feat(gfql): restore lambda support in filter_dict for local-only flows

[lmeyerov]

Summary

Lambda functions in filter_dict (e.g., n({"score": lambda x: (x > 50) & (x % 2 == 0)})) were removed when AST validation was tightened to require JSON-serializable values for the wire protocol and remote execution.

This is a regression for local-only workflows where serialization isn't needed. Lambdas were a convenient way to express compound conditions that don't map cleanly to a single predicate.

RCA

The check at graphistry/compute/ast.py:167 rejects any filter_dict value that isn't an ASTPredicate or JSON-serializable. Lambdas/callables fail this check.

The reason is wire protocol: filter_dict values need to round-trip through JSON for to_json() / from_json() and remote execution via gfql_remote().

Proposed

For local-only execution (g.gfql([...]), not g.gfql_remote([...])), allow callables in filter_dict:

• Accept callable values at AST construction time (skip the JSON-serializable check)
• In filter_by_dict(), apply callables directly: df[df[col].apply(val)]
• Raise at serialization time (to_json()) if a callable is present, not at construction time

This preserves the wire protocol constraint while restoring local usability.

Current workaround

# Instead of: n({"score": lambda x: (x > 50) & (x % 2 == 0)})
# Use query string:
n(query="score > 50 and score % 2 == 0")

[lmeyerov]

Design note: numba UDFs vs lambda restoration

Investigated whether numba-jitted functions could provide a cuDF-compatible lambda path:

• cuDF Series.apply() requires numba-jitted functions (not plain Python lambdas)
• numba UDFs are restricted to numeric operations — no string filtering, no complex logic
• numba is a heavy dependency for marginal benefit

Recommendation: Don't add numba. Instead:

1. Restore plain lambdas for local-only flows — defer the JSON-serializable check from AST construction (ast.py:167) to to_json() time. Lambdas work fine on pandas Series.apply(). Raise at serialization time if someone tries to_json() or gfql_remote() with a lambda in filter_dict.

2. Keep query= strings as the primary cross-backend approach — df.query("score > 50 and score % 2 == 0") works on both pandas and cuDF with no serialization issues.

This gives local users their lambdas back while keeping the wire protocol clean.