Description:
------------
Reopening bug #36711 because it is NOT a documentation problem. Setting 'detect_unicode=Off' is NOT a solution, just a workaround.

In practice, because of this bug, PHK or PHAR packages cannot run on zend-multibyte-enabled environments, unless detect_unicode is turned off. Which makes them unusable in environments running unicode-encoded scripts. As a side effect, it also makes it impossible to include an unicode-encoded script inside a PHAR/PHK package, as it cannot be run.

There is no logical reason to bind the __halt_compiler() feature with the zend-multibyte unicode detection capability. Everything after an __halt_compiler() directive must be considered as binary data and should not be scanned for unicode detection. If this data contains a unicode script, it will be scanned and detected when include()d through the stream wrapper.

My (humble) suggestions to fix the problem:

In zend_multibyte_detect_unicode(), the BOM detection does not have to be modified but, then, the script is scanned for null bytes :

return zend_multibyte_detect_utf_encoding(LANG_SCNG(script_org), LANG_SCNG(script_org_size) TSRMLS_CC);

There, the size should not be LANG_SCNG(script_org_size), but the offset of the __halt_compiler() directive. But I don't know where to find the COMPILER_HALT_OFFSET constant for the script. I even suspect it not to be available at this time...

Another way, if the previous one is not possible, would be to scan for a binary string that cannot correspond to any unicode encoding. This way, PHK and PHAR could insert this string after ther __halt_compiler() directive, and it could be detected by zend_multibyte_detect_utf_encoding() as a stop string. I am ready to implement it if somebody provides a sequence of bytes that cannot be found in any unicode-encoded document.

Reproduce code:
---------------
<?php
echo "OK\n";
__halt_compiler();<null-byte>

Expected result:
----------------
OK

Actual result:
--------------
??????????????????