Phase 2 of the destructor-reentrancy work, building on the error-handler deferral in 22515. Adds ZEND_DTOR_HAZARD_BEGIN/END sections and a deferral gate in zend_objects_store_del: while a section is open, an object that hits refcount zero and has a destructor is pinned and its destructor deferred until the section closes. Applied to three reentrancy use-after-frees where an internal C loop drops a refcount while holding a raw pointer a synchronous __destruct could free: concat (phpGH-20477), zend_hash_clean (phpGH-22061), and unserialize teardown (phpGH-21824), with a regression test for the concat-assign case (phpGH-19999); that test expects NULL because the deferred destructor runs at section exit, after the concat writes the result, and nulls $GLOBALS['a'].

How this differs from arnaud-lb#31: that prototype deferred every destructor through a single global EG(delayed_effects) queue, pinned the dying object with GC_SET_REFCOUNT(obj, 1), and flushed at the VM function-leave helper. It left three problems open: throwing destructors reordered against try/finally and catch (6 XFAILed tests), 67 tests needed a dummy call to create a flush point (a timing change), and GC reconciliation between the RC=1 pin and cycle collection stayed unresolved. This mechanism is scoped instead. BEGIN/END bracket only the specific internal loops that hold the raw pointer, so no user code runs inside a window. That drops the try/finally reordering question (no exception crosses a boundary), drops the timing change (flush is at section exit, not an arbitrary safepoint), uses GC_ADDREF instead of GC_SET_REFCOUNT so it reuses the existing store_del invariants, and closes the GC hole by having gc_possible_root_when_full skip synchronous collection while a section is open.

Performance: the store_del gate is one branch guarded by has_dtor, so destructor-less objects are untouched. Sections arm only on cold paths (concat arms only when an operand needed stringification or the result aliases an operand; hash_clean only when the table has a destructor). Release callgrind over 2M concats shows +1.2 instructions per concat, 0.45% of concat_function, with the common in-place concat paying one predicted-not-taken branch. No global vm_interrupt arming, no JIT may_throw change.