Fixes GH-10134.

The bug

A fiber can be suspended below an internal function. That leaves the internal frame parked on the fiber's stack with the arguments that were pushed for it still live — the frame owns them and only releases them when it returns.

zend_unfinished_execution_gc_ex() returns early for internal frames without reporting those arguments:

if (!ZEND_USER_CODE(EX(func)->common.type)) {
    ZEND_ASSERT(...);
    return NULL;          /* arguments never reported */
}

So the cycle collector never sees them, and any cycle running through such an argument is uncollectable — it is only broken at request shutdown.

Not specific to generators

GH-10134 reports this as a generator problem, because the reporter hit it through Fiber::suspend() inside a generator resumed by Generator::send($fiber) — $fiber lands in the argument slot of the internal Generator::send frame. But nothing about it is generator-specific. With no generator anywhere, a fiber suspended inside array_map() whose array argument holds the fiber behaves identically (second test added here).

How it was diagnosed

zend_fiber_object_gc() is not at fault — it correctly walks the suspended fiber's frames, finds the generator frame, and calls zend_generator_frame_gc() on it. Instrumenting the collector's root processing showed the actual failure:

[root-before]     Fiber rc=2  PURPLE
  reported obj    Fiber rc=2            <- reported once
[root-after-mark] Fiber rc=1  GREY
[root-after-mark] Generator rc=3        <- restored from 0

The Fiber's refcount is 2 but the mark phase discovers only one of those references, so it lands at 1 instead of 0, gc_scan concludes it is externally referenced, and gc_scan_black restores every refcount it had decremented. Dumping the frames shows where the second reference lives:

frame ... INTERNAL func=send nargs=1
   ARG[0] type=8 Fiber 0x...      <- never reported

On double counting

This was the main risk. It does not happen: on ZEND_DO_ICALL / ZEND_DO_FCALL the VM does EX(call) = call->prev_execute_data, so an entered frame is no longer in its caller's EX(call) chain and is therefore not also scanned by zend_unfinished_calls_gc(), which only walks calls that are still being built. Those pending frames are siblings rather than ancestors, so they never show up in the prev_execute_data walk either.

zend_unfinished_execution_gc_ex() has two callers. zend_generator_frame_gc() always passes generator->execute_data, which is user code, so only the fiber path is affected in practice.

Testing

• Both new tests fail on unmodified master and pass with this change — I ran them against a pristine build to confirm they are real regression guards.
• The issue's verbatim script now prints the expected order (Generator dtor, Fiber dtor, Shutdown) instead of Shutdown first, and gc_collect_cycles() returns 1 instead of 0.
• No premature frees, which is the thing to worry about when handing GC additional edges: an object passed as an internal-frame argument that is also still referenced from live scope survives gc_collect_cycles() and remains usable. Checked explicitly.
• Full test suite on this build (--disable-all --enable-debug): 13386 passed, 0 failed, 0 warned (7 expected-fail, the rest skipped for disabled extensions).
• Zend/tests/fibers and Zend/tests/gc under valgrind memcheck: clean.

Built and tested on Linux x86_64, --disable-all --enable-debug, so the extension-heavy parts of the suite were skipped; a normal CI run would cover those.