iconv codecs write an error handler's replacement in a non-initial shift state

[serhiy-storchaka]

Bug report

An error handler's replacement bytes are copied to the output without returning it to the initial shift state first, so in a stateful encoding they are read back as encoded data.

>>> '中😀中'.encode('iso-2022-cn', 'replace')
b'\x1b$)A\x0eVP?\x1b$)A\x0eVP\x0f'
>>> _.decode('iso-2022-cn')
UnicodeDecodeError: 'iso-2022-cn' codec can't decode byte 0x3f in position 7: invalid multibyte sequence
>>> '中😀中'.encode('iso-2022-cn', 'backslashreplace').decode('iso-2022-cn')
'中苷鞍氨娑鞍中'

The built-in codec of the same family shifts back to ASCII first:

>>> '日😀日'.encode('iso2022_jp', 'replace')            # built-in
b'\x1b$BF|\x1b(B?\x1b$BF|\x1b(B'
>>> '日😀日'.encode('iconv:ISO-2022-JP', 'replace')     # iconv
b'\x1b$BF|?\x1b$BF|\x1b(B'

_PyUnicode_EncodeIconv() does reset the shift state after each handled error, but with a NULL output buffer, so the conversion forgets the state while the stream never gets the closing sequence.

This affects the ISO-2022 family and the stateful EBCDIC code pages (CP1364, CP1371, ...). ignore is unaffected, it injects nothing.

Linked PRs

• gh-155004: Fix iconv error handlers in a non-initial shift state #155005