polars chain: forward/reverse varlen with min_hops>1 returns nodes outside the hop window (silent wrong)

[lmeyerov]

Summary

On the polars chain engine, a forward or reverse variable-length pattern with min_hops > 1 returns nodes whose hop distance is outside the requested window. No error is raised — the answer is silently wrong.

Undirected *2..3 correctly declines with a NotImplementedError. Only fwd/rev silently mis-answer.

Repro

import pandas as pd, polars as pl, graphistry
e = pd.DataFrame({"s": ["p0","p1","p2","p1"], "d": ["p1","p2","p4","p0"]})
q = "MATCH (a {id:'p0'})-[*2..3]->(b) RETURN b"

graphistry.edges(e, "s", "d").gfql(q, engine="pandas")
# -> p2, p4                      (correct: distances 2 and 3)

graphistry.edges(pl.from_pandas(e), "s", "d").gfql(q, engine="polars")
# -> p0, p1, p2, p4              (WRONG: p0 is distance 0, p1 is distance 1)

Cause

The polars chain gates a node alias by hop distance only when the auto hop-label column is present. For min_hops > 1 the label is deliberately not requested (those labels come from pandas' layered backward walk, which is not ported), so the alias ends up ungated and every reachable node passes.

Undirected happens to escape because it declines the whole shape earlier.

Scope — measured, not inferred

4-engine differential sweep against pandas as oracle, 144 cases (4 hand graphs + 12 random graphs x 9 patterns), run on dgx-spark under safe_run.sh, RAPIDS 26.02 image.

build	pandas	cudf	polars	polars-gpu
master c355660a	144/144	144/144	79 ok / 49 wrong / 16 NIE	79 / 49 / 16
master + #1746 + #1747	144/144	144/144	114 ok / 14 wrong / 16 NIE	114 / 14 / 16

Set-diff of the mismatch sets: 0 regressions, 70 of 98 fixed. All 28 residual mismatches (14 per polars engine) are this one shape, -[*2..3]->.

So this bug is pre-existing on master and is not introduced by the #1741 stack — that stack removes 71% of the silent-wrong surface and leaves exactly this hole.

Fix direction

Either port the layered backward walk that produces min_hops > 1 labels, or — cheaper and strictly better than today — decline (NIE) the fwd/rev min_hops > 1 + node-alias shape the same way undirected already does, so a wrong answer becomes an honest error.

Related: #1741, #1746, #1747.

[lmeyerov]

Handled in #1747 by honest decline (not silent-wrong). Rather than leave this shape returning wrong rows, #1747 now declines fwd/rev min_hops>1 + node-alias with a NotImplementedError pointing here (adapted from the retired #1742 decline pattern). Differential vs pandas: 30/30 named shapes NIE, 30/30 unnamed run and match pandas, 0 silent-wrong, 0 over-decline; 4-engine sweep now 112 agree / 0 disagree / 32 NIE (was 114/14/16). This issue stays open for the full fix — porting pandas' layered backward walk so min_hops>1 aliases are hop-gated and run natively — but the silent-wrong severity is resolved.