fix(intensity): report first-order features in the slide's own intensity domain

[darkclad]

Report first-order intensity features in the slide's own intensity domain

Closes the CT/Hounsfield defect reported against 0.11.0.post1.dev76: a CT NIfTI read with no
flag set came back in an arbitrary offset domain rather than in Hounsfield units.

The defect

NiftiLoader::unhounsfield read a minimum off the buffer it was loading and shifted the whole
volume by -min whenever that minimum was negative — which every CT's is. Nothing recorded the
shift and nothing undid it, so MEAN, MIN, MAX, MEDIAN, the percentiles and ENERGY were
displaced by an image-dependent amount. --preserve-hu shifted by a different base (the
within-mask minimum), so the two settings disagreed with each other as well as with the truth, and
only the intensity-histogram family inverted anything.

Shift-invariant features (RANGE, VARIANCE, STDDEV) were correct in both modes, which is what
made it easy to miss: the whole feature set was self-consistent, just displaced.

Measured on the reporter's own image and mask, with no flag:

feature	before	after	truth (PyRadiomics / SimpleITK)
3MEAN	2371.901625	-607.098375	-607.098375
3MIN	1966	-1013	-1013
3MAX	3810	831	831
3MEDIAN	2320	-659	-659
3RANGE	1844	1844	1844
3VARIANCE_BIASED	30067.203284	30067.203284	30067.203284
3ENERGY	3.314212e+13	2.335867e+12	2.335867e+12

preserve_hu=True now produces the same numbers instead of differing by 1966.

The fix

The load-time map is recorded per slide and undone on the way out.

• SlideProps carries inten_scale / inten_offset / inten_map, filled by
  Nyxus::record_intensity_domain_map at the end of scan_slide_props. Three maps: the offset map
  (shift by the floored all-pixel minimum, only when the slide holds a negative value), the
  min-max quantization of a real-valued slide into [0, --fpimgdr], and native.
• Both intensity families apply the inverse in report_in_source_domain: affine for the location
  statistics, the scale alone for the dispersions, nothing for the standardized moments and the
  histogram shape, and a recompute from the mapped parts for COV, QCOD and UNIFORMITY_PIU.
  ENERGY is reconstructed from the two sums already on hand, so there is no second pass over the
  pixel cloud. The identity map returns early, which leaves every ordinary non-negative integer
  image byte-identical.
• No loader reads a minimum of its own any more — ImageLoader::open hands each one the recorded
  offset. The offset is driven by the all-pixel minimum rather than the within-mask one:
  offsetting by a within-mask minimum would clamp every off-mask pixel below it and move the ROI's
  own minimum onto grey level 0, which the grey binning reads as background.

Also in scope, because they are the same defect from other loaders

• Signed-integer TIFF no longer clamps negatives to 0 in either the tile or the strip loader; both
  take the offset map. The strip loader had no real-valued handling at all and now shares the same
  map as the tile loader.
• NIfTI scl_slope/scl_inter and DICOM RescaleSlope/RescaleIntercept are applied
  unconditionally rather than only under --preserve-hu, so a scanner writing a non-identity
  rescale is no longer wrong by default. Mask loaders are built with the rescale explicitly off,
  since their pixels are labels.
• A real-valued slide's min-max quantization is inverted too, so its features come back in its own
  float range instead of in quantization steps — what the IH family already did.
• OME-Zarr is recorded as native. Its tile loader copies voxels untouched, so the previous
  derivation was applying a quantization inverse to a float zarr that never happened.
• COVERED_IMAGE_INTENSITY_RANGE scales its ROI range, closing the quantized-vs-source half of the
  domain mismatch. The truncation half stays and is documented in not_covered.md §F.

--preserve-hu changes meaning

It now selects the offset map for a real-valued slide. Integer, DICOM and NIfTI slides take it
regardless, so the flag is no longer what stands between a CT and its Hounsfield units. README.md,
docs/source/supported_formats.rst, docs/source/cmdline_and_examples.rst and the three nyxus.py
docstrings say so.

The name was kept deliberately, and it is now a poor one — a CT user reading --preserve-hu
would reasonably conclude they must pass it, and they must not. Renaming was considered and
declined for this PR; every doc surface therefore leads with the explicit negative ("NOT needed for
CT — Hounsfield units are preserved and reported either way, despite the name") rather than merely
describing the correct behaviour and leaving the inference to the reader. The flag has never
appeared in a tagged release (last tag 0.11.0, Feb 2026; the flag landed on main in July 2026),
so a rename remains available later at no cost to any released API.

Values that move

Every first-order intensity feature on a slide that holds a negative pixel or is real-valued. This
is not a silent bugfix from a user's point of view — the numbers change, and they change to the
correct ones.

Every 3D texture golden is unchanged. The offset base is still the all-pixel minimum and the map
still truncates, so the stored grey levels are byte-identical wherever a slide holds a negative
value. Whether the truncation should become rounding is a separate call, deliberately kept out.

Goldens re-derived

The 3D first-order MATLAB family was pinned in the loader's shifted domain, so all 29 values moved.
Re-derived on GNU Octave 11.3.0 + statistics over the same built-ins the MATLAB generator names,
with the previous MATLAB values reproduced to 13 significant digits on the old domain first as a
control. Every location statistic moved by exactly -1024, every dispersion statistic is unchanged,
and 3COV / 3QCOD / 3UNIFORMITY_PIU / 3ROOT_MEAN_SQUARED were recomputed from their moved
parts. The generator and the audit report carry the new domain.

The comparison changed with them. A relative tolerance is meaningless once a percentile can sit near
zero — 3P01 is 13 on a 2000-wide range, where 2.4 grey levels of binning is 18% of the value and
0.12% of the range — so the binned percentiles are now measured against the voxel range at 2e-3
(worst residual 1.19e-3), a zero reference against the range at 1e-3, and everything else keeps
its relative tier.

Tests

• test_hu_analytic.h rewritten: five cases pinning the recorder's three branches and the forward
  and inverse maps, including the deliberate choice that OME-Zarr is carried natively.
• test_2d_hu_ct_small_pydicom.py now asserts absolute Hounsfield units directly against pydicom
  rather than reconstructing them in the test, plus a new case for the invariant this defect would
  have failed: for a CT read in Hounsfield units, MIN is the ROI's true Hounsfield minimum.
• test_2d_signed_int16_loader_mechanics.py asserts the negatives survive rather than that they
  clamp to 0.
• test_2d_hu_regression.py / test_3d_hu_nifti_regression.py re-pinned to absolute values, and
  each gained a case asserting the flag is no longer what makes the difference.
• test_3d_nifti_mechanics.h supplies the offset the way ImageLoader::open does, and asserts what
  happens without one.

Gates

gate	result
gtest	889 passed / 1 skipped (GPU) of 890
pytest tests/python/	98 passed / 1 skipped, 0 failed
check_coverage.py --check	clean
check_test_names.py --check	0 violations
ASan + UBSan (ubtest1)	0 diagnostics, gtest rc=0, 886 cases (885 passed / 1 GPU skip), checksums verified before the build and again after the run
reporter's own CT pair	all 7 features match PyRadiomics/SimpleITK, both flag settings

Not in scope

Confirmed on the same data but untouched here: 3VOXEL_VOLUME still reports a voxel count
rather than a volume (it ignores voxel spacing), and 3MESH_VOLUME is still byte-identical to
3VOLUME_CONVEXHULL. Both are tracked separately.