ad_plugins_rs/

file_hdf5.rs

use std::path::{Path, PathBuf};

use ad_core_rs::attributes::{NDAttrDataType, NDAttrSource, NDAttrValue, NDAttribute};
use ad_core_rs::codec::{Codec, CodecName};
use ad_core_rs::error::{ADError, ADResult};
use ad_core_rs::finalize::Finalize;
use ad_core_rs::ndarray::{NDArray, NDDataBuffer, NDDataType, NDDimension};
use ad_core_rs::ndarray_pool::NDArrayPool;
use ad_core_rs::plugin::file_base::{FileAttributes, NDFileMode, NDFileWriter};
use ad_core_rs::plugin::file_controller::FilePluginController;
use ad_core_rs::plugin::runtime::{
    NDPluginProcess, ParamChangeResult, ParamUpdate, PluginParamSnapshot, ProcessResult,
};
use parking_lot::Mutex;

use rust_hdf5::H5File;
use rust_hdf5::format::messages::datatype::{ByteOrder, DatatypeMessage};
use rust_hdf5::format::messages::filter::{
    FILTER_BLOSC, FILTER_BSHUF, FILTER_JPEG, FILTER_SZIP, Filter, FilterPipeline,
};
use rust_hdf5::swmr::SwmrFileWriter;

use crate::hdf5_layout::Hdf5Layout;

/// C ADCore compression type enum values (matching NDFileHDF5.h).
const COMPRESS_NONE: i32 = 0;
const COMPRESS_NBIT: i32 = 1;
const COMPRESS_SZIP: i32 = 2;
const COMPRESS_ZLIB: i32 = 3;
const COMPRESS_BLOSC: i32 = 4;
const COMPRESS_BSHUF: i32 = 5;
const COMPRESS_LZ4: i32 = 6;
const COMPRESS_JPEG: i32 = 7;

/// SZIP nearest-neighbor coding option mask (`H5_SZIP_NN_OPTION_MASK`,
/// H5Zpublic.h); C passes this to `H5Pset_szip` (NDFileHDF5.cpp:3372).
const SZIP_NN_OPTION_MASK: u32 = 32;

/// C ADCore BLOSC compressor sub-types.
const BLOSC_LZ: i32 = 0;
const BLOSC_LZ4: i32 = 1;
const BLOSC_LZ4HC: i32 = 2;
const BLOSC_SNAPPY: i32 = 3;
const BLOSC_ZLIB: i32 = 4;
const BLOSC_ZSTD: i32 = 5;

/// Maximum number of extra dimensions (C `MAXEXTRADIMS`).
const MAX_EXTRA_DIMS: usize = 10;

/// Name of the HDF5 attribute that records the NDArray data type ordinal
/// (matches C `NDDataType_t`). `read_file` uses it to recover the exact type.
const DTYPE_ATTR: &str = "NDArrayDataType";

/// Built-in NeXus layout used when no user layout XML is configured, mirroring
/// C `LayoutXML::DEFAULT_LAYOUT` (NDFileHDF5LayoutXML.cpp:43-70) which C loads
/// via `layout.load_xml()` for an empty `HDF5_layoutFilename` (NDFileHDF5.cpp:
/// 3896-3906). It places the detector image at `/entry/instrument/detector/data`
/// inside a full NXentry/NXinstrument/NXdetector/NXcollection/NXdata tree, with
/// an `/entry/data/data` hardlink to the detector dataset.
const DEFAULT_LAYOUT_XML: &str = r#"
<hdf5_layout>
<group name="entry">
  <attribute name="NX_class" source="constant" value="NXentry" type="string"></attribute>
  <group name="instrument">
    <attribute name="NX_class" source="constant" value="NXinstrument" type="string"></attribute>
    <group name="detector">
      <attribute name="NX_class" source="constant" value="NXdetector" type="string"></attribute>
      <dataset name="data" source="detector" det_default="true">
        <attribute name="signal" source="constant" value="1" type="int"></attribute>
      </dataset>
      <group name="NDAttributes">
        <attribute name="NX_class" source="constant" value="NXcollection" type="string"></attribute>
        <dataset name="ColorMode" source="ndattribute" ndattribute="ColorMode"></dataset>
      </group>
    </group>
    <group name="NDAttributes" ndattr_default="true">
      <attribute name="NX_class" source="constant" value="NXcollection" type="string"></attribute>
    </group>
    <group name="performance">
      <dataset name="timestamp"></dataset>
    </group>
  </group>
  <group name="data">
    <attribute name="NX_class" source="constant" value="NXdata" type="string"></attribute>
    <hardlink name="data" target="/entry/instrument/detector/data"></hardlink>
  </group>
</group>
</hdf5_layout>
"#;

/// User-controlled chunk geometry (C `HDF5_*Chunks` params).
#[derive(Clone)]
struct ChunkConfig {
    /// `HDF5_chunkSizeAuto` — when true, ignore the explicit row/col/frame
    /// values and let the writer pick (full-frame spatial, one frame deep).
    auto: bool,
    n_row_chunks: usize,
    n_col_chunks: usize,
    n_frames_chunks: usize,
    /// `HDF5_NDAttributeChunk` — chunk depth for NDAttribute datasets.
    /// `0` means auto (C param default), resolved by `attribute_chunking`.
    ndattr_chunk: usize,
}

impl Default for ChunkConfig {
    fn default() -> Self {
        Self {
            auto: true,
            n_row_chunks: 0,
            n_col_chunks: 0,
            n_frames_chunks: 1,
            ndattr_chunk: 0,
        }
    }
}

/// One extra-dimension entry (C `HDF5_extraDimSizeN` / `HDF5_extraDimNameN`).
#[derive(Clone, Default)]
struct ExtraDim {
    size: usize,
    name: String,
}

/// State for a single open attribute time-series dataset (one per NDAttribute).
/// Mirrors C++ `NDFileHDF5AttributeDataset`: a 1-D extensible dataset holding
/// one numeric (or string) value per frame.
struct AttributeDataset {
    name: String,
    /// The three remaining self-describing strings C writes as HDF5 attributes
    /// on each attribute dataset (NDFileHDF5.cpp:2715, 2785-2788): description,
    /// source, and the source-type string. Captured at create time because the
    /// source NDArray is gone by flush.
    description: String,
    source: String,
    source_type: String,
    data_type: NDAttrDataType,
    /// C `NDFileHDF5AttributeDataset::isUndefined_`
    /// (`typeAsHdf`'s `default:` arm, NDFileHDF5AttributeDataset.cpp:366-370).
    /// Fixed for the dataset's whole life: an attribute carrying no value when
    /// the file opens gets a float32 dataset whose `epicsNAN` fill no frame
    /// ever overwrites, because every element write is gated on
    /// `!isUndefined_` (`:160`).
    undefined: bool,
    /// Raw little-endian bytes accumulated, one element per frame.
    buffer: Vec<u8>,
    frames: usize,
}

/// C `NDAttribute` source-type string (NDAttribute.cpp:49-62), recorded as the
/// `NDAttrSourceType` HDF5 attribute on each NDAttribute dataset.
fn nd_attr_source_type_string(source: &NDAttrSource) -> &'static str {
    match source {
        NDAttrSource::Driver => "NDAttrSourceDriver",
        NDAttrSource::Param { .. } => "NDAttrSourceParam",
        NDAttrSource::EpicsPV(_) => "NDAttrSourceEPICSPV",
        NDAttrSource::Function(_) => "NDAttrSourceFunct",
        NDAttrSource::Constant(_) => "NDAttrSourceConst",
        NDAttrSource::Undefined => "Undefined",
    }
}

impl AttributeDataset {
    fn new(attr: &NDAttribute) -> Self {
        // C decides the HDF5 datatype and the fill value ONCE, from the
        // attribute's type at file-open: `createDataset` calls `typeAsHdf`
        // (NDFileHDF5AttributeDataset.cpp:73-97) and the type came from
        // `ndAttr->getDataType()` when the node was constructed
        // (NDFileHDF5.cpp:2797, :2830). Nothing re-runs it, so a concrete
        // value arriving in a later frame cannot change either. An attribute
        // with no value takes the `default:` arm — `H5T_NATIVE_FLOAT` with an
        // `epicsNAN` fill (`:366-370`) — not the type of whichever value
        // happens to arrive first.
        let undefined = matches!(attr.value, NDAttrValue::Undefined);
        Self {
            name: attr.name.clone(),
            description: attr.description.clone(),
            source: attr.source.source_string().to_string(),
            source_type: nd_attr_source_type_string(&attr.source).to_string(),
            data_type: if undefined {
                NDAttrDataType::Float32
            } else {
                attr.value.data_type()
            },
            undefined,
            buffer: Vec::new(),
            frames: 0,
        }
    }

    /// Element byte width for this attribute's numeric type. Strings are
    /// stored as a fixed-width field (matching C++ `MAX_ATTRIBUTE_STRING_SIZE`).
    fn element_size(&self) -> usize {
        match self.data_type {
            NDAttrDataType::Int8 | NDAttrDataType::UInt8 => 1,
            NDAttrDataType::Int16 | NDAttrDataType::UInt16 => 2,
            NDAttrDataType::Int32 | NDAttrDataType::UInt32 | NDAttrDataType::Float32 => 4,
            NDAttrDataType::Int64 | NDAttrDataType::UInt64 | NDAttrDataType::Float64 => 8,
            NDAttrDataType::String => MAX_ATTRIBUTE_STRING_SIZE,
        }
    }

    /// Append one frame's value, encoding it to the dataset's native type.
    fn push(&mut self, value: &NDAttrValue) {
        let es = self.element_size();
        let mut bytes = vec![0u8; es];
        if self.undefined {
            // C never writes an element into an undefined dataset, so the
            // `epicsNAN` fill installed by `H5Pset_fill_value`
            // (NDFileHDF5AttributeDataset.cpp:106) stands for every frame.
            // The port materialises the dataset whole at flush rather than
            // element-by-element, so it emits the fill explicitly; on disk and
            // to a reader the two are the same bytes.
            bytes.copy_from_slice(&f32::NAN.to_le_bytes());
            self.buffer.extend_from_slice(&bytes);
            self.frames += 1;
            return;
        }
        match self.data_type {
            NDAttrDataType::Int8 => bytes[0] = value.as_i64().unwrap_or(0) as i8 as u8,
            NDAttrDataType::UInt8 => bytes[0] = value.as_i64().unwrap_or(0) as u8,
            NDAttrDataType::Int16 => {
                bytes.copy_from_slice(&(value.as_i64().unwrap_or(0) as i16).to_le_bytes())
            }
            NDAttrDataType::UInt16 => {
                bytes.copy_from_slice(&(value.as_i64().unwrap_or(0) as u16).to_le_bytes())
            }
            NDAttrDataType::Int32 => {
                bytes.copy_from_slice(&(value.as_i64().unwrap_or(0) as i32).to_le_bytes())
            }
            NDAttrDataType::UInt32 => {
                bytes.copy_from_slice(&(value.as_i64().unwrap_or(0) as u32).to_le_bytes())
            }
            NDAttrDataType::Int64 => {
                bytes.copy_from_slice(&(value.as_i64().unwrap_or(0)).to_le_bytes())
            }
            NDAttrDataType::UInt64 => {
                bytes.copy_from_slice(&(value.as_i64().unwrap_or(0) as u64).to_le_bytes())
            }
            NDAttrDataType::Float32 => {
                bytes.copy_from_slice(&(value.as_f64().unwrap_or(0.0) as f32).to_le_bytes())
            }
            NDAttrDataType::Float64 => {
                bytes.copy_from_slice(&(value.as_f64().unwrap_or(0.0)).to_le_bytes())
            }
            NDAttrDataType::String => {
                let s = value.as_string();
                let src = s.as_bytes();
                let n = src.len().min(es - 1);
                bytes[..n].copy_from_slice(&src[..n]);
            }
        }
        self.buffer.extend_from_slice(&bytes);
        self.frames += 1;
    }
}

/// Fixed string field width for string-typed attribute datasets
/// (C++ `MAX_ATTRIBUTE_STRING_SIZE`).
const MAX_ATTRIBUTE_STRING_SIZE: usize = 256;

/// Element marker for a 256-byte fixed-length HDF5 string attribute dataset.
///
/// C `NDFileHDF5AttributeDataset.cpp:321-323` stores a string-valued
/// NDAttribute as a rank-1 dataset whose element datatype is `H5T_C_S1` sized
/// to `MAX_ATTRIBUTE_STRING_SIZE` bytes (null-terminated, ASCII). Implementing
/// `H5Type` so `hdf5_type()` returns that fixed-length string datatype lets the
/// high-level dataset builder emit `H5T_STR_NULLTERM` strings rather than a 2-D
/// `H5T_STD_U8LE` byte array. The element is byte-identical to the raw 256-byte
/// field already accumulated per frame, so the existing byte-oriented write
/// path is reused unchanged.
#[derive(Clone, Copy)]
#[repr(transparent)]
struct FixedStr256([u8; MAX_ATTRIBUTE_STRING_SIZE]);

impl rust_hdf5::types::H5Type for FixedStr256 {
    fn hdf5_type() -> rust_hdf5::format::messages::datatype::DatatypeMessage {
        rust_hdf5::format::messages::datatype::DatatypeMessage::fixed_string(
            MAX_ATTRIBUTE_STRING_SIZE as u32,
        )
    }

    fn element_size() -> usize {
        MAX_ATTRIBUTE_STRING_SIZE
    }
}

/// Write state for one detector-source dataset (C `NDFileHDF5Dataset` held in
/// `detDataMap`). Each `<dataset source="detector">` node gets its own live
/// dataset handle, leading-axis frame counter, and partial chunk band so
/// `detector_data_destination` routing can send frames to different datasets,
/// each extending independently. The on-disk datatype/dataspace/frame shape are
/// shared (set from the first frame) and live on `Hdf5Writer`.
struct DetectorDataset {
    /// Live dataset handle, retained across frames so the leading dimension can
    /// be extended (`H5File::dataset` cannot re-open a dataset in write mode).
    ds: rust_hdf5::H5Dataset,
    /// Frames routed to THIS dataset — its leading-axis index and final extent.
    frame_count: usize,
    /// LE bytes of frames buffered for THIS dataset until a `nFramesChunks`-deep
    /// chunk band fills. With one frame per chunk this holds at most one frame.
    frame_band: Vec<Vec<u8>>,
}

/// Internal handle: either a standard H5File or a SWMR streaming writer.
enum Hdf5Handle {
    Standard {
        file: H5File,
        /// Detector image datasets keyed by C full name (leading slash) — the
        /// C `detDataMap`. Created lazily on the first frame; one entry for the
        /// common single-`<dataset source="detector">` layout.
        detectors: std::collections::HashMap<String, DetectorDataset>,
    },
    Swmr {
        // Boxed: `SwmrFileWriter` is much larger than the `Standard` variant.
        writer: Box<SwmrFileWriter>,
        ds_index: usize,
        /// True only when a compression type was requested but no filter
        /// pipeline could be built for it; false when compression is applied.
        compression_dropped: bool,
        /// `Some` when the streaming dataset is a fixed multi-extra-dimension
        /// grid (`HDF5_nExtraDims >= 1`, uncompressed): frames are placed at
        /// odometer positions via `write_chunk_at`. `None` for the common
        /// single frame-axis case, which streams via `append_frame`.
        grid: Option<SwmrGridLayout>,
    },
}

/// Fixed multi-extra-dimension grid layout for the SWMR write path. The
/// streaming dataset is a `create_grid_dataset` of full shape
/// `[eds[N], …, eds[0], Y, X]`; each frame is written at its odometer chunk
/// position with `SwmrFileWriter::write_chunk_at`, mirroring the standard
/// path's [`Hdf5Writer::flush_band`]. Compressed multi-extra-dim stays
/// collapsed (no compressed grid constructor; backend-blocked).
struct SwmrGridLayout {
    /// Fixed leading axes, outermost-first (`extra_dim_axes`).
    leading: Vec<usize>,
    /// HDF5-order frame dims `[Y, X]` (fastest-varying last).
    frame_dims: Vec<usize>,
    /// Full per-chunk shape `[1, …, 1, rc, cc]` (same rank as the dataset).
    chunk: Vec<usize>,
    /// Element size in bytes.
    elem_size: usize,
}

/// HDF5 file writer using the rust-hdf5 crate.
pub struct Hdf5Writer {
    current_path: Option<PathBuf>,
    handle: Option<Hdf5Handle>,
    /// Total frames written to the file this open (C `nextRecord` across all
    /// detector datasets): drives the create-on-first-frame gate, the
    /// `frame_count()` accessor and the attribute-dataset row count. Per-dataset
    /// leading-axis placement uses each [`DetectorDataset`]'s own `frame_count`.
    frame_count: usize,
    dataset_name: String,
    /// Cached data type of the open primary dataset.
    open_data_type: Option<NDDataType>,
    /// Cached spatial (per-frame) dimensions, fastest-varying last.
    open_frame_dims: Option<Vec<usize>>,
    /// Codec of the open detector dataset(s) when they were created from a
    /// pre-compressed first frame (direct chunk write path). `None` for an
    /// uncompressed file. Mirrors C `NDFileHDF5Dataset::codec`, used by
    /// `verifyChunking` to reject a mid-file codec change.
    open_codec: Option<Codec>,
    // compression
    compression_type: i32,
    z_compress_level: u32,
    szip_num_pixels: u32,
    nbit_precision: u32,
    nbit_offset: u32,
    jpeg_quality: u32,
    blosc_shuffle_type: i32,
    blosc_compressor: i32,
    blosc_compress_level: u32,
    // chunking & layout
    chunk: ChunkConfig,
    /// File write mode of the currently-open file (C `NDFileWriteMode`), and the
    /// configured capture target (C `NDFileNumCapture`, pushed by the controller
    /// before `open_file`). Both feed `attribute_chunking` — C's
    /// `calculateAttributeChunking` resolves the auto (0) chunk from them.
    open_mode: NDFileMode,
    num_capture: usize,
    n_extra_dims: usize,
    extra_dims: [ExtraDim; MAX_EXTRA_DIMS],
    fill_value: f64,
    dim_att_datasets: bool,
    // SWMR
    swmr_mode: bool,
    flush_nth_frame: usize,
    pub swmr_cb_counter: u32,
    // options
    pub store_attributes: bool,
    pub store_performance: bool,
    /// Whether to fsync on file close (durable). Default `true`; set
    /// `AD_HDF5_FSYNC_ON_CLOSE=0` (or `false`/`no`/`off`) to skip the
    /// close-time fsync on the standard (non-SWMR) write path, trading
    /// durability for a faster close (`H5File::close_no_sync`, rust-hdf5
    /// 0.3.2). SWMR is unaffected — its high-level writer exposes no
    /// no-fsync close, and it streams/flushes incrementally regardless.
    fsync_on_close: bool,
    pub total_runtime: f64,
    pub total_bytes: u64,
    /// Per-frame I/O timing rows for the `timestamp` performance dataset.
    /// Each row is the 5 doubles C++ `writePerformanceDataset` records.
    perf_rows: Vec<[f64; 5]>,
    perf_prev: Option<std::time::Instant>,
    perf_first: Option<std::time::Instant>,
    /// Open NDAttribute time-series datasets, keyed by attribute name.
    attr_datasets: Vec<AttributeDataset>,
    /// C `pFileAttributes` (NDFileHDF5.cpp:243, :1437): the writer's own
    /// attribute list, sticky for the life of the file.
    file_attributes: FileAttributes,
    /// Layout XML state.
    layout_filename: Option<PathBuf>,
    layout: Option<Hdf5Layout>,
    pub layout_valid: bool,
    pub layout_error: String,
    /// Full path of the primary image dataset for the currently-open file.
    /// `"data"` (flat root) when no valid layout is loaded; the layout's
    /// `det_default` dataset path (e.g. `entry/instrument/detector/data`)
    /// otherwise. Leading slash stripped — keyed as `rust-hdf5` keys datasets.
    resolved_dataset_path: String,
    /// Group prefix (no leading/trailing slash) for NDAttribute datasets.
    /// Empty when flat; the layout `ndattr_default` group otherwise.
    resolved_ndattr_group: String,
    /// Group prefix for the performance dataset. Empty when flat.
    resolved_perf_group: String,
    /// Live NDAttribute values for the names referenced by layout
    /// `<attribute source="ndattribute">` element-attrs (ADP-79). `first` is
    /// the open-time frame (C `storeOnOpenAttributes` → OnFileOpen/OnFrame),
    /// `last` the most recent frame (C `storeOnCloseAttributes` → OnFileClose).
    /// Empty for the default layout, which declares no such element-attrs.
    ndattr_first_values: std::collections::HashMap<String, NDAttrValue>,
    ndattr_last_values: std::collections::HashMap<String, NDAttrValue>,
    /// Distinct NDAttribute names referenced by element-attrs, captured at
    /// open so per-frame updates need not re-walk the layout.
    ndattr_element_names: Vec<String>,
}

impl Hdf5Writer {
    pub fn new() -> Self {
        Self {
            current_path: None,
            handle: None,
            frame_count: 0,
            dataset_name: "data".to_string(),
            open_data_type: None,
            open_frame_dims: None,
            open_codec: None,
            compression_type: 0,
            z_compress_level: 6,
            szip_num_pixels: 16,
            // C default precision=8 bit, offset=0 (NDFileHDF5.cpp:2340-2341):
            // a default-config N-bit request packs to 8 significant bits.
            nbit_precision: 8,
            nbit_offset: 0,
            jpeg_quality: 90,
            // C default bloscShuffleType=1 (byte shuffle), NDFileHDF5.cpp:2344.
            blosc_shuffle_type: 1,
            blosc_compressor: 0,
            blosc_compress_level: 5,
            chunk: ChunkConfig::default(),
            open_mode: NDFileMode::Single,
            num_capture: 1,
            n_extra_dims: 0,
            extra_dims: Default::default(),
            fill_value: 0.0,
            dim_att_datasets: false,
            swmr_mode: false,
            flush_nth_frame: 0,
            swmr_cb_counter: 0,
            store_attributes: true,
            store_performance: false,
            fsync_on_close: Self::env_fsync_on_close(),
            total_runtime: 0.0,
            total_bytes: 0,
            perf_rows: Vec::new(),
            perf_prev: None,
            perf_first: None,
            attr_datasets: Vec::new(),
            file_attributes: FileAttributes::default(),
            layout_filename: None,
            // No user layout XML by default → C's built-in NeXus DEFAULT_LAYOUT.
            layout: Some(Self::default_layout()),
            layout_valid: false,
            layout_error: String::new(),
            resolved_dataset_path: "data".to_string(),
            resolved_ndattr_group: String::new(),
            resolved_perf_group: String::new(),
            ndattr_first_values: std::collections::HashMap::new(),
            ndattr_last_values: std::collections::HashMap::new(),
            ndattr_element_names: Vec::new(),
        }
    }

    /// Read `AD_HDF5_FSYNC_ON_CLOSE` to decide the close-time fsync policy.
    fn env_fsync_on_close() -> bool {
        Self::parse_fsync_on_close_env(std::env::var("AD_HDF5_FSYNC_ON_CLOSE").ok().as_deref())
    }

    /// Pure parse of the `AD_HDF5_FSYNC_ON_CLOSE` value. `None` (unset) or any
    /// value outside the falsey set keeps the durable default (`true`);
    /// `0` / `false` / `no` / `off` (trimmed, case-insensitive) opts into the
    /// no-fsync fast close on the standard write path.
    fn parse_fsync_on_close_env(v: Option<&str>) -> bool {
        match v {
            Some(s) => !matches!(
                s.trim().to_ascii_lowercase().as_str(),
                "0" | "false" | "no" | "off"
            ),
            None => true,
        }
    }

    pub fn set_dataset_name(&mut self, name: &str) {
        self.dataset_name = name.to_string();
    }

    pub fn set_compression_type(&mut self, v: i32) {
        self.compression_type = v;
    }

    pub fn set_z_compress_level(&mut self, v: u32) {
        self.z_compress_level = v;
    }

    pub fn set_szip_num_pixels(&mut self, v: u32) {
        self.szip_num_pixels = v;
    }

    pub fn set_blosc_shuffle_type(&mut self, v: i32) {
        self.blosc_shuffle_type = v;
    }

    pub fn set_blosc_compressor(&mut self, v: i32) {
        self.blosc_compressor = v;
    }

    pub fn set_blosc_compress_level(&mut self, v: u32) {
        self.blosc_compress_level = v;
    }

    pub fn set_nbit_precision(&mut self, v: u32) {
        self.nbit_precision = v;
    }

    pub fn set_nbit_offset(&mut self, v: u32) {
        self.nbit_offset = v;
    }

    pub fn set_jpeg_quality(&mut self, v: u32) {
        self.jpeg_quality = v;
    }

    pub fn set_store_attributes(&mut self, v: bool) {
        self.store_attributes = v;
    }

    pub fn set_store_performance(&mut self, v: bool) {
        self.store_performance = v;
    }

    pub fn set_swmr_mode(&mut self, v: bool) {
        self.swmr_mode = v;
    }

    pub fn set_flush_nth_frame(&mut self, v: usize) {
        self.flush_nth_frame = v;
    }

    pub fn set_chunk_size_auto(&mut self, v: bool) {
        self.chunk.auto = v;
    }

    pub fn set_n_row_chunks(&mut self, v: usize) {
        self.chunk.n_row_chunks = v;
    }

    pub fn set_n_col_chunks(&mut self, v: usize) {
        self.chunk.n_col_chunks = v;
    }

    pub fn set_n_frames_chunks(&mut self, v: usize) {
        self.chunk.n_frames_chunks = v;
    }

    pub fn set_ndattr_chunk(&mut self, v: usize) {
        // `0` is the auto sentinel (C `HDF5_NDAttributeChunk` default); preserve
        // it rather than clamping to 1, so `attribute_chunking` can resolve it.
        self.chunk.ndattr_chunk = v;
    }

    pub fn set_n_extra_dims(&mut self, v: usize) {
        // C `NDFileHDF5::writeInt32` rejects `HDF5_nExtraDims > MAXEXTRADIMS-1`
        // (NDFileHDF5.cpp:1789). The dataset gains `nExtraDims+1` leading axes
        // (the virtual scan dims plus the innermost frames-per-point axis), so
        // axis i reads `extra_dims[nExtraDims - i]`; the highest index used is
        // `nExtraDims`, which must stay within `extra_dims[0..MAX_EXTRA_DIMS]`.
        self.n_extra_dims = v.min(MAX_EXTRA_DIMS - 1);
    }

    pub fn set_extra_dim_size(&mut self, idx: usize, size: usize) {
        if idx < MAX_EXTRA_DIMS {
            self.extra_dims[idx].size = size;
        }
    }

    pub fn set_extra_dim_name(&mut self, idx: usize, name: &str) {
        if idx < MAX_EXTRA_DIMS {
            self.extra_dims[idx].name = name.to_string();
        }
    }

    pub fn set_fill_value(&mut self, v: f64) {
        self.fill_value = v;
    }

    pub fn set_dim_att_datasets(&mut self, v: bool) {
        self.dim_att_datasets = v;
    }

    /// Parse the built-in [`DEFAULT_LAYOUT_XML`] into an `Hdf5Layout`. The
    /// constant is known-good (covered by a unit test), so a parse failure is a
    /// build bug, not a runtime condition.
    fn default_layout() -> Hdf5Layout {
        Hdf5Layout::parse(DEFAULT_LAYOUT_XML).expect("built-in default layout must parse")
    }

    /// Set the layout XML filename and (re)parse it. Returns whether parsing
    /// succeeded; `layout_error` carries any message (C `HDF5_layoutErrorMsg`).
    pub fn set_layout_filename(&mut self, path: &str) -> bool {
        if path.trim().is_empty() {
            // C loads the built-in DEFAULT_LAYOUT for an empty filename
            // (NDFileHDF5.cpp:3896-3906); `layout_valid` tracks user-file
            // validity, so it stays false here.
            self.layout_filename = None;
            self.layout = Some(Self::default_layout());
            self.layout_valid = false;
            self.layout_error.clear();
            return true;
        }
        let p = PathBuf::from(path);
        match Hdf5Layout::from_file(&p) {
            Ok(layout) => {
                self.layout_filename = Some(p);
                self.layout = Some(layout);
                self.layout_valid = true;
                self.layout_error.clear();
                true
            }
            Err(e) => {
                self.layout_filename = Some(p);
                self.layout = None;
                self.layout_valid = false;
                self.layout_error = e.0;
                false
            }
        }
    }

    pub fn frame_count(&self) -> usize {
        self.frame_count
    }

    /// Trigger a SWMR flush. No-op if not in SWMR mode.
    pub fn flush_swmr(&mut self) {
        if let Some(Hdf5Handle::Swmr { ref mut writer, .. }) = self.handle {
            if writer.flush().is_ok() {
                self.swmr_cb_counter += 1;
            }
        }
    }

    /// Returns true if SWMR is currently active.
    pub fn is_swmr_active(&self) -> bool {
        matches!(self.handle, Some(Hdf5Handle::Swmr { .. }))
    }

    /// Whether a requested SWMR compression type had no buildable pipeline.
    pub fn swmr_compression_dropped(&self) -> bool {
        matches!(
            self.handle,
            Some(Hdf5Handle::Swmr {
                compression_dropped: true,
                ..
            })
        )
    }

