rust-hdf5 0.7.2

#!/usr/bin/env python3
"""The oracle's feature matrix.

Each case names one HDF5 feature, carries an h5py generator that writes the
reference file, and — when rust-hdf5's *public* API can express the same file —
the id of the matching arm in `src/bin/oracle_probe.rs`'s `write` subcommand.

The data every case writes is deliberately formulaic (`arange`-like ramps,
short literal strings) so the rust writer can reproduce it byte for byte
without the two sides sharing a data file.

Only the standard library, numpy and h5py are used.
"""

import pathlib
import shutil

import numpy as np

import h5py
from h5py import h5d, h5p, h5s, h5t

# --------------------------------------------------------------------------
# helpers
# --------------------------------------------------------------------------

N = 8  # default element count for the 1-D dtype ramps


def ramp(dtype, n=N):
    return np.arange(n, dtype=np.dtype(dtype))


def lowlevel_dataset(f, name, tid, sid, data=None, dcpl=None, mtype=None):
    """Create a dataset from a raw TypeID/SpaceID, which h5py cannot express.

    Returns the low-level DatasetID so the caller can write through it.
    """
    dsid = h5d.create(f.id, name.encode("utf-8"), tid, sid, dcpl=dcpl)
    if data is not None:
        dsid.write(h5s.ALL, h5s.ALL, np.ascontiguousarray(data), mtype=mtype)
    return dsid


def chunked_dcpl(chunk, alloc_time=None, layout=h5d.CHUNKED):
    dcpl = h5p.create(h5p.DATASET_CREATE)
    dcpl.set_layout(layout)
    if layout == h5d.CHUNKED:
        dcpl.set_chunk(tuple(chunk))
    if alloc_time is not None:
        dcpl.set_alloc_time(alloc_time)
    return dcpl


class Case:
    def __init__(self, name, group, gen, rust=None, note="", ext_files=(), access=None):
        self.name = name
        self.group = group
        self.gen = gen
        self.rust = rust
        self.note = note
        self.ext_files = ext_files
        # Dataset-access properties both sides open every dataset under —
        # `{"view": "first_missing"|"last_available", "printf_gap": int}`.
        # They are not stored in the file, so they belong to the case, not to
        # the generator: the same bytes describe differently under different
        # ones.
        self.access = access

    def __repr__(self):
        return "Case(%s)" % self.name


# --------------------------------------------------------------------------
# integer widths / signedness / endianness
# --------------------------------------------------------------------------


def _int_case(name, npdtype, rust):
    def gen(path):
        with h5py.File(path, "w") as f:
            f.create_dataset("data", data=ramp(npdtype))

    return Case(name, "dtype-int", gen, rust, "1-D ramp of %s" % npdtype)


INT_CASES = [
    _int_case("int_i8", "i1", "int_i8"),
    _int_case("int_u8", "u1", "int_u8"),
    _int_case("int_i16le", "<i2", "int_i16le"),
    _int_case("int_u16le", "<u2", "int_u16le"),
    _int_case("int_i32le", "<i4", "int_i32le"),
    _int_case("int_u32le", "<u4", "int_u32le"),
    _int_case("int_i64le", "<i8", "int_i64le"),
    _int_case("int_u64le", "<u8", "int_u64le"),
    _int_case("int_i16be", ">i2", "int_i16be"),
    _int_case("int_i32be", ">i4", "int_i32be"),
    _int_case("int_u64be", ">u8", "int_u64be"),
]


# --------------------------------------------------------------------------
# floating point
# --------------------------------------------------------------------------


def _float_case(name, npdtype, rust):
    def gen(path):
        with h5py.File(path, "w") as f:
            f.create_dataset("data", data=ramp(npdtype))

    return Case(name, "dtype-float", gen, rust, "1-D ramp of %s" % npdtype)


def gen_float_specials(path):
    bits = np.array(
        [
            0x7FF8000000000001,  # quiet NaN with a payload
            0x7FF0000000000000,  # +inf
            0xFFF0000000000000,  # -inf
            0x8000000000000000,  # -0.0
            0x0000000000000001,  # smallest denormal
            0x3FF0000000000000,  # 1.0
            0xBFF0000000000000,  # -1.0
            0x0000000000000000,  # +0.0
        ],
        dtype="<u8",
    )
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=bits.view("<f8"))


FLOAT_CASES = [
    _float_case("float_f16le", "<f2", "float_f16le"),
    _float_case("float_f32le", "<f4", "float_f32le"),
    _float_case("float_f64le", "<f8", "float_f64le"),
    _float_case("float_f64be", ">f8", "float_f64be"),
    Case(
        "float_specials",
        "dtype-float",
        gen_float_specials,
        "float_specials",
        "NaN payload, +/-inf, -0.0, denormal — bit patterns must survive",
    ),
]


# --------------------------------------------------------------------------
# strings
# --------------------------------------------------------------------------

STRINGS = ["alpha", "b", "", "delta12"]
UNISTR = ["été", "日本", "", "café"]


def gen_str_fixed_ascii(path):
    with h5py.File(path, "w") as f:
        f.create_dataset(
            "data",
            data=np.array([s.encode("ascii") for s in STRINGS], dtype="S8"),
            dtype=h5py.string_dtype("ascii", 8),
        )


def gen_str_fixed_utf8(path):
    # libhdf5 has no ASCII -> UTF-8 conversion path, so the memory type has to
    # be the UTF-8 fixed string itself and the bytes go through verbatim.
    with h5py.File(path, "w") as f:
        tid = h5t.C_S1.copy()
        tid.set_size(16)
        tid.set_cset(h5t.CSET_UTF8)
        tid.set_strpad(h5t.STR_NULLPAD)
        sid = h5s.create_simple((len(UNISTR),))
        lowlevel_dataset(
            f,
            "data",
            tid,
            sid,
            np.array([s.encode("utf-8") for s in UNISTR], dtype="S16"),
            mtype=tid,
        )


def _fixed_str_pad(strpad):
    # libhdf5 does not re-pad on a same-size same-cset copy, so the padding
    # bytes have to be written explicitly for the declared rule to be what is
    # actually on disk.
    padbyte = b" " if strpad == h5t.STR_SPACEPAD else b"\0"

    def gen(path):
        with h5py.File(path, "w") as f:
            tid = h5t.C_S1.copy()
            tid.set_size(8)
            tid.set_cset(h5t.CSET_ASCII)
            tid.set_strpad(strpad)
            sid = h5s.create_simple((len(STRINGS),))
            data = np.array(
                [s.encode("ascii").ljust(8, padbyte) for s in STRINGS], dtype="S8"
            )
            lowlevel_dataset(f, "data", tid, sid, data, mtype=tid)

    return gen


def gen_str_vlen_ascii(path):
    with h5py.File(path, "w") as f:
        f.create_dataset(
            "data",
            data=np.array(STRINGS, dtype=object),
            dtype=h5py.string_dtype("ascii"),
        )


def gen_str_vlen_utf8(path):
    with h5py.File(path, "w") as f:
        f.create_dataset(
            "data",
            data=np.array(UNISTR, dtype=object),
            dtype=h5py.string_dtype("utf-8"),
        )


