feat(gfql/cypher): searchAny cross-column search predicate — inspector semantics, native on 4 engines

[lmeyerov]

Implements searchAny(entity, term[, opts]) — the streamgl-viz inspector's cross-column table search as a composable openCypher-style WHERE predicate, native on all four engines (pandas/cuDF/polars/polars-gpu), plus Python twins g.search_nodes() / g.search_edges(). Viz-filter L2 (frozen semantics: panel-algebra D2; naming/options via a 4-persona design pass — artifacts under plans/viz-filter-pipeline/research/searchany-persona-pass.md). Stacked on #1679.

Semantics (dtype gate is SEMANTICS, not optimization)

• OR across the entity's columns; case-insensitive substring by default (case-folded, never regex — the common viz call avoids every engine regex limit); regex: true opt-in obeying the same per-engine decline rules as =~ (now enforced on cuDF too: prep + casefold-or-decline before delegating, instead of libcudf crashes or \D→\d inversion).
• Dtype gate: string columns always; integer columns iff the term is a numeric literal (the inspector's /^[0-9.-]+$/ gate); floats/dates/booleans only via the explicit columns: list. Explicit non-string columns are served where stringification provably matches the pandas oracle: bool everywhere (polars canonicalizes 'True'/'False'), float/date on pandas only — polars and cuDF decline them honestly (dgx-probed: pandas str(1e16)='1e+16' vs polars '1e16' vs cuDF '1.0e+16', and cuDF truncates mantissas — 0.1+0.2 → '0.3').
• Options map {caseSensitive, regex, columns} strict-validated: unknown keys, non-boolean flags, malformed column lists, and unbound aliases all error clearly; missing explicit columns raise the same E108 on pandas AND polars (pinned to beat the dtype gate in precedence); $param terms decline with a clear message; null cells never match (incl. astype(str)'s 'nan'/'<NA>' artifacts, masked back out).
• searchAny(...) spelled inside a string literal is data, not a call — the lowering lift matches against a literal/comment-masked copy of the WHERE text (with rescan-after-skip so quote-spanning pseudo-matches can't swallow real calls).
• Nodes and edges independently searchable; composes through AND/OR/NOT (lowered like the pattern-predicate markers); two searchAny in one WHERE compose (pinned).

MATCH (n) WHERE searchAny(n, 'acme') RETURN n
MATCH (n) WHERE searchAny(n, 'alpha') AND searchAny(n, '7') RETURN n

Engineering

Per-column matching reuses the parity-hardened Contains predicate on pandas/cuDF (every cuDF workaround + honest decline from the #1675 review waves inherited) and a fold/any_horizontal lowering on polars (Rust-regex guard on the regex path; module-per-family polars/search.py). Safelisted row op with schema effects; conformance-ledger entries; polars coverage baseline gains a search.py floor (79.0; dgx-measured 86.96 on the CI-verbatim lane). One deliberate base change: the call executor's NIE pass-through extends from polars-only to row-pipeline calls on any engine (is_row_pipeline_call), so kernel-level per-engine declines surface as honest NotImplementedError instead of a wrapped E303 — non-row-pipeline NIEs (e.g. fa2_layout without GPU) still wrap, pinned.

Review + verification

5 adversarial findings-mode waves to convergence (waves 4+5 consecutive clean; artifacts plans/viz-filter-pipeline/waves/pr1680/). Waves fixed 2 BLOCKERs (in-literal lift rewrite; cuDF regex guard bypass) and 5+1 IMPORTANTs (null-cell stringify match, polars E108, $param, polars bool 'true' silent divergence, float repr divergence — each now pinned). dgx GB10 (--gpus all, rapids 26.02): targeted 5-file lane 1703 passed at the converged tip; CI green incl. the polars coverage audit. ruff + mypy clean.

Boundaries (clear errors, all pinned)

searchAny in RETURN/WITH projections (WHERE only); edge-alias searchAny(r, …) declines on polars (pending multi-entity binding-row support); explicit columns beyond string/int/bool decline on polars + cuDF; polars-frame Python twins decline (use the cypher op); regex declines mirror =~.

🤖 Generated with Claude Code

[lmeyerov]

ensure cross-platform not just polars, and cross-test

[lmeyerov]

we should audit our tests:

• what happens when nested structures
• are we appropriatley short circuiting, eg, skipping bool cols and numeric cols for non-numeric str
• ... what happens when pretty printer adds "." and "," to ui? should we search those too?
• true/false variants (True, False, 0, 1, ...) can probably be converted to bool vals for when searching bool cols, what's normal here?

[lmeyerov]

DRY