test(gfql): viz-filter-pipeline acceptance suite + regression benchmark (L3)

[lmeyerov]

The viz-filter-pipeline ACCEPTANCE SUITE + REGRESSION BENCHMARK (viz-filter L3) — the proveout layer over the whole stack (#1675 → #1677 → #1679 → #1680). Frozen semantics: plans/viz-filter-pipeline/research/panel-algebra.md (v1-FINAL). Stacked on #1680.

graphistry/tests/compute/gfql/test_viz_pipeline_conformance.py

• Curated full-panel pipelines (9, each: hand-derived oracle pin + an independent plain-pandas second oracle + assert_parity_or_nie across pandas/cuDF/polars/polars-gpu): node+edge filters, exclusion-dominates composition (AND NOT (...) with three-valued null semantics pinned), (pred OR IS NULL) keep-null leaves, both EXISTS prune-isolated flavors, searchAny composed with filters, deterministic ORDER BY/SKIP/LIMIT paging ascending AND descending — the fixture's row order is deterministically shuffled so an engine that silently drops the sort fails on membership, and the ASC+DESC pins are jointly seed-independent (review wave-3 proof).
• Graph-state prune shapes: exact node and edge pins for GRAPH { MATCH (a)-[e]-(b) } (keep-self: self-loop-only node kept with its edge, mutual pair keeps both directed edges, isolated dropped) + the drop-self variant, on a formula-derived fixture (self-loops, parallel edges, isolated + self-loop-only discriminator nodes, unicode/metachar/null name cells, a Categorical column).
• 40-seed panel-state fuzzer: random panel states → generated single cypher query → pandas result vs an independently computed plain-pandas expectation (zero pandas declines allowed), then 4-engine parity-or-NIE; out-of-dialect combos must decline with clear errors (crash classes fail).
• Case/regex/unicode trick matrix (17 rows): (?i) vs case-sensitive =~, ß/İ full-case-mapping pins, literal-vs-regex metachars through searchAny, anchored =~ vs CONTAINS, null cells, Categorical (decline ok, silent-wrong not).

benchmarks/gfql/viz_filter_pipeline.py

Six streamgl-viz panel scenarios (filters+excludes+ORDER/LIMIT, keep-self GRAPH prune, EXISTS prune, node search, edge search, combined) × --scale {100k,1m,10m} × engines with native-frame-per-engine fairness (one-time conversion excluded from timing; both result frames touched inside the clock), warmup+timed medians, an NIE-tolerant never-crash matrix, --est for safe-run budgeting, upfront output-path validation, and JSON receipts. The ~350ms@800K interactive reference is printed as a comparison column, not asserted.

Review + verification

4 adversarial findings-mode waves to convergence (waves 3+4 consecutive clean; artifacts plans/viz-filter-pipeline/waves/pr1681/) — every curated pin and all 17 trick rows independently recomputed by hand in review; oracle-independence and bench fairness adversarially confirmed. dgx GB10 (--gpus all): suite 31 passed at the converged tip on the final stack; CI green.

Receipts (dgx GB10 — plans/viz-filter-pipeline/receipts/)

• 100k ×4 engines: polars 2–5× pandas across panel scenarios; everything within the interactive reference except pandas combined (1.55×).
• 1m ×4 engines: prune/exists/combined at 1.15–2.3× the 350ms reference on polars/cudf/polars-gpu (pandas 3.4–15.7×); search sub-budget on all engines (cudf search_nodes 36ms).
• 10m ×4 engines: receipt in receipts/ (25M edges; run under the memory-safety harness).

Findings surfaced by this layer (pinned + tracked, exactly its job)

• The same-path WHERE route dedupes parallel edges (one per unordered endpoint pair — a mutual pair may collapse to one direction) — diverges from the panel algebra's E₂ edge multiplicity; pinned at exactly that tolerance with the prune-correctness facts still asserted; repo-debt filed.
• Edge-alias searchAny(e, ...) declines on polars (honest NIE) — gap tracked; explicit non-string/int/bool searchAny columns decline on polars + cuDF (float repr divergence, dgx-probed in feat(gfql/cypher): searchAny cross-column search predicate — inspector semantics, native on 4 engines #1680).
• Pre-existing (inherited notes): cuDF same-path TypeError on the drop-self GRAPH form; bare pattern predicates inside GRAPH bodies silently drop (row_pre_filters unwired in the graph compiler).

🤖 Generated with Claude Code

[lmeyerov]

Superseded by #1696 (same branch, rebased onto master, base=master). This PR was auto-closed when its base branch #1680 merged+deleted, and GitHub won't reopen a closed PR with a deleted base — so the L3 acceptance content lives on in #1696. Ignore this one.