    /// Build a FilterPipeline from the current compression settings.
    fn build_pipeline(&self, element_size: usize) -> Option<FilterPipeline> {
        match self.compression_type {
            COMPRESS_NONE => None,
            COMPRESS_ZLIB => Some(FilterPipeline::deflate(self.z_compress_level)),
            COMPRESS_SZIP => Some(FilterPipeline {
                filters: vec![Filter {
                    id: FILTER_SZIP,
                    flags: 0,
                    // C calls H5Pset_szip(cparms, H5_SZIP_NN_OPTION_MASK, szipNumPixels)
                    // (NDFileHDF5.cpp:3372): nearest-neighbor coding (mask 32), not
                    // entropy coding (mask 4). cd_values[1] is pixels_per_block; the
                    // rust-hdf5 SZIP filter defaults bits_per_pixel/pixels_per_scanline.
                    cd_values: vec![SZIP_NN_OPTION_MASK, self.szip_num_pixels],
                }],
            }),
            COMPRESS_LZ4 => Some(FilterPipeline::lz4()),
            COMPRESS_BSHUF => Some(FilterPipeline {
                // Bitshuffle (HDF5 filter 32008): cd_values are
                // [major_ver, minor_ver, elem_size, block_size, comp_type].
                // comp_type 2 == LZ4, matching ADCore's default bitshuffle.
                filters: vec![Filter {
                    id: FILTER_BSHUF,
                    flags: 0,
                    cd_values: vec![0, 0, element_size as u32, 0, 2],
                }],
            }),
            COMPRESS_BLOSC => {
                let compressor_code = match self.blosc_compressor {
                    BLOSC_LZ => 0,
                    BLOSC_LZ4 => 1,
                    BLOSC_LZ4HC => 2,
                    BLOSC_SNAPPY => 3,
                    BLOSC_ZLIB => 4,
                    BLOSC_ZSTD => 5,
                    _ => 0,
                };
                Some(FilterPipeline {
                    filters: vec![Filter {
                        id: FILTER_BLOSC,
                        flags: 0,
                        cd_values: vec![
                            2,
                            2,
                            element_size as u32,
                            0,
                            self.blosc_compress_level,
                            self.blosc_shuffle_type as u32,
                            compressor_code,
                        ],
                    }],
                })
            }
            COMPRESS_NBIT => {
                // C (NDFileHDF5.cpp:3355-3357) applies N-bit compression by
                // narrowing the dataset DATATYPE — H5Tset_precision /
                // H5Tset_offset — and then registering a parameterless
                // H5Pset_nbit filter; libhdf5's set_local callback derives the
                // nbit `cd_values` parameter tree from that reduced-precision
                // datatype. The filter therefore cannot be expressed by a
                // pipeline alone: it must travel together with a reduced
                // datatype override on the dataset builder. `build_pipeline` is
                // shared with the SWMR streaming path, whose
                // `create_streaming_dataset_chunked_compressed` hardcodes
                // `T::hdf5_type()` and has no datatype-override hook — so N-bit
                // is handled out of band by `nbit_packing` on the standard
                // write path (which CAN override the datatype) and returns
                // `None` here so SWMR keeps degrading N-bit to uncompressed
                // instead of writing an nbit filter over a full-width datatype.
                None
            }
            COMPRESS_JPEG => Some(FilterPipeline {
                filters: vec![Filter {
                    id: FILTER_JPEG,
                    flags: 0,
                    cd_values: vec![self.jpeg_quality],
                }],
            }),
            _ => None,
        }
    }

    /// N-bit packing for the standard write path: the reduced-precision on-disk
    /// datatype override and the matching nbit filter, derived together.
    ///
    /// C narrows `this->datatype` via `H5Tset_precision`/`H5Tset_offset` and
    /// then registers a parameterless `H5Pset_nbit` (NDFileHDF5.cpp:3355-3357);
    /// libhdf5's `set_local` callback reads the reduced datatype to build the
    /// filter `cd_values`. Here `DatasetBuilder::datatype` supplies the reduced
    /// `FixedPoint` datatype and `FilterPipeline::nbit` builds the matching
    /// `cd_values` tree `[nparms, need_not_compress, d_nelmts, NBIT_ATOMIC,
    /// size, order, precision, offset]`, so the dataset is byte-readable by
    /// h5py / libhdf5.
    ///
    /// Returns `Some((datatype, pipeline))` only for `COMPRESS_NBIT` over an
    /// integer (FixedPoint) element type. Float element types do not map to a
    /// reduced-precision fixed-point datatype (N-bit over floats is not an
    /// areaDetector use case), and an out-of-range precision (0, which HDF5's
    /// `H5Tset_precision` rejects) cannot pack — both return `None`, leaving the
    /// frame on the lossless uncompressed path.
    ///
    /// `chunk_nelmts` is the element count of one on-disk chunk (= `d_nelmts`);
    /// `flush_band` always writes a full, zero-padded chunk, so every filtered
    /// write supplies exactly `chunk_nelmts * element_size` bytes — the size
    /// `apply_nbit` requires.
    fn nbit_packing(
        &self,
        data_type: NDDataType,
        chunk_nelmts: usize,
    ) -> Option<(DatatypeMessage, FilterPipeline)> {
        if self.compression_type != COMPRESS_NBIT {
            return None;
        }
        let (size, signed) = match data_type {
            NDDataType::Int8 => (1u32, true),
            NDDataType::UInt8 => (1, false),
            NDDataType::Int16 => (2, true),
            NDDataType::UInt16 => (2, false),
            NDDataType::Int32 => (4, true),
            NDDataType::UInt32 => (4, false),
            NDDataType::Int64 => (8, true),
            NDDataType::UInt64 => (8, false),
            NDDataType::Float32 | NDDataType::Float64 => return None,
        };
        if self.nbit_precision == 0 || self.nbit_precision > size * 8 {
            return None;
        }
        let dt = DatatypeMessage::FixedPoint {
            size,
            byte_order: ByteOrder::LittleEndian,
            signed,
            bit_offset: self.nbit_offset as u16,
            bit_precision: self.nbit_precision as u16,
        };
        let pipeline = FilterPipeline::nbit(&dt, chunk_nelmts);
        Some((dt, pipeline))
    }

    /// Build the HDF5 filter pipeline that matches a **pre-compressed** input
    /// array's codec, so direct chunk write records the filter the compressed
    /// bytes were produced with. Mirrors C `NDFileHDF5::configureCompression`
    /// (NDFileHDF5.cpp:3314-3331), which overrides the writer's own compression
    /// settings from `pArray->codec`. Returns `None` (caller rejects the frame,
    /// as C `verifyChunking` does) for any codec C does not direct-chunk-write:
    /// the only `NDCODEC` values C handles are JPEG, BLOSC, LZ4 and BSLZ4
    /// (Codec.h:12-18); the Rust-only `Zlib`/`LZ4HDF5`/`None` have no C analog.
    ///
    /// `element_size` is the size of one **uncompressed** element (the codec's
    /// `original_data_type`), which the BLOSC/bitshuffle filters record as the
    /// type size.
    fn codec_filter_pipeline(&self, codec: &Codec) -> Option<FilterPipeline> {
        let element_size = codec.original_data_type.element_size() as u32;
        match codec.name {
            CodecName::LZ4 => Some(FilterPipeline::lz4()),
            CodecName::BSLZ4 => Some(FilterPipeline {
                // Bitshuffle (HDF5 filter 32008), matching `build_pipeline`'s
                // BSHUF arm: [major, minor, elem_size, block_size(0=auto),
                // comp_type=2 (LZ4)].
                filters: vec![Filter {
                    id: FILTER_BSHUF,
                    flags: 0,
                    cd_values: vec![0, 0, element_size, 0, 2],
                }],
            }),
            CodecName::Blosc => {
                // C copies the array's own blosc params (level/shuffle/
                // compressor) into the dataset filter (configureCompression
                // NDFileHDF5.cpp:3320-3323), not the writer's defaults.
                Some(FilterPipeline {
                    filters: vec![Filter {
                        id: FILTER_BLOSC,
                        flags: 0,
                        cd_values: vec![
                            2,
                            2,
                            element_size,
                            0,
                            codec.level.max(0) as u32,
                            codec.shuffle.max(0) as u32,
                            codec.compressor.max(0) as u32,
                        ],
                    }],
                })
            }
            // C `configureCompression` does not copy a JPEG quality from the
            // array; the dataset's JPEG filter carries the writer's configured
            // `HDF5_jpegQuality`. Quality is a compression-side parameter, unused
            // when the reader reverses the filter, so direct-chunk-write parity
            // is unaffected by its exact value.
            CodecName::JPEG => Some(FilterPipeline {
                filters: vec![Filter {
                    id: FILTER_JPEG,
                    flags: 0,
                    cd_values: vec![self.jpeg_quality],
                }],
            }),
            CodecName::None | CodecName::Zlib | CodecName::LZ4HDF5 => None,
        }
    }

    /// Chunk dims for the frame (image) axes only — no leading dims.
    ///
    /// For a 2-D frame these are the row/column chunk sizes
    /// (`HDF5_nRowChunks` / `HDF5_nColChunks`; 0, auto, or out-of-range → the
    /// full dimension, matching C++ `NDFileHDF5` chunk-size selection). Other
    /// ranks get one full per-frame tile (no sub-tiling).
    fn frame_chunk_dims(&self, frame_dims: &[usize]) -> Vec<usize> {
        if frame_dims.len() == 2 {
            let y = frame_dims[0].max(1);
            let x = frame_dims[1].max(1);
            vec![
                Self::clamp_chunk(self.chunk.n_row_chunks, y, self.chunk.auto),
                Self::clamp_chunk(self.chunk.n_col_chunks, x, self.chunk.auto),
            ]
        } else {
            frame_dims.iter().map(|&d| d.max(1)).collect()
        }
    }

    /// The fixed leading extra-dimension axes, outermost-first, for the
    /// standard (non-SWMR) write path; `None` when `HDF5_nExtraDims == 0`.
    ///
    /// `HDF5_nExtraDims = N` produces `N+1` leading axes: the `N` virtual scan
    /// dimensions plus the innermost frames-per-point axis. C builds them in
    /// reverse param order (NDFileHDF5.cpp:3182-3215, doc order
    /// `{Nth virtual, …, Y, X, frames-per-point, frame Y, frame X}`): axis `i`
    /// uses `extraDimSize[N - i]`, so outermost is `extraDimSize[N]` and the
    /// innermost leading axis is `extraDimSize[0]` ("N", frames per point).
    fn extra_dim_axes(&self) -> Option<Vec<usize>> {
        if self.n_extra_dims == 0 {
            return None;
        }
        Some(
            (0..=self.n_extra_dims)
                .rev()
                .map(|i| self.extra_dims[i].size.max(1))
                .collect(),
        )
    }

    /// Standard (non-SWMR) dataset shape and chunk geometry for the primary
    /// image dataset, faithful to C `NDFileHDF5::configureDims`.
    ///
    /// Layout, fastest-varying last. With no extra dims it is `[frame, Y, X]`
    /// and the single leading axis is extensible. With `HDF5_nExtraDims = N`
    /// it is `[eds[N], …, eds[1], eds[0], Y, X]` — `N+1` fixed leading axes
    /// created at full configured size, each chunked at 1 (the frame data is
    /// row-major identical to the collapsed form, only the dataspace rank
    /// differs). `close_file` calls `set_extent` to trim any frame-axis
    /// chunk-rounding back to the exact frame shape.
    ///
    /// Returns `(shape, chunk, leading)` where `leading` is `Some([eds…])`
    /// (the fixed leading axes, outermost-first) when extra dims fix the
    /// dataset up front, or `None` when the single leading frame axis is
    /// extended per write.
    fn standard_layout(
        &self,
        frame_dims: &[usize],
    ) -> (Vec<usize>, Vec<usize>, Option<Vec<usize>>) {
        let frame_chunk = self.frame_chunk_dims(frame_dims);
        match self.extra_dim_axes() {
            Some(leading) => {
                let mut shape = leading.clone();
                shape.extend_from_slice(frame_dims);
                // Each leading axis is chunked at 1 (extraDimChunk defaults to
                // 1; per-axis chunking is not plumbed in this port).
                let mut chunk = vec![1usize; leading.len()];
                chunk.extend_from_slice(&frame_chunk);
                (shape, chunk, Some(leading))
            }
            None => {
                let mut shape = vec![1usize];
                shape.extend_from_slice(frame_dims);
                let mut chunk = vec![self.chunk.n_frames_chunks.max(1)];
                chunk.extend_from_slice(&frame_chunk);
                (shape, chunk, None)
            }
        }
    }

    /// Dataset shape and chunk geometry for a pre-compressed (direct chunk
    /// write) detector dataset. Identical leading-axis structure to
    /// `standard_layout`, but every chunk holds exactly **one whole frame**:
    /// the codec compressed each frame as a single unit, so C
    /// `NDFileHDF5Dataset::verifyChunking` (NDFileHDF5Dataset.cpp:191-233)
    /// requires the leading frame-axis chunk == 1 and every frame-axis chunk ==
    /// the full frame dimension (no row/column sub-tiling). A pre-compressed
    /// chunk cannot be split, so `HDF5_nRowChunks`/`nColChunks`/`nFramesChunks`
    /// do not apply here.
    fn compressed_layout(
        &self,
        frame_dims: &[usize],
    ) -> (Vec<usize>, Vec<usize>, Option<Vec<usize>>) {
        let frame_chunk: Vec<usize> = frame_dims.iter().map(|&d| d.max(1)).collect();
        match self.extra_dim_axes() {
            Some(leading) => {
                let mut shape = leading.clone();
                shape.extend_from_slice(frame_dims);
                let mut chunk = vec![1usize; leading.len()];
                chunk.extend_from_slice(&frame_chunk);
                (shape, chunk, Some(leading))
            }
            None => {
                let mut shape = vec![1usize];
                shape.extend_from_slice(frame_dims);
                let mut chunk = vec![1usize];
                chunk.extend_from_slice(&frame_chunk);
                (shape, chunk, None)
            }
        }
    }

    /// Collapsed single-leading-axis layout used by the SWMR streaming path,
    /// whose backend (`SwmrFileWriter`) supports only one extensible frame
    /// axis. With extra dims the leading axis is fixed at the product of the
    /// extra-dim sizes (the N-dimensional structure is recorded as HDF5
    /// attributes); the non-SWMR path uses `standard_layout`, which builds
    /// C's full multi-extra-dimension dataspace.
    ///
    /// Returns `(shape, chunk, extra_dim_extent)` where `extra_dim_extent` is
    /// `Some(total_frames)` when extra dims fix the dataset size up front, or
    /// `None` when the leading frame axis is extended per write.
    fn primary_layout(&self, frame_dims: &[usize]) -> (Vec<usize>, Vec<usize>, Option<usize>) {
        let extra_extent = if self.n_extra_dims > 0 {
            Some(
                (0..self.n_extra_dims)
                    .map(|i| self.extra_dims[i].size.max(1))
                    .product::<usize>(),
            )
        } else {
            None
        };

        let mut shape: Vec<usize> = vec![extra_extent.unwrap_or(1)];
        shape.extend_from_slice(frame_dims);

        let mut chunk = vec![if extra_extent.is_some() {
            1
        } else {
            self.chunk.n_frames_chunks.max(1)
        }];
        chunk.extend_from_slice(&self.frame_chunk_dims(frame_dims));
        (shape, chunk, extra_extent)
    }

    /// Row-major unravel of a flat frame index into per-axis coordinates over
    /// `dims` (outermost first; the innermost axis varies fastest). This is C's
    /// extra-dimension odometer — `NDFileHDF5Dataset::extendDataSet`
    /// (NDFileHDF5Dataset.cpp:137-157) increments the innermost axis first and
    /// carries outward.
    fn unravel(mut idx: usize, dims: &[usize]) -> Vec<usize> {
        let mut coords = vec![0usize; dims.len()];
        for d in (0..dims.len()).rev() {
            let s = dims[d].max(1);
            coords[d] = idx % s;
            idx /= s;
        }
        coords
    }

    /// C++ `NDFileHDF5` chunk-size rule: 0, auto, or a value larger than the
    /// dimension means "chunk the whole dimension"; otherwise the user value.
    fn clamp_chunk(requested: usize, dim: usize, auto: bool) -> usize {
        if auto || requested == 0 || requested > dim {
            dim
        } else {
            requested
        }
    }

    /// Write one chunk band of `chunk[0]` consecutive frames into the primary
    /// dataset at the given `leading_coords` (the chunk-grid coordinates of the
    /// leading axes, outermost-first).
    ///
    /// For the extensible single-axis layout `leading_coords` is `[band_idx]`
    /// and up to `fc = chunk[0]` frames stack along that axis. For the fixed
    /// multi-extra-dimension layout `leading_coords` is the full
    /// `[eds[N], …, eds[0]]` odometer position (each leading chunk is 1, so
    /// `fc == 1` and one frame is written per call).
    ///
    /// The frame is split into `ceil(Y/rc) x ceil(X/cc)` chunk tiles, each
    /// written with `write_chunk_at(leading_coords ++ [row_tile, col_tile], ..)`.
    /// Tiles are `[fc, rc, cc]`; edge tiles and a partial final band (fewer
    /// than `fc` frames) are zero-padded. `close_file`'s `set_extent` trims
    /// the resulting over-extension back to the exact frame shape.
    fn flush_band(
        ds: &rust_hdf5::H5Dataset,
        leading_coords: &[usize],
        frames: &[Vec<u8>],
        frame_dims: &[usize],
        chunk: &[usize],
        elem_size: usize,
    ) -> ADResult<()> {
        for (coords, buf) in
            Self::band_chunk_writes(leading_coords, frames, frame_dims, chunk, elem_size)
        {
            ds.write_chunk_at(&coords, &buf).map_err(|e| {
                ADError::UnsupportedConversion(format!("HDF5 write_chunk_at error: {}", e))
            })?;
        }
        Ok(())
    }

    /// Compute the `(chunk_coords, chunk_bytes)` writes for one band of `frames`
    /// at `leading_coords`. The single source of tile math shared by the
    /// standard path's [`flush_band`](Self::flush_band) (writes each via
    /// `H5Dataset::write_chunk_at`) and the SWMR grid path
    /// ([`write_swmr`](Self::write_swmr), via `SwmrFileWriter::write_chunk_at`).
    ///
    /// `chunk` is the full per-chunk shape `[fc, frame_chunk…]`; `fc = chunk[0]`
    /// frames stack along the first leading axis. A 2-D frame is split into
    /// `ceil(Y/rc) x ceil(X/cc)` tiles `[fc, rc, cc]`; a non-2-D frame is one
    /// chunk per band. Edge tiles and a partial final band (fewer than `fc`
    /// frames) are zero-padded. Coords are in chunk units, row-major:
    /// `leading_coords ++ [row_tile, col_tile]` (or `++ [0; frame_rank]`).
    fn band_chunk_writes(
        leading_coords: &[usize],
        frames: &[Vec<u8>],
        frame_dims: &[usize],
        chunk: &[usize],
        elem_size: usize,
    ) -> Vec<(Vec<usize>, Vec<u8>)> {
        let lead = leading_coords.len();
        let fc = chunk[0];
        // Non-2-D frame: one chunk per band, frames stacked along the first
        // leading axis (a partial band leaves trailing frames zero).
        if frame_dims.len() != 2 {
            let frame_len = frame_dims.iter().product::<usize>() * elem_size;
            let mut buf = vec![0u8; fc * frame_len];
            for (f, fb) in frames.iter().take(fc).enumerate() {
                buf[f * frame_len..f * frame_len + frame_len].copy_from_slice(fb);
            }
            let mut coords = leading_coords.to_vec();
            coords.extend(std::iter::repeat_n(0, frame_dims.len()));
            return vec![(coords, buf)];
        }

        let (y, x) = (frame_dims[0], frame_dims[1]);
        let (rc, cc) = (chunk[lead], chunk[lead + 1]);
        let row_tiles = y.div_ceil(rc);
        let col_tiles = x.div_ceil(cc);
        let mut writes = Vec::with_capacity(row_tiles * col_tiles);
        for ry in 0..row_tiles {
            for cx in 0..col_tiles {
                let mut tile = vec![0u8; fc * rc * cc * elem_size];
                for f in 0..fc {
                    let Some(fb) = frames.get(f) else {
                        break; // partial band: trailing frames stay zero
                    };
                    for r in 0..rc {
                        let sy = ry * rc + r;
                        if sy >= y {
                            break;
                        }
                        for c in 0..cc {
                            let sx = cx * cc + c;
                            if sx >= x {
                                break;
                            }
                            let src = (sy * x + sx) * elem_size;
                            let dst = ((f * rc + r) * cc + c) * elem_size;
                            tile[dst..dst + elem_size].copy_from_slice(&fb[src..src + elem_size]);
                        }
                    }
                }
                let mut coords = leading_coords.to_vec();
                coords.push(ry);
                coords.push(cx);
                writes.push((coords, tile));
            }
        }
        writes
    }

    /// Flush each detector dataset's partial frame band and trim its logical
    /// extent to its own frame count. Called from `close_file`. Every dataset in
    /// `detDataMap` is finalised independently, each at the count of frames the
    /// `detector_data_destination` routing sent it (C closes every
    /// `NDFileHDF5Dataset` it created, not just the default).
    fn finalize_standard_datasets(&mut self) -> ADResult<()> {
        let Some(frame_dims) = self.open_frame_dims.clone() else {
            return Ok(());
        };
        let (_, chunk, leading) = self.standard_layout(&frame_dims);
        let elem_size = self.open_data_type.map(|t| t.element_size()).unwrap_or(1);
        let fc = chunk[0];
        let Some(Hdf5Handle::Standard { detectors, .. }) = self.handle.as_mut() else {
            return Ok(());
        };
        for det in detectors.values_mut() {
            let total = det.frame_count;
            // A partial band only survives the extensible single-axis layout
            // (the fixed multi-dim layout writes one frame per call, `fc == 1`).
            if !det.frame_band.is_empty() {
                let last = total.saturating_sub(1);
                let leading_coords = match &leading {
                    Some(lead) => Self::unravel(last, lead),
                    None => vec![last / fc],
                };
                Self::flush_band(
                    &det.ds,
                    &leading_coords,
                    &det.frame_band,
                    &frame_dims,
                    &chunk,
                    elem_size,
                )?;
                det.frame_band.clear();
            }
            // Trim the logical extent: write_chunk_at rounds dims up to chunk
            // boundaries; set_extent restores the exact frame shape. The
            // extensible axis trims to `total`; the fixed leading axes are
            // already exact and only the frame axes may have over-extended.
            if total > 0 {
                let mut dims = leading.clone().unwrap_or_else(|| vec![total]);
                dims.extend_from_slice(&frame_dims);
                det.ds.set_extent(&dims).map_err(|e| {
                    ADError::UnsupportedConversion(format!("HDF5 set_extent error: {}", e))
                })?;
            }
        }
        Ok(())
    }

    /// Open file in SWMR streaming mode.
    ///
    /// Ordering mirrors C `NDFileHDF5::openFile` (`NDFileHDF5.cpp:173`-`337`):
    /// the file layout tree and datasets are created, then `createHardLinks`
    /// (`NDFileHDF5.cpp:319`-`320`) runs, and only then `startSWMR`
    /// (`NDFileHDF5.cpp:323`-`325`). `SwmrFileWriter` exposes `create_group` /
    /// `create_hard_link` callable before `start_swmr()`; a group or link
    /// created before `start_swmr()` is visible to SWMR readers for the whole
    /// streaming window. So here the image dataset is placed at the layout's
    /// nested `resolved_dataset_path` and the layout `<hardlink>` elements are
    /// materialised before SWMR mode is entered — not on the close path.
    ///
    /// The layout groups are built first because the dataset carries its full
    /// nested path as its name, and rust-hdf5 resolves that path against the
    /// groups that already exist (libhdf5's rule: the default link creation
    /// property list has `H5Pset_create_intermediate_group` off). The parent
    /// group the path names is therefore also the dataset's single placement
    /// owner — nothing re-parents it afterwards.
    fn open_swmr(&mut self, path: &Path, array: &NDArray) -> ADResult<()> {
        let mut swmr = SwmrFileWriter::create(path)
            .map_err(|e| ADError::UnsupportedConversion(format!("SWMR create error: {}", e)))?;
        self.build_swmr_layout_groups(&mut swmr)?;

        let usize_frame_dims: Vec<usize> = array.dims.iter().rev().map(|d| d.size).collect();
        let frame_dims: Vec<u64> = usize_frame_dims.iter().map(|&d| d as u64).collect();

        // Full chunk geometry, `[fc, rc, cc]`: HDF5_nFramesChunks deep and
        // the row/column tile sizes. rust-hdf5 0.2.15
        // `create_streaming_dataset_chunked` band-buffers whole frames and
        // zero-pads the final partial band at close, keeping the logical
        // frame count exact.
        let element_size = array.data.data_type().element_size();
        let pipeline = self.build_pipeline(element_size);
        let chunk: Vec<u64> = {
            let (_, c, _) = self.primary_layout(&usize_frame_dims);
            c.iter().map(|&v| v as u64).collect()
        };

        // With `HDF5_nExtraDims >= 1` and no compression, mirror the standard
        // path's full multi-extra-dimension dataspace: a fixed grid of shape
        // `[eds[N], …, eds[0], Y, X]` filled at odometer positions via
        // `write_chunk_at`. Compressed multi-extra-dim has no grid constructor
        // and stays collapsed to the single-axis streaming layout above.
        let (grid_shape_usize, grid_chunk_usize, grid_leading) =
            self.standard_layout(&usize_frame_dims);
        let use_grid = grid_leading.is_some() && pipeline.is_none();
        let grid_shape: Vec<u64> = grid_shape_usize.iter().map(|&v| v as u64).collect();
        let grid_chunk: Vec<u64> = grid_chunk_usize.iter().map(|&v| v as u64).collect();

        // The streaming dataset is created with its full nested layout path
        // as the dataset name (default flat `data` without a layout). The
        // `SwmrFileWriter` links it into the group that path names — built
        // just above — under just the leaf, while keeping the full name
        // addressable so a layout `<hardlink target="/entry/.../data">`
        // resolves against it.
        let ds_name = self.resolved_dataset_path.clone();

        macro_rules! create_ds {
            ($t:ty) => {
                if use_grid {
                    swmr.create_grid_dataset::<$t>(&ds_name, &grid_shape, &grid_chunk)
                        .map_err(|e| {
                            ADError::UnsupportedConversion(format!(
                                "SWMR create grid dataset error: {}",
                                e
                            ))
                        })
                } else {
                    match pipeline.clone() {
                        Some(pl) => swmr
                            .create_streaming_dataset_chunked_compressed::<$t>(
                                &ds_name,
                                &frame_dims,
                                &chunk,
                                pl,
                            )
                            .map_err(|e| {
                                ADError::UnsupportedConversion(format!(
                                    "SWMR create compressed dataset error: {}",
                                    e
                                ))
                            }),
                        None => swmr
                            .create_streaming_dataset_chunked::<$t>(&ds_name, &frame_dims, &chunk)
                            .map_err(|e| {
                                ADError::UnsupportedConversion(format!(
                                    "SWMR create dataset error: {}",
                                    e
                                ))
                            }),
                    }
                }
            };
        }

        let ds_index = match array.data.data_type() {
            NDDataType::Int8 => create_ds!(i8)?,
            NDDataType::UInt8 => create_ds!(u8)?,
            NDDataType::Int16 => create_ds!(i16)?,
            NDDataType::UInt16 => create_ds!(u16)?,
            NDDataType::Int32 => create_ds!(i32)?,
            NDDataType::UInt32 => create_ds!(u32)?,
            NDDataType::Int64 => create_ds!(i64)?,
            NDDataType::UInt64 => create_ds!(u64)?,
            NDDataType::Float32 => create_ds!(f32)?,
            NDDataType::Float64 => create_ds!(f64)?,
        };

        // Materialise the image dataset's constant attributes and the layout
        // `<hardlink>` elements — all BEFORE `start_swmr()` so SWMR readers
        // see the nested paths and aliases for the whole streaming window.
        // C `NDFileHDF5.cpp:319`-`325`: `createHardLinks` then `startSWMR`.
        self.write_swmr_layout_dataset_attrs(&mut swmr, ds_index)?;
        self.write_swmr_ndarray_default_attrs(&mut swmr, ds_index, array)?;
        self.build_swmr_layout_hardlinks(&mut swmr)?;
        // Open-time `<attribute source="ndattribute">` group and dataset
        // element-attrs must be created before the SWMR lock; close-time ones
        // cannot (HDF5 forbids attribute creation after the lock).
        self.write_swmr_ndattr_element_attrs(&mut swmr, ds_index)?;

        swmr.start_swmr()
            .map_err(|e| ADError::UnsupportedConversion(format!("SWMR start error: {}", e)))?;

        // Compression is applied to SWMR datasets via the filter pipeline
        // above. `compression_dropped` is only set when a compression type was
        // requested but no pipeline could be built for it (an unsupported
        // compressor) — never a silent drop.
        let compression_dropped = self.compression_type != COMPRESS_NONE && pipeline.is_none();
        if compression_dropped {
            eprintln!(
                "NDFileHDF5: WARNING — SWMR mode requested compression type {} \
                 but no filter pipeline could be built for it; the SWMR file \
                 will be written UNCOMPRESSED.",
                self.compression_type
            );
        }

        let grid = if use_grid {
            // `use_grid` implies `grid_leading.is_some()`.
            grid_leading.map(|leading| SwmrGridLayout {
                leading,
                frame_dims: usize_frame_dims.clone(),
                chunk: grid_chunk_usize.clone(),
                elem_size: element_size,
            })
        } else {
            None
        };
        self.handle = Some(Hdf5Handle::Swmr {
            writer: Box::new(swmr),
            ds_index,
            compression_dropped,
            grid,
        });
        self.open_data_type = Some(array.data.data_type());
        self.open_frame_dims = Some(array.dims.iter().rev().map(|d| d.size).collect::<Vec<_>>());
        Ok(())
    }

    /// Build every group node declared in the loaded layout XML against a
    /// `SwmrFileWriter`, the SWMR counterpart of `build_layout_groups`.
    ///
    /// Paths are created parent-first (shortest path-depth first) via the
    /// rust-hdf5 0.2.17 `SwmrFileWriter::create_group` API, which takes the
    /// parent group path and a leaf name. Called from `open_swmr` before
    /// `start_swmr()` so the groups are visible to SWMR readers for the whole
    /// streaming window. No-op when no layout is loaded.
    fn build_swmr_layout_groups(&self, swmr: &mut SwmrFileWriter) -> ADResult<()> {
        let layout = match self.layout.as_ref() {
            Some(l) => l,
            None => return Ok(()),
        };
        fn collect<'a>(
            g: &'a crate::hdf5_layout::LayoutGroup,
            prefix: &str,
            out: &mut Vec<(String, &'a crate::hdf5_layout::LayoutGroup)>,
        ) {
            let here = if prefix.is_empty() {
                g.name.clone()
            } else {
                format!("{}/{}", prefix, g.name)
            };
            out.push((here.clone(), g));
            for sub in &g.groups {
                collect(sub, &here, out);
            }
        }
        let mut nodes = Vec::new();
        for g in &layout.groups {
            collect(g, "", &mut nodes);
        }
        nodes.sort_by_key(|(p, _)| p.matches('/').count());
        let mut created: std::collections::HashSet<String> = std::collections::HashSet::new();
        for (path, node) in &nodes {
            if !created.insert(path.clone()) {
                continue;
            }
            let (parent, leaf) = match path.rsplit_once('/') {
                Some((p, l)) => (format!("/{}", p), l),
                None => ("/".to_string(), path.as_str()),
            };
            swmr.create_group(&parent, leaf).map_err(|e| {
                ADError::UnsupportedConversion(format!("SWMR layout group '{}': {}", path, e))
            })?;
            // C attaches the group's constant <attribute> nodes (NX_class) via
            // writeHdfAttributes (NDFileHDF5.cpp:693-695); the absolute group
            // path is `/<path>`.
            write_swmr_group_constant_attrs(swmr, &format!("/{}", path), &node.attributes)?;
        }
        Ok(())
    }

