…parse

_build_transformer() defined two @DataClass(frozen=True) helpers and the Lark
Transformer subclass INSIDE its own body, so all three types were re-created on
every _parse_expr_cached miss. @DataClass generates __init__/__eq__ as source and
execs it, so each rebuild ran the compiler.

Profiling q7 of the matched graph benchmark (31 cold compiles, 0.649s total)
attributed 0.261s -- 40% of ALL compile time -- to this one function: 434 calls
producing 868 dataclass creations, 0.120s of it in builtins.exec. The whole
LALR(1) parse was 0.151s by comparison. Lowering, not parsing, dominates GFQL
compile, and this was most of lowering.

The class is now built once per process behind lru_cache(maxsize=1); instances are
still created per parse, so nothing stateful is shared and there is no
re-entrancy question to argue. Only type creation is hoisted, and a type is a
function of the code, not of any query -- so it is classified EXEMPT in the
clear-caches enumeration, alongside the Lark parser objects.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YYZRXegrALuXd3NHH5evqx