STRING_CASES = [
    Case("str_fixed_ascii", "dtype-string", gen_str_fixed_ascii, "str_fixed_ascii",
         "8-byte fixed ASCII strings"),
    Case("str_fixed_utf8", "dtype-string", gen_str_fixed_utf8, "str_fixed_utf8",
         "16-byte fixed UTF-8 strings"),
    Case("str_fixed_nullpad", "dtype-string", _fixed_str_pad(h5t.STR_NULLPAD), "str_fixed_nullpad",
         "fixed string with STR_NULLPAD"),
    Case("str_fixed_spacepad", "dtype-string", _fixed_str_pad(h5t.STR_SPACEPAD), "str_fixed_spacepad",
         "fixed string with STR_SPACEPAD"),
    Case("str_vlen_ascii", "dtype-string", gen_str_vlen_ascii, "str_vlen_ascii",
         "variable-length ASCII strings via the global heap"),
    Case("str_vlen_utf8", "dtype-string", gen_str_vlen_utf8, "str_vlen_utf8",
         "variable-length UTF-8 strings via the global heap"),
]


# --------------------------------------------------------------------------
# compound / array / enum / opaque / bitfield / references / vlen
# --------------------------------------------------------------------------

COMPOUND_SIMPLE = np.dtype([("x", "<f4"), ("y", "<f4")])
COMPOUND_NESTED = np.dtype([("a", "<i4"), ("inner", [("u", "<i2"), ("v", "<i2")])])
COMPOUND_STR = np.dtype([("id", "<i4"), ("name", "S8")])
COMPOUND_PAD = np.dtype(
    {"names": ["a", "b"], "formats": ["<i2", "<i4"], "offsets": [0, 4], "itemsize": 12}
)


def gen_compound_simple(path):
    arr = np.zeros(4, dtype=COMPOUND_SIMPLE)
    arr["x"] = np.arange(4, dtype="<f4")
    arr["y"] = np.arange(100, 104, dtype="<f4")
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=arr)


def gen_compound_nested(path):
    arr = np.zeros(4, dtype=COMPOUND_NESTED)
    arr["a"] = np.arange(4, dtype="<i4")
    arr["inner"]["u"] = np.arange(10, 14, dtype="<i2")
    arr["inner"]["v"] = np.arange(20, 24, dtype="<i2")
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=arr)


def gen_compound_with_string(path):
    arr = np.zeros(3, dtype=COMPOUND_STR)
    arr["id"] = np.arange(3, dtype="<i4")
    arr["name"] = [b"aa", b"bbb", b"cccc"]
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=arr)


def gen_compound_padded(path):
    arr = np.zeros(4, dtype=COMPOUND_PAD)
    arr["a"] = np.arange(4, dtype="<i2")
    arr["b"] = np.arange(1000, 1004, dtype="<i4")
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=arr)


def gen_array_dtype(path):
    with h5py.File(path, "w") as f:
        tid = h5t.array_create(h5t.IEEE_F64LE, (2, 3))
        sid = h5s.create_simple((2,))
        data = np.arange(12, dtype="<f8").reshape(2, 2, 3)
        lowlevel_dataset(f, "data", tid, sid, data, mtype=tid)


def gen_enum_i8(path):
    dt = h5py.enum_dtype({"RED": 0, "GREEN": 1, "BLUE": 2}, basetype="i1")
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=np.array([0, 1, 2, 1], dtype="i1"), dtype=dt)


def gen_enum_i32(path):
    dt = h5py.enum_dtype({"LOW": -1, "MID": 0, "HIGH": 1000}, basetype="<i4")
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=np.array([-1, 0, 1000, 0], dtype="<i4"), dtype=dt)


def gen_compound_dtype_v4(path):
    # A v1.12 low bound makes libhdf5 tag the compound datatype message
    # version 4 (H5O_dtype_ver_bounds); its members stay version 1. Nothing
    # else in the matrix produces a datatype message above version 3.
    #
    # Chunked on purpose: a *contiguous* dataset at a v1.10+ bound is dropped
    # from the listing by an unrelated gap (`layout_contiguous_v110`), which
    # would mask what this case is about.
    arr = np.zeros(4, dtype=COMPOUND_SIMPLE)
    arr["x"] = np.arange(4, dtype="<f4")
    arr["y"] = np.arange(100, 104, dtype="<f4")
    with h5py.File(path, "w", libver=("v112", "v112")) as f:
        ds = f.create_dataset("data", (4,), chunks=(4,), dtype=COMPOUND_SIMPLE)
        ds[...] = arr


def gen_opaque(path):
    with h5py.File(path, "w") as f:
        tid = h5t.create(h5t.OPAQUE, 4)
        tid.set_tag(b"raw4")
        sid = h5s.create_simple((3,))
        data = np.frombuffer(bytes(range(12)), dtype="V4")
        lowlevel_dataset(f, "data", tid, sid, data, mtype=tid)


def gen_bitfield(path):
    with h5py.File(path, "w") as f:
        tid = h5t.STD_B8LE.copy()
        sid = h5s.create_simple((4,))
        lowlevel_dataset(
            f, "data", tid, sid, np.array([0x01, 0x80, 0xFF, 0x00], dtype="u1")
        )


def gen_ref_object(path):
    with h5py.File(path, "w") as f:
        f.create_dataset("target", data=ramp("<i4"))
        g = f.create_group("grp")
        refs = np.array([f["target"].ref, g.ref], dtype=h5py.ref_dtype)
        f.create_dataset("refs", data=refs, dtype=h5py.ref_dtype)


def gen_ref_region(path):
    with h5py.File(path, "w") as f:
        t = f.create_dataset("target", data=ramp("<i4"))
        refs = np.array([t.regionref[0:3], t.regionref[4:8]],
                        dtype=h5py.regionref_dtype)
        f.create_dataset("refs", data=refs, dtype=h5py.regionref_dtype)


def gen_vlen_numeric(path):
    dt = h5py.vlen_dtype(np.dtype("<i4"))
    with h5py.File(path, "w") as f:
        ds = f.create_dataset("data", (3,), dtype=dt)
        ds[0] = np.array([1, 2, 3], dtype="<i4")
        ds[1] = np.array([], dtype="<i4")
        ds[2] = np.array([-7], dtype="<i4")


def gen_vlen_bytes(path):
    dt = h5py.vlen_dtype(np.dtype("u1"))
    with h5py.File(path, "w") as f:
        ds = f.create_dataset("data", (3,), dtype=dt)
        ds[0] = np.array([0, 1, 2], dtype="u1")
        ds[1] = np.array([], dtype="u1")
        ds[2] = np.array([255], dtype="u1")


def gen_named_datatype(path):
    with h5py.File(path, "w") as f:
        f["t"] = np.dtype("<i4")
        f.create_dataset("data", data=ramp("<i4"))
        # A dataset created from the committed TypeID stores a *shared*
        # datatype message pointing at /t rather than a datatype of its own.
        sid = h5s.create_simple((N,))
        lowlevel_dataset(f, "shared", f["t"].id, sid, ramp("<i4"))


