fix(gfql): release-blocker canvas — stale memo, polars-gpu target, LazyFrame crash, divergence pins

[lmeyerov]

Stacked on #1873 (next in the line). The owner-directed issue canvas verified 13 open issues empirically at the stack head; this PR fixes the three confirmed blockers and pins the two deferred divergences:

• GFQL: two-hop count returns a STALE answer after in-place frame mutation (degree-count memo setattrs onto the caller's Plottable, keyed by id() — the BLOCKER-1 pattern the same file forbids) #1825 (fixed): two-hop count(*) returned a STALE answer after in-place frame mutation — the equal-domain degree-count memo setattr'd onto the caller keyed by id() (the BLOCKER-1 pattern). Removed, not re-keyed: per-execution it never hits, and cross-call reuse is the fingerprint-validated degree facts' job. String-id mutation pin (the int-id repro is masked by the dense kernel), repeat==one-shot identity.
• GFQL: Cypher OLAP fast paths silently run on CPU under engine='polars-gpu' (bare LazyFrame.collect() bypasses the GPU-or-error contract) #1824 (fixed): fast paths silently collected on CPU under engine='polars-gpu' (they run before the chain route, the only place the GPU target was set). Dispatch now sets the target; all collects go through lazy.collect (GPU-or-error); a non-GPU-executable plan is an honest recorded decline. Source pin forbids bare collects. Needs the dgx GPU run before merge.
• GFQL polars: LazyFrame node input crashes undirected single-hop Cypher (TypeError, not parity/NIE) #1740 (fixed): LazyFrame nodes crashed undirected single-hop (polars joins don't mix eagerness); 2×2 eagerness matrix pinned to pandas parity.
• GFQL: a DANGLING DESTINATION is matched by pandas/cuDF and dropped by polars — silent cross-engine divergence #1808/GFQL polars: HAS_<Label> aggregation shape diverges from pandas on duplicate-id graphs (no disambiguation, no decline) #1739 (pinned, deferred): strict xfails for the degenerate-input divergences (dangling destination; duplicate-id HAS_label aggregation) — pre-existing across releases, fix PRs flip a pin next cycle.

Also verified fixed at head and commented on: #1712, #1753, #1748, #1723, #1713, #1744, #1726 (polars CPU), #1745 (crash class). Closed: #1795. Canvas + probe logs: plans/gfql-benchmark-numbers/.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MF7uRZLKZaD6Q9FGWSmyXi

[lmeyerov]

Boundary-testing pass (owner review concern) updates on this branch:

• GFQL: Cypher OLAP fast paths silently run on CPU under engine='polars-gpu' (bare LazyFrame.collect() bypasses the GPU-or-error contract) #1824 scoped down after real-GPU measurement: the behavior flip exposed 86 polars-gpu tests that were green only via the CPU mislabel (eager arms never collect). Behavior reverted to CPU-served (unchanged from master); collect routing + GPU-scoped NIE plumbing + source pin + strict-xfail disclosure land now; the per-arm GPU-or-decline rewrite is GFQL: Cypher OLAP fast paths silently run on CPU under engine='polars-gpu' (bare LazyFrame.collect() bypasses the GPU-or-error contract) #1824 next cycle. CPU NIE handling tightened: only the GPU target treats NIE as a decline — a CPU-side NIE still surfaces (negative boundary).
• pyg-bench conformance note: the GFQL: two-hop count returns a STALE answer after in-place frame mutation (degree-count memo setattrs onto the caller's Plottable, keyed by id() — the BLOCKER-1 pattern the same file forbids) #1825 memo removal deliberately costs the UNINDEXED q8 cell its warm-repeat advantage; scoreboard G-cells will be re-read after merge (typed-index headline is registry-served, unaffected). tck-gfql runs in this PR's CI.

