`ZipFile.close()` writes wrong offsets after a failed `repack()`

[fedonman]

Bug description:

_ZipRepacker.repack() updates ZipInfo.header_offset for each member before moving that member's bytes, so a failing _copy_bytes() leaves the in-memory offsets describing a layout that was never written. remove() has already set _didModify, so a later normal close() commits a central directory built from those offsets: the caller handles the OSError, closes cleanly, and gets an archive that zipfile itself cannot read.

import io, zipfile

class Flaky(io.BytesIO):
    countdown = None
    def write(self, b):
        if self.countdown is not None:
            self.countdown -= 1
            if self.countdown < 0:
                raise OSError(28, 'No space left on device')
        return super().write(b)

buf = io.BytesIO()
with zipfile.ZipFile(buf, 'w') as zf:
    for c in 'abcd':
        zf.writestr(c + '.txt', c.upper().encode() * 5000)

flaky = Flaky(buf.getvalue())
with zipfile.ZipFile(flaky, 'a') as zf:
    flaky.countdown = 1
    try:
        zf.repack([zf.remove('b.txt')], chunk_size=4096)
    except OSError as exc:
        print('repack raised:', exc)
    flaky.countdown = None        # space freed; the caller closes normally

with zipfile.ZipFile(flaky) as zf:
    print('testzip:', zf.testzip())

repack raised: [Errno 28] No space left on device
testzip: c.txt

Expected: after the caller has handled the OSError, close() should either leave a readable archive or raise, rather than committing offsets that no write produced.

remove() and repack() are new in 3.16, so no released version is affected.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-156434: Restore entry offsets when repack() fails #156435