fix(concat_on_disk): retain aligned-mapping keys under join="outer"

[alejandro-publius]

• Closes concat_on_disk with join='outer' doesn't retain all .obsm fields #2394
• Tests added

Problem

concat_on_disk writes the aligned mappings (.obsm/.varm and .layers) over the intersection of their keys, regardless of join. Under join="outer" this silently drops any mapping key present in only some of the inputs — unlike in-memory anndata.concat, which keeps the union and fills the missing entries. When the retained keys additionally have differing shapes, the dense path raises ValueError: Shapes do not align instead of padding.

import anndata as ad, numpy as np, pandas as pd
from scipy import sparse

a = ad.AnnData(sparse.csr_matrix((3, 5)), obs=pd.DataFrame(index=[f"a{i}" for i in range(3)]),
               obsm={"X_emb": np.zeros((3, 2))})
b = ad.AnnData(sparse.csr_matrix((2, 5)), obs=pd.DataFrame(index=[f"b{i}" for i in range(2)]))
a.write_h5ad("a.h5ad"); b.write_h5ad("b.h5ad")

ad.concat([a, b], join="outer").obsm            # {'X_emb': (5, 2)}  ✅
ad.experimental.concat_on_disk(["a.h5ad", "b.h5ad"], "out.h5ad", join="outer")
ad.read_h5ad("out.h5ad").obsm                   # {}   ❌ X_emb silently dropped

Fix

_write_concat_mappings now mirrors the in-memory outer_concat_aligned_mapping:

• takes the union of keys for join="outer" (intersection for "inner", unchanged);
• generates outer reindexers so each key is padded along its free (feature) axis to the max width;
• fills keys missing from some objects with fill_value (promoting integer dtypes to float where the fill is NaN), by materializing just those keys through the same concat_arrays path the in-memory concat uses.

Keys present in every object that need no reindexing keep the existing memory-efficient streaming path, so the common case is unchanged. As a side effect this also fixes the inner-join case where differing .obsm widths previously raised.

Tests

• test_concat_on_disk_outer_mapping_missing_keys — .obsm with dense / sparse / dataframe values, one key missing from an object and shared keys of differing widths.
• test_concat_on_disk_outer_layers_missing_keys — a .layers key missing from one object with a differing var axis (exercises the reindexed path).

Both assert the on-disk result equals in-memory concat. The full tests/test_concatenate_disk.py suite passes (498 tests).

[codecov]

Codecov Report

❌ Patch coverage is 93.75000% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.86%. Comparing base (c37e213) to head (8f2b03b).
⚠️ Report is 4 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/anndata/experimental/merge.py	93.75%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2558      +/-   ##
==========================================
- Coverage   87.81%   85.86%   -1.96%     
==========================================
  Files          49       49              
  Lines        7799     7833      +34     
==========================================
- Hits         6849     6726     -123     
- Misses        950     1107     +157     

Files with missing lines	Coverage Δ
src/anndata/experimental/merge.py	90.23% <93.75%> (-0.76%)	⬇️

... and 8 files with indirect coverage changes

[ilan-gold]

Why do these need to be brought into memory?

[ilan-gold]

Is it to avoid rewriting _write_concat_sequence to handle missing elements?