GFQL row pipeline: cuDF host-bridge gap in order_by + test placement debt for compute/gfql/row/ (follow-ups from #1362/#1359) #1373

Two related follow-ups surfaced while auditing PR #1362 / #1359 (Comparator + ORDER BY semantics, meta #1353 item #1). Both are non-blocking for #1362 (already merged-ready) but warrant separate work.

1. cuDF host-bridge gap in order_by list-path

While auditing #1362's stringified-list ORDER BY surface for cuDF compatibility, found a latent host-bridge protocol violation at graphistry/compute/gfql/row/pipeline.py:3964.

Setup. parse_stringified_list_series (graphistry/compute/gfql/row/ordering.py:188-219) is intentionally pandas-only — it ast.literal_evals string-encoded lists, which has no cuDF analog. Its docstring documents this and points at the canonical pattern at _gfql_eval_dynamic_list_subscript (pipeline.py:2369-2401).

The gap. The canonical pattern has two halves:

• Producer (pipeline.py:2394-2401): emits a pd.Series.
• Consumer (pipeline.py:2424-2437): detects pd.Series and routes around cudf_df.assign(...) via a fresh pd.DataFrame build.

PR #1362's order_by callsite (pipeline.py:3955-3967) implements the producer half but does work_df = work_df.assign(**{sort_col: parsed}) without the consumer-side to_pandas() bridge that select uses at pipeline.py:3651-3658:

if resolve_engine(EngineAbstract.AUTO, table_df) == Engine.CUDF and any(
    isinstance(value, pd.Series) for value in projected.values()
):
    out_table_df = table_df.to_pandas()
    ...

Reachability today. Unreachable in production because the only column dtype that can carry ast.literal_eval-able list strings is pandas string/object, and current upstream gating ensures work_df is already pandas by the time order_by runs. The bug is not a regression in #1362 and not blocking. It is a latent cuDF break that surfaces if either (a) cuDF general object dtype lands or (b) upstream gating regresses.

Coverage gap. Zero _on_cudf tests cover the list / stringified-list / list-comprehension-key ORDER BY surface (52 _on_cudf tests in test_lowering.py, none on this surface). The 8 pin tests at test_lowering.py:7721-7972 are pandas-only by _mk_graph(pd.DataFrame(...)).

Other latent issues found in the same audit (separate, deferrable):

• ordering.py:306 — only .pivot( call in the whole repo; no established cuDF-safety pattern. Replace with groupby+merge idiom.
• ordering.py:289 — value_str.where(num_mask, None).astype(\"float64\") is cuDF-fragile (None-fill then float cast).
• ordering.py:317 — key_frame[col] = key_frame[col].fillna(\"\") is a pure-functional checklist violation per agents/skills/review/SKILL.md.

Proposed work:

1. Mirror the select consumer-side host-bridge at pipeline.py:3964 (single-line guard).
2. Add at least one _on_cudf pin test that covers stringified-list ORDER BY on a cuDF input (asserting clean raise OR success, whichever matches expected behavior).
3. Tighten parse_stringified_list_series docstring to document the caller contract.
4. (Optional) refactor .pivot( away from build_list_sort_columns:306.
5. (Optional) where(_, None).astype(\"float64\") cuDF fragility.
6. (Optional) key_frame[col] = ... mutation fix.

2. Test placement debt: compute/gfql/row/ has no test mirror

PR #1362 added 6 new tests at graphistry/tests/compute/gfql/cypher/test_lowering.py:7781-7972 exercising the cypher → gfql → row-pipeline integration path, but the actual unit-of-change is in graphistry/compute/gfql/row/{ordering,pipeline}.py. These belong in row-pipeline unit tests.

Layout asymmetry

graphistry/tests/compute/gfql/row/ directory does not exist. Source side has 7 files under row/; test side has zero mirror.

What PR #1362 did

Why this matters

1. Wrong abstraction layer. Bug class is "row pipeline mishandles stringified lists." Cypher → gfql lowering is unaffected. Testing through cypher adds parse + lower + project noise to every assertion.
2. Test file bloat. test_lowering.py is 14,732 LOC; test_row_pipeline_ops.py is 2615 LOC. A tests/compute/gfql/row/ subdir splitting per-file would scale better.
3. Discoverability. A reviewer scanning graphistry/compute/gfql/row/ordering.py for tests has no parallel tests/compute/gfql/row/ to look in.
4. Coupling to cypher. Row-pipeline tests written through cypher implicitly couple to cypher parser shape; tightening cypher syntax later breaks unrelated row-pipeline pins.

Recommended scope

• Phase A (small): relocate the 5 useful tests from test_lowering.py:7781-7972 to test_row_pipeline_ops.py (drop the redundant Python-list one); keep one thin cypher-integration smoke test.
• Phase B (medium): create tests/compute/gfql/row/ subdirectory; split test_row_pipeline_ops.py's 2615 LOC into per-source-file mirrors.
• Phase C (larger): sweep test_lowering.py for other tests that are really row-pipeline tests and relocate.

Cross-refs

• Audit details: plans/1359-orderby-wrong-rows/audits/cudf-flow-audit.md (local).
• Related: GFQL/Cypher: drive down WITH ORDER BY wrong-row tranche (14 cases, #1353 item #1) #1359 (semantic fix, pandas), Meta: Prioritized direct-Cypher wrong-row bugfix tranches (92 cases) #1353 (meta), GFQL/Cypher: ORDER BY on stringified-list columns uses Cypher list-orderability (#1359) #1362 (PR), Vectorized Cypher row-pipeline semantics (pay-as-you-go, cuDF-parity): WITH/UNWIND/ORDER/SKIP/LIMIT #1349 (cuDF-parity for WITH/UNWIND/ORDER/SKIP/LIMIT — natural meta).

Severity buckets

• GFQL/Cypher: ORDER BY on stringified-list columns uses Cypher list-orderability (#1359) #1362 BLOCKER: none.
• IMPORTANT follow-up: (1.1) order_by host-bridge guard at pipeline.py:3964; (1.2) _on_cudf pin test for stringified-list ORDER BY; (2.A) test relocation.
• SUGGESTION: all other audit items.

Status (2026-05-09)

• Progress landed via PR [CHORE] GFQL row tests: relocate #1362 ordering cases into row mirror (#1377) #1383 (merged 2026-05-09): [CHORE] GFQL row tests: relocate #1362 ordering cases into row mirror (#1377) #1383
• Item 2 (test-placement debt) is now partially addressed:
  • graphistry/tests/compute/gfql/row/ now exists
  • new row/test_ordering.py added
  • key ordering/list-string semantics moved from cypher integration to row tests
  • test_lowering.py retains one thin integration smoke
• Remaining scope in this meta-issue is still open, including item 1 (cuDF host-bridge gap) and broader row-module decomposition beyond ordering.