    /// Materialise every `<hardlink>` declared in the loaded layout XML against
    /// a `SwmrFileWriter`, the SWMR counterpart of `build_layout_hardlinks`.
    ///
    /// Uses the rust-hdf5 0.2.17 `SwmrFileWriter::create_hard_link` API. Called
    /// from `open_swmr` after the layout groups and image dataset exist and
    /// before `start_swmr()` — matching C `NDFileHDF5.cpp:319`-`320`
    /// `createHardLinks`, which runs before `startSWMR`. A link created before
    /// `start_swmr()` is visible to SWMR readers for the whole streaming
    /// window. No-op when no layout is loaded.
    fn build_swmr_layout_hardlinks(&self, swmr: &mut SwmrFileWriter) -> ADResult<()> {
        let layout = match self.layout.as_ref() {
            Some(l) => l,
            None => return Ok(()),
        };
        fn collect<'a>(
            g: &'a crate::hdf5_layout::LayoutGroup,
            prefix: &str,
            out: &mut Vec<(String, &'a crate::hdf5_layout::LayoutHardlink)>,
        ) {
            let here = if prefix.is_empty() {
                g.name.clone()
            } else {
                format!("{}/{}", prefix, g.name)
            };
            for hl in &g.hardlinks {
                out.push((here.clone(), hl));
            }
            for sub in &g.groups {
                collect(sub, &here, out);
            }
        }
        let mut links = Vec::new();
        for g in &layout.groups {
            collect(g, "", &mut links);
        }
        for (parent_path, hl) in &links {
            let parent = format!("/{}", parent_path);
            swmr.create_hard_link(&parent, &hl.name, &hl.target)
                .map_err(|e| {
                    ADError::UnsupportedConversion(format!(
                        "SWMR layout hardlink '{}/{}' -> '{}': {}",
                        parent_path, hl.name, hl.target, e
                    ))
                })?;
        }
        Ok(())
    }

    /// Materialise the loaded layout XML's `constant` HDF5 attributes attached
    /// to the primary image dataset against a `SwmrFileWriter`. This mirrors
    /// the standard close path's `layout_ds_attrs` block in
    /// `create_primary_dataset` (e.g. the NeXus `signal=1` marker). Only
    /// `constant`-sourced attributes are materialised; `ndattribute`-sourced
    /// nodes carry per-frame values and are out of scope here. No-op when no
    /// layout is loaded.
    fn write_swmr_layout_dataset_attrs(
        &self,
        swmr: &mut SwmrFileWriter,
        ds_index: usize,
    ) -> ADResult<()> {
        use crate::hdf5_layout::{LayoutDataType, LayoutSource};
        let layout = match self.layout.as_ref() {
            Some(l) => l,
            None => return Ok(()),
        };
        let resolved_ds = self.resolved_dataset_path.as_str();
        let mut attrs: Vec<(String, LayoutDataType, String)> = Vec::new();
        layout.for_each_dataset(|path, d| {
            let full = format!("{}/{}", path, d.name);
            if full.trim_start_matches('/') == resolved_ds {
                for a in &d.attributes {
                    if a.source == LayoutSource::Constant {
                        attrs.push((a.name.clone(), a.data_type, a.value.clone()));
                    }
                }
            }
        });
        for (name, dtype, value) in &attrs {
            match dtype {
                LayoutDataType::Int => {
                    let v: i64 = value.trim().parse().unwrap_or(0);
                    swmr.set_dataset_attr_numeric(ds_index, name, &v)
                }
                LayoutDataType::Float => {
                    let v: f64 = value.trim().parse().unwrap_or(0.0);
                    swmr.set_dataset_attr_numeric(ds_index, name, &v)
                }
                LayoutDataType::String => swmr.set_dataset_attr_string(ds_index, name, value),
            }
            .map_err(|e| {
                ADError::UnsupportedConversion(format!(
                    "SWMR layout dataset attribute '{}': {}",
                    name, e
                ))
            })?;
        }
        Ok(())
    }

    /// Resolve the on-disk dataset/group paths from the loaded layout XML.
    ///
    /// With a valid layout this places the image dataset at the layout's
    /// `det_default` dataset path, NDAttribute datasets under the
    /// `ndattr_default` group, and the performance dataset under the group
    /// holding the `timestamp` dataset — matching C `NDFileHDF5`'s
    /// `/entry/instrument/detector/data` tree. Without a layout the flat
    /// root defaults (`data`, `NDAttributes`, `performance`) are kept.
    ///
    /// All returned paths have the leading `/` stripped, since `rust-hdf5`
    /// keys datasets/groups without a leading slash.
    fn resolve_layout_paths(&mut self) {
        let strip = |s: String| s.trim_start_matches('/').to_string();
        match self.layout.as_ref() {
            Some(layout) => {
                self.resolved_dataset_path = layout
                    .detector_dataset_path()
                    .map(strip)
                    .unwrap_or_else(|| self.dataset_name.clone());
                self.resolved_ndattr_group =
                    layout.ndattr_default_group().map(strip).unwrap_or_default();
                self.resolved_perf_group = layout
                    .dataset_group_path("timestamp")
                    .map(strip)
                    .unwrap_or_default();
            }
            None => {
                self.resolved_dataset_path = self.dataset_name.clone();
                self.resolved_ndattr_group.clear();
                self.resolved_perf_group.clear();
            }
        }
    }

    /// Build every group node declared in the loaded layout XML so that empty
    /// NeXus-style groups (e.g. an `NXdata` placeholder) also exist on disk,
    /// not just the groups implied by the dataset placement. No-op when no
    /// layout is loaded.
    ///
    /// `rust-hdf5` 0.2.15's `create_group` errors on a duplicate path, so each
    /// distinct group path is created exactly once via a created-set; paths
    /// are processed shortest-first so a parent always exists before a child.
    fn build_layout_groups(&self, file: &H5File) -> ADResult<()> {
        let layout = match self.layout.as_ref() {
            Some(l) => l,
            None => return Ok(()),
        };
        fn collect<'a>(
            g: &'a crate::hdf5_layout::LayoutGroup,
            prefix: &str,
            out: &mut Vec<(String, &'a crate::hdf5_layout::LayoutGroup)>,
        ) {
            let here = if prefix.is_empty() {
                g.name.clone()
            } else {
                format!("{}/{}", prefix, g.name)
            };
            out.push((here.clone(), g));
            for sub in &g.groups {
                collect(sub, &here, out);
            }
        }
        let mut nodes = Vec::new();
        for g in &layout.groups {
            collect(g, "", &mut nodes);
        }
        nodes.sort_by_key(|(p, _)| p.matches('/').count());
        let mut created: std::collections::HashSet<String> = std::collections::HashSet::new();
        for (path, node) in &nodes {
            if !created.insert(path.clone()) {
                continue;
            }
            let (parent, leaf) = match path.rsplit_once('/') {
                Some((p, l)) => (p, l),
                None => ("", path.as_str()),
            };
            // The parent path was created earlier (shorter, sorted first).
            let parent_group = if parent.is_empty() {
                None
            } else {
                Some(Self::open_write_group(file, parent)?)
            };
            let group = match parent_group.as_ref() {
                Some(g) => g.create_group(leaf),
                None => file.create_group(leaf),
            }
            .map_err(|e| {
                ADError::UnsupportedConversion(format!("HDF5 layout group '{}': {}", path, e))
            })?;
            // C attaches the group's constant <attribute> nodes (e.g. the NeXus
            // NX_class markers) via writeHdfAttributes (NDFileHDF5.cpp:693-695).
            write_group_constant_attrs(&group, &node.attributes);
        }
        Ok(())
    }

    /// Materialise every `<hardlink>` declared in the loaded layout XML.
    ///
    /// A layout `<hardlink name="..." target="..."/>` inside a `<group>`
    /// declares an HDF5 hard link: an additional name (`name`, a leaf within
    /// the enclosing group) for the object already living at `target` (an
    /// absolute object path). C++ `NDFileHDF5::createHardLinks` walks the
    /// layout after the groups/datasets exist and calls `H5Lcreate_hard`.
    ///
    /// Called from `close_file` for the standard (non-SWMR) close path so that
    /// both the primary image dataset and the per-frame NDAttribute datasets —
    /// any of which a hardlink may target — already exist on disk. No-op when
    /// no layout is loaded.
    ///
    /// `file` is the live `Standard` write-mode HDF5 handle. The SWMR path has
    /// its own counterpart, `build_swmr_layout_hardlinks`, which runs before
    /// `start_swmr()` (C++ `NDFileHDF5.cpp:319`-`325`: `createHardLinks` then
    /// `startSWMR`) so SWMR readers see the links during streaming.
    fn build_layout_hardlinks(&self, file: &H5File) -> ADResult<()> {
        let layout = match self.layout.as_ref() {
            Some(l) => l,
            None => return Ok(()),
        };
        // Collect (parent_group_path, hardlink) for every group in the tree.
        fn collect<'a>(
            g: &'a crate::hdf5_layout::LayoutGroup,
            prefix: &str,
            out: &mut Vec<(String, &'a crate::hdf5_layout::LayoutHardlink)>,
        ) {
            let here = if prefix.is_empty() {
                g.name.clone()
            } else {
                format!("{}/{}", prefix, g.name)
            };
            for hl in &g.hardlinks {
                out.push((here.clone(), hl));
            }
            for sub in &g.groups {
                collect(sub, &here, out);
            }
        }
        let mut links = Vec::new();
        for g in &layout.groups {
            collect(g, "", &mut links);
        }
        for (parent_path, hl) in &links {
            // The enclosing group already exists (created by
            // `build_layout_groups`); re-open it and create the link inside it.
            let parent = Self::open_write_group(file, parent_path)?;
            parent.link(&hl.name, &hl.target).map_err(|e| {
                ADError::UnsupportedConversion(format!(
                    "HDF5 layout hardlink '{}/{}' -> '{}': {}",
                    parent_path, hl.name, hl.target, e
                ))
            })?;
        }
        Ok(())
    }

    /// Re-open an already-created group by full path in write mode. In write
    /// mode `H5Group::group` returns a handle without verification, so this is
    /// a pure handle constructor walking each path segment.
    fn open_write_group(file: &H5File, path: &str) -> ADResult<rust_hdf5::H5Group> {
        let mut current: Option<rust_hdf5::H5Group> = None;
        for seg in path.split('/').filter(|s| !s.is_empty()) {
            let next = match current.as_ref() {
                Some(g) => g.group(seg),
                None => file.root_group().group(seg),
            }
            .map_err(|e| {
                ADError::UnsupportedConversion(format!("HDF5 group reopen '{}': {}", seg, e))
            })?;
            current = Some(next);
        }
        current.ok_or_else(|| ADError::UnsupportedConversion("empty group path".into()))
    }

    /// Create every detector-source dataset on the first frame in standard
    /// mode. Each is an extensible `[nframes, .., Y, X]` array whose leading
    /// dimension later frames extend (C++ `NDFileHDF5Dataset`). C creates one
    /// `NDFileHDF5Dataset` per `<dataset source="detector">` node up front
    /// (`createDatasetDetector`, NDFileHDF5.cpp:1324-1357) and routes frames
    /// among them; the common single-dataset layout produces exactly one. All
    /// datasets share the datatype/dataspace/frame shape taken from this frame.
    fn create_detector_datasets(&mut self, array: &NDArray) -> ADResult<()> {
        let frame_dims: Vec<usize> = array.dims.iter().rev().map(|d| d.size).collect();
        // A pre-compressed input array (codec set) is written verbatim through a
        // matching HDF5 filter via direct chunk write (C
        // `NDFileHDF5Dataset::writeFile`, the `compressionAware` path). Its
        // dataset carries the ORIGINAL element type (`codec.original_data_type`),
        // a whole-frame chunk (`compressed_layout`), and a filter pipeline
        // mirroring the codec rather than the writer's configured compression.
        // An uncompressed array takes the standard tiled-write path unchanged.
        let (ds_data_type, shape, chunk, leading, pipeline) = match array.codec.as_ref() {
            Some(codec) => {
                let pl = self.codec_filter_pipeline(codec).ok_or_else(|| {
                    ADError::UnsupportedConversion(format!(
                        "HDF5 cannot direct-chunk-write codec '{}' \
                         (only jpeg/blosc/lz4/bslz4 are supported)",
                        codec.name.as_str()
                    ))
                })?;
                let (shape, chunk, leading) = self.compressed_layout(&frame_dims);
                (codec.original_data_type, shape, chunk, leading, Some(pl))
            }
            None => {
                let dt = array.data.data_type();
                let (shape, chunk, leading) = self.standard_layout(&frame_dims);
                (
                    dt,
                    shape,
                    chunk,
                    leading,
                    self.build_pipeline(dt.element_size()),
                )
            }
        };
        // N-bit packing couples a reduced-precision datatype override with the
        // nbit filter (C narrows `this->datatype` then `H5Pset_nbit`); only this
        // standard, non-codec path can override the dataset datatype. For a
        // pre-compressed array the codec pipeline already governs the dataset,
        // so N-bit does not apply. `d_nelmts` is the per-chunk element count.
        let (nbit_dt, pipeline): (Option<DatatypeMessage>, Option<FilterPipeline>) =
            if array.codec.is_none() {
                let chunk_nelmts: usize = chunk.iter().product();
                match self.nbit_packing(ds_data_type, chunk_nelmts) {
                    Some((dt, pl)) => (Some(dt), Some(pl)),
                    None => (None, pipeline),
                }
            } else {
                (None, pipeline)
            };
        // Max shape: with extra dims the dataset is created at its full fixed
        // multi-dimensional size (every leading axis chunked at 1, so its
        // ceiling equals its size). Without extra dims the single leading frame
        // axis is extensible (`None`). Every other axis gets headroom to the
        // chunk-aligned ceiling so a `write_chunk_at` edge tile of a
        // non-dividing chunk can extend into it — `close_file`'s `set_extent`
        // trims back to the exact frame shape.
        let max_shape: Vec<Option<usize>> = shape
            .iter()
            .zip(chunk.iter())
            .enumerate()
            .map(|(i, (&s, &c))| {
                if leading.is_none() && i == 0 {
                    None
                } else {
                    Some(s.div_ceil(c) * c)
                }
            })
            .collect();

        // Build the layout group hierarchy (if a layout XML is loaded) before
        // placing the dataset. With no layout this is a no-op and the dataset
        // lands flat at the file root.
        match self.handle {
            Some(Hdf5Handle::Standard { ref file, .. }) => self.build_layout_groups(file)?,
            _ => return Err(ADError::UnsupportedConversion("no HDF5 file open".into())),
        }

        // Enumerate every detector-source dataset (C `detDataMap`), each with
        // the `constant` HDF5 attributes the layout XML attaches to it (e.g. the
        // NeXus `signal=1` marker). Keys are leading-slash full names so the
        // `detector_data_destination` routing value matches C `detDataMap`.
        // Only constant attributes are materialised here; `ndattribute`-sourced
        // nodes carry per-frame values, out of scope for dataset creation.
        let detector_keys: Vec<String> = match self.layout.as_ref() {
            Some(l) => {
                let mut keys = l.detector_dataset_paths();
                if keys.is_empty() {
                    keys.push(format!("/{}", self.resolved_dataset_path));
                }
                keys
            }
            None => vec![format!("/{}", self.resolved_dataset_path)],
        };
        let detector_specs: Vec<(
            String,
            Vec<(String, crate::hdf5_layout::LayoutDataType, String)>,
        )> = detector_keys
            .iter()
            .map(|key| {
                let stripped = key.trim_start_matches('/').to_string();
                let attrs = self
                    .layout
                    .as_ref()
                    .map(|l| {
                        use crate::hdf5_layout::LayoutSource;
                        let mut out = Vec::new();
                        l.for_each_dataset(|path, d| {
                            let full = format!("{}/{}", path, d.name);
                            if full.trim_start_matches('/') == stripped {
                                for a in &d.attributes {
                                    if a.source == LayoutSource::Constant {
                                        out.push((a.name.clone(), a.data_type, a.value.clone()));
                                    }
                                }
                            }
                        });
                        out
                    })
                    .unwrap_or_default();
                (key.clone(), attrs)
            })
            .collect();

        let dtype_ordinal = ds_data_type as i32;
        let fill = self.fill_value;
        let row_chunks = self.chunk.n_row_chunks as i32;
        let col_chunks = self.chunk.n_col_chunks as i32;
        let frame_chunks = self.chunk.n_frames_chunks as i32;
        let n_extra = self.n_extra_dims as i32;
        let extra_meta: Vec<(usize, i32, String)> = (0..self.n_extra_dims)
            .map(|i| {
                (
                    i,
                    self.extra_dims[i].size.max(1) as i32,
                    self.extra_dims[i].name.clone(),
                )
            })
            .collect();

        // C writeDefaultDatasetAttributes (NDFileHDF5.cpp:3695-3719) attaches
        // NDArrayNumDims (scalar int32) and one int32 value per dimension for
        // NDArrayDimOffset/Binning/Reverse to every detector dataset. Dim order
        // is native NDArray order (dims[0]…), not the reversed HDF5 axis order.
        let nd_num_dims = array.dims.len() as i32;
        // writeH5attrInt32 (NDFileHDF5.cpp:1142-1191) writes a single value as a
        // scalar and multiple values as a 1-D int32 array of length ndims. The
        // per-dimension values, native NDArray order.
        let dim_offsets: Vec<i32> = array.dims.iter().map(|d| d.offset as i32).collect();
        let dim_binnings: Vec<i32> = array.dims.iter().map(|d| d.binning as i32).collect();
        let dim_reverses: Vec<i32> = array.dims.iter().map(|d| d.reverse as i32).collect();

        macro_rules! create_ds {
            ($t:ty, $h5file:expr, $ds_group:expr, $ds_name:expr, $constant_attrs:expr) => {{
                let mut builder = match $ds_group.as_ref() {
                    Some(g) => g.new_dataset::<$t>(),
                    None => $h5file.new_dataset::<$t>(),
                }
                .shape(&shape[..])
                .chunk(&chunk[..])
                .max_shape(&max_shape[..])
                // C parity: NDFileHDF5 sets HDF5_fillValue on the dataset
                // creation property list (H5Pset_fill_value). rust-hdf5 0.2.15
                // exposes `DatasetBuilder::fill_value`, which writes it into the
                // DCPL fill-value message so unwritten chunks read back as
                // `fill` rather than zero.
                .fill_value(fill as $t);
                // N-bit: store the reduced-precision datatype (C narrows
                // `this->datatype`). The byte footprint stays `size_of::<$t>()`;
                // the nbit filter packs the significant bits within it.
                if let Some(ref dt) = nbit_dt {
                    builder = builder.datatype(dt.clone());
                }
                if let Some(ref pl) = pipeline {
                    builder = builder.filter_pipeline(pl.clone());
                }
                let ds = builder.create($ds_name.as_str()).map_err(|e| {
                    ADError::UnsupportedConversion(format!("HDF5 dataset error: {}", e))
                })?;
                // Record the exact NDArray data type for lossless read-back.
                let _ = ds
                    .new_attr::<i32>()
                    .shape(())
                    .create(DTYPE_ATTR)
                    .and_then(|a| a.write_numeric(&dtype_ordinal));
                // Also expose the fill value as an attribute for tooling that
                // inspects HDF5_fillValue directly (the DCPL above is the
                // authoritative copy).
                let _ = ds
                    .new_attr::<f64>()
                    .shape(())
                    .create("HDF5_fillValue")
                    .and_then(|a| a.write_numeric(&fill));
                // Record the requested chunk geometry. The on-disk chunk is
                // one frame per chunk (crate limitation); these attributes
                // preserve the user's intent for downstream tooling.
                for (name, val) in [
                    ("HDF5_nRowChunks", row_chunks),
                    ("HDF5_nColChunks", col_chunks),
                    ("HDF5_nFramesChunks", frame_chunks),
                    ("HDF5_nExtraDims", n_extra),
                ] {
                    let _ = ds
                        .new_attr::<i32>()
                        .shape(())
                        .create(name)
                        .and_then(|a| a.write_numeric(&val));
                }
                // Record extra-dimension sizes and names so the flat leading
                // axis can be reshaped into the intended N-D layout.
                for (i, size, name) in &extra_meta {
                    let _ = ds
                        .new_attr::<i32>()
                        .shape(())
                        .create(&format!("HDF5_extraDimSize{}", i))
                        .and_then(|a| a.write_numeric(size));
                    if !name.is_empty() {
                        let s = rust_hdf5::types::VarLenUnicode(name.clone());
                        let _ = ds
                            .new_attr::<rust_hdf5::types::VarLenUnicode>()
                            .shape(())
                            .create(&format!("HDF5_extraDimName{}", i))
                            .and_then(|a| a.write_scalar(&s));
                    }
                }
                // Materialise the layout XML's constant dataset attributes
                // (e.g. NeXus `signal=1`), typed per the XML `type` attribute.
                for (aname, atype, avalue) in $constant_attrs {
                    use crate::hdf5_layout::LayoutDataType;
                    match atype {
                        LayoutDataType::Int => {
                            let v: i64 = avalue.trim().parse().unwrap_or(0);
                            let _ = ds
                                .new_attr::<i64>()
                                .shape(())
                                .create(aname)
                                .and_then(|a| a.write_numeric(&v));
                        }
                        LayoutDataType::Float => {
                            let v: f64 = avalue.trim().parse().unwrap_or(0.0);
                            let _ = ds
                                .new_attr::<f64>()
                                .shape(())
                                .create(aname)
                                .and_then(|a| a.write_numeric(&v));
                        }
                        LayoutDataType::String => {
                            let s = rust_hdf5::types::VarLenUnicode(avalue.clone());
                            let _ = ds
                                .new_attr::<rust_hdf5::types::VarLenUnicode>()
                                .shape(())
                                .create(aname)
                                .and_then(|a| a.write_scalar(&s));
                        }
                    }
                }
                // C parity (writeDefaultDatasetAttributes): NDArrayNumDims plus
                // the per-dimension offset/binning/reverse. writeH5attrInt32
                // (NDFileHDF5.cpp:1142-1191) emits a single dimension as a scalar
                // int32 and multiple dimensions as a 1-D int32 array of length
                // ndims.
                let _ = ds
                    .new_attr::<i32>()
                    .shape(())
                    .create("NDArrayNumDims")
                    .and_then(|a| a.write_numeric(&nd_num_dims));
                for (name, vals) in [
                    ("NDArrayDimOffset", &dim_offsets),
                    ("NDArrayDimBinning", &dim_binnings),
                    ("NDArrayDimReverse", &dim_reverses),
                ] {
                    let _ = if vals.len() == 1 {
                        ds.new_attr::<i32>()
                            .shape(())
                            .create(name)
                            .and_then(|a| a.write_numeric(&vals[0]))
                    } else {
                        ds.new_attr::<i32>()
                            .shape([vals.len()])
                            .create(name)
                            .and_then(|a| a.write_array(vals))
                    };
                }
                ds
            }};
        }

        let h5file = match self.handle {
            Some(Hdf5Handle::Standard { ref file, .. }) => file,
            _ => return Err(ADError::UnsupportedConversion("no HDF5 file open".into())),
        };

        let mut detectors: std::collections::HashMap<String, DetectorDataset> =
            std::collections::HashMap::with_capacity(detector_specs.len());
        for (key, constant_attrs) in &detector_specs {
            // Resolve each dataset's parent group and leaf name. A key is e.g.
            // `/entry/instrument/detector/data` with a layout, or `/data` flat.
            let stripped = key.trim_start_matches('/');
            let (ds_group, ds_name): (Option<rust_hdf5::H5Group>, String) =
                match stripped.rsplit_once('/') {
                    Some((group_path, leaf)) => (
                        Some(Self::open_write_group(h5file, group_path)?),
                        leaf.to_string(),
                    ),
                    None => (None, stripped.to_string()),
                };
            // Dispatch the dataset element type on `ds_data_type`: the original
            // (uncompressed) type for a pre-compressed array — whose `array.data`
            // is the collapsed `U8` byte buffer — and the buffer's own type
            // otherwise.
            let ds = match ds_data_type {
                NDDataType::Int8 => create_ds!(i8, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::UInt8 => create_ds!(u8, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::Int16 => create_ds!(i16, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::UInt16 => create_ds!(u16, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::Int32 => create_ds!(i32, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::UInt32 => create_ds!(u32, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::Int64 => create_ds!(i64, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::UInt64 => create_ds!(u64, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::Float32 => create_ds!(f32, h5file, ds_group, ds_name, constant_attrs),
                NDDataType::Float64 => create_ds!(f64, h5file, ds_group, ds_name, constant_attrs),
            };
            detectors.insert(
                key.clone(),
                DetectorDataset {
                    ds,
                    frame_count: 0,
                    frame_band: Vec::new(),
                },
            );
        }

        if let Some(Hdf5Handle::Standard { detectors: d, .. }) = self.handle.as_mut() {
            *d = detectors;
        }
        self.open_data_type = Some(ds_data_type);
        self.open_frame_dims = Some(frame_dims);
        self.open_codec = array.codec.clone();
        Ok(())
    }

    /// Resolve which detector dataset a frame is written to (C
    /// `NDFileHDF5::writeFile`, NDFileHDF5.cpp:1449-1474). The default is
    /// `defDsetName` — the leading-slash full name of the `det_default`/first
    /// detector dataset (= `resolved_dataset_path`). When the layout's
    /// `<global name="detector_data_destination" ndattribute="X"/>` names an
    /// NDAttribute that the frame carries as a *string* whose value is an
    /// existing detector-dataset key, that key is the destination; an unknown
    /// value falls back to the default. A present-but-non-string attribute is
    /// an error, matching C (`getValue(NDAttrString,…)` → `ND_ERROR`, which
    /// aborts the write).
    fn resolve_destination_key(&self, array: &NDArray) -> ADResult<String> {
        let default = format!("/{}", self.resolved_dataset_path);
        let attr_name = self
            .layout
            .as_ref()
            .and_then(|l| l.detector_data_destination.as_deref())
            .filter(|s| !s.is_empty());
        let Some(attr_name) = attr_name else {
            return Ok(default);
        };
        // C 1451-1452: a missing destination attribute keeps the default with
        // no error (only a present-but-unreadable one aborts).
        let Some(attr) = array.attributes.get(attr_name) else {
            return Ok(default);
        };
        match attr.value.as_string_typed() {
            Some(value) => {
                let known = matches!(
                    &self.handle,
                    Some(Hdf5Handle::Standard { detectors, .. })
                        if detectors.contains_key(value)
                );
                Ok(if known { value.to_string() } else { default })
            }
            None => Err(ADError::UnsupportedConversion(format!(
                "HDF5 detector_data_destination attribute '{}' is not a string",
                attr_name
            ))),
        }
    }

    /// Write a frame in standard (non-SWMR) mode, routing it to its detector
    /// dataset (C `detector_data_destination`) and extending that dataset's
    /// leading dimension. Each detector dataset keeps its own frame counter and
    /// band; the file-wide `frame_count` is the total across all of them.
    fn write_standard(&mut self, array: &NDArray) -> ADResult<()> {
        if self.frame_count == 0 {
            self.create_detector_datasets(array)?;
            self.create_attribute_datasets();
        }

        let frame_dims = self
            .open_frame_dims
            .clone()
            .ok_or_else(|| ADError::UnsupportedConversion("dataset not initialised".into()))?;
        let cur_dims: Vec<usize> = array.dims.iter().rev().map(|d| d.size).collect();
        if cur_dims != frame_dims {
            return Err(ADError::UnsupportedConversion(format!(
                "HDF5 frame shape changed mid-stream: {:?} != {:?}",
                cur_dims, frame_dims
            )));
        }

        // verifyChunking codec-match (C `NDFileHDF5Dataset.cpp:194-200`): the
        // file's codec is fixed when the dataset is created from the first frame.
        // A later frame whose codec differs (including compressed↔uncompressed)
        // cannot be written into it.
        if !codecs_match(self.open_codec.as_ref(), array.codec.as_ref()) {
            return Err(ADError::UnsupportedConversion(format!(
                "HDF5 codec changed mid-stream: dataset codec {:?} != frame codec {:?}",
                self.open_codec.as_ref().map(|c| c.name),
                array.codec.as_ref().map(|c| c.name),
            )));
        }

        let (_shape, chunk, leading) = self.standard_layout(&frame_dims);
        let fc = chunk[0];

        // Resolve the destination dataset (C `detDataMap` routing) and serialize
        // the per-frame payload before borrowing the detector map mutably. A
        // pre-compressed frame becomes one direct-chunk-write byte stream (C
        // `NDFileHDF5Dataset::writeFile`, compressionAware path); an uncompressed
        // frame becomes little-endian pixel bytes for band tiling.
        let dest_key = self.resolve_destination_key(array)?;
        let chunk_bytes = array.codec.as_ref().map(|codec| {
            let total_bytes =
                frame_dims.iter().product::<usize>() * codec.original_data_type.element_size();
            codec_chunk_bytes(codec, total_bytes, array.data.as_u8_slice())
        });
        let frame_le = if chunk_bytes.is_none() {
            Some(nd_buffer_to_le_bytes(&array.data))
        } else {
            None
        };
        let elem_size = array.data.data_type().element_size();

        {
            let Some(Hdf5Handle::Standard { detectors, .. }) = self.handle.as_mut() else {
                return Err(ADError::UnsupportedConversion(
                    "HDF5 detector datasets not initialised".into(),
                ));
            };
            let det = detectors.get_mut(&dest_key).ok_or_else(|| {
                ADError::UnsupportedConversion(format!(
                    "HDF5 destination dataset '{}' not found",
                    dest_key
                ))
            })?;

