gh-148395: Add regression tests for decompressor UAF after MemoryError (CVE-2026-6100)

[Vamsi-klu]

Summary

Adds the missing regression tests for CVE-2026-6100 (gh-148395). The security fix itself (next_in = NULL at the error: labels in bz2 / lzma / zlib decompressors) is already merged on main and most release branches. No production code changes in this PR.

Closes the test gap for gh-148395 (does not replace the still-open 3.12 backport #148503, which is an RM action).

---

What the issue is

CVE-2026-6100: after a MemoryError inside {BZ2,LZMA,_Zlib}Decompressor.decompress(), the static helper left next_in pointing into the caller's released Py_buffer. The next decompress() call could memcpy through a freed pointer (attacker-controlled length and content) — CVSS 9.1.

The fix (8fc66ae and backports) nulls next_in on the error path. No branch shipped a regression test. A future refactor can reintroduce the UAF with a green CI.

---

Why I solved it that way

1. Tests only. Competing with [3.12] gh-148395: Fix a possible UAF in {LZMA,BZ2,_Zlib}Decompressor (GH-148396) #148503 or re-patching the C modules would be wrong — the fix is done; coverage is not.
2. support.nomemtest, not bare _testcapi.set_nomemory, so Py_TRACE_REFS builds skip cleanly (they fatal inside _PyRefchain_Trace on allocation failure).
3. Sweep start in range(0, 40) rather than a hard-coded allocation index — indices are unstable across pymalloc/ASAN/debug/bit-width.
4. bytearray(data) + max_length=0 — bytes can stay alive and hide the UAF; max_length=0 forces the !input_buffer_in_use tail-copy path that historically hit the failing PyMem_Malloc.
5. remove_mem_hooks() in finally — the hook is process-global across RAW/MEM/OBJ; leaking it destroys the rest of the suite.
6. Single-threaded only — not placed under test_free_threading (global allocator hooks + workers = flaky).
7. zlib gated on hasattr(zlib, '_ZlibDecompressor') for older-branch cherry-picks.

I did not fold in the optional sticky-error / success-path hardening from the issue discussion — that is a behavior change for a follow-up.

---

How I did it

File	Test
Lib/test/test_bz2.py	BZ2DecompressorTest.test_decompress_memoryerror_no_dangling_input
Lib/test/test_lzma.py	CompressorDecompressorTestCase.test_decompress_memoryerror_no_dangling_input
Lib/test/test_zlib.py	ZlibDecompressorTest.test_decompress_memoryerror_no_dangling_input

Shape (bz2; others mirror):

@support.nomemtest
def test_decompress_memoryerror_no_dangling_input(self):
    import _testcapi
    data = bz2.compress(b'x' * 4096)
    for start in range(0, 40):
        d = bz2.BZ2Decompressor()
        try:
            _testcapi.set_nomemory(start, start + 1)
            try:
                d.decompress(bytearray(data), max_length=0)
            except MemoryError:
                pass
        finally:
            _testcapi.remove_mem_hooks()
        try:
            out = d.decompress(bytearray(data))
        except (MemoryError, ValueError, EOFError, OSError):
            continue
        self.assertIsInstance(out, bytes)
        self.assertNotIn(b'\x00' * 64, out)  # cheap corruption tripwire

No NEWS entry (test-only; Security NEWS for the CVE already released).

---

Impact

• Locks in the CVE fix against silent regression.
• Negligible CI cost (40 short decompress attempts × 3 modules).
• Safe to backport to 3.15 / 3.14 / 3.13 with the fix. Do not push tests-only onto 3.12/3.11/3.10 security-only branches (and do not entangle [3.12] gh-148395: Fix a possible UAF in {LZMA,BZ2,_Zlib}Decompressor (GH-148396) #148503).

---

Testing plan

• ./python -m test test_bz2 test_lzma test_zlib on a debug build — all three new tests PASS with the fix present.
• Pre-commit / ruff on the test files.
• Ideal watch-it-fail (optional, for reviewer confidence): locally revert the three next_in = NULL lines, rebuild with ASAN (--with-address-sanitizer --without-pymalloc), confirm heap-use-after-free on the memcpy path, restore fix. I did not block this PR on an ASAN cycle; happy to attach traces if a reviewer wants them.
• CI: standard Linux/macOS/Windows test jobs.

---

Everything else

• Related PRs: gh-148395: Fix a possible UAF in {LZMA,BZ2,_Zlib}Decompressor #148396 (main fix, merged), [3.12] gh-148395: Fix a possible UAF in {LZMA,BZ2,_Zlib}Decompressor (GH-148396) #148503 (3.12 backport, RM-blocked — please do not apply the stale NEWS suggestions that drop _ZlibDecompressor).
• Assignee of the CVE work: StanFromIreland — requesting review as SME.
• Backports: needs backport to 3.15, 3.14, 3.13 after merge if desired.

---

Requested reviewers (subject-matter experts)

Could the following SMEs take a look when convenient (I cannot formally request reviews from this fork account):

• StanFromIreland — CVE fix author / issue assignee
• gpshead or security-interested core devs for test shape

Thank you!

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[bedevere-app]

Most changes to Python require a NEWS entry. Add one using the blurb_it web app or the blurb command-line tool.

If this change has little impact on Python users, wait for a maintainer to apply the skip news label instead.

[Vamsi-klu]

Added a Misc/NEWS.d/next/Tests/ blurb for these regression tests so bedevere/news is satisfied. (A Security NEWS entry already shipped with the CVE fix itself; this Tests entry only covers the new coverage.)

For pure test-only changes, skip news would also be appropriate if a maintainer prefers that instead.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cc8494c611

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Make the bz2 test detect the missing pointer reset

When the injected failure reaches the tail-buffer allocation (start == 2 on a --with-pydebug build), removing both bzs->next_in = NULL assignments from _bz2module.c makes the follow-up call raise OSError: Invalid data stream; this handler accepts that result, so the test still passes with the CVE fix reverted. Use a follow-up operation and explicit expected outcome that distinguish a cleared pointer from the dangling-pointer state rather than treating this decompression error as success.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Make the zlib test detect the missing pointer reset

When the injected failure reaches the tail-buffer allocation (start == 2 on a --with-pydebug build), removing both self->zst.next_in = NULL assignments from zlibmodule.c makes the follow-up call raise zlib.error for an incorrect data check; this handler accepts that result, so the test still passes with the CVE fix reverted. Exercise a post-error operation with an explicit safe result that differs when the stale pointer remains instead of allowing this error to satisfy the regression test.

Useful? React with 👍 / 👎.

[picnixz]

We donnot add tests for that as it may depend on many other factors and is CPython specific.

[picnixz]

Also avoid just letting an agent review everytjing and write everything. This is against our AI policy explained in the devguide.