fix: allow categorical index in read_lazy

[ilan-gold]

Fixed a bug reported by Jan. We do have a warning about setting non-string indices, but it is not disallowed, in non-lazy AnnData. So this fix matches the behavior for read_lazy.

• Closes #
• Tests added
• Release note added (or unnecessary)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.57%. Comparing base (c72dfff) to head (2e14d4f).
⚠️ Report is 18 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2062      +/-   ##
==========================================
- Coverage   87.60%   85.57%   -2.03%     
==========================================
  Files          46       46              
  Lines        7052     7059       +7     
==========================================
- Hits         6178     6041     -137     
- Misses        874     1018     +144     

Files with missing lines	Coverage Δ
src/anndata/_io/specs/lazy_methods.py	96.31% <100.00%> (+0.16%)	⬆️

... and 7 files with indirect coverage changes

[flying-sheep]

We do have a warning about setting non-string indices, but it is not disallowed, in non-lazy AnnData

which isn’t a great thing. do we raise an error when writing that or …?

As said: UUIDIndex or (barring that) AnonymousIndex makes sense, RangeIndex or Index(dtype=np.int*) don’t.

[ilan-gold]

which isn’t a great thing. do we raise an error when writing that or …?

We have a warning about it:

anndata/src/anndata/_core/anndata.py

Lines 785 to 800 in c72dfff

	if (
	len(value) > 0
	and not isinstance(value, pd.RangeIndex)
	and infer_dtype(value) not in {"string", "bytes"}
	):
	sample = list(value[: min(len(value), 5)])
	msg = dedent(
	f"""
	AnnData expects .{attr}.index to contain strings, but got values like:
	{sample}
	Inferred to be: {infer_dtype(value)}
	"""
	)
	warnings.warn(msg, stacklevel=2)
	return value

I think this fix falls under the "people can in theory write it to disk, so we should read it" category. If we want to start disallowing this behavior in some scope (writing? in-memory? both?), I think that's a separate conversation.

[flying-sheep]

I see, one more instance of “some old stuff where we weren’t strict enough”. This is some old stuff from #300.

I say we shouldn’t finish enshrining this as a feature. We say we don’t support numeric indices, so we shouldn’t betray that by allowing more and more support to creep in.

Instead we should do something like converting to string on read (for old wrong-written indices) and write (to no longer write wrong indices) – with a warning each. Something that keeps things from breaking but doesn’t allow an unsupported index type to survive a full write-read cycle.