            // Per-dataset leading-axis index. With a fixed extra-dim layout the
            // counter must not exceed the product of the extra-dim sizes.
            let frame_idx = det.frame_count;
            if let Some(ref lead) = leading {
                let total: usize = lead.iter().product();
                if frame_idx >= total {
                    return Err(ADError::UnsupportedConversion(format!(
                        "HDF5 extra-dimension capacity exceeded: frame {} >= {}",
                        frame_idx, total
                    )));
                }
            }

            if let Some(cb) = &chunk_bytes {
                // Direct chunk write. The dataset was created with a whole-frame
                // chunk (`compressed_layout`), so each frame is exactly one chunk
                // and its linear chunk index equals the per-dataset frame index
                // (every leading chunk is 1, whether single-axis or extra-dim).
                // filter_mask = 0: the codec already applied the full pipeline.
                det.ds.write_chunk_raw(frame_idx, cb, 0).map_err(|e| {
                    ADError::UnsupportedConversion(format!("HDF5 direct chunk write error: {}", e))
                })?;
            } else {
                // Uncompressed: frames accumulate in this dataset's band and a
                // full `fc`-deep band is flushed as a grid of write_chunk_at
                // tiles; close_file flushes the partial final band. With a fixed
                // multi-extra-dim layout every leading chunk is 1, so `fc == 1`
                // and each frame is placed at its odometer position via `unravel`.
                let fle = frame_le.expect("frame_le is set whenever chunk_bytes is None");
                det.frame_band.push(fle);
                if det.frame_band.len() >= fc {
                    let leading_coords = match &leading {
                        Some(lead) => Self::unravel(frame_idx, lead),
                        None => vec![frame_idx / fc],
                    };
                    Self::flush_band(
                        &det.ds,
                        &leading_coords,
                        &det.frame_band,
                        &frame_dims,
                        &chunk,
                        elem_size,
                    )?;
                    det.frame_band.clear();
                }
            }
            det.frame_count += 1;
        }

        // Append NDAttribute values for this frame, read out of the sticky
        // file list exactly as C does (`pFileAttributes->find`,
        // NDFileHDF5.cpp:2957). The datasets were built from that same list at
        // open and the list only grows, so the lookup cannot miss; C treats a
        // miss as `continue` (:2959-2963) rather than as a value.
        if self.store_attributes {
            let Self {
                attr_datasets,
                file_attributes,
                ..
            } = self;
            for ad in attr_datasets.iter_mut() {
                if let Some(attr) = file_attributes.get(&ad.name) {
                    ad.push(&attr.value);
                }
            }
        }
        Ok(())
    }

    /// Create one attribute time-series dataset per NDAttribute, preserving
    /// the NDAttrValue numeric type. Mirrors C++ `createAttributeDataset`.
    fn create_attribute_datasets(&mut self) {
        self.attr_datasets.clear();
        if !self.store_attributes {
            return;
        }
        self.attr_datasets = self
            .file_attributes
            .iter()
            .map(AttributeDataset::new)
            .collect();
    }

    /// Effective chunk depth for the NDAttribute datasets and the performance
    /// dataset's leading dimension, mirroring C `calculateAttributeChunking`
    /// (NDFileHDF5.cpp:2869-2920). The `HDF5_NDAttributeChunk` param (0 = auto):
    /// when auto, Single mode chunks at 1; otherwise at the capture target
    /// (`NDFileNumCapture`), or `16*1024` when that is unlimited (≤ 0).
    fn attribute_chunking(&self) -> usize {
        if self.chunk.ndattr_chunk != 0 {
            return self.chunk.ndattr_chunk;
        }
        if self.open_mode == NDFileMode::Single {
            return 1;
        }
        if self.num_capture > 0 {
            self.num_capture
        } else {
            16 * 1024
        }
    }

    /// Reset and seed the NDAttribute element-attr value caches (ADP-79) from
    /// the open-time frame. The distinct referenced names are cached so the
    /// per-frame update (`update_ndattr_element_values`) need not re-walk the
    /// layout. No-op unless the layout declares `<attribute source="ndattribute">`
    /// on a group or dataset (the default NeXus layout declares none).
    fn seed_ndattr_element_values(&mut self) {
        self.ndattr_first_values.clear();
        self.ndattr_last_values.clear();
        self.ndattr_element_names.clear();
        let Some(layout) = self.layout.as_ref() else {
            return;
        };
        let mut names: Vec<String> = Vec::new();
        for e in layout.ndattribute_element_attrs() {
            if !names.contains(&e.ndattribute) {
                names.push(e.ndattribute);
            }
        }
        for name in &names {
            if let Some(attr) = self.file_attributes.get(name) {
                self.ndattr_first_values
                    .insert(name.clone(), attr.value.clone());
                self.ndattr_last_values
                    .insert(name.clone(), attr.value.clone());
            }
        }
        self.ndattr_element_names = names;
    }

    /// Update the last-frame value cache for every referenced element-attr
    /// NDAttribute name (C `pFileAttributes` tracks the most recent frame, so a
    /// `when="OnFileClose"` attribute records the final value). First-frame
    /// values are left untouched — an attribute absent at open stays unwritten
    /// for `OnFileOpen`/`OnFrame`, matching C's open-time `find` miss.
    fn update_ndattr_element_values(&mut self) {
        if self.ndattr_element_names.is_empty() {
            return;
        }
        let Self {
            ndattr_element_names,
            ndattr_last_values,
            file_attributes,
            ..
        } = self;
        for name in ndattr_element_names.iter() {
            if let Some(attr) = file_attributes.get(name) {
                ndattr_last_values.insert(name.clone(), attr.value.clone());
            }
        }
    }

    /// Materialise every layout `<attribute source="ndattribute">` as an HDF5
    /// attribute on its group/dataset (C `storeOnOpenCloseAttribute`). Called at
    /// close on the standard path, when every group and dataset exists: an
    /// `OnFileOpen`/`OnFrame` attribute takes the first-frame value (C writes it
    /// at open, where a later close re-create would fail as a duplicate, so the
    /// open value wins), an `OnFileClose` attribute the last-frame value.
    /// `OnFileWrite` is not materialised by this path, matching C. The HDF5
    /// attribute datatype follows the live NDAttribute value (C `typeNd2Hdf` on
    /// the runtime type); an undefined/absent value is skipped.
    ///
    /// Group attributes are written through the file handle. Dataset attributes
    /// need a live dataset handle: a detector dataset retains its create-time
    /// handle in `detectors` at close, and the lazily-created attribute/
    /// performance datasets are reopened by full path through
    /// `H5File::dataset_writer` (rust-hdf5 0.2.22), so dataset element-attrs are
    /// honoured for every dataset.
    fn flush_ndattr_element_attrs(
        &self,
        file: &H5File,
        detectors: &std::collections::HashMap<String, DetectorDataset>,
    ) -> ADResult<()> {
        use crate::hdf5_layout::LayoutWhen;
        let Some(layout) = self.layout.as_ref() else {
            return Ok(());
        };
        for e in layout.ndattribute_element_attrs() {
            let value = match e.when {
                LayoutWhen::OnFileOpen | LayoutWhen::OnFrame => {
                    self.ndattr_first_values.get(&e.ndattribute)
                }
                LayoutWhen::OnFileClose => self.ndattr_last_values.get(&e.ndattribute),
                LayoutWhen::OnFileWrite => None,
            };
            let Some(value) = value else {
                continue;
            };
            let path = e.element_path.trim_start_matches('/');
            if e.is_dataset {
                // A detector dataset retains its create-time write handle in
                // `detectors` (keyed by leading-slash full name); the lazily-
                // created attribute/performance datasets do not. Both exist on
                // disk by now (close-path ordering), so reopen the latter by
                // full path in write mode (rust-hdf5 0.2.22 `dataset_writer`,
                // which registers names as slash-trimmed full paths).
                if let Some(det) = detectors.get(&e.element_path) {
                    write_ndattr_dataset_attr(&det.ds, &e.attr_name, value);
                } else if let Ok(ds) = file.dataset_writer(path) {
                    write_ndattr_dataset_attr(&ds, &e.attr_name, value);
                }
            } else {
                let group = Self::open_write_group(file, path)?;
                write_ndattr_group_attr(&group, &e.attr_name, value);
            }
        }
        Ok(())
    }

    /// SWMR counterpart of C `writeDefaultDatasetAttributes`
    /// (NDFileHDF5.cpp:3695-3719): attach `NDArrayNumDims` (scalar int32) and the
    /// per-dimension `NDArrayDimOffset`/`NDArrayDimBinning`/`NDArrayDimReverse` to
    /// the single streaming dataset (addressed by `ds_index`). A 1-D array writes
    /// each as a scalar int32, a multi-dim array as a 1-D int32 array of length
    /// ndims (C `writeH5attrInt32`, NDFileHDF5.cpp:1142-1191), native NDArray dim
    /// order. Must run before `start_swmr()` — HDF5 forbids adding attributes
    /// after the SWMR lock.
    fn write_swmr_ndarray_default_attrs(
        &self,
        swmr: &mut SwmrFileWriter,
        ds_index: usize,
        array: &NDArray,
    ) -> ADResult<()> {
        let nd_num_dims = array.dims.len() as i32;
        swmr.set_dataset_attr_numeric(ds_index, "NDArrayNumDims", &nd_num_dims)
            .map_err(|e| {
                ADError::UnsupportedConversion(format!("SWMR NDArrayNumDims attr: {}", e))
            })?;
        let dim_offsets: Vec<i32> = array.dims.iter().map(|d| d.offset as i32).collect();
        let dim_binnings: Vec<i32> = array.dims.iter().map(|d| d.binning as i32).collect();
        let dim_reverses: Vec<i32> = array.dims.iter().map(|d| d.reverse as i32).collect();
        for (name, vals) in [
            ("NDArrayDimOffset", &dim_offsets),
            ("NDArrayDimBinning", &dim_binnings),
            ("NDArrayDimReverse", &dim_reverses),
        ] {
            let res = if vals.len() == 1 {
                swmr.set_dataset_attr_numeric(ds_index, name, &vals[0])
            } else {
                swmr.set_dataset_attr_array(ds_index, name, &[vals.len() as u64], vals)
            };
            res.map_err(|e| ADError::UnsupportedConversion(format!("SWMR {} attr: {}", name, e)))?;
        }
        Ok(())
    }

    /// SWMR counterpart of [`Hdf5Writer::flush_ndattr_element_attrs`], limited
    /// to the open-time set that can be written before `start_swmr()` locks the
    /// file: `OnFileOpen`/`OnFrame` `<attribute source="ndattribute">` nodes,
    /// using the first-frame value. Group attrs address the group by path;
    /// dataset attrs address the single streaming dataset by its index
    /// (`ds_index`) when their element path matches it — rust-hdf5 0.2.22's
    /// `set_dataset_attr_*` makes dataset element-attrs addressable in SWMR.
    /// `OnFileClose` cannot be honoured in SWMR (HDF5 forbids creating
    /// attributes after the SWMR lock — C's close-time `H5Acreate2` fails
    /// identically); that one remains a recorded residual.
    fn write_swmr_ndattr_element_attrs(
        &self,
        swmr: &mut SwmrFileWriter,
        ds_index: usize,
    ) -> ADResult<()> {
        use crate::hdf5_layout::LayoutWhen;
        let Some(layout) = self.layout.as_ref() else {
            return Ok(());
        };
        for e in layout.ndattribute_element_attrs() {
            if !matches!(e.when, LayoutWhen::OnFileOpen | LayoutWhen::OnFrame) {
                continue;
            }
            let Some(value) = self.ndattr_first_values.get(&e.ndattribute) else {
                continue;
            };
            if e.is_dataset {
                // Only the single streaming image dataset exists in SWMR; attach
                // by its known index when the element path resolves to it.
                // `resolved_dataset_path` is stored slash-trimmed.
                if e.element_path.trim_start_matches('/') == self.resolved_dataset_path {
                    write_swmr_ndattr_dataset_attr(swmr, ds_index, &e.attr_name, value)?;
                }
            } else {
                write_swmr_ndattr_group_attr(swmr, &e.element_path, &e.attr_name, value)?;
            }
        }
        Ok(())
    }

    /// Flush accumulated NDAttribute datasets into the open standard file.
    /// Each becomes a chunked, extensible 1-D dataset under `NDAttributes/`.
    fn flush_attribute_datasets(&mut self) -> ADResult<()> {
        if self.attr_datasets.is_empty() {
            return Ok(());
        }
        let chunk_depth = self.attribute_chunking();
        let ndattr_group = self.resolved_ndattr_group.clone();
        // Route each attribute to its XML-declared `<dataset source="ndattribute">`
        // parent group + dataset name (C `find_dset_ndattr`, NDFileHDF5.cpp:2792),
        // falling back to the default ndattr group keyed by the raw attribute name.
        let targets: Vec<(String, String)> = self
            .attr_datasets
            .iter()
            .map(|ad| {
                match self
                    .layout
                    .as_ref()
                    .and_then(|l| l.ndattribute_dataset(&ad.name))
                {
                    Some((g, name)) => (g.trim_start_matches('/').to_string(), name),
                    None => (ndattr_group.clone(), ad.name.clone()),
                }
            })
            .collect();
        let h5file = match self.handle {
            Some(Hdf5Handle::Standard { ref file, .. }) => file,
            _ => return Ok(()),
        };
        // Group handles, cached by path. A non-empty path is a layout group
        // already created by `build_layout_groups` (re-open it); an empty path
        // is the flat `NDAttributes` fallback (no layout) created at the root.
        let mut group_cache: std::collections::HashMap<String, rust_hdf5::H5Group> =
            std::collections::HashMap::new();

        for (ad, (group_path, ds_name)) in self.attr_datasets.iter().zip(targets.iter()) {
            if ad.frames == 0 {
                continue;
            }
            let n = ad.frames;
            // C does not clamp the chunk to the frame count: the attribute
            // dataset is extensible (unlimited dim 0), so the chunk may exceed
            // the current extent (`calculateAttributeChunking` → numCapture).
            let chunk = chunk_depth;

            if !group_cache.contains_key(group_path) {
                let g = if group_path.is_empty() {
                    h5file.create_group("NDAttributes").map_err(|e| {
                        ADError::UnsupportedConversion(format!("HDF5 group error: {}", e))
                    })?
                } else {
                    Self::open_write_group(h5file, group_path)?
                };
                group_cache.insert(group_path.clone(), g);
            }
            let group = &group_cache[group_path];

            macro_rules! create_attr_ds {
                ($t:ty) => {{
                    let es = std::mem::size_of::<$t>();
                    let ds = group
                        .new_dataset::<$t>()
                        .shape(&[n])
                        .chunk(&[chunk])
                        .max_shape(&[None])
                        .create(ds_name)
                        .map_err(|e| {
                            ADError::UnsupportedConversion(format!(
                                "HDF5 attribute dataset error: {}",
                                e
                            ))
                        })?;
                    // One chunk holds `chunk` consecutive frames; write each
                    // chunk's whole byte span (zero-padded for the trailing
                    // partial chunk, as rust-hdf5 requires full-chunk writes).
                    write_chunked_buffer(&ds, &ad.buffer, chunk * es)?;
                    ds
                }};
            }

            let ds = match ad.data_type {
                NDAttrDataType::Int8 => create_attr_ds!(i8),
                NDAttrDataType::UInt8 => create_attr_ds!(u8),
                NDAttrDataType::Int16 => create_attr_ds!(i16),
                NDAttrDataType::UInt16 => create_attr_ds!(u16),
                NDAttrDataType::Int32 => create_attr_ds!(i32),
                NDAttrDataType::UInt32 => create_attr_ds!(u32),
                NDAttrDataType::Int64 => create_attr_ds!(i64),
                NDAttrDataType::UInt64 => create_attr_ds!(u64),
                NDAttrDataType::Float32 => create_attr_ds!(f32),
                NDAttrDataType::Float64 => create_attr_ds!(f64),
                NDAttrDataType::String => {
                    // C stores a string attribute as a rank-1 `[n]` dataset of a
                    // fixed 256-byte `H5T_C_S1` string, not a 2-D byte array
                    // (NDFileHDF5AttributeDataset.cpp:321-323, rank_=1). The
                    // accumulated buffer already holds one 256-byte field per
                    // frame, so the element datatype carries the width and the
                    // dataset is 1-D.
                    let es = MAX_ATTRIBUTE_STRING_SIZE;
                    let ds = group
                        .new_dataset::<FixedStr256>()
                        .shape([n])
                        .chunk(&[chunk])
                        .max_shape(&[None])
                        .create(&ad.name)
                        .map_err(|e| {
                            ADError::UnsupportedConversion(format!(
                                "HDF5 attribute dataset error: {}",
                                e
                            ))
                        })?;
                    write_chunked_buffer(&ds, &ad.buffer, chunk * es)?;
                    ds
                }
            };

            // C attaches up to four self-describing string attributes to every
            // NDAttribute dataset (NDFileHDF5.cpp:2817-2822), each written only
            // when non-empty.
            write_ndattr_descriptors(&ds, ad);
        }
        Ok(())
    }

    /// Write the `timestamp` performance dataset (`[nframes, 5]` doubles)
    /// into the open standard file. Mirrors C++ `writePerformanceDataset`.
    fn flush_performance_dataset(&mut self) -> ADResult<()> {
        if !self.store_performance || self.perf_rows.is_empty() {
            return Ok(());
        }
        let n = self.perf_rows.len();
        let mut flat: Vec<f64> = Vec::with_capacity(n * 5);
        for row in &self.perf_rows {
            flat.extend_from_slice(row);
        }
        // f64 doubles serialized explicitly little-endian to match the LE
        // datatype `rust-hdf5` records (write_chunk copies bytes verbatim).
        let raw: Vec<u8> = flat.iter().flat_map(|v| v.to_le_bytes()).collect();

        // C `writePerformanceDataset` chunks `[chunking, 5]` where `chunking`
        // is the same `calculateAttributeChunking` value (NDFileHDF5.cpp:2645-2647),
        // not one row per chunk.
        let chunking = self.attribute_chunking();
        let perf_group = self.resolved_perf_group.clone();
        let h5file = match self.handle {
            Some(Hdf5Handle::Standard { ref file, .. }) => file,
            _ => return Ok(()),
        };
        // With a valid layout the performance group (the group holding the
        // `timestamp` dataset) was already created by `build_layout_groups`;
        // re-open it. Without a layout, fall back to a flat `performance`
        // group at the file root.
        let group = if perf_group.is_empty() {
            h5file
                .create_group("performance")
                .map_err(|e| ADError::UnsupportedConversion(format!("HDF5 group error: {}", e)))?
        } else {
            Self::open_write_group(h5file, &perf_group)?
        };
        let ds = group
            .new_dataset::<f64>()
            .shape([n, 5])
            .chunk(&[chunking, 5])
            .max_shape(&[None, Some(5)])
            .create("timestamp")
            .map_err(|e| {
                ADError::UnsupportedConversion(format!("HDF5 performance dataset error: {}", e))
            })?;
        // One chunk spans `chunking` rows of 5 doubles; the dataset is
        // extensible so the final partial band is zero-padded to a full chunk.
        write_chunked_buffer(&ds, &raw, chunking * 5 * 8)?;
        Ok(())
    }

    /// Write a frame in SWMR mode.
    fn write_swmr(&mut self, array: &NDArray) -> ADResult<()> {
        // A pre-compressed frame cannot be direct-chunk-written in SWMR mode:
        // `SwmrFileWriter` exposes only `append_frame`, which runs the supplied
        // bytes through the dataset's filter pipeline. Feeding already-compressed
        // bytes through it would re-compress them into garbage, so the frame is
        // rejected (faithful to C returning `asynError` when it cannot write
        // pre-compressed data). The standard (non-SWMR) write path direct-chunk-
        // writes such frames; SWMR + pre-compressed needs a raw-chunk append API
        // the SWMR backend does not yet provide.
        if array.codec.is_some() {
            return Err(ADError::UnsupportedConversion(
                "HDF5 SWMR mode cannot write a pre-compressed array (no raw-chunk \
                 append in the SWMR backend); use standard capture mode for \
                 direct chunk write"
                    .into(),
            ));
        }

        // This frame's 0-based odometer index (`frame_count` increments only
        // after `write_swmr` returns); read before the mutable handle borrow.
        let frame_idx = self.frame_count;

        let (writer, ds_index, grid) = match self.handle {
            Some(Hdf5Handle::Swmr {
                ref mut writer,
                ds_index,
                ref grid,
                ..
            }) => (writer, ds_index, grid),
            _ => return Err(ADError::UnsupportedConversion("no SWMR writer open".into())),
        };

        // The SWMR streaming dataset declares a little-endian element type and
        // both `append_frame` and `write_chunk_at` copy the supplied `&[u8]`
        // verbatim; serialize to LE explicitly (see `nd_buffer_to_le_bytes`) so
        // the file is portable.
        let frame_bytes = nd_buffer_to_le_bytes(&array.data);
        match grid {
            // Fixed multi-extra-dimension grid: place this frame at its odometer
            // chunk position(s) via `write_chunk_at`, mirroring the standard
            // path's `flush_band` (one frame per leading position, `fc == 1`).
            Some(g) => {
                let leading_coords = Self::unravel(frame_idx, &g.leading);
                for (coords, tile) in Self::band_chunk_writes(
                    &leading_coords,
                    std::slice::from_ref(&frame_bytes),
                    &g.frame_dims,
                    &g.chunk,
                    g.elem_size,
                ) {
                    let coords_u64: Vec<u64> = coords.iter().map(|&c| c as u64).collect();
                    writer
                        .write_chunk_at(ds_index, &coords_u64, &tile)
                        .map_err(|e| {
                            ADError::UnsupportedConversion(format!(
                                "SWMR write_chunk_at error: {}",
                                e
                            ))
                        })?;
                }
            }
            None => {
                writer.append_frame(ds_index, &frame_bytes).map_err(|e| {
                    ADError::UnsupportedConversion(format!("SWMR append error: {}", e))
                })?;
            }
        }

        // Periodic flush
        let count = self.frame_count + 1; // will be incremented after return
        if self.flush_nth_frame > 0 && count % self.flush_nth_frame == 0 {
            writer
                .flush()
                .map_err(|e| ADError::UnsupportedConversion(format!("SWMR flush error: {}", e)))?;
        }
        Ok(())
    }

    /// Record one frame's I/O timing into the performance buffer.
    fn record_performance(&mut self, write_duration: f64, frame_bytes: usize) {
        let now = std::time::Instant::now();
        let first = *self.perf_first.get_or_insert(now);
        let runtime = now.duration_since(first).as_secs_f64();
        let period = match self.perf_prev {
            Some(prev) => now.duration_since(prev).as_secs_f64(),
            None => write_duration,
        };
        self.perf_prev = Some(now);
        let fb = frame_bytes as f64;
        let inst_speed = if period > 0.0 { fb / period } else { 0.0 };
        let avg_speed = if runtime > 0.0 {
            (self.perf_rows.len() as f64 + 1.0) * fb / runtime
        } else {
            0.0
        };
        self.perf_rows
            .push([write_duration, period, runtime, inst_speed, avg_speed]);
    }
}

/// Whether two NDArray codecs are equal for the purpose of C
/// `NDFileHDF5Dataset::verifyChunking` (NDFileHDF5Dataset.cpp:194-200), which
/// compares `pArray->codec != this->codec`. C `Codec_t::operator!=` (Codec.h)
/// equates name, level, shuffle and compressor; `original_data_type` is not part
/// of the codec identity (it travels with the array's element type). Two absent
/// codecs (an uncompressed file) also match.
fn codecs_match(a: Option<&Codec>, b: Option<&Codec>) -> bool {
    match (a, b) {
        (None, None) => true,
        (Some(x), Some(y)) => {
            x.name == y.name
                && x.level == y.level
                && x.shuffle == y.shuffle
                && x.compressor == y.compressor
        }
        _ => false,
    }
}

/// Build the on-disk HDF5 chunk byte stream for one pre-compressed frame,
/// matching C `NDFileHDF5Dataset::writeFile` (NDFileHDF5Dataset.cpp:291-329).
/// `payload` is the codec's compressed output (this port stores it in the
/// array's collapsed `U8` buffer); `uncompressed_total_bytes` is the size of one
/// uncompressed frame (C `NDArrayInfo::totalBytes`).
///
/// - **LZ4**: this port's `compress_lz4` emits a raw LZ4 block with no header
///   (C `pArray->pData` likewise), so prepend the 16-byte big-endian header the
///   HDF5 LZ4 filter expects — uncompressed size (u64), block size = uncompressed
///   size (u32; frames assumed < 1 GiB, as C does), compressed size (u32) — then
///   the block.
/// - **BSLZ4**: this port's `compress_bslz4` emits the headerless canonical
///   bitshuffle+LZ4 stream (C `pArray->pData` likewise — the per-block
///   `[u32 nbytes_BE]` headers are part of the stream, but the chunk-level header
///   is not), so prepend the 12-byte big-endian header the HDF5 bitshuffle filter
///   expects — uncompressed size (u64) and the bitshuffle block size **in bytes**
///   (u32 = `block_elems * elem_size`; C hardcodes its 8192 default,
///   NDFileHDF5Dataset.cpp:316-328) — then the stream.
/// - **BLOSC / JPEG**: this port's codecs already emit the exact on-disk filter
///   format (blosc and jpeg streams are self-describing), so the payload is
///   written verbatim — the same bytes C produces after its per-codec header step.
fn codec_chunk_bytes(codec: &Codec, uncompressed_total_bytes: usize, payload: &[u8]) -> Vec<u8> {
    match codec.name {
        CodecName::LZ4 => {
            let mut out = Vec::with_capacity(16 + payload.len());
            out.extend_from_slice(&(uncompressed_total_bytes as u64).to_be_bytes());
            out.extend_from_slice(&(uncompressed_total_bytes as u32).to_be_bytes());
            out.extend_from_slice(&(payload.len() as u32).to_be_bytes());
            out.extend_from_slice(payload);
            out
        }
        CodecName::BSLZ4 => {
            let elem_size = codec.original_data_type.element_size();
            let block_bytes = crate::codec::bshuf_default_block_size(elem_size) * elem_size;
            let mut out = Vec::with_capacity(12 + payload.len());
            out.extend_from_slice(&(uncompressed_total_bytes as u64).to_be_bytes());
            out.extend_from_slice(&(block_bytes as u32).to_be_bytes());
            out.extend_from_slice(payload);
            out
        }
        _ => payload.to_vec(),
    }
}

/// Serialize an NDArray data buffer to **little-endian** bytes.
///
/// `rust-hdf5` 0.2.15 records every numeric datatype message as little-endian
/// (`Endianness::LittleEndian`) and its only chunked-write API, `write_chunk`,
/// copies the supplied `&[u8]` verbatim into the chunk with no byte-swap.
/// `NDDataBuffer::as_u8_slice()` returns the buffer in *host* byte order, so
/// feeding it directly into a typed chunked dataset is correct only on a
/// little-endian host. This helper makes the on-disk bytes match the declared
/// LE datatype on every host: on LE it is a verbatim copy, on BE it swaps each
/// element. Used for every typed-dataset chunk write where no typed
/// chunked-write path exists in the crate.
fn nd_buffer_to_le_bytes(buf: &NDDataBuffer) -> Vec<u8> {
    match buf {
        NDDataBuffer::I8(v) => v.iter().map(|&x| x as u8).collect(),
        NDDataBuffer::U8(v) => v.clone(),
        NDDataBuffer::I16(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::U16(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::I32(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::U32(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::I64(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::U64(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::F32(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
        NDDataBuffer::F64(v) => v.iter().flat_map(|&x| x.to_le_bytes()).collect(),
    }
}

/// Write `buffer` into a chunked dataset, one `chunk_bytes`-sized chunk at a
/// time at consecutive linear indices. The trailing partial chunk is
/// zero-padded to a full chunk, which `rust-hdf5`'s `write_chunk` requires.
fn write_chunked_buffer(
    ds: &rust_hdf5::H5Dataset,
    buffer: &[u8],
    chunk_bytes: usize,
) -> ADResult<()> {
    let n_chunks = buffer.len().div_ceil(chunk_bytes.max(1));
    for c in 0..n_chunks {
        let start = c * chunk_bytes;
        let end = ((c + 1) * chunk_bytes).min(buffer.len());
        let slice = &buffer[start..end];
        if slice.len() == chunk_bytes {
            ds.write_chunk(c, slice)
        } else {
            let mut padded = vec![0u8; chunk_bytes];
            padded[..slice.len()].copy_from_slice(slice);
            ds.write_chunk(c, &padded)
        }
        .map_err(|e| ADError::UnsupportedConversion(format!("HDF5 chunk write error: {}", e)))?;
    }
    Ok(())
}

/// Attach the four self-describing string HDF5 attributes C writes on each
/// NDAttribute dataset — `NDAttrName`, `NDAttrDescription`, `NDAttrSourceType`,
/// `NDAttrSource` (NDFileHDF5.cpp:2715, 2817-2822) — each only when non-empty
/// (C `if (strlen <= 0) continue`). Written as scalar strings, consistent with
/// the port's other string dataset attributes.
fn write_ndattr_descriptors(ds: &rust_hdf5::H5Dataset, ad: &AttributeDataset) {
    for (aname, avalue) in [
        ("NDAttrName", ad.name.as_str()),
        ("NDAttrDescription", ad.description.as_str()),
        ("NDAttrSourceType", ad.source_type.as_str()),
        ("NDAttrSource", ad.source.as_str()),
    ] {
        if avalue.is_empty() {
            continue;
        }
        let s = rust_hdf5::types::VarLenUnicode(avalue.to_string());
        let _ = ds
            .new_attr::<rust_hdf5::types::VarLenUnicode>()
            .shape(())
            .create(aname)
            .and_then(|a| a.write_scalar(&s));
    }
}

/// Write a live NDAttribute value as an HDF5 attribute on a standard-mode
/// group (C `storeOnOpenCloseAttribute` group branch). The datatype follows
/// the runtime NDAttribute value (C `typeNd2Hdf`); a string writes a scalar
/// string, an `Undefined` value is skipped. Errors are non-fatal, mirroring
/// C's per-attribute warn-and-skip.
fn write_ndattr_group_attr(group: &rust_hdf5::H5Group, attr_name: &str, value: &NDAttrValue) {
    macro_rules! num {
        ($v:expr) => {{
            let _ = group.set_attr_numeric(attr_name, &$v);
        }};
    }
    match value {
        NDAttrValue::Int8(v) => num!(*v),
        NDAttrValue::UInt8(v) => num!(*v),
        NDAttrValue::Int16(v) => num!(*v),
        NDAttrValue::UInt16(v) => num!(*v),
        NDAttrValue::Int32(v) => num!(*v),
        NDAttrValue::UInt32(v) => num!(*v),
        NDAttrValue::Int64(v) => num!(*v),
        NDAttrValue::UInt64(v) => num!(*v),
        NDAttrValue::Float32(v) => num!(*v),
        NDAttrValue::Float64(v) => num!(*v),
        NDAttrValue::String(s) => {
            let _ = group.set_attr_string(attr_name, s);
        }
        NDAttrValue::Undefined => {}
    }
}

