fix(gfql): decline undirected varlen + node alias on polars chain (honest NIE, #1741)

[lmeyerov]

Summary

Fixes the silent-wrong-results half of #1741: on engine='polars'/'polars-gpu', MATCH (a {..})-[:T*1..2]-(f) RETURN f.id returned rows pandas (the parity oracle) correctly excludes.

• Root cause: undirected multi-hop on the polars chain lacks pandas' backtrack avoidance. The node/edge sets still agree (reachability), but the named-alias flag on nodes after the undirected varlen edge tags nodes whose only path bounces back over a single edge — e.g. the seed at *1..2. Cypher RETURN projects those flags → silent wrong rows.
• Fix (interim, integrity-first): decline the affected shape — undirected non-single-hop edge + any node alias in the chain — with a NotImplementedError referencing GFQL polars chain: undirected bounded varlen (-[:T*1..k]-) silently diverges from pandas — missing backtrack avoidance includes the seed #1741, per the parity-or-honest-NIE contract. Unnamed undirected varlen chains keep native execution (their parity is pinned by a new test).
• The real backtrack-avoidance port (which also unlocks LDBC IC6/IC11 on polars) stays tracked in GFQL polars chain: undirected bounded varlen (-[:T*1..k]-) silently diverges from pandas — missing backtrack avoidance includes the seed #1741.

Tests

• TestUndirectedVarlenAliasNIE: NIE on the named shape; the exact GFQL polars chain: undirected bounded varlen (-[:T*1..k]-) silently diverges from pandas — missing backtrack avoidance includes the seed #1741 cypher repro now declines instead of returning a wrong row set (pandas oracle result pinned); unnamed undirected varlen keeps native parity.
• Full polars chain suite: 994 passed locally.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y6dQEcjdazEnzuvuwf73ZL

[lmeyerov]

Reviewer pushback accepted — my first diagnosis was wrong, and the real fix is smaller than I claimed

@lmeyerov asked why we raise instead of implementing it. Re-investigated, and the framing in this PR's description ("missing backtrack avoidance") is incorrect. Direct A/B:

• hop(undirected, hops=2) and hop(undirected, min_hops=1, max_hops=2) from seed p0 return identical node sets on pandas and polars: [p0, p1, p2, p4]. Both engines include the seed reached by bouncing back over one edge. Neither implements trail semantics in the traversal.
• The final chain _nodes also match on both engines.
• The only divergence is the node-alias flag: pandas flags f on [p1, p2, p4], polars on [p0, p1, p2, p4]. RETURN f.id projects that flag → the one bogus row.

What pandas actually does (compute/chain.py ~452-503): for a named ASTNode whose previous op is an ASTEdge, it narrows the alias with a hop-distance gate — label_mask &= hop_col.notna() & (hop_col >= min_hop) & (hop_col <= max_hop), with min/max from the edge's output_min_hops/min_hops/hops. The seed carries hop 0, fails >= 1, and loses the alias. It is not backtrack avoidance at all. _apply_node_names in the polars chain has no hop gating whatsoever — that is the entire bug.

Why it isn't a one-liner today: the gate needs per-node hop distances (label_node_hops), which the polars chain currently NIEs for undirected varlen. So the real fix is:

1. Emit per-node hop distances for undirected bounded varlen in the polars hop — the layered BFS already knows each layer's distance, it just doesn't emit the column. This also removes an existing NIE.
2. Port pandas' gate into _apply_node_names (~20 lines).
3. Drop this PR's decline; its differential test flips from "raises" to "matches the pandas oracle" and becomes the parity gate.

Proposed disposition: keep this open only as stop-the-bleeding (master currently returns silently wrong rows, which is worse than an error). If (1)+(2) land promptly, close this unmerged and ship the real fix instead. Happy to go straight to the real fix if you'd rather not carry the interim decline at all — say the word.

Full diagnosis with the A/B evidence is recorded in plans/gfql-benchmark-numbers/plan.md so it survives context loss.

[lmeyerov]

Superseded by the real fix: #1746 (native label_node_hops in the polars hop) + #1747 (the hop-distance alias gate, stacked on it). The repro now matches pandas exactly through the Cypher surface — {p1, p2} on both engines — instead of declining.

Leaving this open only until #1747 merges, since master currently returns silently wrong rows for this shape and a decline is strictly better than that. Once #1747 lands I'll close this unmerged and flip its test from "raises" to "matches the pandas oracle".

[lmeyerov]

Superseded by #1747 — will close once that lands

This PR declines (NotImplementedError) the undirected varlen + node-alias shape. #1747 now implements the pandas hop-distance alias gate instead of declining it, so that exact shape returns the correct rows (MATCH (a)-[*1..2]-(b) no longer over-includes the backtracked-to seed) — verified 4-engine against the pandas oracle. Implementing beats raising here.

Keeping this open only as a safety net until #1747 merges; it will be closed at that point. No further work planned on it.

[lmeyerov]

Closing now (not waiting for #1747 to merge): this PR touches the same two files (chain.py + test_engine_polars_chain.py) and the same undirected-varlen-alias shape as #1747, with the opposite approach — declining (NIE) what #1747 correctly implements. Two open PRs conflicting on the same lines is confusing and offers no benefit: #1747 is reviewed, cross-engine-verified (0 regressions), and CI-green. Reopen if #1747 is ever rejected. Superseded by #1747 (refs #1741).