[Hendrik-code]

Stacked on #77 (fix/clearer-cli-messages), whose --input validation this builds on. The diff
shown here is just the incremental change; merge #77 first and the base switches to main automatically.

Problem

Plain .nii is what dcm2niix writes without -z y, and what a good share of public datasets ship.
SPINEPS could not read it at all:

• the CLI appended .nii.gz to whatever it was given, turning scan.nii into scan.nii.nii.gz,
• BIDS_FILE.file is keyed by extension, so the three img_ref.file["nii.gz"] lookups raised KeyError
  for a .nii input, and
• process_dataset's query filtered on filter_filetype("nii.gz"), so .nii scans were invisible to
  dataset mode.

A user with a .nii dataset had to gzip all of it first.

Fix

• A shared input_image_path(img_ref) in seg_utils resolves nii.gz or nii, used at all three
  lookups, with NIFTI_FILE_TYPES as the single source of truth.
• process_dataset passes both extensions to filter_filetype.
• --input and spineps.segment() accept both; the extension is still appended when --input is given
  without one. A genuinely unreadable extension (scan.mha) still gets the clear error from fix: make CLI messages name flags that exist and causes users can act on #77, and
  that PR's test is retargeted accordingly.
• Outputs are unchanged: always .nii.gz, since get_changed_path(file_type="nii.gz") sets them.

Tests

unit_tests/test_plain_nii_input.py: input_image_path on both extensions, the compatibility check on a
.nii, a full segment_image run on sub-01_T2w.nii that asserts all four outputs land on disk as
.nii.gz, and run_sample not rejecting a .nii. The file does not even import on the base branch.

Full suite: 233 passed, ruff clean.

🤖 Generated with Claude Code

[Hendrik-code]

No CI here on purpose: the tests workflow only triggers for pull requests whose base is main, and this one targets fix/clearer-cli-messages (#77). GitHub retargets it to main when #77 merges, and CI runs then. Verified locally in the meantime: 233 tests pass on this branch, and 284 pass in a tree with all ten PRs of this round merged together.