COMPOSITE_CASES = [
    Case("compound_simple", "dtype-composite", gen_compound_simple, "compound_simple",
         "two f32 members, no padding"),
    Case("compound_nested", "dtype-composite", gen_compound_nested, "compound_nested",
         "compound member inside a compound"),
    Case("compound_with_string", "dtype-composite", gen_compound_with_string,
         "compound_with_string", "fixed string member"),
    Case("compound_padded", "dtype-composite", gen_compound_padded, "compound_padded",
         "member offsets with gaps and trailing padding"),
    Case("compound_dtype_v4", "dtype-composite", gen_compound_dtype_v4,
         "compound_dtype_v4", "version-4 datatype message (libver v1.12 bounds)"),
    Case("array_dtype", "dtype-composite", gen_array_dtype, "array_dtype",
         "H5T_ARRAY element type (2x3 f64)"),
    Case("enum_i8", "dtype-composite", gen_enum_i8, "enum_i8", "3-member i8 enum"),
    Case("enum_i32", "dtype-composite", gen_enum_i32, "enum_i32",
         "i32 enum with a negative member"),
    Case("opaque", "dtype-composite", gen_opaque, "opaque", "H5T_OPAQUE with a tag"),
    Case("bitfield", "dtype-composite", gen_bitfield, "bitfield",
         "H5T_BITFIELD (STD_B8LE)"),
    Case("ref_object", "dtype-composite", gen_ref_object, "ref_object",
         "object references to a dataset and a group"),
    Case("ref_region", "dtype-composite", gen_ref_region, "ref_region",
         "dataset region references"),
    Case("vlen_numeric", "dtype-composite", gen_vlen_numeric, "vlen_numeric",
         "variable-length i32 sequences"),
    Case("vlen_bytes", "dtype-composite", gen_vlen_bytes, "vlen_bytes",
         "variable-length u8 sequences"),
    Case("named_datatype", "dtype-composite", gen_named_datatype, "named_datatype",
         "committed datatype object, and a dataset that shares it"),
]


# --------------------------------------------------------------------------
# layouts and chunk indexes
# --------------------------------------------------------------------------


def gen_layout_compact(path):
    with h5py.File(path, "w") as f:
        dcpl = h5p.create(h5p.DATASET_CREATE)
        dcpl.set_layout(h5d.COMPACT)
        sid = h5s.create_simple((16,))
        lowlevel_dataset(
            f, "data", h5t.STD_I32LE, sid, ramp("<i4", 16), dcpl=dcpl
        )


def gen_layout_contiguous(path):
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=ramp("<i4", 16))


def gen_layout_contiguous_v110(path):
    with h5py.File(path, "w", libver=("v110", "v110")) as f:
        f.create_dataset("data", data=ramp("<i4", 16))


def gen_layout_chunked_v110(path):
    with h5py.File(path, "w", libver=("v110", "v110")) as f:
        ds = f.create_dataset("data", (16,), chunks=(16,), dtype="<i4")
        ds[...] = ramp("<i4", 16)


def gen_chunkidx_btree1(path):
    with h5py.File(path, "w", libver="earliest") as f:
        ds = f.create_dataset(
            "data", (8,), maxshape=(None,), chunks=(4,), dtype="<i4"
        )
        ds[...] = ramp("<i4")


def gen_chunkidx_single(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset("data", (8,), chunks=(8,), dtype="<i4")
        ds[...] = ramp("<i4")


def gen_chunkidx_implicit(path):
    with h5py.File(path, "w", libver="latest") as f:
        dcpl = chunked_dcpl((4,), alloc_time=h5d.ALLOC_TIME_EARLY)
        sid = h5s.create_simple((16,))
        lowlevel_dataset(f, "data", h5t.STD_I32LE, sid, ramp("<i4", 16), dcpl=dcpl)


def gen_chunkidx_farray(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset("data", (16,), chunks=(4,), dtype="<i4")
        ds[...] = ramp("<i4", 16)


def gen_chunkidx_earray(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset(
            "data", (16,), maxshape=(None,), chunks=(4,), dtype="<i4"
        )
        ds[...] = ramp("<i4", 16)


def gen_chunkidx_earray_unlim_inner(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset(
            "data", (4, 4), maxshape=(4, None), chunks=(2, 2), dtype="<i4"
        )
        ds[...] = ramp("<i4", 16).reshape(4, 4)


def gen_layout_contiguous_v108(path):
    with h5py.File(path, "w", libver=("v108", "v108")) as f:
        f.create_dataset("data", data=ramp("<i4", 16))


def gen_layout_chunked_v108(path):
    with h5py.File(path, "w", libver=("v108", "v108")) as f:
        ds = f.create_dataset("data", (16,), chunks=(16,), dtype="<i4")
        ds[...] = ramp("<i4", 16)


def gen_chunkidx_earray_dim1(path):
    # The extensible dimension is the fastest-changing one, so the chunk
    # coordinate the index is keyed on is not the first.
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset(
            "data", (4, 4), maxshape=(4, None), chunks=(2, 4), dtype="<i4"
        )
        ds[...] = ramp("<i4", 16).reshape(4, 4)


def gen_external_storage(path):
    raw = path.parent / (path.stem + "_ext.raw")
    raw.write_bytes(ramp("<i4", 16).tobytes())
    with h5py.File(path, "w") as f:
        f.create_dataset(
            "data", shape=(16,), dtype="<i4", external=[(raw.name, 0, 64)]
        )


def gen_vds(path):
    src = path.parent / (path.stem + "_src.h5")
    with h5py.File(src, "w") as g:
        g.create_dataset("src", data=ramp("<i4", 16))
    layout = h5py.VirtualLayout(shape=(16,), dtype="<i4")
    layout[...] = h5py.VirtualSource(src.name, "src", shape=(16,))
    with h5py.File(path, "w") as f:
        f.create_virtual_dataset("vds", layout)


def gen_external_unlimited(path):
    """The last EFL slot sized H5O_EFL_UNLIMITED over an unlimited dataspace.

    `H5D__efl_construct` requires exactly this pairing: a dataset whose
    dataspace can grow has no finite reservation that could cover it, so its
    last — and here only — external slot must be unlimited.
    """
    raw = path.parent / (path.stem + "_ext.raw")
    raw.write_bytes(ramp("<i4", 16).tobytes())
    with h5py.File(path, "w") as f:
        f.create_dataset(
            "data",
            shape=(16,),
            maxshape=(None,),
            dtype="<i4",
            external=[(raw.name, 0, h5py.h5f.UNLIMITED)],
        )


def gen_vds_unlim(path):
    """A mapping unlimited on both sides: the extent follows the source.

    `H5D__virtual_set_extent_unlim` clips the virtual selection against what
    the source actually holds, so the dataset created one row tall reads back
    as tall as its source is.
    """
    src = path.parent / (path.stem + "_src.h5")
    with h5py.File(src, "w") as g:
        g.create_dataset(
            "src", data=ramp("<i4", 20).reshape(10, 2), maxshape=(None, 2),
            chunks=(5, 2),
        )
    layout = h5py.VirtualLayout(shape=(1, 2), dtype="<i4", maxshape=(None, 2))
    vsrc = h5py.VirtualSource(src.name, "src", shape=(1, 2), maxshape=(None, 2))
    layout[: h5s.UNLIMITED, :] = vsrc[: h5s.UNLIMITED, :]
    with h5py.File(path, "w") as f:
        f.create_virtual_dataset("vds", layout)


def gen_vds_printf_unlim(path):
    """A printf-pattern source name over an unlimited virtual selection.

