gfql polars: the plain single-hop chain branch ignores prune_to_endpoints (pandas keeps only the arrival side)

[lmeyerov]

Summary

Route-level sibling found while hoisting the chain admission predicates: the pandas/cuDF chain fast path declines prune_to_endpoints=True ("prune keeps only the arrival side → full path"), but the polars chain's plain single-hop branch admits the shape and returns both endpoints plus the edges, so polars and pandas disagree.

Repro

import pandas as pd, polars as pl, graphistry
from graphistry.compute.ast import n, e_forward
nodes = pd.DataFrame({"key": [1, 2, 3, 4, 5], "id": [10, 20, 30, 40, 50]})
edges = pd.DataFrame({"s": [1, 1, 2, 3, 3, 4], "d": [2, 3, 3, 1, 1, 5], "eid": range(6)})
ops = [n({"key": 1}), e_forward(prune_to_endpoints=True), n()]
graphistry.nodes(nodes, "key").edges(edges, "s", "d", "eid").gfql(ops)                       # nodes [2, 3], no edges
graphistry.nodes(pl.from_pandas(nodes), "key").edges(pl.from_pandas(edges), "s", "d", "eid").gfql(ops, engine="polars")  # nodes [1, 2, 3], edges [0, 1]

Unseeded [n(), e_forward(prune_to_endpoints=True), n()]: pandas [1, 2, 3, 5], polars [1, 2, 3, 4, 5].

Expected

Either the polars plain branch declines prune_to_endpoints like the pandas gate (polars_plain_single_hop_admits is the single place to do it once the predicate refactor lands) or it implements the arrival-side contract; both engines must agree.

Pins

A strict-xfail parity pin lives in graphistry/tests/compute/gfql/lazy/engine/polars/test_chain_admission.py over the shared route corpus and flips when this is fixed.

[lmeyerov]

Scope update from the route harness (#2061): the divergence is not only the plain single-hop branch. With every hot path declined (GFQL_ROUTES_OFF=all, pygraphistry's test-side route switch), the polars general traversal returns the same rows as the plain branch, so the polars engine as a whole ignores prune_to_endpoints while pandas keeps only the arrival side.

Probe on the issue's frames ([n({"key": 1}), e_forward(prune_to_endpoints=True), n()]):

path	nodes	edges
pandas lane	[2, 3]	[]
pandas general	[2, 3]	[]
polars lane	[1, 2, 3]	[(1,2), (1,3)]
polars general	[1, 2, 3]	[(1,2), (1,3)]

Unseeded [n(), e_forward(prune_to_endpoints=True), n()]: pandas general [1, 2, 3, 5], polars general [1, 2, 3, 4, 5] (plus the (4,5) edge).

So declining the shape in polars_plain_single_hop_admits would only move the wrong answer from the lane to the full path. The fix is the arrival-side contract in the polars chain itself (or a typed decline until it exists). The harness pins this as a strict expected failure on polars-plain keyed by tag #2053 (routes corpus + the pandas _BYPASS_SHAPES prune rows), so it flips when fixed.