This script completely self-contained reproducing script. But as I 
mention before, I can't make it smaller because it break 
reproducibility.

Looks like something in the re2c stuff that's causing it to overread.

This is being caused because of mis-use of mmap().  We are currently relying on mmap to pad the end of our mmap'd file with zeros for detection of EOF in the scanner and scanning ahead.  We specifically add ZEND_MMAP_AHEAD to the len passed to mmap in zend_stream_fixup():

/*  *buf[size] is zeroed automatically by the kernel */
*buf = mmap(0, size + ZEND_MMAP_AHEAD, PROT_READ, MAP_PRIVATE, fileno(file_handle->handle.fp), 0);
 
But AFAIK mmap does not support this usage of the len parameter, as it's a limit rather than able to extend the mmap region.  This appears to work under most cases as mmap will pad zeroes up to PAGESIZE.  This error will occur anytime we use mmap in this way on a file that is not ZEND_MMAP_AHEAD bytes less than PAGESIZE and therefore attempt to access a byte over PAGESIZE.

It will be easy to fix the mmap calls, however this will break the re2c scanner.  Originally for the EOF checks I was going to re-implement YYFILL to malloc additional space for the scanner after EOF, this may be an option to correct this.

This bug has been fixed in CVS.

Snapshots of the sources are packaged every three hours; this change
will be in the next snapshot. You can grab the snapshot at
http://snaps.php.net/.
 
Thank you for the report, and for helping us make PHP better.