Raise OverflowError when use_single_float cannot represent a value

[dylanpulver]

Fuzzing the Cython packer against the pure Python fallback over 4000 random objects, packb disagreed on 38 of them. Every disagreement was the same case: use_single_float=True with a value above FLT_MAX.

# C extension
>>> msgpack.packb(1e300, use_single_float=True).hex()
'ca7f800000'          # +inf

# MSGPACK_PUREPYTHON=1
>>> msgpack.packb(1e300, use_single_float=True)
OverflowError: float too large to pack with f format

Two things kept the C path quiet. _packer.pyx cast the value with <float>o before the call, so msgpack_pack_float only ever saw a narrowed float. And msgpack_pack_float discarded the return value of PyFloat_Pack4, which is what reports the overflow.

Passing the double through and checking that return value covers both. PyFloat_Pack4 backs struct.pack('>f', ...), so the exception type and message now match the fallback exactly.

Out-of-range integers are handled this way today: packb(2**64) raises PackOverflowError from either packer, pinned by test_limits.py::test_integer. The new test sits next to it.

Reverting the source makes the new test fail on the C extension. It passes either way under MSGPACK_PUREPYTHON=1, where it pins existing behaviour. After the change the same fuzz run reports 0 of 4000 disagreements on packb.

Infinity is representable in single precision and packs unchanged, which the test also covers.

[dylanpulver]

The one red job (3.15t on windows-latest) does not come from this change.

Its step Test (C extension) in parallel under free-threading passed, and that is the path this PR touches. The job then spent 46 minutes inside Test (pure Python fallback) in parallel under free-threading and was killed there.

That stall reproduces on e94e1dc with no patch applied. MSGPACK_PUREPYTHON=1 pytest --parallel-threads=auto --iterations=20 test parks on test_case.py::test_array16 on an 18 core machine. At --parallel-threads=4 the same suite finishes in about 25 seconds on both e94e1dc and this branch. auto scales with runner CPU count, so the cost of test_array16 and test_array32 under the pure Python packer grows with the size of the runner.

A re-run should come back green.

[Copilot]

Pull request overview

Aligns C-extension single-float packing with the Python fallback by raising OverflowError for unrepresentable finite values.

Changes:

• Preserves double precision until PyFloat_Pack4.
• Propagates packing overflow errors.
• Tests float boundaries and infinities.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
msgpack/_packer.pyx	Passes doubles to single-float packing.
msgpack/pack_template.h	Checks and propagates PyFloat_Pack4 errors.
test/test_limits.py	Covers FLT_MAX, overflow, and infinities.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.