[lmeyerov]

Verification bundle — all green at final head 184fdee3f:

• CI Tests workflow: success (workflow-level conclusion, not check-run histogram) — includes tck-gfql and the polars lanes
• dgx GPU sweep (arm64 Grace, cudf 26.02): 8,602 passed / 0 failed — the 86 polars-gpu failures from the unscoped GFQL: Cypher OLAP fast paths silently run on CPU under engine='polars-gpu' (bare LazyFrame.collect() bypasses the GPU-or-error contract) #1824 flip are gone after scoping to groundwork+disclosure
• pyg-bench conformance, measured: the GFQL: two-hop count returns a STALE answer after in-place frame mutation (degree-count memo setattrs onto the caller's Plottable, keyed by id() — the BLOCKER-1 pattern the same file forbids) #1825 memo removal is board-neutral — unindexed q8 repeat-median 13.60 ms (byte-identical to the release re-lane receipt; the fused lazy lane had already superseded the memo) and typed-index 3.55 ms (headline holds). No cell moves, no re-receipting needed.
• Boundary pass (owner review): GFQL polars: LazyFrame node input crashes undirected single-hop Cypher (TypeError, not parity/NIE) #1740 re-fixed at the class level (chain-entry eagerness normalization; 3 undirected shapes × 4 eagerness combos pinned) after negative-boundary probes showed varlen/multi-hop still crashing on the seam patch; GFQL: Cypher OLAP fast paths silently run on CPU under engine='polars-gpu' (bare LazyFrame.collect() bypasses the GPU-or-error contract) #1824 NIE-decline scoped to the GPU target only (a CPU-side NIE still surfaces); strict-xfail pins for GFQL: a DANGLING DESTINATION is matched by pandas/cuDF and dropped by polars — silent cross-engine divergence #1808/GFQL polars: HAS_<Label> aggregation shape diverges from pandas on duplicate-id graphs (no disambiguation, no decline) #1739/GFQL: Cypher OLAP fast paths silently run on CPU under engine='polars-gpu' (bare LazyFrame.collect() bypasses the GPU-or-error contract) #1824.

Test-amplification round-001 (owner-directed) is running around the touched seams + error classes; any additions land as follow-up commits on this branch before merge.

🤖 Generated with Claude Code

[lmeyerov]

Final: all gates green at head 17d362008.

• CI Tests workflow: success (tck-gfql + polars lanes included); type-hygiene guard reports the branch 7 files below baseline
• dgx GPU sweep (arm64 Grace, cudf 26.02): 8,664 passed / 0 failed, including the ~60 new amplification tests' cudf params
• pyg-bench: GFQL: two-hop count returns a STALE answer after in-place frame mutation (degree-count memo setattrs onto the caller's Plottable, keyed by id() — the BLOCKER-1 pattern the same file forbids) #1825 fix measured board-neutral (13.60 ms unindexed q8 = the receipt; 3.55 ms typed-index)

Amplification round-001 (owner-directed) is folded in: ~600 probes across 3 error-class agents → 25 confirmed instances → 7 fixed here (incl. 3 HIGH: nan-clean stale-verdict silent-wrong rows, direct-hop lazy crash class ×16 shapes, WITH…MATCH re-entry self-contradictory NIE), 5 deferred behind issues #1879–#1882 with strict-xfail pins. Provenance audit: nearly every fixable bug shipped in 0.58.0 (released mid-campaign, 2026-07-21); three are pre-campaign-ancient; zero introduced by this stack. Artifacts: plans/gfql-release-amplification/round-001/.

🤖 Generated with Claude Code

[lmeyerov]

This PR was NOT merged by anyone — GitHub auto-marked it. During branch-ref repair, this PR's head briefly pointed at the base branch's own commit (0b0a6154a), and when the base was pushed to that same commit GitHub detected head⊆base and auto-closed as MERGED (mergeCommit is literally the base's own commit — no stack content was merged). Master is untouched; the full stack lives on fix/gfql-release-blockers @ 45332aa6c and continues in a successor PR.