Ship prebuilt abi3 R-tree wheels so users don't need a C compiler

[tlambert03]

Closes #56 for the R-tree half.

spatial_graph JIT-compiles its extensions at runtime via witty, so every end
user needs a working C compiler. That's what's been breaking Windows users of
motile_tracker (via tracksdata).

The graph half genuinely can't be prebuilt — node/edge attribute dtypes are only
known when you construct the graph. The R-tree is different: its compile
parameters are just an item dtype, a coordinate dtype, the dimensionality, and
the tree class. That's a small, enumerable set, so those variants are now
compiled ahead of time and shipped in the wheels. witty remains as the
fallback for anything not prebuilt, and for the graph.

This does it inside spatial_graph rather than extracting a standalone
rtree-nd package, since the split turned out not to be necessary for the goal.

What ships

Wheels are cp311-abi3 — one per platform, covering every supported
CPython, verified end to end (a wheel built on 3.11 is installed and tested on
3.12/3.13/3.14 in CI). 8 wheels per release: manylinux + musllinux × x86_64 /
aarch64, win_amd64 + win32, macOS x86_64 + arm64. ~2.7 MB each, 32 prebuilt
modules:

classes	PointRTree, LineRTree
item dtypes	int64, uint64 (int64[2] / uint64[2] for LineRTree)
coord dtypes	float32, float64
dimensions	2, 3, 4, 5

tracksdata only ever constructs PointRTree(item_dtype="int64", coord_dtype="float32", dims=N) with N of 2–4, so it is fully covered.

How it works

RTree.__init__ looks up a prebuilt module first and falls back to witty if
there isn't one:

• _rtree/_codegen.py — the variant matrix and the pyx renderer. setup.py and
  the JIT path both call build_wrapper(), so the two can't drift.
• _rtree/_naming.py — deterministic module names, e.g.
  rtree_int64_float_d3_df3f6198. Dtypes are canonicalised through
  DType.base_c_type (so int/int64 and float32/float collapse to one
  module) and the digest covers the C/pyx a subclass injects — so an RTree
  subclass with custom code can never be served a module built from different
  code, and falls through to JIT instead.
• _rtree/_prebuilt/ — populated at build time; empty in a source checkout.

The compiled modules have no numpy C-API dependency (the wrappers use plain
buffer-protocol memoryviews), so the wheels need no numpy ABI pinning.

Why 3.11, and why 3.10 had to go

Py_buffer and the PyBUF_* flags only entered the limited API in 3.11 —
before that they live in cpython/object.h, which is excluded under
Py_LIMITED_API. Cython's typed memoryviews compile straight down to
PyObject_GetBuffer/PyBuffer_Release, so 3.11 is a hard floor for a
stable-ABI build.

This is the one user-visible consequence: requires-python is now >=3.11,
so 3.10 users will stay on the last release. (3.10 reaches EOL in October.)
Everything else should be invisible — same API, same behaviour, witty and
CT3 are still unconditional dependencies. What changes is that they no longer
get invoked for the common cases, which is what actually required a compiler.

Other changes

• Build backend hatchling → setuptools. Hatchling has no first-party
  C-extension support, so this otherwise meant hand-driving build_ext from a
  custom hook. hatch-vcs → setuptools-scm is like-for-like.
• cibuildwheel for release wheels. uv build alone emits linux_x86_64,
  which PyPI rejects, so the previous deploy path could not have shipped
  compiled wheels at all.
• CI tests the built wheel, not an editable install (uv sync --no-editable), and SPATIAL_GRAPH_REQUIRE_PREBUILT=1 makes a build that
  silently degrades to pure Python fail instead of going green.
• import spatial_graph no longer pulls in the compilation toolchain: ~35 ms,
  down from ~88 ms.
• Sdist installs still work without a compiler — setup.py probes for one and
  falls back to a pure-Python wheel that JIT-compiles on first use, exactly as
  before.

Testing

tests/test_prebuilt.py blocks witty.compile_cython and constructs all 32
declared variants, asserting the real property directly: prebuilt variants
must not invoke the compiler. That runs on every OS and Python in the matrix.
Plus: dtype aliases share a module, unlisted combinations fall back to JIT, and
subclasses with custom C aren't served a prebuilt module.

Not included (deliberate)

• int32/uint32 item dtypes — those still JIT.
• The cross-version abi3 check runs against the Linux artifact only; other
  platforms' wheels are tested on 3.11 by cibuildwheel.
• No [jit] extra — witty/CT3 stay required, they're just not invoked.

Follow-up

The conda-forge feedstock will keep working via the JIT path, but won't deliver
the benefit until it's converted from noarch to per-platform builds.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.71%. Comparing base (caab135) to head (06df3db).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #57      +/-   ##
==========================================
+ Coverage   95.11%   95.71%   +0.59%     
==========================================
  Files          12       14       +2     
  Lines         389      443      +54     
==========================================
+ Hits          370      424      +54     
  Misses         19       19              

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[codspeed-hq]

Merging this PR will improve performance by 29.94%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 1 improved benchmark
✅ 5 untouched benchmarks

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_roi_query_performance[100000]	1.9 ms	1.4 ms	+29.94%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing tlambert03:prebuilt-abi3-proto (06df3db) with main (e655613)}