/// Write a live NDAttribute value as an HDF5 attribute on a standard-mode
/// dataset, via a live dataset handle. rust-hdf5 0.2.17 cannot reopen a
/// dataset by name while the file is in write mode (`H5File::dataset` errors
/// with "cannot open a dataset by name in write mode"), so the caller supplies
/// the handle obtained at create time. C `storeOnOpenCloseAttribute` dataset
/// branch; `Undefined` is skipped, errors are non-fatal.
fn write_ndattr_dataset_attr(ds: &rust_hdf5::H5Dataset, attr_name: &str, value: &NDAttrValue) {
    macro_rules! num {
        ($v:expr, $t:ty) => {{
            let v: $t = $v;
            let _ = ds
                .new_attr::<$t>()
                .shape(())
                .create(attr_name)
                .and_then(|a| a.write_numeric(&v));
        }};
    }
    match value {
        NDAttrValue::Int8(v) => num!(*v, i8),
        NDAttrValue::UInt8(v) => num!(*v, u8),
        NDAttrValue::Int16(v) => num!(*v, i16),
        NDAttrValue::UInt16(v) => num!(*v, u16),
        NDAttrValue::Int32(v) => num!(*v, i32),
        NDAttrValue::UInt32(v) => num!(*v, u32),
        NDAttrValue::Int64(v) => num!(*v, i64),
        NDAttrValue::UInt64(v) => num!(*v, u64),
        NDAttrValue::Float32(v) => num!(*v, f32),
        NDAttrValue::Float64(v) => num!(*v, f64),
        NDAttrValue::String(s) => {
            let sv = rust_hdf5::types::VarLenUnicode(s.clone());
            let _ = ds
                .new_attr::<rust_hdf5::types::VarLenUnicode>()
                .shape(())
                .create(attr_name)
                .and_then(|a| a.write_scalar(&sv));
        }
        NDAttrValue::Undefined => {}
    }
}

/// SWMR counterpart of `write_ndattr_group_attr` for group element-attrs:
/// write a live NDAttribute value as a group HDF5 attribute, addressed by the
/// group's absolute path. The datatype follows the runtime NDAttribute value
/// (C `typeNd2Hdf`); `Undefined` is skipped.
fn write_swmr_ndattr_group_attr(
    swmr: &mut SwmrFileWriter,
    group_path: &str,
    attr_name: &str,
    value: &NDAttrValue,
) -> ADResult<()> {
    let res = match value {
        NDAttrValue::Int8(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::UInt8(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::Int16(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::UInt16(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::Int32(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::UInt32(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::Int64(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::UInt64(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::Float32(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::Float64(v) => swmr.set_group_attr_numeric(group_path, attr_name, v),
        NDAttrValue::String(s) => swmr.set_group_attr_string(group_path, attr_name, s),
        NDAttrValue::Undefined => return Ok(()),
    };
    res.map_err(|e| {
        ADError::UnsupportedConversion(format!(
            "SWMR ndattribute group attribute '{}/{}': {}",
            group_path, attr_name, e
        ))
    })
}

/// SWMR counterpart of [`write_ndattr_dataset_attr`]: write a live NDAttribute
/// value as a dataset HDF5 attribute, addressed by the dataset's index (the
/// only handle `SwmrFileWriter` exposes for datasets). The datatype follows the
/// runtime NDAttribute value (C `typeNd2Hdf`); `Undefined` is skipped. Must be
/// called before `start_swmr()` — HDF5 forbids creating attributes after the
/// SWMR lock.
fn write_swmr_ndattr_dataset_attr(
    swmr: &mut SwmrFileWriter,
    ds_index: usize,
    attr_name: &str,
    value: &NDAttrValue,
) -> ADResult<()> {
    let res = match value {
        NDAttrValue::Int8(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::UInt8(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::Int16(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::UInt16(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::Int32(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::UInt32(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::Int64(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::UInt64(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::Float32(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::Float64(v) => swmr.set_dataset_attr_numeric(ds_index, attr_name, v),
        NDAttrValue::String(s) => swmr.set_dataset_attr_string(ds_index, attr_name, s),
        NDAttrValue::Undefined => return Ok(()),
    };
    res.map_err(|e| {
        ADError::UnsupportedConversion(format!(
            "SWMR ndattribute dataset attribute '{}': {}",
            attr_name, e
        ))
    })
}

/// Write a layout group's `constant`-sourced `<attribute>` nodes (e.g. the
/// NeXus `NX_class` markers) to an open standard-mode group, typed per the XML
/// `type` attribute. `ndattribute`-sourced group attributes carry per-frame
/// values and are written separately by [`write_ndattr_group_attr`]. Errors are
/// non-fatal.
fn write_group_constant_attrs(
    group: &rust_hdf5::H5Group,
    attrs: &[crate::hdf5_layout::LayoutAttribute],
) {
    use crate::hdf5_layout::{LayoutDataType, LayoutSource};
    for a in attrs {
        if a.source != LayoutSource::Constant {
            continue;
        }
        let _ = match a.data_type {
            LayoutDataType::Int => {
                group.set_attr_numeric(&a.name, &a.value.trim().parse::<i64>().unwrap_or(0))
            }
            LayoutDataType::Float => {
                group.set_attr_numeric(&a.name, &a.value.trim().parse::<f64>().unwrap_or(0.0))
            }
            LayoutDataType::String => group.set_attr_string(&a.name, &a.value),
        };
    }
}

/// SWMR counterpart of [`write_group_constant_attrs`]: write a layout group's
/// `constant` `<attribute>` nodes against a `SwmrFileWriter`, addressing the
/// group by its absolute path.
fn write_swmr_group_constant_attrs(
    swmr: &mut SwmrFileWriter,
    group_path: &str,
    attrs: &[crate::hdf5_layout::LayoutAttribute],
) -> ADResult<()> {
    use crate::hdf5_layout::{LayoutDataType, LayoutSource};
    for a in attrs {
        if a.source != LayoutSource::Constant {
            continue;
        }
        match a.data_type {
            LayoutDataType::Int => swmr.set_group_attr_numeric(
                group_path,
                &a.name,
                &a.value.trim().parse::<i64>().unwrap_or(0),
            ),
            LayoutDataType::Float => swmr.set_group_attr_numeric(
                group_path,
                &a.name,
                &a.value.trim().parse::<f64>().unwrap_or(0.0),
            ),
            LayoutDataType::String => swmr.set_group_attr_string(group_path, &a.name, &a.value),
        }
        .map_err(|e| {
            ADError::UnsupportedConversion(format!(
                "SWMR layout group attribute '{}/{}': {}",
                group_path, a.name, e
            ))
        })?;
    }
    Ok(())
}

impl Default for Hdf5Writer {
    fn default() -> Self {
        Self::new()
    }
}

impl NDFileWriter for Hdf5Writer {
    fn open_file(&mut self, path: &Path, mode: NDFileMode, array: &NDArray) -> ADResult<()> {
        self.current_path = Some(path.to_path_buf());
        self.open_mode = mode;
        self.frame_count = 0;
        self.total_runtime = 0.0;
        self.total_bytes = 0;
        self.swmr_cb_counter = 0;
        self.open_data_type = None;
        self.open_frame_dims = None;
        self.open_codec = None;
        self.perf_rows.clear();
        self.perf_prev = None;
        self.perf_first = None;
        self.attr_datasets.clear();
        // C `openFile` clears `pFileAttributes` and copies this frame's list
        // into it (NDFileHDF5.cpp:243-252); every attribute read for the rest
        // of the file goes through it, not through the frame.
        self.file_attributes.open(array);
        // Resolve where image/attribute/performance datasets land for this
        // file: the loaded layout XML tree, or the flat root default.
        self.resolve_layout_paths();
        // Seed the open-time NDAttribute values for any layout
        // `<attribute source="ndattribute">` element-attrs (ADP-79).
        self.seed_ndattr_element_values();

        if self.swmr_mode && mode == NDFileMode::Stream {
            self.open_swmr(path, array)
        } else {
            let h5file = H5File::create(path)
                .map_err(|e| ADError::UnsupportedConversion(format!("HDF5 create error: {}", e)))?;
            self.handle = Some(Hdf5Handle::Standard {
                file: h5file,
                detectors: std::collections::HashMap::new(),
            });
            Ok(())
        }
    }

    fn set_num_capture(&mut self, n: usize) {
        self.num_capture = n;
    }

    fn write_file(&mut self, array: &NDArray) -> ADResult<()> {
        let start = std::time::Instant::now();

        // C `writeFile` merges this frame into `pFileAttributes` before it
        // writes anything (NDFileHDF5.cpp:1420-1437).
        self.file_attributes.frame(array);

        let is_swmr = matches!(self.handle, Some(Hdf5Handle::Swmr { .. }));
        if is_swmr {
            self.write_swmr(array)?;
        } else {
            self.write_standard(array)?;
        }
        // Track the latest NDAttribute values for `when="OnFileClose"`
        // element-attrs (ADP-79); no-op unless the layout declares any.
        self.update_ndattr_element_values();
        self.frame_count += 1;

        let elapsed = start.elapsed().as_secs_f64();
        let frame_bytes = array.data.as_u8_slice().len();
        if self.store_performance {
            self.total_runtime += elapsed;
            self.total_bytes += frame_bytes as u64;
            self.record_performance(elapsed, frame_bytes);
        }
        Ok(())
    }

    fn read_file(&mut self) -> ADResult<NDArray> {
        // The image dataset lives at the layout-resolved path (flat `data`
        // by default, or the nested layout path). Resolve it so read-back
        // tracks the same placement as the write path.
        self.resolve_layout_paths();
        let dataset_path = self.resolved_dataset_path.clone();
        let path = self
            .current_path
            .as_ref()
            .ok_or_else(|| ADError::UnsupportedConversion("no file open".into()))?;

        let h5file = H5File::open(path)
            .map_err(|e| ADError::UnsupportedConversion(format!("HDF5 open error: {}", e)))?;

        let ds = h5file
            .dataset(&dataset_path)
            .map_err(|e| ADError::UnsupportedConversion(format!("HDF5 dataset error: {}", e)))?;

        let shape = ds.shape();
        let dims: Vec<NDDimension> = shape.iter().rev().map(|&s| NDDimension::new(s)).collect();
        let element_size = ds.element_size();

        // Prefer the exact data type recorded at write time.
        let recorded: Option<NDDataType> = ds
            .attr(DTYPE_ATTR)
            .ok()
            .and_then(|a| a.read_numeric::<i32>().ok())
            .and_then(|v| NDDataType::from_ordinal(v as u8));

        let data_type = recorded.unwrap_or(match element_size {
            1 => NDDataType::UInt8,
            2 => NDDataType::UInt16,
            4 => NDDataType::Float32,
            8 => NDDataType::Float64,
            other => {
                return Err(ADError::UnsupportedConversion(format!(
                    "unsupported HDF5 element size {}",
                    other
                )));
            }
        });

        macro_rules! read_typed {
            ($t:ty, $variant:ident) => {{
                let data = ds.read_raw::<$t>().map_err(|e| {
                    ADError::UnsupportedConversion(format!("HDF5 read error: {}", e))
                })?;
                let mut arr = NDArray::new(dims, data_type);
                arr.data = NDDataBuffer::$variant(data);
                return Ok(arr);
            }};
        }

        match data_type {
            NDDataType::Int8 => read_typed!(i8, I8),
            NDDataType::UInt8 => read_typed!(u8, U8),
            NDDataType::Int16 => read_typed!(i16, I16),
            NDDataType::UInt16 => read_typed!(u16, U16),
            NDDataType::Int32 => read_typed!(i32, I32),
            NDDataType::UInt32 => read_typed!(u32, U32),
            NDDataType::Int64 => read_typed!(i64, I64),
            NDDataType::UInt64 => read_typed!(u64, U64),
            NDDataType::Float32 => read_typed!(f32, F32),
            NDDataType::Float64 => read_typed!(f64, F64),
        }
    }

    /// Close the open file.
    ///
    /// Every step below is fallible and each one runs against the open handle,
    /// so the release is owned by a finalizer rather than by the tail of the
    /// function: dropping `handle` closes the H5 file id and its dataset ids on
    /// every exit path, and HDF5 finalises the file on that drop. Without it a
    /// failing flush leaked one file id per attempt and left the file
    /// unfinalised under its `.tmp` name.
    fn close_file(&mut self) -> ADResult<()> {
        let mut closing = Finalize::new(self, |w: &mut Self| {
            w.handle = None;
            w.current_path = None;
        });
        closing.run(|w| {
            match w.handle {
                Some(Hdf5Handle::Standard { .. }) => {
                    // Flush each detector dataset's partial frame band and trim its
                    // extent, then emit the accumulated attribute and performance
                    // datasets before the file is finalised.
                    w.finalize_standard_datasets()?;
                    w.flush_attribute_datasets()?;
                    w.flush_performance_dataset()?;
                    // Materialise layout `<hardlink>` elements last, once every
                    // dataset a link may target exists on disk.
                    match w.handle {
                        Some(Hdf5Handle::Standard {
                            ref file,
                            ref detectors,
                        }) => {
                            w.build_layout_hardlinks(file)?;
                            // Every group and dataset now exists, so layout
                            // `<attribute source="ndattribute">` element-attrs can
                            // be attached with their open/close values (ADP-79).
                            w.flush_ndattr_element_attrs(file, detectors)?;
                        }
                        _ => unreachable!("handle is Standard in this arm"),
                    }
                    // Finalize the file. Dropping the `H5File` finalizes durably
                    // (with fsync); when `AD_HDF5_FSYNC_ON_CLOSE` opts out, use the
                    // no-fsync fast close instead (rust-hdf5 0.3.2 `close_no_sync`).
                    // Detector dataset handles are released only after finalize,
                    // preserving the prior file-before-datasets drop order.
                    if let Some(Hdf5Handle::Standard { file, detectors }) = w.handle.take() {
                        if w.fsync_on_close {
                            drop(file);
                        } else {
                            file.close_no_sync().map_err(|e| {
                                ADError::UnsupportedConversion(format!(
                                    "HDF5 close_no_sync error: {}",
                                    e
                                ))
                            })?;
                        }
                        drop(detectors);
                    }
                }
                Some(Hdf5Handle::Swmr { .. }) => {
                    // The layout group tree, the nested dataset placement and the
                    // layout `<hardlink>` elements were all materialised in
                    // `open_swmr` before `start_swmr()` (C `NDFileHDF5.cpp:319`-
                    // `325`: `createHardLinks` then `startSWMR`), so SWMR readers
                    // see them for the whole streaming window. Closing the writer
                    // only finalises the streamed frames.
                    if let Some(Hdf5Handle::Swmr { mut writer, .. }) = w.handle.take() {
                        // Commit any chunk writes made since the last periodic flush
                        // before closing. `SwmrFileWriter::close` finalizes the
                        // extensible-array index of the `append_frame` path but does
                        // not flush the fixed-array index that the grid path's
                        // `write_chunk_at` records, so a grid file's tail frames
                        // would otherwise read back as fill.
                        writer.flush().map_err(|e| {
                            ADError::UnsupportedConversion(format!(
                                "SWMR flush-on-close error: {}",
                                e
                            ))
                        })?;
                        writer.close().map_err(|e| {
                            ADError::UnsupportedConversion(format!("SWMR close error: {}", e))
                        })?;
                    }
                }
                None => {}
            }
            Ok(())
        })
    }

    fn supports_multiple_arrays(&self) -> bool {
        true
    }
}

// ============================================================
// Processor
// ============================================================

/// Param indices for HDF5-specific params.
#[derive(Default)]
struct Hdf5ParamIndices {
    compression_type: Option<usize>,
    z_compress_level: Option<usize>,
    szip_num_pixels: Option<usize>,
    nbit_precision: Option<usize>,
    nbit_offset: Option<usize>,
    jpeg_quality: Option<usize>,
    blosc_shuffle_type: Option<usize>,
    blosc_compressor: Option<usize>,
    blosc_compress_level: Option<usize>,
    store_attributes: Option<usize>,
    store_performance: Option<usize>,
    total_runtime: Option<usize>,
    total_io_speed: Option<usize>,
    swmr_mode: Option<usize>,
    swmr_flush_now: Option<usize>,
    swmr_running: Option<usize>,
    swmr_cb_counter: Option<usize>,
    swmr_supported: Option<usize>,
    flush_nth_frame: Option<usize>,
    chunk_size_auto: Option<usize>,
    n_row_chunks: Option<usize>,
    n_col_chunks: Option<usize>,
    n_frames_chunks: Option<usize>,
    ndattr_chunk: Option<usize>,
    n_extra_dims: Option<usize>,
    extra_dim_size: [Option<usize>; MAX_EXTRA_DIMS],
    extra_dim_name: [Option<usize>; MAX_EXTRA_DIMS],
    fill_value: Option<usize>,
    dim_att_datasets: Option<usize>,
    layout_filename: Option<usize>,
    layout_valid: Option<usize>,
    layout_error_msg: Option<usize>,
}

/// HDF5 file processor wrapping `FilePluginController<Hdf5Writer>`.
pub struct Hdf5FileProcessor {
    ctrl: Mutex<FilePluginController<Hdf5Writer>>,
    hdf5_params: Hdf5ParamIndices,
}

impl Hdf5FileProcessor {
    pub fn new() -> Self {
        Self {
            ctrl: Mutex::new(FilePluginController::new(Hdf5Writer::new())),
            hdf5_params: Hdf5ParamIndices::default(),
        }
    }

    pub fn set_dataset_name(&mut self, name: &str) {
        self.ctrl.lock().writer.set_dataset_name(name);
    }
}

/// Register all HDF5-specific params.
fn register_hdf5_params(
    base: &mut asyn_rs::port::PortDriverBase,
) -> asyn_rs::error::AsynResult<()> {
    use asyn_rs::param::ParamType;
    base.create_param("HDF5_SWMRFlushNow", ParamType::Int32)?;
    base.create_param("HDF5_chunkSizeAuto", ParamType::Int32)?;
    base.create_param("HDF5_nRowChunks", ParamType::Int32)?;
    base.create_param("HDF5_nColChunks", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize2", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize3", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize4", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize5", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize6", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize7", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize8", ParamType::Int32)?;
    base.create_param("HDF5_chunkSize9", ParamType::Int32)?;
    base.create_param("HDF5_nFramesChunks", ParamType::Int32)?;
    base.create_param("HDF5_NDAttributeChunk", ParamType::Int32)?;
    base.create_param("HDF5_chunkBoundaryAlign", ParamType::Int32)?;
    base.create_param("HDF5_chunkBoundaryThreshold", ParamType::Int32)?;
    base.create_param("HDF5_nExtraDims", ParamType::Int32)?;
    base.create_param("HDF5_extraDimSizeN", ParamType::Int32)?;
    base.create_param("HDF5_extraDimNameN", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSizeX", ParamType::Int32)?;
    base.create_param("HDF5_extraDimNameX", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSizeY", ParamType::Int32)?;
    base.create_param("HDF5_extraDimNameY", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize3", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName3", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize4", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName4", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize5", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName5", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize6", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName6", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize7", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName7", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize8", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName8", ParamType::Octet)?;
    base.create_param("HDF5_extraDimSize9", ParamType::Int32)?;
    base.create_param("HDF5_extraDimName9", ParamType::Octet)?;
    base.create_param("HDF5_storeAttributes", ParamType::Int32)?;
    base.create_param("HDF5_storePerformance", ParamType::Int32)?;
    base.create_param("HDF5_totalRuntime", ParamType::Float64)?;
    base.create_param("HDF5_totalIoSpeed", ParamType::Float64)?;
    base.create_param("HDF5_flushNthFrame", ParamType::Int32)?;
    base.create_param("HDF5_compressionType", ParamType::Int32)?;
    base.create_param("HDF5_nbitsPrecision", ParamType::Int32)?;
    base.create_param("HDF5_nbitsOffset", ParamType::Int32)?;
    base.create_param("HDF5_szipNumPixels", ParamType::Int32)?;
    base.create_param("HDF5_zCompressLevel", ParamType::Int32)?;
    base.create_param("HDF5_bloscShuffleType", ParamType::Int32)?;
    base.create_param("HDF5_bloscCompressor", ParamType::Int32)?;
    base.create_param("HDF5_bloscCompressLevel", ParamType::Int32)?;
    base.create_param("HDF5_jpegQuality", ParamType::Int32)?;
    base.create_param("HDF5_dimAttDatasets", ParamType::Int32)?;
    base.create_param("HDF5_layoutErrorMsg", ParamType::Octet)?;
    base.create_param("HDF5_layoutValid", ParamType::Int32)?;
    base.create_param("HDF5_layoutFilename", ParamType::Octet)?;
    base.create_param("HDF5_SWMRSupported", ParamType::Int32)?;
    base.create_param("HDF5_SWMRMode", ParamType::Int32)?;
    base.create_param("HDF5_SWMRRunning", ParamType::Int32)?;
    base.create_param("HDF5_SWMRCbCounter", ParamType::Int32)?;
    base.create_param("HDF5_posRunning", ParamType::Int32)?;
    base.create_param("HDF5_posNameDimN", ParamType::Octet)?;
    base.create_param("HDF5_posNameDimX", ParamType::Octet)?;
    base.create_param("HDF5_posNameDimY", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim3", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim4", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim5", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim6", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim7", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim8", ParamType::Octet)?;
    base.create_param("HDF5_posNameDim9", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDimN", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDimX", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDimY", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim3", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim4", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim5", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim6", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim7", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim8", ParamType::Octet)?;
    base.create_param("HDF5_posIndexDim9", ParamType::Octet)?;
    base.create_param("HDF5_fillValue", ParamType::Float64)?;
    base.create_param("HDF5_extraDimChunkX", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunkY", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk3", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk4", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk5", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk6", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk7", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk8", ParamType::Int32)?;
    base.create_param("HDF5_extraDimChunk9", ParamType::Int32)?;
    Ok(())
}

impl Default for Hdf5FileProcessor {
    fn default() -> Self {
        Self::new()
    }
}

/// Names of the `HDF5_extraDimSizeN..9` params in slot order.
const EXTRA_DIM_SIZE_PARAMS: [&str; MAX_EXTRA_DIMS] = [
    "HDF5_extraDimSizeN",
    "HDF5_extraDimSizeX",
    "HDF5_extraDimSizeY",
    "HDF5_extraDimSize3",
    "HDF5_extraDimSize4",
    "HDF5_extraDimSize5",
    "HDF5_extraDimSize6",
    "HDF5_extraDimSize7",
    "HDF5_extraDimSize8",
    "HDF5_extraDimSize9",
];

/// Names of the `HDF5_extraDimNameN..9` params in slot order.
const EXTRA_DIM_NAME_PARAMS: [&str; MAX_EXTRA_DIMS] = [
    "HDF5_extraDimNameN",
    "HDF5_extraDimNameX",
    "HDF5_extraDimNameY",
    "HDF5_extraDimName3",
    "HDF5_extraDimName4",
    "HDF5_extraDimName5",
    "HDF5_extraDimName6",
    "HDF5_extraDimName7",
    "HDF5_extraDimName8",
    "HDF5_extraDimName9",
];

impl NDPluginProcess for Hdf5FileProcessor {
    fn process_array(&self, array: &NDArray, _pool: &NDArrayPool) -> ProcessResult {
        let mut ctrl = self.ctrl.lock();
        let was_swmr = ctrl.writer.is_swmr_active();
        let mut result = ctrl.process_array(array);
        let is_swmr = ctrl.writer.is_swmr_active();

        // SWMR running status changed
        if was_swmr != is_swmr {
            if let Some(idx) = self.hdf5_params.swmr_running {
                result
                    .param_updates
                    .push(ParamUpdate::int32(idx, if is_swmr { 1 } else { 0 }));
            }
        }

        // SWMR callback counter
        if is_swmr {
            if let Some(idx) = self.hdf5_params.swmr_cb_counter {
                result
                    .param_updates
                    .push(ParamUpdate::int32(idx, ctrl.writer.swmr_cb_counter as i32));
            }
        }

        // Performance stats
        if ctrl.writer.store_performance {
            if let Some(idx) = self.hdf5_params.total_runtime {
                result
                    .param_updates
                    .push(ParamUpdate::float64(idx, ctrl.writer.total_runtime));
            }
            if let Some(idx) = self.hdf5_params.total_io_speed {
                let speed = if ctrl.writer.total_runtime > 0.0 {
                    ctrl.writer.total_bytes as f64 / ctrl.writer.total_runtime / 1_000_000.0
                } else {
                    0.0
                };
                result.param_updates.push(ParamUpdate::float64(idx, speed));
            }
        }

        result
    }

    fn plugin_type(&self) -> &str {
        "NDFileHDF5"
    }

    /// C `NDFileHDF5` passes `compressionAware = true` to the file driver
    /// (NDFileHDF5.cpp:2268): it accepts pre-compressed input arrays and writes
    /// their bytes verbatim through a matching HDF5 filter via direct chunk
    /// write, rather than having the framework decompress them first. The
    /// standard (non-SWMR) write path implements this; see `write_standard`.
    fn compression_aware(&self) -> bool {
        true
    }

    /// C `NDPluginFile.cpp:948` (base of every file writer) sets
    /// `NDArrayCallbacks = 0`: file plugins write to disk, not downstream.
    fn does_array_callbacks(&self) -> bool {
        false
    }

    fn register_params(
        &mut self,
        base: &mut asyn_rs::port::PortDriverBase,
    ) -> asyn_rs::error::AsynResult<()> {
        self.ctrl.lock().register_params(base)?;
        register_hdf5_params(base)?;
        self.hdf5_params.compression_type = base.find_param("HDF5_compressionType");
        self.hdf5_params.z_compress_level = base.find_param("HDF5_zCompressLevel");
        self.hdf5_params.szip_num_pixels = base.find_param("HDF5_szipNumPixels");
        self.hdf5_params.nbit_precision = base.find_param("HDF5_nbitsPrecision");
        self.hdf5_params.nbit_offset = base.find_param("HDF5_nbitsOffset");
        self.hdf5_params.jpeg_quality = base.find_param("HDF5_jpegQuality");
        self.hdf5_params.blosc_shuffle_type = base.find_param("HDF5_bloscShuffleType");
        self.hdf5_params.blosc_compressor = base.find_param("HDF5_bloscCompressor");
        self.hdf5_params.blosc_compress_level = base.find_param("HDF5_bloscCompressLevel");
        self.hdf5_params.store_attributes = base.find_param("HDF5_storeAttributes");
        self.hdf5_params.store_performance = base.find_param("HDF5_storePerformance");
        self.hdf5_params.total_runtime = base.find_param("HDF5_totalRuntime");
        self.hdf5_params.total_io_speed = base.find_param("HDF5_totalIoSpeed");
        self.hdf5_params.swmr_mode = base.find_param("HDF5_SWMRMode");
        self.hdf5_params.swmr_flush_now = base.find_param("HDF5_SWMRFlushNow");
        self.hdf5_params.swmr_running = base.find_param("HDF5_SWMRRunning");
        self.hdf5_params.swmr_cb_counter = base.find_param("HDF5_SWMRCbCounter");
        self.hdf5_params.swmr_supported = base.find_param("HDF5_SWMRSupported");
        self.hdf5_params.flush_nth_frame = base.find_param("HDF5_flushNthFrame");
        self.hdf5_params.chunk_size_auto = base.find_param("HDF5_chunkSizeAuto");
        self.hdf5_params.n_row_chunks = base.find_param("HDF5_nRowChunks");
        self.hdf5_params.n_col_chunks = base.find_param("HDF5_nColChunks");
        self.hdf5_params.n_frames_chunks = base.find_param("HDF5_nFramesChunks");
        self.hdf5_params.ndattr_chunk = base.find_param("HDF5_NDAttributeChunk");
        self.hdf5_params.n_extra_dims = base.find_param("HDF5_nExtraDims");
        for i in 0..MAX_EXTRA_DIMS {
            self.hdf5_params.extra_dim_size[i] = base.find_param(EXTRA_DIM_SIZE_PARAMS[i]);
            self.hdf5_params.extra_dim_name[i] = base.find_param(EXTRA_DIM_NAME_PARAMS[i]);
        }
        self.hdf5_params.fill_value = base.find_param("HDF5_fillValue");
        self.hdf5_params.dim_att_datasets = base.find_param("HDF5_dimAttDatasets");
        self.hdf5_params.layout_filename = base.find_param("HDF5_layoutFilename");
        self.hdf5_params.layout_valid = base.find_param("HDF5_layoutValid");
        self.hdf5_params.layout_error_msg = base.find_param("HDF5_layoutErrorMsg");

