chore: hypothesis property tests for fill-value handling

[maxrjones]

What this PR adds

Property-based test coverage for the fill-sentinel-values issue cluster, plus a zarr-spec compliance test that asserts virtualizarr's correctness independently of xarray's HDF5-style _FillValue encoding.

Three test modules under virtualizarr/tests/test_parsers/:

• test_zarr_spec_compliance.py — primary correctness check. Asserts ZarrParser produces metadata matching zarr-python's view of the source, no xarray in the loop. 7/7 passing.
• test_hdf/test_hdf_fill_value_equivalence.py — HDFParser interop check against xarray+h5netcdf. Hypothesis-driven dtype × fill matrix + curated @examples.
• test_zarr_fill_value_equivalence.py — ZarrParser interop check against xr.open_zarr.

Plus _fill_value_common.py — shared infrastructure (sentinels, hypothesis profiles, two-layer assertion helpers with attribution).

Hypothesis is added to the dev dependency group with >=6.100. Module-level pytestmark = pytest.mark.hypothesis_tests allows skipping via pytest -m "not hypothesis_tests". VIRTUALIZARR_HYPOTHESIS_PROFILE=ci shrinks the random-draw count for fast CI runs.

What this PR doesn't change

No production code. No parser fixes. This is test infrastructure only. Failing tests are the to-fix list, tracked in https://github.com/NASA-IMPACT/veda-odd#371.

Why it matters

Test failures are attributed by category, so reviewers and maintainers can tell what to fix and where:

Attribution	Meaning
BothEnginesFailedIdenticallyError	Both engines fail identically — upstream issue (e.g. xarray's FillValueCoder, tracked at pydata/xarray#11332), not virtualizarr
observed (virtualizarr) failed; reference succeeded	Virtualizarr-specific bug — fix lives in this repo
Plain assert_identical diff	Real correctness gap — investigate

This separates virtualizarr-side bugs from upstream xarray gaps cleanly. For example, #989's _FillValue=0.0-as-JSON-number case shows up as BothEnginesFailedIdenticallyError in the equivalence module — and the spec-compliance module passes on the same fixture, proving the gap is upstream (pydata/xarray#11332).

Current results

31 passed, 10 failed in ~6s (ci profile)

The 10 failures decompose as:

• 2 TestCompoundDtype — parser-side TypeError at hdf.py:364 (new finding; tracked in roadmap as Cluster A).
• 5 TestBasicEquivalence::test_equivalence_curated[*] (HDF) — string-dtype _FillValue. Improved attribution surfaced this is a parser-side encode error, not the downstream decode error previously documented.
• 1 TestBasicEquivalence::test_equivalence_random (HDF) — random draws hitting the same string-dtype cluster.
• 1 TestBasicEquivalence::test_equivalence_curated[float64-nan-fill] (Zarr) — upstream xarray decode gap, pydata/xarray#11332 (also user-reported as FillValue decoding unexpected fill_value: null (on version 2.6.1) #989).
• 1 TestBasicEquivalence::test_equivalence_random (Zarr) — random draws hitting the same xarray decode gap.

Each failure points at one of the clusters in the roadmap.

Notable findings surfaced while building this

• ZarrParser get_metadata fails for LocalStore Zarr with dtype uint8 #811 doesn't reproduce in current main. The curated example for the original bug report (uint8 with no explicit fill_value) opens cleanly through both engines. Worth a separate verification but suggests it was silently fixed.
• vlen-string _FillValue is virtualizarr-specific, not upstream. The improved attribution showed the failure is an encode error at parser manifest construction, not a decode error in xarray.

How to run

pixi run -e test pytest virtualizarr/tests/test_parsers/test_zarr_spec_compliance.py
pixi run -e test pytest virtualizarr/tests/test_parsers/test_hdf/test_hdf_fill_value_equivalence.py
pixi run -e test pytest virtualizarr/tests/test_parsers/test_zarr_fill_value_equivalence.py

# Skip hypothesis-driven tests (faster everyday CI):
pixi run -e test pytest -m "not hypothesis_tests"

# Smaller hypothesis matrix:
VIRTUALIZARR_HYPOTHESIS_PROFILE=ci pixi run -e test pytest ...

Refs

• Addresses #347 (hypothesis-driven HDF test suite).
• Roadmap and full fix sequencing: https://github.com/NASA-IMPACT/veda-odd#371.
• Companion: PR Correctly handle HDF5 fillvalue for string dtype arrays. #988 (fix/problem_fillvalues) — partial Phase 1 fix (currently not in this branch's history; this PR sits on main).

Acceptance criteria:

• Closes #xxxx
• Tests added
• Tests passing
• No test coverage regression
• Full type hint coverage
• Changes are documented in docs/releases.md
• New functions/methods are listed in an appropriate *.md file under docs/api
• New functionality has documentation