[dexhunter]

TL;DR

• Make visit_selection_sets() in collect_fields.py faster, which groups the fields of each selection set before execution.
• Skip the @skip/@include check for selections without directives, and scan directives in one plain loop.
• Wide-query benchmark: 73.9 ms -> 63.1 ms, 14.6% less time.
• PYTHONPATH=src python -m pytest tests/execution/test_collect_fields.py -q: 5 passed; full suite: 3,821 passed, 18 skipped.

Refs #315

What & Why

Before, the loop called should_include_node() for every selection, in three separate branches.
That function built two generator expressions to find @skip and @include, even when the selection had no directives.

Now the loop checks selection.directives once, before the type dispatch.
It calls should_include_node() only when directives are present.
should_include_node() finds the first @skip and the first @include in one plain loop.

The output stays identical.
A selection without directives was always included, so skipping the call changes no result.
The check order stays the same: @skip first, then @include, then the fragment type condition.

Results

Situation	Before	After	Change
collect_fields on 4,000 plain fields	4458.7 µs	4113.4 µs	7.7% less time
collect_fields on 4,000 fields, every 10th aliased	4531.0 µs	4176.0 µs	7.8% less time
collect_fields on 4,000 fields with @include(if: $t)	10988.9 µs	8593.5 µs	21.8% less time
collect_fields on 400 fields with @skip(if: true)	653.3 µs	518.4 µs	20.7% less time
collect_fields on 40 inline fragments with 100 fields each	4561.1 µs	4165.7 µs	8.7% less time
collect_fields on 400 @defer inline fragments	1440.1 µs	1294.3 µs	10.1% less time
collect_fields on 400 named fragment spreads	960.0 µs	889.9 µs	7.3% less time
collect_fields on a 10-field query	4.8 µs	4.0 µs	16.6% less time
test_execute_introspection_query (repo benchmark)	243.1 ms	241.2 ms	No change (within noise)

The collect_fields rows are medians of 31 rounds, then the median of 3 alternating processes per version.
The introspection row is the median of 5 alternating pytest-benchmark runs per version.
They ran on CPython 3.12.13, Linux x86-64, AMD EPYC 9R14, on one dedicated core.
Full execution benchmarks do not change, because field collection is a small share of their time.

Validation

PYTHONPATH=src python -m pytest tests/execution/test_collect_fields.py -q
PYTHONPATH=src python -m pytest tests -q
PYTHONPATH=src python -m pytest tests/benchmarks/test_introspection_from_schema.py --benchmark-enable --benchmark-min-rounds=50

The collect-fields tests passed (5 tests), and the full test suite passed (3,821 passed, 18 skipped).
The wide-query timing calls the public collect_fields() and collect_subfields() 14 times on a 4,000-field selection set.
Its output was compared with the unchanged code, including field order, defer usages, errors and forbidden directives.
ruff check and ruff format --check pass on the changed file.

Why this is safe and easy to maintain

The patch changes one loop and one helper in a single file and adds no dependencies.
It removes two of the three should_include_node() calls and one of the two directive scans.
Grouped field order, FieldDetails values, defer usages, raised errors and forbidden-directive lists are unchanged.
tests/execution/test_collect_fields.py and the execution tests for @skip, @include, @defer and fragments cover this code.

AI assistance

Codex gpt-6.1-sol generated the candidate changes, tracked with Weco Observe.
Claude Code ran the automated checks and prepared this description.

Weco share link: https://dashboard.weco.ai/share/LnYj9kTcZUOUYikZsvxmmnif2otC4CA9

[codspeed]

Merging this PR will not alter performance

✅ 20 untouched benchmarks

---

_{Comparing dexhunter:perf/collect-fields-directive-check (ddd3d49) with main (1d505c1)}