fix: merge strategies for concat_on_disk

[ilan-gold]

I'm putting this in the soonest upcoming milestone because

a) I think it's a bug that we didn't have any sort of warning or anything about this and
b) It's in experimental and this isn't even an API change

As for the nature of the changes, I still need to look into why read_elem_lazy produces different results in tokenization for dask dense than read_elem on the off-axis. Otherwise, this could be fully lazy Bug resolved

• Closes Concatenating on disk doesn't respect the merge argument. #2110 and part of concat_on_disk fails to write alternative axis mapping and uns #1854
• Tests added
• Release note added (or unnecessary)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.62%. Comparing base (5080856) to head (c95898f).
⚠️ Report is 2 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2122      +/-   ##
==========================================
- Coverage   85.70%   85.62%   -0.08%     
==========================================
  Files          46       46              
  Lines        7078     7083       +5     
==========================================
- Hits         6066     6065       -1     
- Misses       1012     1018       +6     

Files with missing lines	Coverage Δ
src/anndata/_core/merge.py	85.11% <100.00%> (+0.08%)	⬆️
src/anndata/experimental/merge.py	87.73% <100.00%> (+0.53%)	⬆️

... and 3 files with indirect coverage changes

[ilan-gold]

So you might wonder why completely ignoring an argument and also writing to the wrong key for a given function wouldn't produce an error previously. Crazily enough it was that

1. alt_mapping was always {} here because merge is always None in the tests
2. We then write a dataframe (because not {} is True) to the wrong key alt_axis_name instead of f"{alt_axis_name}m" but its index is correct
3. We never actually checked the off-axis dataframe in the tests because merge was None on the groundtruth as well for anndata.concat so all of the other columns were dropped in the groundtruth concatenation

[flying-sheep]

I’m confused: isn’t the behavior of this PR right now

• raises NotImplementedError when set to True
• {obs,var}p are merged when it’s set to false?

doesn’t seem to make sense to me!

Shouldn’t this PR instead

• document this as working
• remove the NotImplementedError
• wrap the call to _write_alt_pairwise in if pairwise?

[ilan-gold]

No, I wanted to match the behavior of anndata.concat here so this PR raises NotImplementedError only for the pairwise mappings along the concatenation axis, but just like anndata.concat automatically merges the off-axis.

[flying-sheep]

Ah I see _write_alt_pairwise is about the off-axis. Makes sense!

But the rest is still valid:

anndata/src/anndata/experimental/merge.py

Lines 580 to 582 in 1d4cb87

	if pairwise:
	msg = "pairwise concatenation not yet implemented"
	raise NotImplementedError(msg)

… means that “[pairwise] is ignored when True.” is a lie. It’s not ignored, it throws an error.

[ilan-gold]

I've updated the message now to hopefully be clearer