gfql/cypher: resolve Lark ambiguity between where_predicates and expr (grammar-level follow-up to #1125) #1194

Context

PR #1193 (issue #1125) hardened WHERE label narrowing by replacing a fragile regex in `binder.py` with AST-derived narrowing plus a text-split fallback in `parser.py::generic_where_clause`.

The text-split fallback exists because of a Lark grammar ambiguity: `WHERE n:Admin AND n:Active` can match BOTH the structured rule `where_predicates: where_predicate ("AND"i where_predicate)*` (parser.py:127) AND the generic expression path `where_clause: "WHERE"i expr -> generic_where_clause` (parser.py:126). Lark's ambiguity resolution prefers the generic path, so the AND-joined label predicates land in `WhereClause.expr` as raw text rather than as structured `WhereClause.predicates`.

The current PR works around this at the transformer level — `generic_where_clause` re-parses the raw text and lifts it back into structured predicates via `_split_top_level_and_terms` + `_BARE_LABEL_PREDICATE_RE`. This is correct and well-tested, but architecturally backward: the grammar already declares the structured path we want; we should let it win.

Proposed follow-up

Fix the ambiguity at the grammar layer so AND-joined bare label predicates naturally route to `where_predicates`:

1. Investigate why `expr` currently wins the ambiguity contest — likely because `bare_label_predicate_expr` is reachable from `?predicate` inside `?and_expr`.
2. Either:
   • Remove the `bare_label_predicate_expr` alternative from the `expr` path, forcing label predicates to go through `where_predicates` exclusively, OR
   • Raise the priority of `where_predicates` (Lark rule priority annotation), OR
   • Introduce an explicit `where_clause` disambiguator rule.
3. Once the grammar disambiguates correctly, the text-split logic in `generic_where_clause` becomes dead code for AND-joined labels — simplify to the single-label case (still a valid generic fallback) or remove entirely.

Tests to keep green

• All 10 label-narrowing tests in `test_binder.py` (single, double, triple AND, multi-alias, multi-label, lowercase-and, mixed label+property, XOR/OR/NOT conservative, string-literal false-positive guards).
• All 4 parser-level AST tests in `test_parser.py`.

Constraints

• Must not regress existing `where_predicates` behavior (structured comparison predicates, IS NULL, CONTAINS, etc.).
• Must keep string-literal false-positive protection (the `fullmatch` invariant in the current fallback).

Priority

Low-to-medium. The current text-level workaround is correct and tested; this is an architectural cleanup that removes a duplicate implementation of "split on top-level AND" and lets the grammar be the source of truth.