RobustUnpacker: release all memoryviews of the resync buffer before resizing it

[ThomasWaldmann]

Fixes the "Tests (pypy-3.11-nightly, ubuntu-26.04)" CI job, which fails since the PyPy nightly of 2026-10-11 (PyPy 8.0.2-alpha0, rev 544bd2f768f9): 12 tests in archive_test.py (test_missing_chunk, test_corrupt_chunk, test_resync_*) die with

BufferError: Existing exports of data: object cannot be re-sized

at del self._buffered_data[:drop_len] in RobustUnpacker._search().

Cause

_search() scans the buffer inside with memoryview(self._buffered_data) as data: and called valid_msgpacked_dict(data[offset:]) for every offset; that function sliced the key out of the view as well. CPython frees these unreferenced child views at once, so the bytearray has no exports left when the with block releases the parent view. PyPy has no refcounting, the child views live until the GC runs. PyPy 8.0.0 did not count child views as exports of the bytearray, the new nightly does (as CPython always did), so the resize after the with block fails.

Upstream background

The nightly behaviour is a deliberate PyPy fix, not a regression: PyPy 8.0.0 dropped the bytearray's export when the parent view was released even while child slices were still alive, which could segfault (pypy/pypy#5613, fixed by pypy/pypy@49f894f0f541 "keep a memoryview's export alive while its slices and casts use it", followed by pypy/pypy@f61a05f029e5 "refactor buffer exports to use a BufferExport class"). CPython 3.11 raises the same BufferError for this pattern if the child view is still referenced. PyPy's docs (pypy/pypy@225bffc76d6d, for pypy/pypy#5612) now say that views are kept alive until the GC frees them, so views must be released explicitly (view.release() or with memoryview(buf) as view:) before the viewed object is resized. For _search() that means the child slices data[offset:] need releasing too, or must not be created at all; the with on the parent view alone is not enough.

Fix

valid_msgpacked_dict(d, offset=0) only indexes d and never slices it, and _search() passes (data, offset) instead of a slice. With the already explicitly released slice in _try_item_start(), no unreleased view of _buffered_data exists when it is resized. The algorithm and its linear-time behaviour (#10459) are unchanged.

Verification

• PyPy 8.0.2-alpha0 nightly (544bd2f768f9, macOS arm64): archive_test.py 12 failed / 80 passed before, 93 passed after.
• PyPy 8.0.0 release: 93 passed.
• CPython 3.11: archive_test.py 93 passed, archiver/check_cmd_test.py 120 passed (binary variants skipped).
• CPython scan of 2 MiB garbage after resync(): 0.179 s before, 0.148 s after (no slice object per offset).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.29%. Comparing base (dc473d2) to head (4da7021).
⚠️ Report is 25 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #10517      +/-   ##
==========================================
+ Coverage   89.24%   89.29%   +0.04%     
==========================================
  Files         104      104              
  Lines       19927    19941      +14     
  Branches     3124     3127       +3     
==========================================
+ Hits        17784    17806      +22     
+ Misses       1483     1478       -5     
+ Partials      660      657       -3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.