Force bool to float32, add explicit error message

[tjgalvin]

Beth, of GASKAP-OH fame, pointed out two issues.

The first was that in some circumstances a bool mask array is passed for fits.writeto, which subsequently raises an error. I don't know how long this one has been around for - I feel like I should have seen it long ago. We may have overlooked something when removing the logic from flint to eye-patch. An easy fix anyway.

The next was around how the gaussian kernels are constructed in the beam erosion stage. It is possible for the scale to be so small that a gaussian kernel can not be created. The sigma width is so small that it is well within a single pixel, which then breaks the coordinate system we evaluate the kernel over. In this case I have added a more informative error as there is really nothing I think we ought to be doing here. This error would only be tripped for a very small scale, like --beam-shape-erode-scales 1, otherwise no change.

[tjgalvin]

I also

• added a logger sub-module like our other packages
• copied over the flint precommit yaml. This one was complaining about an older version of python not being compatible with the type alias we are using

[AlecThomson]

"Consider removing it."

What does "it" mean here? Can we make this message more explicit on what to change?

[AlecThomson]

Why are these removed?

[AlecThomson]

Suggested change

	files: flint|tests
	files: eye_patch|tests

Bad copy-paste

[AlecThomson]

@tjgalvin - can we revert the pre-commit changes? I'd rather address more explicitly, I think its a numpy / mypy / python version clash.

We're also running to an issue I've hit bunch lately with my CD.yml

The fix is to bump to hynek/build-and-inspect-python-package@v3.0.1 Lemme know if you like me to sort that

[AlecThomson]

@tjgalvin heads up — I've pushed to beth-fix on Alec's behalf, since the pre-commit revert he asked about is tangled up with the CI fix. Nothing of yours was dropped, and no history was rewritten (it's a merge, so your checkout stays valid). Three things:

1. Reverted .pre-commit-config.yaml to main's version. The copied flint config regressed a few things in this repo — most importantly files: flint|tests, which points mypy at a package that doesn't exist here, so it was checking nothing. It also dropped the exclude for the copier answers, downgraded blacken-docs to black 24, commented out the prettier and pygrep hooks, switched ruff-check to the deprecated ruff alias, and dropped the numpy/capn-crunch/pytest-stub deps mypy needs to resolve this package's types.

2. The version clash is fixed explicitly in #15, which is merged into this branch so it can go green now. You were right that it's a numpy / mypy / python clash — the specifics:

requires-python is >=3.11, but [tool.mypy] python_version was still "3.10", left over from the cookie template. The Format job runs on 3.x (3.14), where numpy resolves to 2.5.x, and numpy 2.5's stubs use PEP 695 type statements that mypy won't parse below 3.12. Hence the error you were seeing.

Rather than raise the target to 3.12, the hook is pinned to language_version: python3.11, so numpy resolves to 2.4.x whose stubs parse at 3.11 — mypy now actually verifies the minimum version we claim to support.

3. Annotated CustomFormatter.format. Once mypy points at eye_patch again, disallow_untyped_defs applies, so it needed record: logging.LogRecord) -> str.

On the bool mask

Your diagnosis was right and it's a real bug — mypy caught it independently the moment it targeted a supported Python:

masking.py:910: error: Argument "mask" to "beam_shape_erode" has incompatible type
ndarray[..., dtype[numpy.bool]]; expected ndarray[..., dtype[floating]]

One refinement: #15 moves the cast to where the flood-fill mask is created rather than at fits.writeto. reverse_negative_flood_fill returns bool, so the bool array was already wrong one call earlier, going into beam_shape_erode. Casting at the source fixes both, and since beam_shape_erode takes its return dtype from its input, it also removes the bool return that function could previously produce.

Your .astype(np.float32) at the write is kept. It's no longer what saves us from the bool, but the multi-scale path of beam_shape_erode returns float64, so it's what keeps the mask float32 on disk.

Your kernel-too-small ValueError is untouched — nice catch, that one would have been baffling to debug from the linspace failure alone.

---

Generated by Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 88.88889% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
eye_patch/masking.py	25.00%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!

[AlecThomson]

Brought main into this branch — #15 (CI/CD fix), #13 (actions bump) and #8 (pre-commit hook bumps) have all landed, so it had gone stale and had a conflict.

One conflict, in tests/test_masking.py. This branch held an earlier iteration of the flood-fill tests that subsequently landed on main via #15. Resolved in favour of main's version, which is a strict superset — the same test, now annotated, plus the fixture and beam-erosion test added there. Nothing from here was lost: the PR never touched the tests itself, so all of that content had come from my earlier merges. A merge commit, so no history rewritten and your checkout stays valid.

The diff is now just your three changes, which is what it should have been all along:

eye_patch/logging.py	the logging sub-module
masking.py	.astype(np.float32) at the write
masking.py	the kernel-too-small ValueError

Worth calling out that the logger change fixes a bug nobody had noticed: masking.py had

logger = logging.getLogger("__name__")

That's the string "__name__", not the __name__ variable — which is why the CI logs read INFO __name__:masking.py:780. Every module sharing that name shared one logger. Your sub-module replaces it properly.

22 tests pass and pre-commit is clean locally; CI is re-running on the merge.

---

Generated by Claude Code