feat(gfql): polars-gpu = GPU target of the lazy Polars engine (collect-once) [redo of #1654]

[lmeyerov]

Summary

GPU execution target of the lazy Polars engine (engine='polars-gpu') — the redo of the per-op GPU PR (#1654, which was a perf regression) as a thin target of the lazy engine on #1648. engine='polars' and engine='polars-gpu' are now one lazy engine, two targets: build a single deferred pl.LazyFrame plan per single-hop and collect_all it once on CPU or GPU.

Why this replaces #1654

Benchmark showed per-op eager GPU collect was a regression (each op re-transfers tables H2D). collect-once is the fix: single-hop GPU is now a 2.84× win @1m (vs eager) with CPU parity. The win flows from the lazy engine on #1648; this PR just selects the GPU target.

Design (contained)

• Engine.POLARS_GPU='polars-gpu', explicit opt-in only (AUTO never selects it); frames stay pl.DataFrame (treated like POLARS in all frame ops). Extends POLARS_ENGINES (introduced on feat(gfql): native lazy Polars engine — collect-once traversals + cypher row pipeline #1648) to include the GPU target — so every engine-aware helper covers it for free.
• Dispatch wraps the lazy hop/chain in target_mode(GPU) (the lazy/ framework's collect/collect_all already run on the active target). engine='polars' (CPU) is byte-for-byte unchanged.
• raise_on_fail=False — GPU-incapable nodes stay on CPU in Polars (no pandas bridge; NO-CHEATING). Uses the cudf-polars in-memory executor (executor="in-memory") — faster + more stable than the default streaming engine="gpu" for in-device-memory GFQL results.

Also here: opt-in CPU streaming collect

GFQL_POLARS_CPU_STREAMING=1 runs the polars-CPU lazy collects on the streaming executor (~1.04–1.11× faster on large multi-hop traversals, parity-identical), default off because small/interactive sizes regress (~0.86× from streaming overhead). No change to default behavior.

Honest scope

Single-hop GPU wins; the chain-level GPU win currently dilutes (a chain runs forward+backward = 2 hop collects + eager _combine_*). Fusing those + moving combine onto the target is the next benchmark-driven opt on this PR.

Validation (dgx, RAPIDS --gpus all)

• Parity engine='polars-gpu' == engine='polars': test_engine_polars_gpu.py (skips without cudf_polars).
• Full graphistry/tests/compute/gfql/ polars suite green; CPU engine='polars' is byte-for-byte unchanged from feat(gfql): native lazy Polars engine — collect-once traversals + cypher row pipeline #1648 (the GPU target adds only the 2 GPU/streaming commits on top).

Stacks on #1648 (lazy Polars engine) → #1652 (general opts) → master. Supersedes #1654 (per-op GPU).

⚠️ This branch was force-rewritten on 2026-06-28 (restack moved the CPU conformance fixes down to #1648). If you have it checked out: git fetch && git reset --hard origin/dev/gfql-lazy-gpu.

🤖 Generated with Claude Code