[PierreRaybaut]

Reading of real SPEC and XDI files as the reference readers do, and validation corpora. Depends on DataLab-Platform/Sigima#109 (format registration API) and DataLab-Platform/SigimaX#23 (legacy NeXus axes, mapping axes on the first dimensions); DataLab side: DataLab-Platform/DataLab#444.

SPEC

• MCA spectra grouped by device (@A1, @A2...), each device with its own #@CALIB/#@CHANN. Before, devices were mixed and the device number was read as a first channel, shifting every spectrum by one channel (silx has the same behaviour).
• Labels separated by single spaces are split according to #N, or to the width of the data when #N is wrong (Bluesky). Repeated labels get a suffix (I0_1), as with spec2nexus.
• Empty #O0 lines (Bluesky) no longer crash. None values are NaN. Other @ lines and their continuations are ignored and counted.
• A scan that cannot be read is skipped with the error, in the import and in File > Open, instead of stopping everything.
• MCA spectra are imported as one image per device and scan (one spectrum per row, calibrated channels along X, X column along Y if strictly monotonic). Mesh scans are imported as images of the two motors (optional), with NaN for missing points. The scan header (#T, #M, #Q, #G, #C, #MD) is kept as SPEC ... metadata.

XDI

• UTF-8 free text, with a Latin-1 fallback.
• Invalid data tables are refused with their line, as the XDI library does. Invalid header lines are recorded (XDI problems). Comments keep their inner spaces.
• Every header field and the comments are kept as object metadata.
• Angles are converted to energies with Mono.d_spacing. Two-dimensional data (Outer.value) gives one signal per block.

NeXus (SOLEIL)

• The "Scan counters" mapping now also applies to the detectors of time scans, thanks to mapping axes describing the first dimensions.
• Step scans need no mapping: SigimaX reads their legacy axes. An integration test checks it.

Formats

• Formats are removed and registered again with Sigima's public API (unregister_format/register_format) instead of editing the registry's private list.

Validation (data next to the code, not in Sigima)

• tests/data/spec: extracts of real files from spec2nexus (CC BY 4.0, reductions listed) and PyMca (MIT), at pinned commits with SHA-256. They cover several MCA devices, mesh, single-space labels, Bluesky, Sardana, interrupted scans and repeated numbers. They are compared with the values read by spec2nexus (silx for the files spec2nexus cannot read), frozen in references.json by tests/validation/make_spec_corpus.py.
• tests/data/xdi: the test files of the XDI library (public domain), checked against BadFiles.txt.
• 202 tests passed; Ruff clean. On the full spec2nexus corpus (24 files), no crash, and the same values as silx except None/nan, which silx reads as 0.

Documentation

README, CHANGELOG (first version, unreleased), CONTRIBUTING (corpora and how to regenerate the references).