Foundation: bench-gate catches AOT performance regressions (#36196)

[PurHur]

Summary

• Adds script/bench-gate.sh + script/bench-gate.php — compares AOT/Zend wall-time ratios (best-of-3, same job) and LLVM IR line counts against benchmarks/BASELINE.json
• Commits baseline with ratio targets from benchmarks/README.md (2026-07-27 table); IR counts from pinned php-compiler:22.04-dev
• Wires as tier 3 in compiler-gate.yml for lib/JIT/**, lib/AOT/**, ext/standard/** path triggers

Gate behavior (intentionally red)

The gate fails on today's master for fibo(30) — proof it bites before the typed-int boxing fix (#36189 / #31980):

bench-gate: PERFORMANCE REGRESSION
  fibo(30).ratio: 0.3013 exceeds baseline 0.1687 by 78.6% (limit +30%)
  mandelbrot.ratio: 118.2140 exceeds baseline 1.1192 by 10462.0% (limit +30%)

mandelbrot also regressed (separate issue). Gate goes green after perf fixes + --update.

Test plan

• ./script/phpunit.sh --filter BenchGateTest — 4/4 OK
• make dev-verify-fast — OK (91s)
• ./script/bench-gate.sh — exit 1, fibo(30) ratio regression reported (as designed)
• make north-star5-verify-fast — not run (tooling-only change)

Closes #36196

Made with Cursor