        // Report SWMR as always supported
        if let Some(idx) = self.hdf5_params.swmr_supported {
            base.set_int32_param(idx, 0, 1)?;
        }
        Ok(())
    }

    fn on_param_change(&self, reason: usize, params: &PluginParamSnapshot) -> ParamChangeResult {
        let mut ctrl = self.ctrl.lock();
        // -- compression params --
        if Some(reason) == self.hdf5_params.compression_type {
            ctrl.writer.set_compression_type(params.value.as_i32());
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.z_compress_level {
            ctrl.writer
                .set_z_compress_level(params.value.as_i32() as u32);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.szip_num_pixels {
            ctrl.writer
                .set_szip_num_pixels(params.value.as_i32() as u32);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.blosc_shuffle_type {
            ctrl.writer.set_blosc_shuffle_type(params.value.as_i32());
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.blosc_compressor {
            ctrl.writer.set_blosc_compressor(params.value.as_i32());
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.blosc_compress_level {
            ctrl.writer
                .set_blosc_compress_level(params.value.as_i32() as u32);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.nbit_precision {
            ctrl.writer.set_nbit_precision(params.value.as_i32() as u32);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.nbit_offset {
            ctrl.writer.set_nbit_offset(params.value.as_i32() as u32);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.jpeg_quality {
            ctrl.writer.set_jpeg_quality(params.value.as_i32() as u32);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.store_attributes {
            ctrl.writer.set_store_attributes(params.value.as_i32() != 0);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.store_performance {
            ctrl.writer
                .set_store_performance(params.value.as_i32() != 0);
            return ParamChangeResult::updates(vec![]);
        }
        // -- chunking params --
        if Some(reason) == self.hdf5_params.chunk_size_auto {
            ctrl.writer.set_chunk_size_auto(params.value.as_i32() != 0);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.n_row_chunks {
            ctrl.writer
                .set_n_row_chunks(params.value.as_i32().max(0) as usize);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.n_col_chunks {
            ctrl.writer
                .set_n_col_chunks(params.value.as_i32().max(0) as usize);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.n_frames_chunks {
            ctrl.writer
                .set_n_frames_chunks(params.value.as_i32().max(0) as usize);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.ndattr_chunk {
            // `0` (auto) is valid; only negatives are coerced away.
            ctrl.writer
                .set_ndattr_chunk(params.value.as_i32().max(0) as usize);
            return ParamChangeResult::updates(vec![]);
        }
        // -- extra dimensions --
        if Some(reason) == self.hdf5_params.n_extra_dims {
            ctrl.writer
                .set_n_extra_dims(params.value.as_i32().max(0) as usize);
            return ParamChangeResult::updates(vec![]);
        }
        for i in 0..MAX_EXTRA_DIMS {
            if Some(reason) == self.hdf5_params.extra_dim_size[i] {
                ctrl.writer
                    .set_extra_dim_size(i, params.value.as_i32().max(1) as usize);
                return ParamChangeResult::updates(vec![]);
            }
            if Some(reason) == self.hdf5_params.extra_dim_name[i] {
                ctrl.writer
                    .set_extra_dim_name(i, params.value.as_string().unwrap_or(""));
                return ParamChangeResult::updates(vec![]);
            }
        }
        if Some(reason) == self.hdf5_params.fill_value {
            ctrl.writer.set_fill_value(params.value.as_f64());
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.dim_att_datasets {
            ctrl.writer.set_dim_att_datasets(params.value.as_i32() != 0);
            return ParamChangeResult::updates(vec![]);
        }
        // -- layout XML --
        if Some(reason) == self.hdf5_params.layout_filename {
            let path = params.value.as_string().unwrap_or("").to_string();
            ctrl.writer.set_layout_filename(&path);
            let mut updates = vec![];
            if let Some(idx) = self.hdf5_params.layout_valid {
                updates.push(ParamUpdate::int32(
                    idx,
                    if ctrl.writer.layout_valid { 1 } else { 0 },
                ));
            }
            if let Some(idx) = self.hdf5_params.layout_error_msg {
                updates.push(ParamUpdate::Octet {
                    reason: idx,
                    addr: 0,
                    value: ctrl.writer.layout_error.clone(),
                });
            }
            return ParamChangeResult::updates(updates);
        }
        // -- SWMR params --
        if Some(reason) == self.hdf5_params.swmr_mode {
            ctrl.writer.set_swmr_mode(params.value.as_i32() != 0);
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.swmr_flush_now {
            if params.value.as_i32() != 0 {
                ctrl.writer.flush_swmr();
                let mut updates = vec![];
                if let Some(idx) = self.hdf5_params.swmr_cb_counter {
                    updates.push(ParamUpdate::int32(idx, ctrl.writer.swmr_cb_counter as i32));
                }
                return ParamChangeResult::updates(updates);
            }
            return ParamChangeResult::updates(vec![]);
        }
        if Some(reason) == self.hdf5_params.flush_nth_frame {
            ctrl.writer
                .set_flush_nth_frame(params.value.as_i32().max(0) as usize);
            return ParamChangeResult::updates(vec![]);
        }
        ctrl.on_param_change(reason, params)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use ad_core_rs::attributes::{NDAttrSource, NDAttrValue, NDAttribute};
    use rust_hdf5::format::messages::filter::FILTER_LZ4;
    use std::sync::atomic::{AtomicU32, Ordering};

    static TEST_COUNTER: AtomicU32 = AtomicU32::new(0);

    fn temp_path(prefix: &str) -> PathBuf {
        let n = TEST_COUNTER.fetch_add(1, Ordering::Relaxed);
        std::env::temp_dir().join(format!(
            "adcore_test_{}_{}_{}.h5",
            std::process::id(),
            prefix,
            n
        ))
    }

    #[test]
    fn test_write_single_frame() {
        let path = temp_path("hdf5_single");
        let mut writer = Hdf5Writer::new();

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt8,
        );
        if let NDDataBuffer::U8(ref mut v) = arr.data {
            for i in 0..16 {
                v[i] = i as u8;
            }
        }

        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        // Single-frame standard mode: dataset is [1, 4, 4].
        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![1, 4, 4]);
        let data: Vec<u8> = ds.read_raw().unwrap();
        assert_eq!(data[0], 0);
        assert_eq!(data[15], 15);
        drop(h5);

        let mut reader = Hdf5Writer::new();
        reader.current_path = Some(path.clone());
        let read_arr = reader.read_file().unwrap();
        assert_eq!(read_arr.dims.len(), 3);
        assert_eq!(read_arr.dims[2].size, 1); // leading frame dim

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_parse_fsync_on_close_env() {
        // Unset keeps the durable default.
        assert!(Hdf5Writer::parse_fsync_on_close_env(None));
        // The falsey set opts into the no-fsync fast close.
        for v in ["0", "false", "no", "off", "FALSE", "Off", "  no  "] {
            assert!(
                !Hdf5Writer::parse_fsync_on_close_env(Some(v)),
                "{v:?} should disable fsync-on-close"
            );
        }
        // Anything else — including empty and truthy values — stays durable.
        for v in ["", "1", "true", "yes", "on", "durable"] {
            assert!(
                Hdf5Writer::parse_fsync_on_close_env(Some(v)),
                "{v:?} should keep fsync-on-close"
            );
        }
    }

    #[test]
    fn test_no_fsync_close_writes_valid_file() {
        // The no-fsync fast close (`H5File::close_no_sync`) must still finalize
        // a complete, readable file — it only skips the durability fsync, not
        // the header/superblock writes.
        let path = temp_path("hdf5_no_fsync");
        let mut writer = Hdf5Writer::new();
        writer.fsync_on_close = false;

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt8,
        );
        if let NDDataBuffer::U8(ref mut v) = arr.data {
            for i in 0..16 {
                v[i] = i as u8;
            }
        }

        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![1, 4, 4]);
        let data: Vec<u8> = ds.read_raw().unwrap();
        assert_eq!(data[0], 0);
        assert_eq!(data[15], 15);
        drop(h5);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_write_multiple_frames() {
        let path = temp_path("hdf5_multi");
        let mut writer = Hdf5Writer::new();

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt8,
        );
        // Mark each frame distinctly so we can verify per-frame placement.
        for f in 0..3u8 {
            if let NDDataBuffer::U8(ref mut v) = arr.data {
                for x in v.iter_mut() {
                    *x = f;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        assert!(writer.supports_multiple_arrays());
        assert_eq!(writer.frame_count(), 3);

        let data = std::fs::read(&path).unwrap();
        assert_eq!(&data[0..8], b"\x89HDF\r\n\x1a\n");

        // Single extensible dataset [3, 4, 4] — NOT one dataset per frame.
        let h5 = H5File::open(&path).unwrap();
        let names = h5.dataset_names();
        assert!(names.contains(&"entry/instrument/detector/data".to_string()));
        assert!(
            !names.contains(&"data_1".to_string()),
            "must not write per-frame datasets"
        );
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(
            ds.shape(),
            vec![3, 4, 4],
            "rank/shape must be [nframes,Y,X]"
        );
        let raw: Vec<u8> = ds.read_raw().unwrap();
        assert_eq!(raw.len(), 3 * 4 * 4);
        // Frame 0 all zeros, frame 1 all ones, frame 2 all twos.
        assert_eq!(raw[0], 0);
        assert_eq!(raw[16], 1);
        assert_eq!(raw[32], 2);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_sub_frame_chunking() {
        // nRowChunks/nColChunks that divide the frame produce a sub-frame
        // chunk grid written via write_chunk_at tiles; the dataset shape
        // stays exactly [N, Y, X] (no padding) and the data round-trips.
        let path = temp_path("hdf5_subchunk");
        let mut writer = Hdf5Writer::new();
        writer.set_chunk_size_auto(false); // honor explicit chunk sizes
        writer.set_n_row_chunks(4); // Y = 8 → 2 row tiles
        writer.set_n_col_chunks(4); // X = 8 → 2 col tiles

        let mut arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        for f in 0..3u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for (i, x) in v.iter_mut().enumerate() {
                    *x = f * 1000 + i as u16;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![3, 8, 8], "shape must not be chunk-padded");
        assert_eq!(
            ds.chunk_dims(),
            Some(vec![1, 4, 4]),
            "chunk grid must be the sub-frame tile size"
        );
        let raw: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(raw.len(), 3 * 64);
        for f in 0..3u16 {
            for i in 0..64usize {
                assert_eq!(
                    raw[f as usize * 64 + i],
                    f * 1000 + i as u16,
                    "frame {} elem {}",
                    f,
                    i
                );
            }
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_sub_frame_chunking_with_compression() {
        // Sub-frame chunk tiles must round-trip through a filter pipeline:
        // each write_chunk_at tile is compressed independently.
        let path = temp_path("hdf5_subchunk_zlib");
        let mut writer = Hdf5Writer::new();
        writer.set_chunk_size_auto(false);
        writer.set_n_row_chunks(4);
        writer.set_n_col_chunks(4);
        writer.set_compression_type(COMPRESS_ZLIB);

        let mut arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        for f in 0..2u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for (i, x) in v.iter_mut().enumerate() {
                    *x = f * 100 + i as u16;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![2, 8, 8]);
        assert_eq!(ds.chunk_dims(), Some(vec![1, 4, 4]));
        let raw: Vec<u16> = ds.read_raw().unwrap();
        for f in 0..2u16 {
            for i in 0..64usize {
                assert_eq!(raw[f as usize * 64 + i], f * 100 + i as u16);
            }
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_non_dividing_chunk_is_honored_and_extent_trimmed() {
        // A chunk size that does not divide the frame is honored as-is;
        // write_chunk_at rounds the extent up, and close_file's set_extent
        // trims the dataset shape back to the exact [N, Y, X].
        let path = temp_path("hdf5_subchunk_nd");
        let mut writer = Hdf5Writer::new();
        writer.set_chunk_size_auto(false); // honor explicit chunk sizes
        writer.set_n_row_chunks(3); // Y = 8, 8 % 3 != 0 → honored
        writer.set_n_col_chunks(4); // X = 8 → honored

        let mut arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for (i, x) in v.iter_mut().enumerate() {
                *x = i as u16;
            }
        }
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![2, 8, 8], "extent trimmed, not padded");
        assert_eq!(ds.chunk_dims(), Some(vec![1, 3, 4]));
        let raw: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(raw.len(), 2 * 64);
        for i in 0..64usize {
            assert_eq!(raw[i], i as u16);
            assert_eq!(raw[64 + i], i as u16);
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_n_frames_chunks_band() {
        // HDF5_nFramesChunks groups frames into a multi-frame chunk band; the
        // logical frame count stays exact even when the last band is partial.
        let path = temp_path("hdf5_framechunks");
        let mut writer = Hdf5Writer::new();
        writer.set_chunk_size_auto(false);
        writer.set_n_frames_chunks(2); // 2 frames per chunk band

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        // 5 frames → bands [0,1], [2,3], [4] (partial).
        for f in 0..5u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for (i, x) in v.iter_mut().enumerate() {
                    *x = f * 1000 + i as u16;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![5, 4, 4], "exact frame count, no padding");
        assert_eq!(ds.chunk_dims(), Some(vec![2, 4, 4]));
        let raw: Vec<u16> = ds.read_raw().unwrap();
        for f in 0..5u16 {
            for i in 0..16usize {
                assert_eq!(raw[f as usize * 16 + i], f * 1000 + i as u16);
            }
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_frames_chunks_with_sub_frame_tiles() {
        // Full chunk geometry: nFramesChunks AND sub-frame row/col tiling at
        // once — exercises the complete flush_band [fc, rc, cc] tile grid
        // with a partial final band.
        let path = temp_path("hdf5_full_chunk");
        let mut writer = Hdf5Writer::new();
        writer.set_chunk_size_auto(false);
        writer.set_n_frames_chunks(2); // 2 frames per band
        writer.set_n_row_chunks(4); // Y = 8 → 2 row tiles
        writer.set_n_col_chunks(4); // X = 8 → 2 col tiles

        let mut arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        // 3 frames → band [0,1] full, band [2] partial; 2x2 tiles each.
        for f in 0..3u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for (i, x) in v.iter_mut().enumerate() {
                    *x = f * 1000 + i as u16;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![3, 8, 8], "exact frame count");
        assert_eq!(ds.chunk_dims(), Some(vec![2, 4, 4]));
        let raw: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(raw.len(), 3 * 64);
        for f in 0..3u16 {
            for i in 0..64usize {
                assert_eq!(
                    raw[f as usize * 64 + i],
                    f * 1000 + i as u16,
                    "frame {} elem {}",
                    f,
                    i
                );
            }
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_attribute_datasets() {
        let path = temp_path("hdf5_attr_ds");
        let mut writer = Hdf5Writer::new();

        let mk = |exposure: f64, count: i32| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "exposure",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Float64(exposure),
            ));
            arr.attributes.add(NDAttribute::new_static(
                "count",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Int32(count),
            ));
            arr
        };

        let a0 = mk(0.5, 10);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk(0.75, 20)).unwrap();
        writer.write_file(&mk(1.25, 30)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        // One HDF5 dataset per NDAttribute, under NDAttributes/, [nframes].
        let exp = h5
            .dataset("entry/instrument/NDAttributes/exposure")
            .unwrap();
        assert_eq!(exp.shape(), vec![3]);
        let exp_vals: Vec<f64> = exp.read_raw().unwrap();
        assert_eq!(exp_vals, vec![0.5, 0.75, 1.25]);

        let cnt = h5.dataset("entry/instrument/NDAttributes/count").unwrap();
        assert_eq!(cnt.shape(), vec![3]);
        // Numeric type preserved: i32, not stringified.
        let cnt_vals: Vec<i32> = cnt.read_raw().unwrap();
        assert_eq!(cnt_vals, vec![10, 20, 30]);

        std::fs::remove_file(&path).ok();
    }

    /// ADP-102. C's per-frame attribute record comes out of `pFileAttributes`,
    /// which `openFile` clears (NDFileHDF5.cpp:243) and every `writeFile`
    /// merges into (`:1437`) through `NDAttributeList::copy`
    /// (NDAttributeList.cpp:185-205) — a copy that adds and overwrites but
    /// never removes. An attribute that stops arriving therefore keeps its last
    /// value for the rest of the file, and C re-writes that value each frame.
    /// Reading the frame's own list instead turned a drop-out into a 0.
    #[test]
    fn attribute_that_drops_out_keeps_its_last_value() {
        let path = temp_path("hdf5_attr_sticky");
        let mut writer = Hdf5Writer::new();

        let mk = |exposure: Option<f64>| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "count",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Int32(7),
            ));
            if let Some(v) = exposure {
                arr.attributes.add(NDAttribute::new_static(
                    "exposure",
                    "",
                    NDAttrSource::Driver,
                    NDAttrValue::Float64(v),
                ));
            }
            arr
        };

        let a0 = mk(Some(0.5));
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        // Present and changed: the new value wins.
        writer.write_file(&mk(Some(0.75))).unwrap();
        // Absent: the previous value stands, it does not become 0.
        writer.write_file(&mk(None)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let exp = h5
            .dataset("entry/instrument/NDAttributes/exposure")
            .unwrap();
        assert_eq!(exp.shape(), vec![3]);
        let exp_vals: Vec<f64> = exp.read_raw().unwrap();
        assert_eq!(exp_vals, vec![0.5, 0.75, 0.75]);

        std::fs::remove_file(&path).ok();
    }

    /// ADP-102, the other side: an attribute first seen after open joins the
    /// sticky list but has no dataset, because C builds `attrList` once from
    /// `pFileAttributes` at the first frame (`createAttributeDataset`,
    /// NDFileHDF5.cpp:2782-2857) and never revisits it.
    #[test]
    fn attribute_appearing_after_open_creates_no_dataset() {
        let path = temp_path("hdf5_attr_late");
        let mut writer = Hdf5Writer::new();

        let mut a0 = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
        a0.attributes.add(NDAttribute::new_static(
            "count",
            "",
            NDAttrSource::Driver,
            NDAttrValue::Int32(7),
        ));
        let mut a1 = a0.clone();
        a1.attributes.add(NDAttribute::new_static(
            "late",
            "",
            NDAttrSource::Driver,
            NDAttrValue::Int32(1),
        ));

        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&a1).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        assert_eq!(
            h5.dataset("entry/instrument/NDAttributes/count")
                .unwrap()
                .shape(),
            vec![2]
        );
        assert!(
            h5.dataset("entry/instrument/NDAttributes/late").is_err(),
            "an attribute absent at open gets no dataset"
        );

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_adp78_undefined_attribute_dataset_is_float32_nan() {
        // ADP-78. C fixes an attribute dataset's HDF5 datatype and fill value
        // once, at file-open, from `ndAttr->getDataType()` (NDFileHDF5.cpp:2797,
        // :2830 → `typeAsHdf`, NDFileHDF5AttributeDataset.cpp:73-97). An
        // attribute with no value takes the `default:` arm — `H5T_NATIVE_FLOAT`
        // with an `epicsNAN` fill (`:366-370`) — and every element write is
        // gated on `!isUndefined_` (`:160`), so the NaN fill stands for every
        // frame even after a concrete value shows up. The port used to take
        // `NDAttrValue::Undefined.data_type()`, which is `Int32`, and wrote a
        // measured-looking 0 for the frames that carried nothing.
        let path = temp_path("hdf5_attr_undef");
        let mut writer = Hdf5Writer::new();

        let mk = |undef_value: NDAttrValue| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "novalue",
                "",
                NDAttrSource::Driver,
                undef_value,
            ));
            arr
        };

        // The frame that opens the file carries no value for the attribute.
        let a0 = mk(NDAttrValue::Undefined);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        // A concrete value arriving later must not retype the dataset, and must
        // not land in it: C never re-runs `typeAsHdf` and never lifts the gate.
        writer.write_file(&mk(NDAttrValue::Int32(7))).unwrap();
        writer.write_file(&mk(NDAttrValue::Int32(9))).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/NDAttributes/novalue").unwrap();
        assert_eq!(ds.shape(), vec![3]);
        // Float32, not the Int32 the first concrete value would imply.
        let vals: Vec<f32> = ds.read_raw().unwrap();
        assert_eq!(vals.len(), 3);
        for (i, v) in vals.iter().enumerate() {
            assert!(
                v.is_nan(),
                "element {} is {} — a reader cannot tell it from a measured value; C leaves epicsNAN",
                i,
                v
            );
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_attribute_dataset_chunk_matches_capture_target() {
        // C `calculateAttributeChunking` (NDFileHDF5.cpp:2869-2920): with the
        // default auto chunk param (0) in a non-Single mode, the attribute
        // dataset chunks at NDFileNumCapture, NOT the actual frame count — the
        // dataset is extensible so the chunk may exceed the current extent.
        // Capture target 10, only 3 frames written: chunk dim 0 must be 10,
        // extent 3; values still round-trip through the single padded chunk.
        let path = temp_path("hdf5_attr_chunk_cap");
        let mut writer = Hdf5Writer::new();
        writer.set_num_capture(10);

        let mk = |c: i32| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "count",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Int32(c),
            ));
            arr
        };

        let a0 = mk(1);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk(2)).unwrap();
        writer.write_file(&mk(3)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/NDAttributes/count").unwrap();
        assert_eq!(ds.shape(), vec![3]);
        assert_eq!(ds.chunk_dims(), Some(vec![10]));
        let vals: Vec<i32> = ds.read_raw().unwrap();
        assert_eq!(vals, vec![1, 2, 3]);
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_attribute_dataset_chunk_single_mode_is_one() {
        // Auto chunk param (0) in Single mode resolves to 1
        // (`calculateAttributeChunking`: fileWriteMode == NDFileModeSingle → 1),
        // regardless of the capture target.
        let path = temp_path("hdf5_attr_chunk_single");
        let mut writer = Hdf5Writer::new();
        writer.set_num_capture(64);

        let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
        arr.attributes.add(NDAttribute::new_static(
            "count",
            "",
            NDAttrSource::Driver,
            NDAttrValue::Int32(7),
        ));
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/NDAttributes/count").unwrap();
        assert_eq!(ds.chunk_dims(), Some(vec![1]));
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_string_attribute_dataset_is_rank1_fixed_string() {
        // C stores a string-valued NDAttribute as a rank-1 `[n]` dataset of a
        // fixed 256-byte H5T_C_S1 string (NDFileHDF5AttributeDataset.cpp:321-323,
        // rank_=1), NOT a 2-D `[n,256]` uint8 array. Verify rank 1, a 256-byte
        // string element, and that the null-terminated bytes round-trip.
        let path = temp_path("hdf5_attr_str");
        let mut writer = Hdf5Writer::new();

        let mk = |mode_name: &str| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "ColorMode",
                "",
                NDAttrSource::Driver,
                NDAttrValue::String(mode_name.to_string()),
            ));
            arr
        };

        let a0 = mk("Mono");
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk("RGB1")).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        // The default layout declares ColorMode as a `<dataset source="ndattribute">`
        // in the detector subgroup, so it is routed there, not the default group.
        let ds = h5
            .dataset("entry/instrument/detector/NDAttributes/ColorMode")
            .unwrap();
        // Rank-1 [2] of a 256-byte fixed-length string element, not [2,256] u8.
        assert_eq!(ds.shape(), vec![2]);
        assert_eq!(ds.element_size(), MAX_ATTRIBUTE_STRING_SIZE);

        let frames: Vec<FixedStr256> = ds.read_raw().unwrap();
        assert_eq!(frames.len(), 2);
        let decode = |f: &FixedStr256| -> String {
            f.0.iter()
                .take_while(|&&b| b != 0)
                .map(|&b| b as char)
                .collect()
        };
        assert_eq!(decode(&frames[0]), "Mono");
        assert_eq!(decode(&frames[1]), "RGB1");
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_ndattribute_dataset_routed_to_declared_group() {
        // C `find_dset_ndattr` (NDFileHDF5.cpp:2792) places an NDAttribute that
        // matches a `<dataset source="ndattribute">` declaration into that
        // dataset's parent group; the default layout declares ColorMode in the
        // detector subgroup. An NDAttribute with no declaration falls back to
        // the `ndattr_default` group. Verify both in a single file.
        let path = temp_path("hdf5_attr_route");
        let mut writer = Hdf5Writer::new();

        let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
        arr.attributes.add(NDAttribute::new_static(
            "ColorMode",
            "",
            NDAttrSource::Driver,
            NDAttrValue::String("Mono".to_string()),
        ));
        arr.attributes.add(NDAttribute::new_static(
            "count",
            "",
            NDAttrSource::Driver,
            NDAttrValue::Int32(7),
        ));

        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        // Declared → routed into the detector subgroup.
        assert!(
            h5.dataset("entry/instrument/detector/NDAttributes/ColorMode")
                .is_ok()
        );
        // Undeclared → default ndattr group.
        assert!(h5.dataset("entry/instrument/NDAttributes/count").is_ok());
        // ColorMode must NOT also appear in the default group.
        assert!(
            h5.dataset("entry/instrument/NDAttributes/ColorMode")
                .is_err()
        );
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_ndattribute_element_attr_open_close_values() {
        // C `storeOnOpenCloseAttribute` (NDFileHDF5.cpp:553-632) writes a
        // layout `<attribute source="ndattribute">` as an HDF5 attribute on its
        // group/dataset: `when="OnFileOpen"` (and the default `OnFrame`) take
        // the first-frame value, `when="OnFileClose"` the last. Cover a group
        // string attribute and a dataset numeric attribute, both phases.
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_layout_elem_attr_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="instrument">
                  <group name="detector">
                    <attribute name="FirstColorMode" source="ndattribute" ndattribute="ColorMode" when="OnFileOpen"/>
                    <attribute name="LastColorMode" source="ndattribute" ndattribute="ColorMode" when="OnFileClose"/>
                    <attribute name="DefaultColorMode" source="ndattribute" ndattribute="ColorMode"/>
                    <dataset name="data" source="detector" det_default="true">
                      <attribute name="GainAtOpen" source="ndattribute" ndattribute="Gain" when="OnFileOpen"/>
                      <attribute name="GainAtClose" source="ndattribute" ndattribute="Gain" when="OnFileClose"/>
                    </dataset>
                  </group>
                  <group name="NDAttributes" ndattr_default="true"/>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_elem_attr");
        let mut writer = Hdf5Writer::new();
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let mk = |mode: &str, gain: i32| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "ColorMode",
                "",
                NDAttrSource::Driver,
                NDAttrValue::String(mode.to_string()),
            ));
            arr.attributes.add(NDAttribute::new_static(
                "Gain",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Int32(gain),
            ));
            arr
        };

        let a0 = mk("Mono", 10);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk("RGB1", 20)).unwrap();
        writer.write_file(&mk("Bayer", 30)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();

        // Group string element-attributes on /entry/instrument/detector.
        let mut grp = h5.root_group();
        for seg in ["entry", "instrument", "detector"] {
            grp = grp.group(seg).unwrap();
        }
        // OnFileOpen and the default (OnFrame, open wins) take the first value.
        assert_eq!(grp.attr_string("FirstColorMode").unwrap(), "Mono");
        assert_eq!(grp.attr_string("DefaultColorMode").unwrap(), "Mono");
        // OnFileClose takes the last value.
        assert_eq!(grp.attr_string("LastColorMode").unwrap(), "Bayer");

        // Dataset numeric element-attributes on the detector dataset.
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(
            ds.attr("GainAtOpen")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            10
        );
        assert_eq!(
            ds.attr("GainAtClose")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            30
        );

        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_ndattr_element_attr_on_lazily_created_dataset() {
        // C storeOnOpenCloseAttribute attaches a layout `<attribute
        // source="ndattribute">` to ANY element the layout names, not just the
        // detector dataset. A lazily-created NDAttribute dataset has no retained
        // create-time handle, so its element-attrs are attached by reopening the
        // dataset by name in write mode (rust-hdf5 0.2.22 dataset_writer). Here
        // a `GainTrace` NDAttribute dataset (sourced from Gain) carries two
        // element-attrs sourced from ColorMode (open=first, close=last value).
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_layout_lazy_ds_attr_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="instrument">
                  <group name="detector">
                    <dataset name="data" source="detector" det_default="true"/>
                    <dataset name="GainTrace" source="ndattribute" ndattribute="Gain">
                      <attribute name="ModeAtOpen" source="ndattribute" ndattribute="ColorMode" when="OnFileOpen"/>
                      <attribute name="ModeAtClose" source="ndattribute" ndattribute="ColorMode" when="OnFileClose"/>
                    </dataset>
                  </group>
                  <group name="NDAttributes" ndattr_default="true"/>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_lazy_ds_attr");
        let mut writer = Hdf5Writer::new();
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let mk = |mode: &str, gain: i32| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "ColorMode",
                "",
                NDAttrSource::Driver,
                NDAttrValue::String(mode.to_string()),
            ));
            arr.attributes.add(NDAttribute::new_static(
                "Gain",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Int32(gain),
            ));
            arr
        };

        let a0 = mk("Mono", 10);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk("RGB1", 20)).unwrap();
        writer.write_file(&mk("Bayer", 30)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        // The element-attrs are attached to the lazily-created NDAttribute
        // dataset (not a detector dataset, so reopened by name at close).
        let ds = h5.dataset("entry/instrument/detector/GainTrace").unwrap();
        assert_eq!(
            ds.attr("ModeAtOpen").unwrap().read_string().unwrap(),
            "Mono"
        );
        assert_eq!(
            ds.attr("ModeAtClose").unwrap().read_string().unwrap(),
            "Bayer"
        );
        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_swmr_ndattr_element_attr_on_streaming_dataset() {
        // In SWMR mode an open-time `<attribute source="ndattribute">` on the
        // streaming dataset is attached before the SWMR lock via
        // set_dataset_attr_* (rust-hdf5 0.2.22 addresses SWMR dataset attributes
        // by index). OnFileClose stays impossible in SWMR (HDF5 forbids
        // post-lock attribute creation; C's close-time H5Acreate2 fails too).
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_layout_swmr_ds_attr_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="instrument">
                  <group name="detector">
                    <dataset name="data" source="detector" det_default="true">
                      <attribute name="ModeAtOpen" source="ndattribute" ndattribute="ColorMode" when="OnFileOpen"/>
                    </dataset>
                  </group>
                  <group name="NDAttributes" ndattr_default="true"/>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_swmr_ds_attr");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let mk = |mode: &str| {
            let mut arr = NDArray::new(
                vec![NDDimension::new(4), NDDimension::new(4)],
                NDDataType::UInt16,
            );
            arr.attributes.add(NDAttribute::new_static(
                "ColorMode",
                "",
                NDAttrSource::Driver,
                NDAttrValue::String(mode.to_string()),
            ));
            arr
        };

        let a0 = mk("Mono");
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk("RGB1")).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(
            ds.attr("ModeAtOpen").unwrap().read_string().unwrap(),
            "Mono"
        );
        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_swmr_ndarray_default_attrs_on_streaming_dataset() {
        // C writeDefaultDatasetAttributes (NDFileHDF5.cpp:3695-3719) attaches
        // NDArrayNumDims plus the per-dimension NDArrayDimOffset/Binning/Reverse
        // to every detector dataset. In SWMR mode they are written on the single
        // streaming dataset before start_swmr() locks the file, addressed by
        // dataset index (rust-hdf5 0.2.22). writeH5attrInt32
        // (NDFileHDF5.cpp:1142-1191): 1-D array => scalar int32, multi-dim => a
        // 1-D int32 array of length ndims, native dim order.