    The two features only exist together: `H5D_virtual_check_mapping_post`
    refuses a `%b` in a source name unless the virtual selection is unlimited
    and the source selection is not, which is the mapping that says "one source
    file per block, as many as are there". Three blocks are present, so the
    extent `H5D__virtual_set_extent_unlim` gives the dataset is three rows.
    """
    stem = path.stem
    for b in range(3):
        with h5py.File(path.parent / ("%s_b%d.h5" % (stem, b)), "w") as g:
            g.create_dataset("data", data=ramp("<i4", 4) + 10 * b)
    layout = h5py.VirtualLayout(shape=(1, 4), dtype="<i4", maxshape=(None, 4))
    vsrc = h5py.VirtualSource("%s_b%%b.h5" % stem, "data", shape=(4,))
    layout[: h5s.UNLIMITED, :] = vsrc
    with h5py.File(path, "w") as f:
        f.create_virtual_dataset("vds", layout)


def gen_vds_printf_gap(path):
    """A printf mapping over blocks 0, 1 and 3 — block 2 is not written.

    What the extent is depends entirely on the dataset access properties the
    open names, and the file records none of them: with the default gap of 0
    the scan stops at block 2 and the dataset is two rows tall, and the cases
    reading this same file with a gap look past it to block 3 and see four,
    the third filled (`H5D__virtual_set_extent_unlim`, H5Dvirtual.c:1519).
    """
    stem = path.stem
    for b in (0, 1, 3):
        with h5py.File(path.parent / ("%s_b%d.h5" % (stem, b)), "w") as g:
            g.create_dataset("data", data=ramp("<i4", 4) + 10 * b)
    layout = h5py.VirtualLayout(shape=(1, 4), dtype="<i4", maxshape=(None, 4))
    vsrc = h5py.VirtualSource("%s_b%%b.h5" % stem, "data", shape=(4,))
    layout[: h5s.UNLIMITED, :] = vsrc
    with h5py.File(path, "w") as f:
        f.create_virtual_dataset("vds", layout, fillvalue=-7)


def gen_vds_view_trail(path):
    """A mapping unlimited on both sides whose stride is wider than its block.

    `H5S_hyper_get_clip_extent_match` takes `incl_trail` from the view
    (H5Dvirtual.c:1447-1451), and it only changes the answer when the last
    mapped block is followed by a gap: three source rows under stride 3,
    block 2 give a two-row extent under `H5D_VDS_LAST_AVAILABLE` and a
    three-row one under `H5D_VDS_FIRST_MISSING`.
    """
    with h5py.File(path, "w") as f:
        f.create_dataset(
            "src", data=ramp("<i4", 6).reshape(3, 2), maxshape=(None, 2),
            chunks=(1, 2),
        )
        vsid = h5s.create_simple((1, 2), (h5s.UNLIMITED, 2))
        vsid.select_hyperslab((0, 0), (h5s.UNLIMITED, 1), stride=(3, 1), block=(2, 2))
        ssid = h5s.create_simple((1, 2), (h5s.UNLIMITED, 2))
        ssid.select_hyperslab((0, 0), (h5s.UNLIMITED, 1), stride=(3, 1), block=(2, 2))
        dcpl = h5p.create(h5p.DATASET_CREATE)
        # `H5Pset_virtual` pokes the layout straight into the list
        # (H5Pdcpl.c:2146) and so leaves the allocation time at the contiguous
        # default, while `H5Pset_layout` resets it to the layout's own
        # (H5Pdcpl.c:1762-1782). h5py's VDS path always makes this call
        # (`VirtualLayout.__init__`, vds.py:174), so every virtual dataset
        # written through an API rather than by hand is incremental.
        dcpl.set_layout(h5d.VIRTUAL)
        dcpl.set_fill_value(np.array(-9, dtype="<i4"))
        dcpl.set_virtual(vsid, b".", b"/src", ssid)
        lowlevel_dataset(f, "vds", h5t.STD_I32LE, vsid, dcpl=dcpl)


def gen_vds_split(path):
    """A mapping whose two selections decompose into different box counts.

    The virtual side is two 1x4 blocks — rows 0 and 2 of a 4x4 dataset — and
    the source side one `H5S_SEL_ALL` over a 2x4 one, so there is no
    positional pairing of boxes to be had. `H5S_select_project_intersection`
    (H5Sselect.c:2402) runs one selection iterator per side and matches the
    two element streams off one against one; the only thing it asks of the
    pair is that the element counts agree, which is also the only thing
    `H5D_virtual_check_mapping_pre` checks when the mapping is created
    (H5Dvirtual.c:254-257).
    """
    with h5py.File(path, "w") as f:
        f.create_dataset("src", data=ramp("<i4", 8).reshape(2, 4))
        vsid = h5s.create_simple((4, 4))
        vsid.select_hyperslab((0, 0), (2, 1), stride=(2, 1), block=(1, 4))
        ssid = h5s.create_simple((2, 4))
        ssid.select_all()
        dcpl = h5p.create(h5p.DATASET_CREATE)
        dcpl.set_layout(h5d.VIRTUAL)
        dcpl.set_fill_value(np.array(-9, dtype="<i4"))
        dcpl.set_virtual(vsid, b".", b"/src", ssid)
        lowlevel_dataset(f, "vds", h5t.STD_I32LE, vsid, dcpl=dcpl)


def gen_chunkidx_btree2(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset(
            "data", (4, 4), maxshape=(None, None), chunks=(2, 2), dtype="<i4"
        )
        ds[...] = ramp("<i4", 16).reshape(4, 4)


LAYOUT_CASES = [
    Case("layout_compact", "layout", gen_layout_compact, "layout_compact",
         "compact layout — data inside the object header"),
    Case("layout_contiguous", "layout", gen_layout_contiguous, "layout_contiguous",
         "contiguous layout"),
    Case("layout_contiguous_v110", "layout", gen_layout_contiguous_v110,
         "layout_contiguous_v110",
         "contiguous layout under v1.10 bounds — data layout message v4"),
    Case("layout_chunked_v110", "layout", gen_layout_chunked_v110,
         "layout_chunked_v110",
         "chunked layout under v1.10 bounds — the control for the case above"),
    Case("chunkidx_btree1", "layout", gen_chunkidx_btree1, "chunkidx_btree1",
         "layout v3 + version-1 B-tree chunk index (libver earliest)"),
    Case("chunkidx_single", "layout", gen_chunkidx_single, "chunkidx_single",
         "single-chunk index"),
    Case("chunkidx_implicit", "layout", gen_chunkidx_implicit, "chunkidx_implicit",
         "implicit index — fixed shape, early allocation, no filter"),
    Case("chunkidx_farray", "layout", gen_chunkidx_farray, "chunkidx_farray",
         "fixed-array index"),
    Case("chunkidx_earray", "layout", gen_chunkidx_earray, "chunkidx_earray",
         "extensible-array index — one unlimited dimension"),
    Case("chunkidx_earray_unlim_inner", "layout", gen_chunkidx_earray_unlim_inner,
         "chunkidx_earray_unlim_inner",
         "extensible-array index — the unlimited dimension is dim 1, not dim 0"),
    Case("chunkidx_btree2", "layout", gen_chunkidx_btree2, "chunkidx_btree2",
         "version-2 B-tree index — two unlimited dimensions"),
    Case("layout_contiguous_v108", "layout", gen_layout_contiguous_v108,
         "layout_contiguous_v108",
         "contiguous layout under v1.8 bounds — the v1.10 pair's control"),
    Case("layout_chunked_v108", "layout", gen_layout_chunked_v108,
         "layout_chunked_v108", "chunked layout under v1.8 bounds"),
    Case("chunkidx_earray_dim1", "layout", gen_chunkidx_earray_dim1,
         "chunkidx_earray_dim1",
         "extensible-array index whose unlimited dimension is not the first"),
    Case("external_storage", "layout", gen_external_storage, "external_storage",
         "contiguous data held in an external raw file",
         ext_files=("_ext.raw",)),
    Case("vds", "layout", gen_vds, "vds",
         "virtual dataset mapped onto a dataset in a sibling file",
         ext_files=("_src.h5",)),
    Case("external_unlimited", "layout", gen_external_unlimited,
         "external_unlimited",
         "H5O_EFL_UNLIMITED on the last external slot of an unlimited dataset",
         ext_files=("_ext.raw",)),
    Case("vds_unlim", "layout", gen_vds_unlim, "vds_unlim",
         "virtual dataset whose mapping is unlimited on both sides — the "
         "extent comes from the source",
         ext_files=("_src.h5",)),
    Case("vds_printf_gap", "layout", gen_vds_printf_gap, "vds_printf_gap",
         "printf mapping with block 2 missing, read at the default printf "
         "gap of 0 — the extent stops at the gap"),
    Case("vds_printf_gap_1", "layout", gen_vds_printf_gap, "vds_printf_gap",
         "the same file read with H5Pset_virtual_printf_gap(1) — block 3 is "
         "reached and block 2 reads as the fill value",
         access={"printf_gap": 1}),
    Case("vds_printf_gap_first_missing", "layout", gen_vds_printf_gap, "vds_printf_gap",
         "the same file under H5D_VDS_FIRST_MISSING with a gap of 2, which "
         "H5D__virtual_init forces back to 0 (H5Dvirtual.c:2182-2188)",
         access={"view": "first_missing", "printf_gap": 2}),
    Case("vds_view_trail", "layout", gen_vds_view_trail, "vds_view_trail",
         "unlimited mapping whose stride exceeds its block, read at the "
         "default H5D_VDS_LAST_AVAILABLE view"),
    Case("vds_view_trail_first_missing", "layout", gen_vds_view_trail, "vds_view_trail",
         "the same file under H5D_VDS_FIRST_MISSING — the extent runs on to "
         "where the next block would start",
         access={"view": "first_missing"}),
    Case("vds_split", "layout", gen_vds_split, "vds_split",
         "a same-file mapping whose virtual and source selections decompose "
         "into different numbers of boxes"),
    Case("vds_printf_unlim", "layout", gen_vds_printf_unlim, "vds_printf_unlim",
         "printf-pattern source name over an unlimited virtual selection — "
         "one source file per block",
         ext_files=("_b0.h5", "_b1.h5", "_b2.h5")),
]


