[perf] Add single-process hotspot profiler + performance methodology

[TomasVotruba]

What

Adds the tooling and written method behind the recent batch of measured perf PRs (#8705-#8714), so the next perf pass starts from data, not guesses.

scripts/rector-hotspot-profile.php

Standalone single-process hotspot profiler. Runs rector process --dry-run over a real project under the Xdebug profiler with parallel disabled (parallel workers each write their own cachegrind and make the numbers useless), parses the cachegrind, and ranks functions two ways:

• by call count - memoization candidates (a function hit many times on repeating inputs is where a cache might pay)
• by self cost - where CPU actually burns

php scripts/rector-hotspot-profile.php --project=/path/to/a/real/app --filter='^(Rector|PHPStan)'

Standalone: drives the binary as a subprocess, no autoloader of its own. Docblock documents the two traps found while building it - small corpora are dominated by container bootstrap, and most type/scope cost lives in PHPStan\* not Rector\*.

CLAUDE.md - Performance work section

Short methodology: profile a single-process real run to find the hot spot, qualify a cache candidate by confirming inputs actually repeat and picking a version-safe key (AST mutates to a fixpoint, so node-identity/file-path keys go stale), prove it with CPU-time and call-count A/B rather than wall time, and gate on the full suite plus the order-dependent fixpoint tests.

Why

Keeps the measurement discipline from the perf batch reproducible instead of tribal knowledge.

[TomasVotruba]

LGMT 👍