        // 1-D streaming array: all four attributes present and scalar.
        let path = temp_path("hdf5_swmr_dimattr_1d");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        let mut d = NDDimension::new(8);
        d.offset = 3;
        d.binning = 2;
        d.reverse = true;
        let arr = NDArray::new(vec![d], NDDataType::UInt16);
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();
        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        let geti = |n: &str| -> i32 { ds.attr(n).unwrap().read_numeric().unwrap() };
        assert_eq!(geti("NDArrayNumDims"), 1);
        assert_eq!(geti("NDArrayDimOffset"), 3);
        assert_eq!(geti("NDArrayDimBinning"), 2);
        assert_eq!(geti("NDArrayDimReverse"), 1);
        std::fs::remove_file(&path).ok();

        // 2-D streaming array: NDArrayNumDims=2; the Dim* attributes are 1-D
        // int32 arrays of length 2 in native dim order (dims[0] then dims[1]).
        let path = temp_path("hdf5_swmr_dimattr_2d");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        let mut d0 = NDDimension::new(4);
        d0.offset = 3;
        d0.binning = 2;
        d0.reverse = true;
        let mut d1 = NDDimension::new(8);
        d1.offset = 5;
        d1.binning = 4;
        d1.reverse = false;
        let arr = NDArray::new(vec![d0, d1], NDDataType::UInt16);
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();
        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        let numdims: i32 = ds.attr("NDArrayNumDims").unwrap().read_numeric().unwrap();
        assert_eq!(numdims, 2);
        let read_i32_arr = |n: &str| -> Vec<i32> {
            let raw = ds.attr(n).unwrap().read_raw().unwrap();
            assert_eq!(raw.len(), 2 * 4, "{n} must be a 2-element int32 array");
            raw.chunks_exact(4)
                .map(|b| i32::from_le_bytes([b[0], b[1], b[2], b[3]]))
                .collect()
        };
        assert_eq!(read_i32_arr("NDArrayDimOffset"), vec![3, 5]);
        assert_eq!(read_i32_arr("NDArrayDimBinning"), vec![2, 4]);
        assert_eq!(read_i32_arr("NDArrayDimReverse"), vec![1, 0]);
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_ndattr_descriptor_attributes_written() {
        // C attaches NDAttrName/NDAttrDescription/NDAttrSourceType/NDAttrSource
        // string HDF5 attributes (non-empty only) to every NDAttribute dataset
        // (NDFileHDF5.cpp:2715, 2817-2822).
        let path = temp_path("hdf5_attr_desc");
        let mut writer = Hdf5Writer::new();

        let mk = |v: f64| {
            let mut arr = NDArray::new(vec![NDDimension::new(4)], NDDataType::UInt8);
            arr.attributes.add(NDAttribute::new_static(
                "AcquireTime",
                "exposure time",
                NDAttrSource::EpicsPV("13SIM1:cam1:AcquireTime_RBV".to_string()),
                NDAttrValue::Float64(v),
            ));
            arr
        };

        let a0 = mk(0.1);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk(0.2)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5
            .dataset("entry/instrument/NDAttributes/AcquireTime")
            .unwrap();
        let read = |n: &str| ds.attr(n).unwrap().read_string().unwrap();
        assert_eq!(read("NDAttrName"), "AcquireTime");
        assert_eq!(read("NDAttrDescription"), "exposure time");
        assert_eq!(read("NDAttrSourceType"), "NDAttrSourceEPICSPV");
        assert_eq!(read("NDAttrSource"), "13SIM1:cam1:AcquireTime_RBV");
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_detector_dataset_ndarray_dim_attributes() {
        // C writeDefaultDatasetAttributes (NDFileHDF5.cpp:3695-3719) attaches
        // NDArrayNumDims (scalar) and the per-dimension NDArrayDimOffset/Binning/
        // Reverse. writeH5attrInt32 (NDFileHDF5.cpp:1142-1191) emits a single
        // dimension as a scalar int32 and multiple dimensions as a 1-D int32
        // array of length ndims, in native dim order.

        // 1-D array: all four attributes present and scalar, in native order.
        let path = temp_path("hdf5_dimattr_1d");
        let mut writer = Hdf5Writer::new();
        let mut d = NDDimension::new(8);
        d.offset = 3;
        d.binning = 2;
        d.reverse = true;
        let arr = NDArray::new(vec![d], NDDataType::UInt16);
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();
        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        let geti = |n: &str| -> i32 { ds.attr(n).unwrap().read_numeric().unwrap() };
        assert_eq!(geti("NDArrayNumDims"), 1);
        assert_eq!(geti("NDArrayDimOffset"), 3);
        assert_eq!(geti("NDArrayDimBinning"), 2);
        assert_eq!(geti("NDArrayDimReverse"), 1);
        std::fs::remove_file(&path).ok();

        // 2-D array: NDArrayNumDims present (=2); the Dim* attributes are 1-D
        // int32 arrays of length 2, in native dim order. Distinct per-dim values
        // verify the order is dims[0] then dims[1] (not reversed HDF5 axes).
        let path = temp_path("hdf5_dimattr_2d");
        let mut writer = Hdf5Writer::new();
        let mut d0 = NDDimension::new(4);
        d0.offset = 3;
        d0.binning = 2;
        d0.reverse = true;
        let mut d1 = NDDimension::new(8);
        d1.offset = 5;
        d1.binning = 4;
        d1.reverse = false;
        let arr = NDArray::new(vec![d0, d1], NDDataType::UInt16);
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();
        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        let numdims: i32 = ds.attr("NDArrayNumDims").unwrap().read_numeric().unwrap();
        assert_eq!(numdims, 2);
        // The array attribute is a 1-D simple dataspace of length ndims: read the
        // raw int32 LE bytes back (4 bytes/elem => 8 bytes proves an array, not a
        // scalar). h5py/libhdf5 reads the same shape and values.
        let read_i32_arr = |n: &str| -> Vec<i32> {
            let raw = ds.attr(n).unwrap().read_raw().unwrap();
            assert_eq!(raw.len(), 2 * 4, "{n} must be a 2-element int32 array");
            raw.chunks_exact(4)
                .map(|b| i32::from_le_bytes([b[0], b[1], b[2], b[3]]))
                .collect()
        };
        assert_eq!(read_i32_arr("NDArrayDimOffset"), vec![3, 5]);
        assert_eq!(read_i32_arr("NDArrayDimBinning"), vec![2, 4]);
        assert_eq!(read_i32_arr("NDArrayDimReverse"), vec![1, 0]);
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_fill_value_recorded_on_dataset() {
        // The configured HDF5_fillValue reaches the DCPL via rust-hdf5 0.2.15's
        // `DatasetBuilder::fill_value`; it is also mirrored as a dataset
        // attribute for tooling. Verify both the attribute and that an
        // unwritten region of a fill-valued dataset reads back as `fill`.
        let path = temp_path("hdf5_fill");
        let mut writer = Hdf5Writer::new();
        writer.set_fill_value(7.5);

        let arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        let fv: f64 = ds.attr("HDF5_fillValue").unwrap().read_numeric().unwrap();
        assert_eq!(fv, 7.5);
        std::fs::remove_file(&path).ok();

        // Direct DCPL check: a fixed-shape dataset created with fill_value and
        // never written reads back the fill value, not zero.
        let path2 = temp_path("hdf5_fill_dcpl");
        {
            let f = H5File::create(&path2).unwrap();
            let _ = f
                .new_dataset::<i32>()
                .shape(&[8][..])
                .fill_value(42i32)
                .create("unwritten")
                .unwrap();
        }
        let h5b = H5File::open(&path2).unwrap();
        let vals: Vec<i32> = h5b.dataset("unwritten").unwrap().read_raw().unwrap();
        assert_eq!(vals, vec![42i32; 8]);
        std::fs::remove_file(&path2).ok();
    }

    #[test]
    fn test_performance_dataset() {
        let path = temp_path("hdf5_perf");
        let mut writer = Hdf5Writer::new();
        writer.set_store_performance(true);

        let arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ts = h5
            .dataset("entry/instrument/performance/timestamp")
            .unwrap();
        assert_eq!(ts.shape(), vec![2, 5]);
        let vals: Vec<f64> = ts.read_raw().unwrap();
        assert_eq!(vals.len(), 10);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_performance_dataset_chunk_matches_capture_target() {
        // C `writePerformanceDataset` chunks the timestamp dataset `[chunking,5]`
        // (NDFileHDF5.cpp:2645-2647) where `chunking` is the same
        // `calculateAttributeChunking` value (the capture target in non-Single
        // mode), not one row per chunk. Capture target 8, 2 frames written: the
        // chunk's leading dim must be 8, extent 2; doubles still round-trip.
        let path = temp_path("hdf5_perf_chunk");
        let mut writer = Hdf5Writer::new();
        writer.set_store_performance(true);
        writer.set_num_capture(8);

        let arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ts = h5
            .dataset("entry/instrument/performance/timestamp")
            .unwrap();
        assert_eq!(ts.shape(), vec![2, 5]);
        assert_eq!(ts.chunk_dims(), Some(vec![8, 5]));
        let vals: Vec<f64> = ts.read_raw().unwrap();
        assert_eq!(vals.len(), 10);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_roundtrip_all_types() {
        macro_rules! roundtrip {
            ($name:expr, $dt:expr, $variant:ident, $ty:ty, $vals:expr) => {{
                let path = temp_path($name);
                let mut writer = Hdf5Writer::new();
                let mut arr = NDArray::new(vec![NDDimension::new(4)], $dt);
                if let NDDataBuffer::$variant(ref mut v) = arr.data {
                    let src: Vec<$ty> = $vals;
                    v.copy_from_slice(&src);
                }
                writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
                writer.write_file(&arr).unwrap();
                writer.close_file().unwrap();

                let mut reader = Hdf5Writer::new();
                reader.current_path = Some(path.clone());
                let r = reader.read_file().unwrap();
                assert_eq!(r.data.data_type(), $dt, "type for {}", $name);
                if let NDDataBuffer::$variant(ref v) = r.data {
                    let src: Vec<$ty> = $vals;
                    assert_eq!(v, &src, "values for {}", $name);
                } else {
                    panic!("wrong buffer variant for {}", $name);
                }
                std::fs::remove_file(&path).ok();
            }};
        }

        roundtrip!("rt_i8", NDDataType::Int8, I8, i8, vec![-1, 0, 1, 127]);
        roundtrip!("rt_u8", NDDataType::UInt8, U8, u8, vec![0, 1, 200, 255]);
        roundtrip!(
            "rt_i16",
            NDDataType::Int16,
            I16,
            i16,
            vec![-32768, -1, 1, 32767]
        );
        roundtrip!(
            "rt_u16",
            NDDataType::UInt16,
            U16,
            u16,
            vec![0, 1, 40000, 65535]
        );
        roundtrip!(
            "rt_i32",
            NDDataType::Int32,
            I32,
            i32,
            vec![i32::MIN, -1, 1, i32::MAX]
        );
        roundtrip!(
            "rt_u32",
            NDDataType::UInt32,
            U32,
            u32,
            vec![0, 1, 3_000_000_000, u32::MAX]
        );
        roundtrip!(
            "rt_i64",
            NDDataType::Int64,
            I64,
            i64,
            vec![i64::MIN, -1, 1, i64::MAX]
        );
        roundtrip!(
            "rt_u64",
            NDDataType::UInt64,
            U64,
            u64,
            vec![0, 1, 9_000_000_000, u64::MAX]
        );
        roundtrip!(
            "rt_f32",
            NDDataType::Float32,
            F32,
            f32,
            vec![-1.5, 0.0, 2.25, 3.75]
        );
        roundtrip!(
            "rt_f64",
            NDDataType::Float64,
            F64,
            f64,
            vec![-1.5, 0.0, 2.25, 3.75]
        );
    }

    #[test]
    fn test_deflate_compressed_write() {
        let path = temp_path("hdf5_deflate");
        let mut writer = Hdf5Writer::new();
        writer.set_compression_type(COMPRESS_ZLIB);
        writer.set_z_compress_level(6);

        let mut arr = NDArray::new(
            vec![NDDimension::new(64), NDDimension::new(64)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for i in 0..v.len() {
                v[i] = (i % 256) as u16;
            }
        }

        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        // Assert compression by COMPARING against an uncompressed baseline of
        // the same frame, not against a fixed byte threshold. Since rust-hdf5
        // 0.2.25 the NeXus group string attributes (NX_class, …) are stored as
        // variable-length UTF-8 in the global heap, so the file's constant
        // metadata overhead now dwarfs this 8 KiB frame — an absolute
        // "< 8192" check no longer reflects whether the data chunk was
        // compressed (h5dump confirms the chunk itself is ~412 B, ~20:1).
        // Writing the identical array uncompressed cancels that shared
        // overhead, isolating the deflate saving on the data chunk.
        let raw_path = temp_path("hdf5_deflate_raw");
        {
            let mut raw_writer = Hdf5Writer::new();
            raw_writer.set_compression_type(COMPRESS_NONE);
            raw_writer
                .open_file(&raw_path, NDFileMode::Single, &arr)
                .unwrap();
            raw_writer.write_file(&arr).unwrap();
            raw_writer.close_file().unwrap();
        }
        let compressed_size = std::fs::metadata(&path).unwrap().len();
        let raw_size = std::fs::metadata(&raw_path).unwrap().len();
        assert!(
            compressed_size + 4096 < raw_size,
            "deflate must shrink the data chunk: compressed={compressed_size} raw={raw_size}"
        );

        let h5file = H5File::open(&path).unwrap();
        let ds = h5file.dataset("entry/instrument/detector/data").unwrap();
        let data: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 64 * 64);
        assert_eq!(data[0], 0);
        assert_eq!(data[255], 255);
        assert_eq!(data[256], 0);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_lz4_compressed_write() {
        let path = temp_path("hdf5_lz4");
        let mut writer = Hdf5Writer::new();
        writer.set_compression_type(COMPRESS_LZ4);

        let mut arr = NDArray::new(
            vec![NDDimension::new(32), NDDimension::new(32)],
            NDDataType::UInt8,
        );
        if let NDDataBuffer::U8(ref mut v) = arr.data {
            for i in 0..v.len() {
                v[i] = (i % 4) as u8;
            }
        }

        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5file = H5File::open(&path).unwrap();
        let ds = h5file.dataset("entry/instrument/detector/data").unwrap();
        let data: Vec<u8> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 32 * 32);
        assert_eq!(data[0], 0);
        assert_eq!(data[3], 3);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_bitshuffle_compressed_write() {
        let path = temp_path("hdf5_bshuf");
        let mut writer = Hdf5Writer::new();
        writer.set_compression_type(COMPRESS_BSHUF);

        let mut arr = NDArray::new(
            vec![NDDimension::new(64), NDDimension::new(64)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for i in 0..v.len() {
                v[i] = (i % 8) as u16;
            }
        }

        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5file = H5File::open(&path).unwrap();
        let ds = h5file.dataset("entry/instrument/detector/data").unwrap();
        let data: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 64 * 64);
        assert_eq!(data[0], 0);
        assert_eq!(data[9], 1);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_szip_uses_nearest_neighbor_mask() {
        // C uses H5_SZIP_NN_OPTION_MASK (32); the SZIP filter's cd_values[0]
        // must declare NN coding, and the data must round-trip.
        let mut writer = Hdf5Writer::new();
        writer.set_compression_type(COMPRESS_SZIP);
        let pipeline = writer.build_pipeline(1).expect("szip pipeline");
        assert_eq!(pipeline.filters[0].cd_values[0], SZIP_NN_OPTION_MASK);

        let path = temp_path("hdf5_szip_nn");
        let mut arr = NDArray::new(
            vec![NDDimension::new(64), NDDimension::new(64)],
            NDDataType::UInt8,
        );
        if let NDDataBuffer::U8(ref mut v) = arr.data {
            for (i, e) in v.iter_mut().enumerate() {
                *e = (i % 13) as u8;
            }
        }
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5file = H5File::open(&path).unwrap();
        let ds = h5file.dataset("entry/instrument/detector/data").unwrap();
        let data: Vec<u8> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 64 * 64);
        assert_eq!(data[20], (20 % 13) as u8);
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_blosc_default_shuffle_is_byte_shuffle() {
        // C default bloscShuffleType=1 (byte shuffle), NDFileHDF5.cpp:2344;
        // BLOSC cd_values[5] carries the shuffle type.
        let mut writer = Hdf5Writer::new();
        writer.set_compression_type(COMPRESS_BLOSC);
        let pipeline = writer.build_pipeline(2).expect("blosc pipeline");
        assert_eq!(pipeline.filters[0].cd_values[5], 1);
    }

    #[test]
    fn test_nbit_packs_to_reduced_precision_datatype() {
        // C's N-bit codec (NDFileHDF5.cpp:3355-3357) narrows the dataset
        // datatype (H5Tset_precision/H5Tset_offset) and registers a
        // parameterless H5Pset_nbit filter. The standard write path reproduces
        // that with rust-hdf5 0.2.22: `DatasetBuilder::datatype` stores a
        // reduced-precision `FixedPoint` and `FilterPipeline::nbit` packs to it,
        // so the file is byte-readable by h5py/libhdf5. `build_pipeline` still
        // returns None for N-bit because the SWMR streaming builder cannot
        // override the datatype — packing is applied out of band by the standard
        // path only.
        let mut writer = Hdf5Writer::new();
        writer.set_compression_type(COMPRESS_NBIT);
        writer.set_nbit_precision(10);
        writer.set_nbit_offset(0);
        assert!(
            writer.build_pipeline(2).is_none(),
            "N-bit packing is applied via a datatype override, not build_pipeline"
        );

        let path = temp_path("hdf5_nbit_packed");
        let mut arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for (i, x) in v.iter_mut().enumerate() {
                *x = (i as u16 * 7) & 0x3FF; // within the 10-bit range
            }
            v[0] = 0xFFFF; // above 10 bits: must pack to the low 10 bits (0x3FF)
        }
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5file = H5File::open(&path).unwrap();
        let ds = h5file.dataset("entry/instrument/detector/data").unwrap();

        // The on-disk datatype carries the reduced precision (C-observable
        // narrower datatype) within the unchanged 2-byte footprint.
        match ds.datatype().unwrap() {
            DatatypeMessage::FixedPoint {
                size,
                bit_precision,
                bit_offset,
                signed,
                ..
            } => {
                assert_eq!(size, 2, "byte footprint unchanged");
                assert_eq!(bit_precision, 10, "precision narrowed to 10 bits");
                assert_eq!(bit_offset, 0);
                assert!(!signed, "u16 is unsigned");
            }
            other => panic!("expected reduced-precision FixedPoint, got {other:?}"),
        }

        let data: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 8 * 8);
        // Real bit-packing: the over-range first element truncates to 10 bits.
        assert_eq!(data[0], 0x3FF, "0xFFFF must pack to the low 10 bits");
        for i in 1..data.len() {
            assert_eq!(
                data[i],
                (i as u16 * 7) & 0x3FF,
                "in-range value {i} round-trips"
            );
        }
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_chunk_geometry_recorded() {
        // Requested row/col chunk geometry is recorded as dataset attributes
        // (the on-disk chunk is one frame per chunk — crate limitation).
        let path = temp_path("hdf5_chunkgeom");
        let mut writer = Hdf5Writer::new();
        writer.set_chunk_size_auto(false);
        writer.set_n_row_chunks(4);
        writer.set_n_col_chunks(2);
        writer.set_n_frames_chunks(3);

        let mut arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for i in 0..v.len() {
                v[i] = i as u16;
            }
        }

        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![2, 8, 8]);
        // Data still round-trips correctly through the per-frame chunks.
        let data: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 2 * 64);
        for i in 0..64usize {
            assert_eq!(data[i], i as u16, "frame0 element {}", i);
            assert_eq!(data[64 + i], i as u16, "frame1 element {}", i);
        }
        // Requested geometry preserved as attributes.
        assert_eq!(
            ds.attr("HDF5_nRowChunks")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            4
        );
        assert_eq!(
            ds.attr("HDF5_nColChunks")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            2
        );
        assert_eq!(
            ds.attr("HDF5_nFramesChunks")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            3
        );

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_extra_dimensions_layout() {
        // HDF5_nExtraDims=2 builds 2+1 fixed leading axes. With the param sizes
        // extraDimSizeN(=eds[0])=2, extraDimSizeX(=eds[1])=3, extraDimSizeY(=
        // eds[2])=4 the dataspace is rank-5 `{Y, X, N, frameY, frameX}` =
        // [4,3,2,4,4] — C `NDFileHDF5::configureDims` order
        // (NDFileHDF5.cpp:3121-3230; docs/ADCore/NDFileHDF5.rst:379-380). The
        // frame data is row-major identical to the collapsed [24,4,4] form
        // (the innermost leading axis "N" varies fastest), so frame `f` lands
        // at flat leading index `f`.
        let path = temp_path("hdf5_extradims");
        let mut writer = Hdf5Writer::new();
        writer.set_n_extra_dims(2);
        writer.set_extra_dim_size(0, 2); // N: frames per point
        writer.set_extra_dim_size(1, 3); // X
        writer.set_extra_dim_size(2, 4); // Y
        writer.set_extra_dim_name(0, "n");
        writer.set_extra_dim_name(1, "scanX");
        writer.set_extra_dim_name(2, "scanY");

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        for f in 0..24u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for x in v.iter_mut() {
                    *x = f;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        // Rank-5 multi-extra-dimension dataspace: [Y, X, N, frameY, frameX].
        assert_eq!(ds.shape(), vec![4, 3, 2, 4, 4]);
        let data: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 24 * 16);
        for f in 0..24usize {
            for i in 0..16usize {
                assert_eq!(data[f * 16 + i], f as u16, "frame {} elem {}", f, i);
            }
        }
        // Extra-dim sizes/names still recorded as recovery attributes.
        assert_eq!(
            ds.attr("HDF5_nExtraDims")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            2
        );
        assert_eq!(
            ds.attr("HDF5_extraDimSize0")
                .unwrap()
                .read_numeric::<i32>()
                .unwrap(),
            2
        );
        assert_eq!(
            ds.attr("HDF5_extraDimName0")
                .unwrap()
                .read_string()
                .unwrap(),
            "n"
        );

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_extra_dimensions_one_virtual() {
        // HDF5_nExtraDims=1 builds 1+1 fixed leading axes:
        // {X, N, frameY, frameX}. With eds[0]=N=2, eds[1]=X=3 the dataspace is
        // rank-4 [3,2,Y,X] (NDFileHDF5.rst:377-378). 6 frames, value==frame.
        let path = temp_path("hdf5_extradims_1");
        let mut writer = Hdf5Writer::new();
        writer.set_n_extra_dims(1);
        writer.set_extra_dim_size(0, 2); // N: frames per point
        writer.set_extra_dim_size(1, 3); // X

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        for f in 0..6u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for x in v.iter_mut() {
                    *x = f;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![3, 2, 4, 4]);
        let data: Vec<u16> = ds.read_raw().unwrap();
        assert_eq!(data.len(), 6 * 16);
        for f in 0..6usize {
            for i in 0..16usize {
                assert_eq!(data[f * 16 + i], f as u16, "frame {} elem {}", f, i);
            }
        }

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_extra_dimensions_grid_layout() {
        // SWMR mirror of `test_extra_dimensions_layout`: HDF5_nExtraDims=2 with
        // eds[0]=2,eds[1]=3,eds[2]=4 must build the same rank-5 fixed grid
        // [Y,X,N,frameY,frameX] = [4,3,2,4,4] as the standard path (C
        // configureDims), not the old collapsed single leading axis. Each frame
        // is placed at its odometer chunk position via write_chunk_at; frame `f`
        // lands at flat leading index `f` (innermost "N" varies fastest).
        let path = temp_path("hdf5_swmr_extradims");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        writer.set_n_extra_dims(2);
        writer.set_extra_dim_size(0, 2); // N: frames per point
        writer.set_extra_dim_size(1, 3); // X
        writer.set_extra_dim_size(2, 4); // Y

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        for f in 0..24u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for x in v.iter_mut() {
                    *x = f;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let mut reader = rust_hdf5::swmr::SwmrFileReader::open(&path).unwrap();
        assert_eq!(
            reader
                .dataset_shape("entry/instrument/detector/data")
                .unwrap(),
            vec![4, 3, 2, 4, 4]
        );
        let data: Vec<u16> = reader
            .read_dataset("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(data.len(), 24 * 16);
        for f in 0..24usize {
            for i in 0..16usize {
                assert_eq!(data[f * 16 + i], f as u16, "frame {} elem {}", f, i);
            }
        }
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_extra_dimensions_grid_subtiled() {
        // SWMR grid with frame sub-tiling: HDF5_nExtraDims=1 (rank-4 [3,2,4,4])
        // and HDF5_nRowChunks=2 splits each 4-row frame into two row tiles, so
        // the grid write path emits two write_chunk_at tiles per frame. The
        // reassembled pixels must still be row-major-correct. 6 frames fill the
        // 3x2 grid exactly; value==frame.
        let path = temp_path("hdf5_swmr_extradims_tiled");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        writer.set_n_extra_dims(1);
        writer.set_extra_dim_size(0, 2); // N
        writer.set_extra_dim_size(1, 3); // X
        writer.set_n_row_chunks(2); // split the 4-row frame into 2 tiles

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        for f in 0..6u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                // Distinct per-element values prove tile reassembly, not just a
                // constant: value = frame*100 + (row*4 + col).
                for (i, x) in v.iter_mut().enumerate() {
                    *x = f * 100 + i as u16;
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let mut reader = rust_hdf5::swmr::SwmrFileReader::open(&path).unwrap();
        assert_eq!(
            reader
                .dataset_shape("entry/instrument/detector/data")
                .unwrap(),
            vec![3, 2, 4, 4]
        );
        let data: Vec<u16> = reader
            .read_dataset("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(data.len(), 6 * 16);
        for f in 0..6usize {
            for i in 0..16usize {
                assert_eq!(
                    data[f * 16 + i],
                    f as u16 * 100 + i as u16,
                    "frame {} elem {}",
                    f,
                    i
                );
            }
        }
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_extra_dimensions_grid_partial_fill() {
        // SWMR grid written with fewer frames than the grid capacity: the fixed
        // extent stays [3,2,4,4] and the unwritten odometer positions read back
        // as the fill value (0). Boundary: partial scan, capacity 6, write 4.
        let path = temp_path("hdf5_swmr_extradims_partial");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        writer.set_n_extra_dims(1);
        writer.set_extra_dim_size(0, 2); // N
        writer.set_extra_dim_size(1, 3); // X

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        for f in 0..4u16 {
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for x in v.iter_mut() {
                    *x = f + 1; // non-zero so fill (0) is distinguishable
                }
            }
            if f == 0 {
                writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
            }
            writer.write_file(&arr).unwrap();
        }
        writer.close_file().unwrap();

        let mut reader = rust_hdf5::swmr::SwmrFileReader::open(&path).unwrap();
        assert_eq!(
            reader
                .dataset_shape("entry/instrument/detector/data")
                .unwrap(),
            vec![3, 2, 4, 4]
        );
        let data: Vec<u16> = reader
            .read_dataset("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(data.len(), 6 * 16);
        for f in 0..6usize {
            let expected = if f < 4 { f as u16 + 1 } else { 0 };
            for i in 0..16usize {
                assert_eq!(data[f * 16 + i], expected, "frame {} elem {}", f, i);
            }
        }
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_streaming() {
        let path = temp_path("hdf5_swmr");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        writer.set_flush_nth_frame(2);

        let arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::Float32,
        );

        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap(); // should trigger flush
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        assert_eq!(writer.frame_count(), 3);

        // Read back via SwmrFileReader
        let mut reader = rust_hdf5::swmr::SwmrFileReader::open(&path).unwrap();
        let shape = reader
            .dataset_shape("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(shape[0], 3); // 3 frames
        assert_eq!(shape[1], 8);
        assert_eq!(shape[2], 8);

        let data: Vec<f32> = reader
            .read_dataset("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(data.len(), 3 * 8 * 8);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_compression_is_applied() {
        // rust-hdf5 0.2.15 exposes a filtered SWMR dataset constructor, so
        // SWMR + compression produces a genuinely compressed file — the
        // compression is NOT dropped, and the data round-trips.
        let path = temp_path("hdf5_swmr_comp");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        writer.set_compression_type(COMPRESS_ZLIB);

        let arr = NDArray::new(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataType::UInt16,
        );
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        assert!(
            !writer.swmr_compression_dropped(),
            "SWMR+ZLIB must apply compression, not drop it"
        );
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        // The compressed SWMR dataset round-trips.
        let mut reader = rust_hdf5::swmr::SwmrFileReader::open(&path).unwrap();
        let shape = reader
            .dataset_shape("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(shape, vec![2, 8, 8]);
        let data: Vec<u16> = reader
            .read_dataset("entry/instrument/detector/data")
            .unwrap();
        assert_eq!(data.len(), 2 * 8 * 8);

        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_layout_xml_param() {
        // Valid and invalid layout XML drive layout_valid / layout_error.
        let mut writer = Hdf5Writer::new();
        let dir = std::env::temp_dir();
        let good = dir.join(format!("adcore_layout_good_{}.xml", std::process::id()));
        std::fs::write(
            &good,
            r#"<hdf5_layout><group name="entry"><dataset name="data" source="detector" det_default="true"/></group></hdf5_layout>"#,
        )
        .unwrap();
        assert!(writer.set_layout_filename(good.to_str().unwrap()));
        assert!(writer.layout_valid);
        assert!(writer.layout_error.is_empty());

        let bad = dir.join(format!("adcore_layout_bad_{}.xml", std::process::id()));
        std::fs::write(&bad, r#"<not_a_layout/>"#).unwrap();
        assert!(!writer.set_layout_filename(bad.to_str().unwrap()));
        assert!(!writer.layout_valid);
        assert!(!writer.layout_error.is_empty());

        std::fs::remove_file(&good).ok();
        std::fs::remove_file(&bad).ok();
    }

    #[test]
    fn test_layout_xml_places_dataset_in_nested_tree() {
        // A valid layout XML must place the image dataset at the layout's
        // det_default path (C ADCore /entry/instrument/detector/data),
        // NDAttributes under the ndattr_default group, and the performance
        // dataset under the group holding the `timestamp` dataset — NOT flat
        // at the file root.
        let dir = std::env::temp_dir();
        let layout = dir.join(format!("adcore_layout_nested_{}.xml", std::process::id()));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="instrument">
                  <group name="detector">
                    <dataset name="data" source="detector" det_default="true">
                      <attribute name="signal" source="constant" value="1" type="int"/>
                    </dataset>
                  </group>
                  <group name="NDAttributes" ndattr_default="true"/>
                  <group name="performance">
                    <dataset name="timestamp"/>
                  </group>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_layout_nested");
        let mut writer = Hdf5Writer::new();
        writer.set_store_performance(true);
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let mk = |fill: f64| {
            let mut arr = NDArray::new(
                vec![NDDimension::new(4), NDDimension::new(4)],
                NDDataType::UInt16,
            );
            arr.attributes.add(NDAttribute::new_static(
                "exposure",
                "",
                NDAttrSource::Driver,
                NDAttrValue::Float64(fill),
            ));
            arr
        };

        let a0 = mk(0.5);
        writer.open_file(&path, NDFileMode::Stream, &a0).unwrap();
        writer.write_file(&a0).unwrap();
        writer.write_file(&mk(0.75)).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let names = h5.dataset_names();
        // Image dataset at the nested layout path, NOT flat `data`.
        assert!(
            names.contains(&"entry/instrument/detector/data".to_string()),
            "image dataset must be at the nested layout path; got {:?}",
            names
        );
        assert!(
            !names.contains(&"data".to_string()),
            "must not also write a flat-root `data` dataset"
        );
        let img = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(img.shape(), vec![2, 4, 4]);
        // Layout constant attribute materialised.
        assert_eq!(
            img.attr("signal").unwrap().read_numeric::<i64>().unwrap(),
            1
        );
        // NDAttribute dataset under the ndattr_default group.
        assert!(
            names.contains(&"entry/instrument/NDAttributes/exposure".to_string()),
            "NDAttribute dataset must be under the layout ndattr group; got {:?}",
            names
        );
        // Performance dataset under the layout's performance group.
        assert!(
            names.contains(&"entry/instrument/performance/timestamp".to_string()),
            "performance dataset must be under the layout group; got {:?}",
            names
        );