# --------------------------------------------------------------------------
# filters
# --------------------------------------------------------------------------


def _filter_case(name, rust, note, **kw):
    def gen(path):
        with h5py.File(path, "w", libver="latest") as f:
            ds = f.create_dataset("data", (64,), chunks=(16,), dtype="<i4", **kw)
            ds[...] = ramp("<i4", 64)

    return Case(name, "filter", gen, rust, note)


FILTER_CASES = [
    _filter_case("filter_deflate", "filter_deflate", "deflate level 6",
                 compression="gzip", compression_opts=6),
    _filter_case("filter_shuffle", "filter_shuffle", "shuffle only",
                 shuffle=True),
    _filter_case("filter_fletcher32", "filter_fletcher32", "fletcher32 checksum",
                 fletcher32=True),
    _filter_case("filter_deflate_shuffle", "filter_deflate_shuffle",
                 "shuffle then deflate",
                 compression="gzip", compression_opts=6, shuffle=True),
    _filter_case("filter_scaleoffset", "filter_scaleoffset",
                 "scale-offset, library-computed minimum bits",
                 scaleoffset=0),
    _filter_case("filter_szip_ec", "filter_szip_ec",
                 "szip entropy coding, 8 pixels per block",
                 compression="szip", compression_opts=("ec", 8)),
    _filter_case("filter_szip_nn", "filter_szip_nn",
                 "szip nearest neighbour, 16 pixels per block",
                 compression="szip", compression_opts=("nn", 16)),
]


# --------------------------------------------------------------------------
# fill values
# --------------------------------------------------------------------------


def gen_fill_default(path):
    with h5py.File(path, "w", libver="latest") as f:
        f.create_dataset("data", (16,), chunks=(4,), dtype="<i4")


def gen_fill_set_int(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset("data", (16,), chunks=(4,), dtype="<i4", fillvalue=-1)
        ds[0:4] = ramp("<i4", 4)


def gen_fill_set_float_nan(path):
    with h5py.File(path, "w", libver="latest") as f:
        f.create_dataset(
            "data", (16,), chunks=(4,), dtype="<f8", fillvalue=np.float64("nan")
        )


FILL_CASES = [
    Case("fill_default", "fillvalue", gen_fill_default, "fill_default",
         "default (zero) fill, nothing written"),
    Case("fill_set_int", "fillvalue", gen_fill_set_int, "fill_set_int",
         "user-defined integer fill, first chunk written"),
    Case("fill_set_float_nan", "fillvalue", gen_fill_set_float_nan,
         "fill_set_float_nan", "user-defined NaN fill"),
]


# --------------------------------------------------------------------------
# dataspaces
# --------------------------------------------------------------------------


def gen_space_scalar(path):
    with h5py.File(path, "w") as f:
        f.create_dataset("data", data=np.int32(42))


def gen_space_null(path):
    with h5py.File(path, "w") as f:
        f["data"] = h5py.Empty("<i4")


def gen_space_zerosized(path):
    with h5py.File(path, "w") as f:
        f.create_dataset("data", (0,), dtype="<i4")


def gen_space_unlimited_resized(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset("data", (4,), maxshape=(None,), chunks=(4,), dtype="<i4")
        ds[...] = ramp("<i4", 4)
        ds.resize((12,))
        ds[4:12] = ramp("<i4", 8) + 100


SPACE_CASES = [
    Case("space_scalar", "dataspace", gen_space_scalar, "space_scalar",
         "scalar (rank 0) dataspace"),
    Case("space_null", "dataspace", gen_space_null, "space_null",
         "NULL dataspace — no elements at all"),
    Case("space_zerosized", "dataspace", gen_space_zerosized, "space_zerosized",
         "simple dataspace with a zero-length dimension"),
    Case("space_unlimited_resized", "dataspace", gen_space_unlimited_resized,
         "space_unlimited_resized", "unlimited maxshape, grown after creation"),
]


# --------------------------------------------------------------------------
# groups, links
# --------------------------------------------------------------------------


def gen_groups_nested(path):
    with h5py.File(path, "w") as f:
        g = f.create_group("a")
        h = g.create_group("b")
        h.create_group("c")
        h.create_dataset("leaf", data=ramp("<i4"))
        f.create_dataset("top", data=ramp("<i4"))


def gen_link_hard(path):
    with h5py.File(path, "w") as f:
        f.create_dataset("orig", data=ramp("<i4"))
        f["alias"] = f["orig"]


def gen_link_soft(path):
    with h5py.File(path, "w") as f:
        f.create_dataset("orig", data=ramp("<i4"))
        f["alias"] = h5py.SoftLink("/orig")


def gen_link_external(path):
    target = path.parent / (path.stem + "_ext.h5")
    with h5py.File(target, "w") as g:
        g.create_dataset("payload", data=ramp("<i4"))
    with h5py.File(path, "w") as f:
        f.create_dataset("orig", data=ramp("<i4"))
        f["ext"] = h5py.ExternalLink(target.name, "/payload")


def gen_link_external_read(path):
    """A master file whose payload lives entirely in a sibling.

