fix: list batch summaries, file set paths and record set ids in a stable order

[renato-umeton]

came up in review of #152: handler batches arrive in rglob order, which is whatever the fs gives us, so the same folder can bake differently on mac (APFS) vs linux (ext4). #152 tests use single value batches on purpose to dodge this, so this is the actual fix.

before, same 3 dicoms, two orders:
3 DICOM files (512x512): MR (1), CT (2)
3 DICOM files (512x512): CT (2), MR (1)
after, always CT (2), MR (1).

sorted per handler, same as image_handler does for format_counts (#140 also gives wsi flavors a fixed order). sorting discovery itself would renumber every FileObject id and reshuffle all goldens, and the e2e tests already compare via _discovery_independent, so discovery stays as is.

fixed:

• dicom modality counts sorted, files w/ no modality pinned last as "unknown"
• nifti dtypes sorted
• fhir chunk + bundle FileSet includes sorted by path
• the stored forms of a listed path (x.ndjson + x.ndjson.gz) and linked dupes riding w/ a FileSet were in rglob order, now sorted by path
• record set ids: lab results.csv + lab@results.csv both sanitize to lab_results and the __2 went to whichever came first. make_record_set_ids (csv, json, fhir) and wfdb now allocate in path order like allocate_record_set_ids already does

already stable: image/OME, parquet, soft/spreadsheet/hdf5 ids, dataset description, references, scan report, duplicate primary pick. FileObject ids and record set order still follow discovery on purpose.

also 9dd8250 + 2ba97dc: dicom counted files w/ no Modality under the key "unknown", so a file whose Modality literally reads "unknown" got its count overwritten by the missing ones (3 files, counts summed to 2). missing ones now sit under None, still last, counts add up to num_files
they still print as "unknown" so normal batches read the same, no golden changes. only if some file says unknown (any case) do the missing ones print as "no modality", e.g. "CT (1), unknown (1), no modality (1)"

ties in the fhir type vote are also order dependent, left for a follow up PR.

each fix has a test that reverses the batch or discovery and fails on main. no golden changed. merges clean w/ #152, #140, #148.

@slobentanzer pls take a look when u get a chance.. thx!

[slobentanzer]

Hi @renato-umeton,
good catch, I just recommend minor efficiency changes:

1. Duplicated code in utils.py. The new disambiguate_in_path_order repeats the four lines at the end of allocate_record_set_ids (the order = sorted(...) block at utils.py:229-232). allocate_record_set_ids should call the new helper, so path-order allocation lives in one place.
2. Pick one label for DICOM files with no Modality. Right now those files print as "unknown", or as "no modality" if another file in the batch says "unknown". So the label depends on the other files. The PR keeps "unknown" so that goldens stay the same, but no golden or existing test contains "unknown (". Could instead (simpler): always print "no modality", remove the clash check, and remove the uppercase UNKNOWN test.
3. Repeated test setup. The same monkeypatch that reverses scan.discover_files appears five times in four test files. It should be one fixture in tests/helpers.py or conftest.py.
4. Local import in test_nifti_handler.py. test_a_mixed_dtype_bake_describes_itself_one_way imports bake inside the function, with a comment saying this avoids a merge conflict with #152. Once both are merged it should be a top-level import and the comment should go. Follow-up issue?

[renato-umeton]

thx @slobentanzer, all good points, done:

1. allocate_record_set_ids now gets its bases from disambiguate_in_path_order, so the bases come from one place. the derived ids loop still walks path order (1 line, needed for __2). same ids, goldens untouched 28347af
2. files w/ no Modality always print "no modality". clash check + UNKNOWN test gone; None key kept so a real "unknown" still counts on its own and counts add up. checked, no test or golden here or in feat: digital pathology whole-slide image formats (SVS, NDPI, SCN, BIF, QPTIFF, DICOM WSI) #140 test: raise line coverage of croissant_baker to 96% #148 fix: drop source.extract from dicom and nifti header fields #152 test: refresh the rai golden and compare it in full #155 has "unknown (" e9d87a5
3. one reverse_discovery fixture in tests/conftest.py (where dataset lives), used by the 5 copies 4231a58 and by the ones already on main (golden, ome, spreadsheet) d0b81fd, w/ readable ids (as-walked / reversed) 152feae
4. opened tests: top level bake import in test_nifti_handler + last reversed discovery copies #159 for 4, plus the 2 copies still coming in w/ feat: digital pathology whole-slide image formats (SVS, NDPI, SCN, BIF, QPTIFF, DICOM WSI) #140 and test: refresh the rai golden and compare it in full #155

trial merged w/ #140 #148 #152 #155, all clean + green. thx!

[slobentanzer]

@renato-umeton looks good, thanks :)