One BIDS intent engine for dcm2niix front ends #516

Problem

dcm2niix converts each series of a DICOM session on its own. This makes conversion
fast and deterministic, but dcm2niix cannot use information across series to find
the purpose of each one. For example, a reverse phase-encoding (PEpolar) fieldmap
can be identical to a DWI acquisition.

Today three separate tools try to find that purpose:

• reproinx.py (Python), maintained by @neurolabusc
• the JavaScript script behind BIDSvue, maintained by @neurolabusc
• the dcm2niix web app in this repository (apps/dcm2niix, PR Add dcm2niix app (DICOM to BIDS) #382), which
  suggests a type for each series and lets the user confirm it

Each tool needs constant changes as sequences, DICOM and BIDS change. Three code
bases triple that work, and they can disagree on the same data.

Proposal

Make one engine that takes dcm2niix's per-series outputs (JSON sidecars,
.bval/.bvec, image dimensions) and returns a BIDS plan: datatype, suffix,
entities, IntendedFor, and B0FieldIdentifier/B0FieldSource.

• Write it in Rust, so the same code runs as a native command line and as
  WebAssembly in the browser. This is how other tools in this repository work
  (see exes/).
• Use ReproIn protocol names (https://npacore.github.io/reproin-namer/) when
  they are present, and use rules from the metadata when they are not.
• Keep the user's review step in the web app. The engine suggests, and the user
  confirms.

Open questions

1. Where does the code live: in this repository (exes/ and packages/), in a
   separate repository, or in the dcm2niix repository next to reproinx.py?
2. Should the engine replace reproinx.py and the BIDSvue script, or start as a
   third implementation that they can move to later?
3. Which BIDS datatypes come first? The web app now does anat (T1w, T2w, FLAIR),
   dwi and PEpolar fieldmaps. BIDSvue also handles physiological data and MRS.

Regression data

The BIDSvue tutorials include validation DICOM datasets
with advanced cases (physiological data, MRS and others). Pin them on the
Hugging Face dataset and use them as regression tests for the engine, in the
same way as the pinned neurolabusc/dcm_qa_philips_dwi data that the dcm2niix
app's browser tests use now.

Context: #382 (comment) (comment by @neurolabusc, 2026-10-20)

@neurolabusc, what do you think of this plan, and where would you like the code to live?