[TomasVotruba]

Proof of concept

AbstractRector::enterNode() runs for every node that matches a rule's getNodeTypes() (the traverser already filters by node type via getVisitorsForNode()). On each call it recomputed values that are constant for the whole file:

• getFile() -> getFilePath()
• is_a($this, HTMLAverseRectorInterface) + File::containsHTML()
• Skipper::matchSkip($this, $filePath) (depends only on rule + file path, never on the node)

For a file with thousands of nodes and many applicable rules, these ran per node x rule although nothing changed within the file.

This PR caches them once per file (keyed on the File instance) and leaves only the genuinely per-node work in the hot path: shouldSkipCurrentNode(), the clone probe on a skip match, refactor(), and post-processing.

Behavior

No behavior change. matchSkip, HTMLAverse and file path are all file-level invariants, so caching them per file is equivalent. Core test suite green (the only failures are the pre-existing tests/Bin/RectorTest subprocess cases, unrelated).

Benchmark - honest result

Interleaved A/B, cold cache each run, single process, broad rule set (all PHP sets + deadCode + codeQuality + codingStyle + typeDeclarations) over src/:

round	baseline	cached	delta
1	67.25s	67.82s	+0.85%
2	68.43s	65.54s	-4.22%
3	69.16s	70.40s	+1.79%
4	69.27s	67.25s	-2.92%

Average ~ -1%, and the sign flips between rounds. The saving is real but within measurement noise at the full-run level: the hoisted calls are cheap next to parsing, PHPStan scope resolution, refactor() bodies and printing, and matchSkip's resolve() is already memoized.

Opening as a proof of concept. The refactor is clean and behavior-neutral; the question is whether the per-node cleanup is worth keeping on its own merit given the macro benchmark is flat.