feat(build): support nanobind 3, widening the bound to <4

[Ravenwater]

Closes #85.

The <3 cap added in #84 — when nanobind 3.0.0 released and broke every wheel build — was a hold, not a fix. This is the port, and it keeps both majors working rather than trading one for the other.

nanobind	result
2.15.0	1374 passed, 3 skipped
3.0.0	1374 passed, 3 skipped

Both are clean configure → full build → link → full suite. The link matters: when #85 was filed, the build had never got past the failing object, so link and runtime compatibility were explicitly unverified. They are now.

The change

nanobind 3 removed the free nb::detail::keep_alive(nurse, patient) and routes the operation through a backend slot that must be handed the extension's context pointer. That pointer is available to extension code — NB_CTX is only unavailable inside nanobind's own build, under NB_BUILD — so this is a respelling, not a redesign.

mtl5_ndarray.cpp now spells it once, in keep_view_alive(), selecting on NB_VERSION_MAJOR. That's why the bound moves to <4 instead of jumping to >=3.

The obvious simplification is wrong

Worth stating plainly, because the compiler suggests it and a reviewer will think it:

replace the three manual calls with the declarative nb::keep_alive<0, 1> annotation

That compiles, and it passes the suite as it existed. It is still wrong. The annotation applies to every return of a .def, and all three sites return a view on one path and a copy — or, for a fully-integer index, a scalar — on another:

method	aliasing path	other path
__getitem__	slice → view	all-int index → scalar
reshape	C-contiguous → view	otherwise → copy
ravel	C-contiguous → view	otherwise → copy

Pinning a large source array to an independent copy of itself is a memory-retention bug. The pin has to stay per-path, so the manual call stays.

Two lifetime-test gaps, closed

Both would have survived a broken keep-alive, which is why the port needed them:

test_a_view_of_an_OWNING_array_keeps_it_alive — the existing lifetime test derives its views from asarray(), where the view also carries the NumPy owner object. The buffer survives on that reference alone, so the keep-alive is never load-bearing there. An array from copy() owns its memory with no second owner — that is the configuration where dropping the parent is a genuine use-after-free. Covers all three manual sites.

test_a_copy_does_not_pin_its_source — asserts the aliasing/copying split directly, via refcount:

view path:  getrefcount(src) 2 -> 3     (pinned)
copy path:  getrefcount(nc)  2 -> 2     (not pinned)

This is the test that fails if someone later folds the manual calls into the annotation.

Version alignment

python/CMakeLists.txt's FetchContent fallback moves v2.15.0 → v3.0.0 to stay in step with what the bound resolves. #84's lesson was that letting those two drift is exactly how a source change passes locally and fails every CI job.

The upper bound stays. nanobind 3.0.0 broke every build the day it released precisely because there wasn't one.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

• Run on-demand review

_{On-demand reviews are free for the next 24 days. After that, they cost $0.25 per reviewed file.}

Or wait 23 minutes for your next included review.

View limit details

Limit details: You’ve used the included review currently available. Your 74 included PR review attempts over the past 7 days set your current allowance at 1 review per hour.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 12767848-7d99-4b7e-b589-f5179db182b8

📥 Commits

Reviewing files that changed from the base of the PR and between 28f809c and 3e62bac.

📒 Files selected for processing (5)

• CHANGELOG.md
• pyproject.toml
• python/CMakeLists.txt
• python/src/mtl5_ndarray.cpp
• tests/test_ndarray.py

---

_{Comment @coderabbitai help to get the list of available commands.}