[PierreRaybaut]

Problem

_is_uniform() (introduced in 589aafa, part of #13, unreleased 1.1.0) accepted a deviation from even spacing of 4 * eps * max|x|, i.e. 4 to 8 ULP of the axis data type. A genuinely non-uniform float32 axis with a large origin was therefore flattened to uniform coordinates, silently altering the spatial calibration of imported NeXus images and mapped (HDF5 profile) images.

Reproduced case: float32 axis [1e6, 1e6 + 1, 1e6 + 2.25, 1e6 + 3] (steps 1, 1.25, 0.75; deviation 4 ULP = 25 % of the step) was imported as uniform, the third coordinate becoming 1e6 + 2.

Fix

• An image gets uniform coordinates only if both axes deviate from even spacing by at most 2.5 ULP of their stored data type (sigimax.h5.nexus.UNIFORM_AXIS_TOLERANCE), evaluated at the largest axis magnitude (float64 ULP for non-float axes). Otherwise the stored coordinates are kept exactly.
• The maximum deviation of each declared axis converted to uniform coordinates is recorded in experimental_source conversions, e.g. axis 'x': uniform coordinates (max deviation 0.000195, 0.4 ULP of float32). Lossless conversions (zero deviation, dimensions without axis) are not recorded.
• Non-finite axis values still lead to explicit coordinates (NaN-safe comparison).

Rationale for 2.5 ULP

Measured on 3000 random evenly spaced float32 axes: linspace, decimation and bin edges → centers deviate by at most 1.0 ULP; float32 arithmetic (x0 + i * dx) by at most 2.1 ULP. A step-relative guard (e.g. ≤ 1 % of the step) was considered and rejected: it would have turned about 15 % of genuinely uniform float32 axes (small step relative to the origin, e.g. 8000 eV / 0.1 eV) into non-uniform coordinates, which Sigima binning and resampling refuse.

Known limit: a genuinely non-uniform axis deviating by ≤ 2.5 ULP of its storage type cannot be distinguished from a uniform axis computed in that type; it is converted, but its deviation is now recorded.

Tests

test_nxdata_image_axis_precision:

• reproduced case → explicit coordinates, stored values preserved exactly, nothing recorded;
• float32 linspace(8000, 8100, 1001) → uniform, deviation recorded (0.4 ULP);
• float32 arithmetic axis -50 + i * 0.37 (1.87 ULP) → uniform, guarding against an over-tight criterion.

The new test fails on the previous implementation. sigimax/tests/hdf5: 309 passed; Ruff and Pylint clean; Sphinx gettext build with -W OK.