    Every dataset is reached only by crossing a link, so the `resolved` field
    is the whole content check: a reader that lists the links but never opens
    the other file matches on `target` and diverges here. The two dangling
    links pin the other half — a target file that is not there and a target
    object that is not there both have to say so rather than read something.
    """
    target = path.parent / (path.stem + "_data.h5")
    with h5py.File(target, "w") as g:
        g.create_dataset("top", data=ramp("<f8"))
        g.create_group("deep").create_dataset("inner", data=ramp("<i2"))
    with h5py.File(path, "w") as f:
        f["direct"] = h5py.ExternalLink(target.name, "/top")
        f["nested"] = h5py.ExternalLink(target.name, "/deep/inner")
        f["gone_object"] = h5py.ExternalLink(target.name, "/absent")
        f["gone_file"] = h5py.ExternalLink("no_such_file.h5", "/top")


def gen_link_nonascii(path):
    """Non-ASCII link names in a file written at the earliest bound.

    h5py encodes a name it is handed as `str` to ASCII when it can and to
    UTF-8 when it cannot, and puts the result in the lcpl's character encoding
    (`CommonStateObject._e`). `H5G_obj_insert` reads that back off the link and
    converts a symbol-table group to link messages the moment it is not ASCII
    (`obj_lnk->cset != H5T_CSET_ASCII`, H5Gobj.c:514) — the same branch an
    external link takes.

    So the root here loses its symbol table over two non-ASCII names while its
    ASCII siblings come along as link messages, and every group whose own
    children are ASCII-named keeps its symbol table. The superblock stays at
    version 0 throughout: the conversion is per group, not per file.
    """
    with h5py.File(path, "w", libver="earliest") as f:
        f.create_dataset("데이터", data=ramp("<i4"))
        f.create_group("plain")
        f.create_group("그룹").create_dataset("inner", data=ramp("<i4", 4))
        f.create_group("ascii_only").create_dataset("inner", data=ramp("<i4", 4))


def gen_links_dense(path):
    # v1.8 bounds, not "latest": dense link storage needs the v1.8 group
    # format, and stopping there keeps the v1.10 layout message out of the
    # file so this case isolates link storage.
    with h5py.File(path, "w", libver=("v108", "v108")) as f:
        g = f.create_group("g", track_order=True)
        for i in range(12):
            g.create_dataset("d%02d" % i, data=np.array([i], dtype="<i4"))


def gen_track_order(path):
    # Creation-order tracking adds a second index (a v2 B-tree keyed on
    # creation order) beside the name index, on both links and attributes.
    with h5py.File(path, "w", track_order=True) as f:
        for name in ("zebra", "apple", "mango"):
            f.create_group(name)
        for i, key in enumerate(("zeta", "alpha", "mu")):
            f.attrs.create(key, np.int32(i))
        g = f.create_group("g", track_order=True)
        g.create_dataset("data", data=ramp("<i4"))
        g.attrs.create("second", np.int32(2))
        g.attrs.create("first", np.int32(1))


def gen_group_storage_modern_root(path):
    """A symbol-table group, holding children, under a link-message root.

    `track_order` migrates the group that asks for it and nothing else, so one
    h5py call writes a root using link messages over a child still using the
    legacy symbol table. A reader that walks a group the way its parent is
    stored lists `legacy` and finds none of its children.
    """
    with h5py.File(path, "w", track_order=True) as f:
        legacy = f.create_group("legacy")
        legacy.create_dataset("a", data=ramp("<i4"))
        legacy.create_group("inner").create_dataset("c", data=ramp("<i2"))


def gen_group_storage_legacy_root(path):
    """The same mismatch the other way up: a link-message group, holding
    children, under a symbol-table root."""
    with h5py.File(path, "w") as f:
        f.create_group("legacy").create_dataset("a", data=ramp("<i4"))
        modern = f.create_group("modern", track_order=True)
        modern.create_dataset("b", data=ramp("<f8"))
        modern.create_group("inner").create_dataset("c", data=ramp("<i2"))


LINK_CASES = [
    Case("groups_nested", "group", gen_groups_nested, "groups_nested",
         "three levels of nested groups plus an empty leaf group"),
    Case("link_hard", "link", gen_link_hard, "link_hard",
         "two names for one object"),
    Case("link_soft", "link", gen_link_soft, "link_soft", "soft link to /orig"),
    Case("link_external", "link", gen_link_external, "link_external",
         "external link into a sibling file",
         ext_files=("_ext.h5",)),
    Case("link_external_read", "link", gen_link_external_read,
         "link_external_read",
         "datasets read through external links, plus a dangling object and a "
         "dangling file",
         ext_files=("_data.h5",)),
    Case("link_nonascii", "link", gen_link_nonascii, "link_nonascii",
         "non-ASCII link names at the earliest bound — the root converts to "
         "link messages, the ASCII-named subgroups keep their symbol tables"),
    Case("links_dense", "link", gen_links_dense, "links_dense",
         "12 links in one group — dense link storage (fractal heap + v2 B-tree)"),
    Case("track_order", "group", gen_track_order, "track_order",
         "creation-order indices on links and attributes"),
    Case("group_storage_modern_root", "group", gen_group_storage_modern_root,
         "group_storage_modern_root",
         "symbol-table group with children under a link-message root"),
    Case("group_storage_legacy_root", "group", gen_group_storage_legacy_root,
         "group_storage_legacy_root",
         "link-message group with children under a symbol-table root"),
]


# --------------------------------------------------------------------------
# attributes
# --------------------------------------------------------------------------


def gen_attr_scalar_num(path):
    with h5py.File(path, "w") as f:
        ds = f.create_dataset("data", data=ramp("<i4"))
        ds.attrs.create("gain", np.float64(2.5))
        ds.attrs.create("count", np.int32(7))


def gen_attr_array_num(path):
    with h5py.File(path, "w") as f:
        ds = f.create_dataset("data", data=ramp("<i4"))
        ds.attrs.create("offsets", np.arange(4, dtype="<i4"))
        ds.attrs.create("matrix", np.arange(6, dtype="<f8").reshape(2, 3))


def gen_attr_string(path):
    with h5py.File(path, "w") as f:
        ds = f.create_dataset("data", data=ramp("<i4"))
        ds.attrs.create("units", "volt", dtype=h5py.string_dtype("utf-8"))
        f.create_group("g").attrs.create(
            "NX_class", "NXdetector", dtype=h5py.string_dtype("utf-8")
        )


def gen_attrs_dense(path):
    # v1.8 bounds for the same reason as links_dense: dense attribute storage
    # arrives with v1.8, and staying there isolates it from the v1.10 layout
    # message.
    with h5py.File(path, "w", libver=("v108", "v108")) as f:
        ds = f.create_dataset("data", data=ramp("<i4"))
        for i in range(12):
            ds.attrs.create("a%02d" % i, np.int32(i))


def gen_attrs_dense_group(path):
    # The same phase change on a group and on the root group, where the
    # attributes share their object header with the link messages rather than
    # with a dataset's layout.
    with h5py.File(path, "w", libver=("v108", "v108")) as f:
        g = f.create_group("g")
        for i in range(12):
            g.attrs.create("g%02d" % i, np.int32(i))
        for i in range(12):
            f.attrs.create("r%02d" % i, np.int32(i))
        f.create_dataset("data", data=ramp("<i4"))


