent pandas 3.1 nullable-ID routes-off issue tracked by #2254. This does not fail the paired scalar-carry cases on e131. TCK [BUG] memoization fails on df's with lists #212 remains a draft.
• F10: candidate-only pandas/Polars boundary runs pass positive, partial, and empty row oracles with traced same_path reentry and carry dispatch. The same-output control used 3 warmups/7 samples; pooled medians were 127.983/127.978 ms for positive and 90.988/95.610 ms for empty (base/candidate). This is bounded context, not a speed or statistical-equivalence claim.
• cuDF remains unverified: the focused local test skips without CUDA. DGX 26.02 admitted the exact-source test under an 8 GiB RMM guard, then skipped after a 24-byte std::bad_alloc while foreign VLLM::EngineCore held 60,103 MiB. No cuDF parity claim.

Identity and review state

• Base: a649e05f20a2457a024722b78b6f99d3ab0aeb46
• Head: e131383a34cda7985db223440c25a32203d05081
• Tree: 6cefa818e73d92700242d3ed1fa1d75d520b047b
• Planner wave6 review: F9 paired TCK and F10 evidence accepted for product review. Actual cuDF execution and final planner review remain open gates; this PR stays open and unmerged.