Evaluate lazy-ghost initialization for ReflectionConstructorInvocation::proceed()

[lisachenko]

Summary

Follow-up to #606 (closed by #612, which implemented native lazy objects in Container and CachedAspectLoader). Item 3 of that issue — rewriting ReflectionConstructorInvocation::proceed() (src/Aop/Framework/ReflectionConstructorInvocation.php ~65-77, currently newInstanceWithoutConstructor() + manual getConstructor()?->invoke(...)) via the PHP 8.4 lazy-ghost initializer sequence — was deliberately skipped and needs its own ticket so it isn't lost.

Why it was skipped in #612

• A lazy ghost there would either need immediate forced initialization (no gain over the explicit sequence) or would defer constructor side effects, changing observable behavior around getThis() and Before advices.
• newLazyGhost() rejects classes the current sequence handles fine (internal classes, property-less classes).

What this ticket asks

Design a behavior-preserving adoption (or conclude and document that the explicit sequence is the right mechanism). If adopted, it is also the natural groundwork for the gap noted in docs/php84-limitations.md: no join point exists today for the moment a lazy proxy materializes.

[lisachenko]

Evaluation plan (decision first, no code lands until it's made)

Candidate outcomes

• A — keep the explicit sequence (newInstanceWithoutConstructor() + ctor invoke), document as deliberate.
• B — ghost with immediate initialization: semantically identical; only justified if the initializer closure becomes the substrate for a future materialization() join point (the lazy-object gap noted in docs/php84-limitations.md).
• C — deferred ghost, opt-in via a new Features flag: proceed() returns an uninitialized ghost; the constructor + advice chain run at first real use. Only variant with a user-visible benefit (creation-heavy paths pay construction only for objects actually used) — and the only one that changes observable behavior.

Evaluation steps (spike branch, ~0.5–1 day, nothing merged)

1. Compatibility matrix by experiment on PHP 8.4/8.5/8.6: internal classes and subclasses, property-less classes (the pre-initialized gotcha found in PHP 8.4 modernization batch: native lazy objects, property hooks, str_*/array_any + latent strpos bug fixes #612), non-public constructors (ReflectionMethod::invoke handles them; newLazyGhost() initializers cannot call them directly), readonly classes (8.4 vs 8.5 differ), hooked properties, #[\SensitiveParameter] ctors → which shapes must fall back per mode.
2. Semantic-diff inventory: run existing ReflectionConstructorInvocation + functional initialization() tests against B and C prototypes; enumerate every difference — getThis() in After/Around advices (touching a ghost triggers init: can that re-enter the advice chain?), ctor exceptions surfacing later in unrelated stack traces (C), spl_object_id/WeakMap identity, serialization/clone, interplay with Doctrine proxies ([Known-issue] Doctrine entities could not be woven  #327).
3. Micro-benchmark (10⁶ instantiations, plain + side-effecting ctors, 8.4/8.5): explicit vs ghost-immediate vs ghost-deferred (never-touched and touched). Establishes B's overhead and C's actual win.
4. Join-point sketch (half a page): is B/C's initializer the right place for a materialization() join point? If yes, B stops being pure overhead and becomes groundwork.
5. Decide against explicit criteria:
   • adopt C only if: opt-in flag, zero diffs with flag off, fallback matrix fully covered, and a real benchmark win (≥20% on the creation path);
   • adopt B only if the join-point sketch shows it is needed as groundwork;
   • otherwise A, and only then land the documentation change.

Deliverable: matrix + diff inventory + benchmark numbers + recommendation, posted here for the final call.

---

Generated by Claude Code