def gen_attr_on_root(path):
    with h5py.File(path, "w") as f:
        f.attrs.create("title", "root", dtype=h5py.string_dtype("utf-8"))
        f.attrs.create("version", np.int64(3))
        f.create_dataset("data", data=ramp("<i4"))


def gen_attr_ref_object(path):
    # An attribute whose value is object references, on all three kinds of
    # object header: a dataset's, a group's and the root's. The value is part
    # of the header message, so unlike a reference dataset's elements it cannot
    # be stamped in after the header is written.
    with h5py.File(path, "w") as f:
        ds = f.create_dataset("data", data=ramp("<i4"))
        g = f.create_group("grp")
        ds.attrs.create("neighbours", np.array([ds.ref, g.ref], dtype=h5py.ref_dtype))
        g.attrs.create("owner", ds.ref, dtype=h5py.ref_dtype)
        f.attrs.create("entry", np.array([g.ref, ds.ref], dtype=h5py.ref_dtype))


def gen_attr_large(path):
    # One attribute past the 64 KiB object-header message limit: the value
    # spills to dense storage no matter how few attributes there are.
    with h5py.File(path, "w", libver=("v108", "v108")) as f:
        ds = f.create_dataset("data", data=ramp("<i4"))
        ds.attrs.create("big", np.arange(25600, dtype="<i4"))


ATTR_CASES = [
    Case("attr_scalar_num", "attribute", gen_attr_scalar_num, "attr_scalar_num",
         "scalar f64 and i32 attributes"),
    Case("attr_array_num", "attribute", gen_attr_array_num, "attr_array_num",
         "1-D and 2-D numeric attributes"),
    Case("attr_string", "attribute", gen_attr_string, "attr_string",
         "vlen UTF-8 string attributes on a dataset and a group"),
    Case("attrs_dense", "attribute", gen_attrs_dense, "attrs_dense",
         "12 attributes — dense attribute storage"),
    Case("attrs_dense_group", "attribute", gen_attrs_dense_group,
         "attrs_dense_group",
         "12 attributes on a group and on the root — dense storage"),
    Case("attr_on_root", "attribute", gen_attr_on_root, "attr_on_root",
         "attributes on the root group"),
    Case("attr_large", "attribute", gen_attr_large, "attr_large",
         "single 100 KiB attribute — dense storage forced by size, not count"),
    Case("attr_ref_object", "attribute", gen_attr_ref_object, "attr_ref_object",
         "object references stored in attributes of a dataset, a group and "
         "the root"),
]


# --------------------------------------------------------------------------
# library version bounds / superblock
# --------------------------------------------------------------------------


def _libver_case(name, libver, rust, note):
    def gen(path):
        with h5py.File(path, "w", libver=libver) as f:
            f.create_dataset("data", data=ramp("<i4"))
            f.create_group("g")

    return Case(name, "superblock", gen, rust, note)


def gen_userblock(path):
    """A 512-byte userblock in front of the superblock.

    The block is filled with text afterwards, as an application that keeps a
    script or a header there would: the reader has to find the superblock at
    512 rather than at 0, and must not mistake the block's bytes for metadata.
    """
    with h5py.File(path, "w", userblock_size=512) as f:
        f.create_dataset("data", data=ramp("<i4"))
        f.create_group("g")
    prefix = b"#!/bin/sh\n# userblock\n"
    with open(path, "r+b") as fh:
        fh.write(prefix + b"#" * (512 - len(prefix) - 1) + b"\n")


def _reopen_append_case(name, libver, rust, note):
    """Create under one bound, reopen under the default fapl, append.

    `H5F__super_init` is the only place a superblock version is decided, and
    open never re-decides it; `H5F__super_read` instead raises the file's low
    library bound to match the version it finds — v2 to `H5F_LIBVER_V18`, v3
    to `H5F_LIBVER_V110`. So the appended dataset is written in whatever
    generation the file already has, not the one the default fapl would have
    picked, and the superblock version comes out of the reopen unchanged.

    The appended dataset is chunked because that is where the two generations
    differ most visibly: a version-3 layout message has the version-1 B-tree
    and nothing else, while a version-4 one picks among the v1.10 indexes.
    """
    def gen(path):
        with h5py.File(path, "w", libver=libver) as f:
            f.create_dataset("data", data=ramp("<i4"))
            f.create_group("g")
        with h5py.File(path, "a") as f:
            f.create_dataset("appended", data=ramp("<i4", 16).reshape(4, 4),
                             chunks=(2, 4))

    return Case(name, "superblock", gen, rust, note)


LIBVER_CASES = [
    _libver_case("libver_earliest", "earliest", "libver_earliest",
                 "libver earliest — superblock v0, symbol-table groups"),
    _libver_case("libver_v108", ("v108", "v108"), "libver_v108",
                 "libver v1.8 bounds"),
    _libver_case("libver_v110", ("v110", "v110"), "libver_v110",
                 "libver v1.10 bounds"),
    _libver_case("libver_latest", "latest", "libver_latest",
                 "libver latest — superblock v3, new-style groups"),
    # Superblock v1 is not reachable from h5py 3.15: it is produced only by a
    # non-default B-tree K value (H5Pset_sym_k / H5Pset_istore_k), and neither
    # is wrapped on PropFCID. v0, v2 and v3 are covered by the four cases
    # above; the user block below is the remaining v0 variant.
    Case("userblock", "superblock", gen_userblock, "userblock",
         "512-byte userblock — the superblock, and every address, is based at 512"),
    _reopen_append_case(
        "reopen_append_earliest", "earliest", "reopen_append_earliest",
        "classic file reopened under the default fapl — stays superblock v0, "
        "the appended chunked dataset takes the version-1 B-tree"),
    _reopen_append_case(
        "reopen_append_v108", ("v108", "v108"), "reopen_append_v108",
        "v1.8 file reopened under the default fapl — stays superblock v2, "
        "the appended chunked dataset takes the version-1 B-tree"),
    _reopen_append_case(
        "reopen_append_latest", "latest", "reopen_append_latest",
        "v1.10+ file reopened under the default fapl — stays superblock v3, "
        "the appended chunked dataset takes a v1.10 index"),
]


# --------------------------------------------------------------------------
# SWMR and bulk
# --------------------------------------------------------------------------


def gen_swmr_created(path):
    with h5py.File(path, "w", libver="latest") as f:
        ds = f.create_dataset("stream", (0, 4), maxshape=(None, 4), chunks=(1, 4),
                              dtype="<f4")
        f.swmr_mode = True
        for i in range(8):
            ds.resize((i + 1, 4))
            ds[i, :] = np.arange(i * 4, i * 4 + 4, dtype="<f4")
            ds.flush()


def gen_large_multi_mb(path):
    with h5py.File(path, "w", libver="latest") as f:
        data = np.arange(512 * 512, dtype="<f8").reshape(512, 512)
        f.create_dataset("big", data=data, chunks=(64, 512))


def gen_fsm_persist(path):
    """A persisting free-space-manager file, reopened and appended to.

    `fs_persist` is what makes the free-space managers on-disk structures
    rather than in-memory bookkeeping: the close writes an H5FS header and
    section-info block per allocation type and names them in the file-space
    info message. Nothing is freed by the create alone, so the reopen and
    append are the half that matters — superseding the superblock extension
    and the root group's object header is what puts sections in a manager,
    and `#freespace` is `tracked` only if they were written back.
    """
    with h5py.File(path, "w", fs_strategy="fsm", fs_persist=True, fs_threshold=1) as f:
        f.create_dataset("data", data=ramp("<i4"))
        f.create_dataset("bulk", data=ramp("<i4", 256))
        f.create_group("g")
    with h5py.File(path, "a") as f:
        del f["bulk"]
        f.create_dataset("appended", data=ramp("<i4"))


def gen_fsm_persist_page(path):
    """[`gen_fsm_persist`] under paged aggregation.

