Answer invalidation checks from a per-holder index of contained node keys

[ondrejmirtes]

Ported from the resolve-type-rewrite-2 branch. ScopeOps::shouldInvalidateExpression() re-scanned the stored holder's whole subtree (after the substring/compositional-key pre-filters) on every invalidation check. The scan establishes a property of the holder's expression alone, so ExpressionTypeHolder now lazily indexes the node keys of its sub-expressions once — holders are shared across scope copies, so the single scan amortizes — and the check becomes one array lookup, exact by construction. $this invalidations keep the finder (they also match self/static/parent and the current class name, whose resolution depends on the asking scope). 2.2.x's newer pre-filter gates (the requireMoreCharacters variable refinement, keyMayHideSubExpressions) are kept.

The native twins mirror the same split; the ExpressionTypeHolder twin exposes the same lazily-cached getContainedNodeKeys(), key-order identical.

Validation: strict build, smoke (incl. new differential coverage for the index + memoization), side-by-side.php, signature-parity.php, full test suite green with and without the active extension, byte-identical analysis output over src/Analyser src/Rules with the extension on vs off.

Honest benchmark verdict: neutral on 2.2.x — 12 interleaved ABBA pairs over src/Analyser src/Rules (no turbo): A=51.54s vs B=51.54s user CPU (±0.03% in both independent rounds). 2.2.x's pre-filters already absorb almost all checks before the subtree scan, so the index has nothing left to win here. The value is convergence: the resolve-type-rewrite-2 branch measured −2.5% user CPU from this change (its engine performs ~2.3M invalidation checks per self-analysis run), carries it as a twin-divergence, and every rebase pays for the conflict surface. If neutral-CPU-for-convergence isn't worth it, holding this until the branch merge is a fine alternative.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DaBZjgksga4c5s6Q9FniY7

[staabm]

are these developer machine local path expected ?

[ondrejmirtes]

Probably shouldn't be commited at all

[ondrejmirtes]

For the record: the branch no longer needs this either. The root cause of its 3.2× invalidation-check volume was virtual-node keys defeating the compositional-substring shortcut (plus a missing caller-level prefilter in its invalidateExpressionEntries), not intrinsic load — same events, same holder population as 2.2.x. With the compositional-gate fixes landed there (f65dfce), calls dropped 2.41M → 420k and containment checks 1.16M → 308k — below 2.2.x's 629k, where this PR measured dead neutral — so the index was dropped from the branch too (5ee84cd) and ExpressionTypeHolder/ScopeOps converged back to the 2.2.x shapes.