Reduce AST visitor bookkeeping without changing traversal behavior #292

Walking the GitHub schema with empty visitor callbacks spends time creating stack entries and empty edit lists.
The same loop also repeats type checks and checks action values when callbacks return None.

The repository benchmarks show these median times for a local change:

The parallel visitor case has a smaller gain because it spends more time in callbacks.
Pickle decoding does not use visit and shows no measurable change.

The proposed change uses plain tuple stack entries and creates edit lists only when needed.
It also reuses array state and skips action checks after None results.
The proposal keeps callback order, callback arguments, and immutable edit behavior unchanged.

This is Python-specific bookkeeping, with no new GraphQL behavior or API changes.
The complete native test suite passed: 3,610 tests, with 18 skipped.
Ruff and mypy passed for the changed module.

Environment: base commit 679be84cb93e421b04aee4a1b8f7ada34aed7d49, CPython 3.12.13, Linux x86-64, AMD EPYC 9R14.
Measurements used pytest 9.0.2 and pytest-benchmark 5.3.0, with at least 50 rounds per case on one isolated slot.

Reproduce on the base commit and proposed change:

PYTHONPATH=src python -m pytest tests/benchmarks/test_visit.py tests/benchmarks/test_serialization.py::test_pickle_large_query_decode --benchmark-enable --benchmark-min-rounds=50

AI assistance: Codex gpt-6.1-sol generated candidates tracked with Weco Observe; Codex helped investigate and validate this proposal.

Co-Authored-By: Aiden aiden@weco.ai