    `H5F_FSPACE_STRATEGY_PAGE` packs everything smaller than the file-space
    page into pages of one kind and page-aligns everything else
    (`H5MF__alloc_pagefs`), so the same edits leave a different free-space
    manager set: sections that never cross a page boundary, and — under sec2,
    which declares no `H5FD_FEAT_PAGED_AGGR` — at most the three managers
    `H5MF__alloc_to_fs_type` can reach.
    """
    with h5py.File(path, "w", fs_strategy="page", fs_persist=True, fs_threshold=1) as f:
        f.create_dataset("data", data=ramp("<i4"))
        f.create_dataset("bulk", data=ramp("<i4", 256))
        f.create_group("g")
    with h5py.File(path, "a") as f:
        del f["bulk"]
        f.create_dataset("appended", data=ramp("<i4"))


def gen_fsm_page_size(path):
    """[`gen_fsm_persist_page`] on a page size that is not the library default.

    `H5Pset_file_space_page_size` takes anything from 512
    (`H5F_FILE_SPACE_PAGE_SIZE_MIN`) to 1 GiB with no power-of-two
    requirement, and the size decides which side of `H5MF__alloc_to_fs_type`
    every request falls on. At 512 the `bulk` dataset's raw data is larger
    than a page and is page-aligned as `H5F_MEM_PAGE_GENERIC`, where at the
    4096-byte default the same bytes are packed into a page — so this is a
    different manager set, not the same file with a different number in the
    message.
    """
    with h5py.File(path, "w", fs_strategy="page", fs_persist=True,
                   fs_threshold=1, fs_page_size=512) as f:
        f.create_dataset("data", data=ramp("<i4"))
        f.create_dataset("bulk", data=ramp("<i4", 256))
        f.create_group("g")
    with h5py.File(path, "a") as f:
        del f["bulk"]
        f.create_dataset("appended", data=ramp("<i4"))


MISC_CASES = [
    Case("swmr_created", "swmr", gen_swmr_created, "swmr_created",
         "file created through the SWMR writer path and appended frame by frame"),
    Case("fsm_persist", "freespace", gen_fsm_persist, "fsm_persist",
         "persisting FSM_AGGR file reopened and appended — the freed blocks "
         "must come back as free-space manager sections"),
    Case("fsm_persist_page", "freespace", gen_fsm_persist_page, "fsm_persist_page",
         "persisting PAGE file reopened and appended — the freed blocks must "
         "come back as page-shaped free-space manager sections"),
    Case("fsm_page_size", "freespace", gen_fsm_page_size, "fsm_page_size",
         "persisting PAGE file on a 512-byte file-space page — the non-default "
         "size must reach the message and shape the allocation"),
    Case("large_multi_mb", "bulk", gen_large_multi_mb, "large_multi_mb",
         "2 MiB chunked f64 dataset — payload compared by SHA-256"),
]


# --------------------------------------------------------------------------
# checked-in fixtures
#
# Some file-level features have no h5py binding at all, so the reference file
# cannot be written from Python. Those come from a C generator run against the
# pinned libhdf5 (`tests/fixtures/gen_*.sh`), are checked in, and are copied
# into the run directory here. h5py still reads them, so direction A compares
# exactly as it does for a generated case.
# --------------------------------------------------------------------------

FIXTURE_DIR = pathlib.Path(__file__).resolve().parent.parent / "tests" / "fixtures"


def _fixture_case(name, fixture, generator, group, note, rust=None):
    def gen(path):
        src = FIXTURE_DIR / fixture
        if not src.exists():
            raise FileNotFoundError(
                "%s is missing; regenerate it with tests/fixtures/%s"
                % (src, generator)
            )
        shutil.copyfile(src, path)

    # `rust` is None where the public API cannot ask for the file at all;
    # where it can, the arm mirrors the C generator rather than an h5py one.
    return Case(name, group, gen, rust, note)


def _fixture_append_case(name, fixture, generator, group, note, rust):
    """A checked-in fixture libhdf5 then reopens and appends to.

    The create is the half h5py cannot express — there is no binding for
    `H5Pset_shared_mesg_index` — but the *append* is an ordinary `'a'` open,
    so the reference is a genuine libhdf5 reopen of a file with a
    shared-message table. The rust arm creates its own equivalent and reopens
    that, which is the only shape the comparison can take: nothing on the
    Python side can hand the rust writer a file it did not create.
    """
    def gen(path):
        src = FIXTURE_DIR / fixture
        if not src.exists():
            raise FileNotFoundError(
                "%s is missing; regenerate it with tests/fixtures/%s"
                % (src, generator)
            )
        shutil.copyfile(src, path)
        with h5py.File(path, "a") as f:
            f.create_dataset("appended", data=ramp("<i4", 8))

    return Case(name, group, gen, rust, note)


FIXTURE_CASES = [
    _fixture_case(
        "sohm_list", "sohm_list.h5", "gen_sohm.sh", "sohm",
        "shared datatype/dataspace/attribute messages, list index "
        "(H5Pset_shared_mesg_index) + a committed datatype",
        rust="sohm_list",
    ),
    _fixture_case(
        "sohm_btree", "sohm_btree.h5", "gen_sohm.sh", "sohm",
        "the same file with the shared-message index forced to a v2 B-tree",
        rust="sohm_btree",
    ),
    _fixture_append_case(
        "sohm_list_append", "sohm_list.h5", "gen_sohm.sh", "sohm",
        "a file with a shared-message list index reopened and appended to — "
        "the table is laid out whole, so the append replaces it",
        rust="sohm_list_append",
    ),
    _fixture_append_case(
        "sohm_btree_append", "sohm_btree.h5", "gen_sohm.sh", "sohm",
        "the same reopen over a v2 B-tree index",
        rust="sohm_btree_append",
    ),
    _fixture_case(
        "ochk_root", "ochk_root.h5", "gen_ochk.sh", "objectheader",
        "root group whose object header spills into two continuation chunks",
        rust="ochk_root",
    ),
    _fixture_case(
        "vds_late_layout", "vds_late_layout.h5", "gen_vds_late_layout.sh",
        "layout",
        "virtual datasets built by H5Pset_virtual with and without a prior "
        "H5Pset_layout — the pairs differ only in allocation time "
        "(H5Pdcpl.c:2146 pokes the layout past H5P__set_layout's default)",
    ),
]


# --------------------------------------------------------------------------

ALL_CASES = (
    INT_CASES
    + FLOAT_CASES
    + STRING_CASES
    + COMPOSITE_CASES
    + LAYOUT_CASES
    + FILTER_CASES
    + FILL_CASES
    + SPACE_CASES
    + LINK_CASES
    + ATTR_CASES
    + LIBVER_CASES
    + MISC_CASES
    + FIXTURE_CASES
)


def by_name(name):
    for c in ALL_CASES:
        if c.name == name:
            return c
    raise KeyError(name)


if __name__ == "__main__":
    print("%d cases" % len(ALL_CASES))
    for c in ALL_CASES:
        print("  %-24s %-16s rust=%s" % (c.name, c.group, c.rust or "-"))