Three places dispatch on the page map entry of a pointer, and reach the
large-run case by elimination, with the assumption written down as a
comment rather than checked:

    if (EXPECTED(info & ZEND_MM_IS_SRUN)) {
        ...
    } else /* if (info & ZEND_MM_IS_LRUN) */ {

The assumption does not always hold: when ZEND_MM_IS_FRUN is 0 and
zend_mm_free_pages_ex() zeroes chunk->map[page_num], a pointer to a
large run that has already been freed has info == 0, so it fails the SRUN test,
and falls into the large-run branch. There, ZEND_MM_LRUN_PAGES(0) is 0,
and the three callers quietly degrade:

 - zend_mm_free_heap() frees a run of zero pages, i.e. a double free of a
   large block is accepted and does nothing at all.
 - zend_mm_size() reports a block size of 0.
 - zend_mm_realloc_heap() takes old_size 0 and reallocates from there.

A large-block double free or a use of a freed pointer is silently absorbed by
the allocator instead of being a hard failure. This commit promotes the comment
to a real ZEND_MM_CHECK() in all three. The value is already in a register at
that point, so it costs a test and a branch.

This was checked under GDB by allocating a large block, freeing it, and then
reusing the pointer. Before, _efree() returned normally,
_zend_mem_block_size() returned 0 and _erealloc() returned a new pointer.
After this commit, each of the three aborts with "zend_mm_heap corrupted".

Amusingly, the two comments naming ZEND_MM_IS_LARGE_RUN referred to a
macro that does not exist: the real name is ZEND_MM_IS_LRUN.