        // Read-back resolves the nested dataset path.
        drop(h5);
        let mut reader = Hdf5Writer::new();
        assert!(reader.set_layout_filename(layout.to_str().unwrap()));
        reader.current_path = Some(path.clone());
        let read_arr = reader.read_file().unwrap();
        assert_eq!(read_arr.dims.len(), 3);

        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    /// Layout XML declaring two `<dataset source="detector">` nodes plus a
    /// `<global name="detector_data_destination">`. C `NDFileHDF5` creates
    /// every detector dataset up front (`detDataMap`) and routes each frame to
    /// the one named by the destination NDAttribute, defaulting to the
    /// `det_default` dataset for an absent or unknown value
    /// (NDFileHDF5.cpp:1449-1519). Frames must land in the right dataset and
    /// each dataset must extend to exactly the count it received.
    #[test]
    fn test_detector_data_destination_routes_by_attribute() {
        let dir = std::env::temp_dir();
        let layout = dir.join(format!("adcore_layout_multidet_{}.xml", std::process::id()));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <global name="detector_data_destination" ndattribute="dest"/>
              <group name="entry">
                <dataset name="data1" source="detector" det_default="true"/>
                <dataset name="data2" source="detector"/>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_multidet_route");
        let mut writer = Hdf5Writer::new();
        // Focus the test on image routing; no attribute time-series datasets.
        writer.store_attributes = false;
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        // Each frame is a uniform 2x2 UInt16 whose value identifies it, with an
        // optional `dest` string attribute selecting the destination dataset.
        let mk = |val: u16, dest: Option<&str>| {
            let mut arr = NDArray::new(
                vec![NDDimension::new(2), NDDimension::new(2)],
                NDDataType::UInt16,
            );
            if let NDDataBuffer::U16(ref mut v) = arr.data {
                for p in v.iter_mut() {
                    *p = val;
                }
            }
            if let Some(d) = dest {
                arr.attributes.add(NDAttribute::new_static(
                    "dest",
                    "",
                    NDAttrSource::Driver,
                    NDAttrValue::String(d.to_string()),
                ));
            }
            arr
        };

        // f0: no dest        -> default data1
        // f1: /entry/data2   -> data2
        // f2: /entry/data2   -> data2
        // f3: /nonexistent   -> unknown, falls back to default data1
        // f4: /entry/data1   -> explicit default data1
        let f0 = mk(10, None);
        writer.open_file(&path, NDFileMode::Stream, &f0).unwrap();
        writer.write_file(&f0).unwrap();
        writer.write_file(&mk(11, Some("/entry/data2"))).unwrap();
        writer.write_file(&mk(12, Some("/entry/data2"))).unwrap();
        writer.write_file(&mk(13, Some("/nonexistent"))).unwrap();
        writer.write_file(&mk(14, Some("/entry/data1"))).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let names = h5.dataset_names();
        assert!(
            names.contains(&"entry/data1".to_string())
                && names.contains(&"entry/data2".to_string()),
            "both detector datasets must exist; got {:?}",
            names
        );

        // data1 received f0, f3, f4 (in write order); data2 received f1, f2.
        let d1 = h5.dataset("entry/data1").unwrap();
        assert_eq!(d1.shape(), vec![3, 2, 2], "default dataset extent");
        let v1: Vec<u16> = d1.read_raw().unwrap();
        assert_eq!(
            v1,
            vec![10, 10, 10, 10, 13, 13, 13, 13, 14, 14, 14, 14],
            "default dataset must hold the default-routed frames in order"
        );

        let d2 = h5.dataset("entry/data2").unwrap();
        assert_eq!(d2.shape(), vec![2, 2, 2], "routed dataset extent");
        let v2: Vec<u16> = d2.read_raw().unwrap();
        assert_eq!(
            v2,
            vec![11, 11, 11, 11, 12, 12, 12, 12],
            "routed dataset must hold only the frames addressed to it"
        );

        drop(h5);
        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    /// A present-but-non-string `detector_data_destination` attribute aborts the
    /// write, matching C `getValue(NDAttrString,…)` returning `ND_ERROR` →
    /// `asynError` (NDFileHDF5.cpp:1465-1471).
    #[test]
    fn test_detector_data_destination_non_string_attribute_errors() {
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_layout_multidet_err_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <global name="detector_data_destination" ndattribute="dest"/>
              <group name="entry">
                <dataset name="data1" source="detector" det_default="true"/>
                <dataset name="data2" source="detector"/>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_multidet_err");
        let mut writer = Hdf5Writer::new();
        writer.store_attributes = false;
        assert!(writer.set_layout_filename(layout.to_str().unwrap()));

        let mut arr = NDArray::new(
            vec![NDDimension::new(2), NDDimension::new(2)],
            NDDataType::UInt16,
        );
        arr.attributes.add(NDAttribute::new_static(
            "dest",
            "",
            NDAttrSource::Driver,
            NDAttrValue::Float64(2.0),
        ));

        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        assert!(
            writer.write_file(&arr).is_err(),
            "a non-string destination attribute must abort the write"
        );
        writer.close_file().ok();

        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_layout_hardlink_is_materialised() {
        // Regression for BUG 2: a `<hardlink>` declared in the layout XML must
        // produce a real HDF5 hard link in the written file. C ADCore
        // `NDFileHDF5::createHardLinks` walks the layout and calls
        // `H5Lcreate_hard`; without that, files written from a layout with a
        // `<hardlink>` silently lack the link.
        let dir = std::env::temp_dir();
        let layout = dir.join(format!("adcore_layout_hardlink_{}.xml", std::process::id()));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="data">
                  <dataset name="data" source="detector" det_default="true"/>
                  <hardlink name="data_alias" target="/entry/data/data"/>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_layout_hardlink");
        let mut writer = Hdf5Writer::new();
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let names = h5.dataset_names();
        // The primary dataset at its layout path.
        assert!(
            names.contains(&"entry/data/data".to_string()),
            "image dataset must exist at the layout path; got {:?}",
            names
        );
        // The hard link is an additional name resolving to the same object.
        assert!(
            names.contains(&"entry/data/data_alias".to_string()),
            "layout <hardlink> must be materialised as a hard link; got {:?}",
            names
        );
        // The link shares the target object: same shape, readable as a dataset.
        let alias = h5.dataset("entry/data/data_alias").unwrap();
        let orig = h5.dataset("entry/data/data").unwrap();
        assert_eq!(alias.shape(), orig.shape());

        drop(h5);
        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    /// D2: a close whose finalisation fails partway must still release the H5
    /// file id and its dataset ids. The broken `<hardlink>` stands in for the
    /// ENOSPC flush failure: it fails `build_layout_hardlinks`, one of the five
    /// fallible steps that used to sit ahead of `handle.take()`.
    #[test]
    fn close_file_releases_the_handle_when_finalisation_fails() {
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_layout_broken_hardlink_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="data">
                  <dataset name="data" source="detector" det_default="true"/>
                  <hardlink name="broken" target="/no_such_object"/>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_close_fail_releases_handle");
        let mut writer = Hdf5Writer::new();
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();

        assert!(
            writer.close_file().is_err(),
            "a hardlink to a missing object must fail the close"
        );
        assert!(
            writer.handle.is_none(),
            "a failed close must still release the H5 file id and its dataset ids"
        );
        assert!(writer.current_path.is_none());
        // The writer is reusable: a second close is a no-op rather than another
        // run at the same failing finalisation.
        writer.close_file().unwrap();

        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_swmr_layout_hardlink_is_materialised() {
        // A `<hardlink>` declared in the layout XML must also be materialised
        // for SWMR-mode files. C ADCore `NDFileHDF5.cpp:319`-`325` calls
        // `createHardLinks` before `startSWMR()`, so the link is committed by
        // `start_swmr()` and visible to SWMR readers for the whole streaming
        // window. The rust-hdf5 0.2.17 `SwmrFileWriter::create_hard_link` API
        // is called from `open_swmr` before `start_swmr()` — no close-path
        // re-open pass.
        //
        // SWMR mode now places the image dataset at the layout's nested
        // `det_default` path (`/entry/data/data`), exactly like standard mode;
        // the layout hardlink targets that nested path.
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_swmr_layout_hardlink_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="data">
                  <dataset name="data" source="detector" det_default="true"/>
                  <hardlink name="data_alias" target="/entry/data/data"/>
                </group>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_swmr_layout_hardlink");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for (i, x) in v.iter_mut().enumerate() {
                *x = i as u16;
            }
        }
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        assert!(
            writer.is_swmr_active(),
            "writer must be in SWMR mode for this test"
        );
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let names = h5.dataset_names();
        // The primary SWMR dataset at its nested layout path.
        assert!(
            names.contains(&"entry/data/data".to_string()),
            "SWMR image dataset must exist at the nested layout path; got {:?}",
            names
        );
        // The hard link materialised under the layout group.
        assert!(
            names.contains(&"entry/data/data_alias".to_string()),
            "SWMR layout <hardlink> must be materialised as a hard link; got {:?}",
            names
        );
        // The link shares the target object: same shape, readable as a dataset.
        let alias = h5.dataset("entry/data/data_alias").unwrap();
        let orig = h5.dataset("entry/data/data").unwrap();
        assert_eq!(alias.shape(), orig.shape());
        assert_eq!(orig.shape(), vec![2, 4, 4]);

        drop(h5);
        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_swmr_layout_nested_dataset_placement() {
        // SWMR mode must place the image dataset at the layout's nested
        // `det_default` path — mirroring C `NDFileHDF5` createTree
        // (`NDFileHDF5.cpp:638`) which builds the group tree and creates the
        // detector dataset inside it. The nested dataset, the layout
        // `<hardlink>`, and a constant dataset attribute must all be visible
        // to a `SwmrFileReader` reading the file back.
        let dir = std::env::temp_dir();
        let layout = dir.join(format!(
            "adcore_swmr_layout_nested_{}.xml",
            std::process::id()
        ));
        std::fs::write(
            &layout,
            r#"<hdf5_layout>
              <group name="entry">
                <group name="instrument">
                  <group name="detector">
                    <dataset name="data" source="detector" det_default="true">
                      <attribute name="signal" source="constant" value="1" type="int"/>
                    </dataset>
                    <hardlink name="data_alias" target="/entry/instrument/detector/data"/>
                  </group>
                </group>
                <group name="empty_placeholder"/>
              </group>
            </hdf5_layout>"#,
        )
        .unwrap();

        let path = temp_path("hdf5_swmr_layout_nested");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        assert!(
            writer.set_layout_filename(layout.to_str().unwrap()),
            "layout XML must parse: {}",
            writer.layout_error
        );

        let mut arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt16,
        );
        if let NDDataBuffer::U16(ref mut v) = arr.data {
            for (i, x) in v.iter_mut().enumerate() {
                *x = (i * 3) as u16;
            }
        }
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        assert!(
            writer.is_swmr_active(),
            "writer must be in SWMR mode for this test"
        );
        writer.write_file(&arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        // Read back via the SWMR reader — these are the exact paths a live
        // reader attaching during the streaming window would resolve.
        let mut reader = rust_hdf5::swmr::SwmrFileReader::open(&path).unwrap();
        let names = reader.dataset_names();
        // Image dataset at the nested layout path, NOT flat `data`.
        assert!(
            names.contains(&"entry/instrument/detector/data".to_string()),
            "SWMR image dataset must live at the nested layout path; got {:?}",
            names
        );
        assert!(
            !names.contains(&"data".to_string()),
            "SWMR image dataset must NOT remain at the flat root; got {:?}",
            names
        );
        // The empty placeholder group exists.
        assert!(
            reader.has_group("entry/empty_placeholder"),
            "empty layout group must be materialised; groups {:?}",
            reader.group_paths()
        );
        // The layout `<hardlink>` resolves to the nested dataset.
        assert!(
            names.contains(&"entry/instrument/detector/data_alias".to_string()),
            "SWMR layout <hardlink> must resolve to the nested dataset; got {:?}",
            names
        );
        let nested = reader
            .dataset_shape("entry/instrument/detector/data")
            .unwrap();
        let alias = reader
            .dataset_shape("entry/instrument/detector/data_alias")
            .unwrap();
        assert_eq!(nested, vec![2, 4, 4]);
        assert_eq!(alias, nested, "hardlink alias must share the target shape");
        // The data round-trips through both names.
        let via_nested: Vec<u16> = reader
            .read_dataset("entry/instrument/detector/data")
            .unwrap();
        let via_alias: Vec<u16> = reader
            .read_dataset("entry/instrument/detector/data_alias")
            .unwrap();
        assert_eq!(via_nested, via_alias);
        assert_eq!(via_nested.len(), 2 * 4 * 4);
        // The constant layout dataset attribute and the C-parity NDArray default
        // attributes (writeDefaultDatasetAttributes, NDFileHDF5.cpp:3695-3719) all
        // materialised before start_swmr().
        let attr_names = reader
            .dataset_attr_names("entry/instrument/detector/data")
            .unwrap();
        for expected in [
            "signal",
            "NDArrayNumDims",
            "NDArrayDimOffset",
            "NDArrayDimBinning",
            "NDArrayDimReverse",
        ] {
            assert!(
                attr_names.iter().any(|n| n == expected),
                "streaming dataset must carry the {expected} attribute; got {attr_names:?}",
            );
        }

        drop(reader);
        std::fs::remove_file(&path).ok();
        std::fs::remove_file(&layout).ok();
    }

    #[test]
    fn test_default_layout_parses_and_resolves_nexus_paths() {
        // Guards DEFAULT_LAYOUT_XML + the `default_layout()` expect(): the
        // built-in layout must parse and resolve C's NeXus placements
        // (NDFileHDF5LayoutXML.cpp:43-70).
        let layout = Hdf5Writer::default_layout();
        assert_eq!(
            layout.detector_dataset_path().as_deref(),
            Some("/entry/instrument/detector/data")
        );
        assert_eq!(
            layout.ndattr_default_group().as_deref(),
            Some("/entry/instrument/NDAttributes")
        );
        assert_eq!(
            layout.dataset_group_path("timestamp").as_deref(),
            Some("/entry/instrument/performance")
        );
    }

    #[test]
    fn test_no_layout_uses_default_nexus_layout() {
        // With no layout file the writer loads C's built-in DEFAULT_LAYOUT
        // (NDFileHDF5LayoutXML.cpp:43-70): the detector image lands at
        // /entry/instrument/detector/data inside the NeXus tree, with an
        // /entry/data/data hardlink to it — not a flat-root `data`.
        let path = temp_path("hdf5_default_layout");
        let mut writer = Hdf5Writer::new();
        let arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt8,
        );
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        // Detector dataset at the NeXus path, not flat root.
        let det = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(det.shape(), vec![1, 4, 4]);
        assert!(
            !h5.dataset_names().contains(&"data".to_string()),
            "must not write a flat-root `data`; got {:?}",
            h5.dataset_names()
        );
        // The NXdata hardlink /entry/data/data resolves to the same image.
        let linked = h5.dataset("entry/data/data").unwrap();
        assert_eq!(linked.shape(), vec![1, 4, 4]);
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_default_layout_group_nx_class_attributes() {
        // C attaches NX_class markers to the default-layout NeXus groups via
        // writeHdfAttributes (NDFileHDF5.cpp:693-695): NXentry/NXinstrument/
        // NXdetector/NXcollection/NXdata.
        let path = temp_path("hdf5_nxclass");
        let mut writer = Hdf5Writer::new();
        let arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::UInt8,
        );
        writer.open_file(&path, NDFileMode::Single, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let nx = |g: &str| {
            let mut grp = h5.root_group();
            for seg in g.split('/') {
                grp = grp.group(seg).unwrap();
            }
            grp.attr_string("NX_class").unwrap()
        };
        assert_eq!(nx("entry"), "NXentry");
        assert_eq!(nx("entry/instrument"), "NXinstrument");
        assert_eq!(nx("entry/instrument/detector"), "NXdetector");
        assert_eq!(nx("entry/instrument/NDAttributes"), "NXcollection");
        assert_eq!(nx("entry/instrument/detector/NDAttributes"), "NXcollection");
        assert_eq!(nx("entry/data"), "NXdata");
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_default_layout_group_nx_class_attributes() {
        // The SWMR path materialises the same NX_class group markers
        // (build_swmr_layout_groups → write_swmr_group_constant_attrs).
        let path = temp_path("hdf5_swmr_nxclass");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        let arr = NDArray::new(
            vec![NDDimension::new(4), NDDimension::new(4)],
            NDDataType::Float32,
        );
        writer.open_file(&path, NDFileMode::Stream, &arr).unwrap();
        writer.write_file(&arr).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let nx = |g: &str| {
            let mut grp = h5.root_group();
            for seg in g.split('/') {
                grp = grp.group(seg).unwrap();
            }
            grp.attr_string("NX_class").unwrap()
        };
        assert_eq!(nx("entry"), "NXentry");
        assert_eq!(nx("entry/instrument/detector"), "NXdetector");
        assert_eq!(nx("entry/data"), "NXdata");
        std::fs::remove_file(&path).ok();
    }

    // ---- ADP-99: direct chunk write of pre-compressed NDArrays -----------

    /// An uncompressed UInt16 frame with a deterministic ramp, ready to feed a
    /// codec. `dims = [y, x]` (HDF5 fastest axis last).
    fn ramp_u16(y: usize, x: usize) -> NDArray {
        let data: Vec<u16> = (0..(y * x) as u16).collect();
        NDArray::with_data(
            vec![NDDimension::new(y), NDDimension::new(x)],
            NDDataBuffer::U16(data),
        )
    }

    fn u16_pixels(arr: &NDArray) -> Vec<u16> {
        match &arr.data {
            NDDataBuffer::U16(v) => v.clone(),
            _ => unreachable!("expected U16 buffer"),
        }
    }

    #[test]
    fn test_codec_chunk_bytes_lz4_prepends_hdf5_header() {
        // C NDFileHDF5Dataset::writeFile (NDFileHDF5Dataset.cpp:299-314): a raw
        // LZ4 block gets a 16-byte big-endian header — uncompressed size (u64),
        // block size = uncompressed size (u32), compressed size (u32) — then the
        // block bytes.
        let codec = Codec {
            name: CodecName::LZ4,
            compressed_size: 4,
            level: 0,
            shuffle: 0,
            compressor: 0,
            original_data_type: NDDataType::UInt16,
        };
        let payload = [0xAAu8, 0xBB, 0xCC, 0xDD];
        let out = codec_chunk_bytes(&codec, 32, &payload);
        let mut expect = Vec::new();
        expect.extend_from_slice(&32u64.to_be_bytes()); // uncompressed size
        expect.extend_from_slice(&32u32.to_be_bytes()); // block size
        expect.extend_from_slice(&4u32.to_be_bytes()); // compressed size
        expect.extend_from_slice(&payload);
        assert_eq!(out, expect);
    }

    #[test]
    fn test_codec_chunk_bytes_verbatim_for_self_describing() {
        // BLOSC and JPEG emit self-describing streams, so the chunk is written
        // verbatim — no extra header. (LZ4 and BSLZ4 get a chunk header; see
        // test_direct_chunk_write_lz4_roundtrips / test_codec_chunk_bytes_bslz4_header.)
        for name in [CodecName::Blosc, CodecName::JPEG] {
            let codec = Codec {
                name,
                compressed_size: 3,
                level: 0,
                shuffle: 0,
                compressor: 0,
                original_data_type: NDDataType::UInt8,
            };
            let payload = [1u8, 2, 3];
            assert_eq!(codec_chunk_bytes(&codec, 99, &payload), payload.to_vec());
        }
    }

    #[test]
    fn test_codecs_match() {
        let a = Codec {
            name: CodecName::LZ4,
            compressed_size: 1,
            level: 0,
            shuffle: 0,
            compressor: 0,
            original_data_type: NDDataType::UInt8,
        };
        assert!(codecs_match(None, None));
        assert!(!codecs_match(Some(&a), None));
        assert!(!codecs_match(None, Some(&a)));
        assert!(codecs_match(Some(&a), Some(&a.clone())));
        let mut diff = a.clone();
        diff.compressor = 1;
        assert!(!codecs_match(Some(&a), Some(&diff)));
        // original_data_type is NOT part of codec identity (C Codec_t::operator!=).
        let mut other_type = a.clone();
        other_type.original_data_type = NDDataType::Float64;
        assert!(codecs_match(Some(&a), Some(&other_type)));
    }

    #[test]
    fn test_codec_filter_pipeline_ids() {
        let w = Hdf5Writer::new();
        let mk = |name| Codec {
            name,
            compressed_size: 0,
            level: 5,
            shuffle: 1,
            compressor: 2,
            original_data_type: NDDataType::UInt16,
        };
        assert_eq!(
            w.codec_filter_pipeline(&mk(CodecName::LZ4))
                .unwrap()
                .filters[0]
                .id,
            FILTER_LZ4
        );
        assert_eq!(
            w.codec_filter_pipeline(&mk(CodecName::BSLZ4))
                .unwrap()
                .filters[0]
                .id,
            FILTER_BSHUF
        );
        let blosc = w.codec_filter_pipeline(&mk(CodecName::Blosc)).unwrap();
        assert_eq!(blosc.filters[0].id, FILTER_BLOSC);
        // C copies the array's own blosc level/shuffle/compressor into the
        // dataset filter (configureCompression NDFileHDF5.cpp:3320-3323).
        assert_eq!(blosc.filters[0].cd_values[4], 5, "level");
        assert_eq!(blosc.filters[0].cd_values[5], 1, "shuffle");
        assert_eq!(blosc.filters[0].cd_values[6], 2, "compressor");
        assert_eq!(
            w.codec_filter_pipeline(&mk(CodecName::JPEG))
                .unwrap()
                .filters[0]
                .id,
            FILTER_JPEG
        );
        // Codecs C does not direct-chunk-write are rejected (None pipeline).
        assert!(w.codec_filter_pipeline(&mk(CodecName::None)).is_none());
        assert!(w.codec_filter_pipeline(&mk(CodecName::Zlib)).is_none());
        assert!(w.codec_filter_pipeline(&mk(CodecName::LZ4HDF5)).is_none());
    }

    #[test]
    fn test_compression_aware_true() {
        // C NDFileHDF5.cpp:2268 passes compressionAware=true.
        assert!(Hdf5FileProcessor::new().compression_aware());
    }

    #[test]
    fn test_direct_chunk_write_lz4_roundtrips() {
        let path = temp_path("hdf5_dcw_lz4");
        let mut writer = Hdf5Writer::new();
        let orig = ramp_u16(4, 5);
        let expect = u16_pixels(&orig);
        let comp = crate::codec::compress_lz4(&orig);
        assert!(comp.codec.is_some());
        writer.open_file(&path, NDFileMode::Single, &comp).unwrap();
        writer.write_file(&comp).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        // HDF5 axes are the NDArray dims reversed (fastest-varying last), so a
        // [4, 5] frame is stored as [1, 5, 4]; the flat pixel buffer order is
        // unchanged and round-trips through the reversed filter intact.
        assert_eq!(ds.shape(), vec![1, 5, 4]);
        assert!(ds.is_chunked());
        let read = ds.read_raw::<u16>().unwrap();
        assert_eq!(
            read, expect,
            "LZ4 direct-chunk-write must reverse to the original pixels"
        );
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_direct_chunk_write_blosc_roundtrips() {
        let path = temp_path("hdf5_dcw_blosc");
        let mut writer = Hdf5Writer::new();
        let orig = ramp_u16(4, 5);
        let expect = u16_pixels(&orig);
        let comp = crate::codec::compress_blosc(&orig, &crate::codec::BloscConfig::default());
        assert_eq!(comp.codec.as_ref().unwrap().name, CodecName::Blosc);
        writer.open_file(&path, NDFileMode::Single, &comp).unwrap();
        writer.write_file(&comp).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![1, 5, 4]);
        let read = ds.read_raw::<u16>().unwrap();
        assert_eq!(read, expect, "BLOSC direct-chunk-write must round-trip");
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_direct_chunk_write_bslz4_roundtrips() {
        // `compress_bslz4` emits the canonical bitshuffle+LZ4 on-disk format
        // (byte-for-byte the libhdf5/h5py/C-areaDetector bytes — locked by
        // codec::tests::test_bitshuffle_matches_c_reference_vector), stored
        // verbatim with the 12-byte chunk header (test_codec_chunk_bytes_bslz4_header).
        // rust-hdf5 0.2.21's bitshuffle reverse filter is canonical (LSB-first),
        // so it reads those bytes back to the original pixels. Use a whole-block
        // frame: u16 blocks are 4096 elems, so 128*128 = 16384 = 4 full blocks
        // with no partial tail.
        let path = temp_path("hdf5_dcw_bslz4");
        let mut writer = Hdf5Writer::new();
        let orig = ramp_u16(128, 128);
        let expect = u16_pixels(&orig);
        let comp = crate::codec::compress_bslz4(&orig);
        assert_eq!(comp.codec.as_ref().unwrap().name, CodecName::BSLZ4);
        writer.open_file(&path, NDFileMode::Single, &comp).unwrap();
        writer.write_file(&comp).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![1, 128, 128]);
        assert!(ds.is_chunked());
        assert_eq!(ds.element_size(), 2, "dataset keeps the original u16 type");
        let read = ds.read_raw::<u16>().unwrap();
        assert_eq!(read, expect, "BSLZ4 direct-chunk-write must round-trip");
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_codec_chunk_bytes_bslz4_header() {
        // The HDF5 bitshuffle filter expects a 12-byte big-endian chunk header ahead of the
        // canonical stream: uncompressed total bytes (u64) and the block size in bytes (u32 =
        // block_elems * elem_size, C hardcodes the 8192 default), per
        // `NDFileHDF5Dataset::writeFile` (`NDFileHDF5Dataset.cpp:316-328`).
        let codec = Codec {
            name: CodecName::BSLZ4,
            compressed_size: 0,
            level: 0,
            shuffle: 0,
            compressor: 0,
            original_data_type: NDDataType::UInt16,
        };
        let payload = [0xDEu8, 0xAD, 0xBE, 0xEF];
        let total = 128 * 128 * 2; // one u16 frame
        let out = codec_chunk_bytes(&codec, total, &payload);
        assert_eq!(&out[0..8], &(total as u64).to_be_bytes());
        // u16 default block is 4096 elems => 8192 bytes (matches C's 8192).
        assert_eq!(&out[8..12], &8192u32.to_be_bytes());
        assert_eq!(&out[12..], &payload, "canonical stream follows verbatim");
    }

    #[test]
    fn test_direct_chunk_write_lz4_multiframe_extends_leading_axis() {
        let path = temp_path("hdf5_dcw_lz4_multi");
        let mut writer = Hdf5Writer::new();
        let frames: Vec<NDArray> = (0..3u16)
            .map(|f| {
                let data: Vec<u16> = (0..20u16).map(|i| i + f * 100).collect();
                NDArray::with_data(
                    vec![NDDimension::new(4), NDDimension::new(5)],
                    NDDataBuffer::U16(data),
                )
            })
            .collect();
        let comp: Vec<NDArray> = frames.iter().map(crate::codec::compress_lz4).collect();
        writer
            .open_file(&path, NDFileMode::Stream, &comp[0])
            .unwrap();
        for c in &comp {
            writer.write_file(c).unwrap();
        }
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![3, 5, 4], "three compressed frames stacked");
        let read = ds.read_raw::<u16>().unwrap();
        let mut expect = Vec::new();
        for f in 0..3u16 {
            for i in 0..20u16 {
                expect.push(i + f * 100);
            }
        }
        assert_eq!(read, expect, "every frame's pixels recovered in order");
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_direct_chunk_write_jpeg_records_chunked_dataset() {
        // JPEG is lossy and the rust-hdf5 reader has no JPEG reverse filter, so
        // pixels cannot round-trip; verify the dataset is created with the right
        // original type/shape and chunked (one whole frame per chunk). The JPEG
        // filter id is covered by test_codec_filter_pipeline_ids.
        let path = temp_path("hdf5_dcw_jpeg");
        let mut writer = Hdf5Writer::new();
        let mono: Vec<u8> = (0..64u16).map(|i| (i * 3) as u8).collect();
        let src = NDArray::with_data(
            vec![NDDimension::new(8), NDDimension::new(8)],
            NDDataBuffer::U8(mono),
        );
        let comp = crate::codec::compress_jpeg(&src, 90).expect("jpeg encode");
        assert_eq!(comp.codec.as_ref().unwrap().name, CodecName::JPEG);
        writer.open_file(&path, NDFileMode::Single, &comp).unwrap();
        writer.write_file(&comp).unwrap();
        writer.close_file().unwrap();

        let h5 = H5File::open(&path).unwrap();
        let ds = h5.dataset("entry/instrument/detector/data").unwrap();
        assert_eq!(ds.shape(), vec![1, 8, 8]);
        assert!(ds.is_chunked());
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_codec_change_mid_stream_errors() {
        // C verifyChunking rejects a frame whose codec differs from the dataset's
        // (NDFileHDF5Dataset.cpp:194-200). Here the file is opened compressed and
        // a later uncompressed frame must be refused.
        let path = temp_path("hdf5_codec_change");
        let mut writer = Hdf5Writer::new();
        let f0 = crate::codec::compress_lz4(&ramp_u16(4, 5));
        let f1_plain = ramp_u16(4, 5); // no codec
        writer.open_file(&path, NDFileMode::Stream, &f0).unwrap();
        writer.write_file(&f0).unwrap();
        assert!(
            writer.write_file(&f1_plain).is_err(),
            "an uncompressed frame must be rejected for a compressed dataset"
        );
        writer.close_file().ok();
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_swmr_rejects_compressed_array() {
        // SWMR mode has no raw-chunk append API, so a pre-compressed frame is
        // rejected rather than re-compressed into garbage (documented residual).
        let path = temp_path("hdf5_swmr_compressed");
        let mut writer = Hdf5Writer::new();
        writer.set_swmr_mode(true);
        let comp = crate::codec::compress_lz4(&ramp_u16(4, 5));
        writer.open_file(&path, NDFileMode::Stream, &comp).unwrap();
        assert!(
            writer.write_file(&comp).is_err(),
            "SWMR mode cannot direct-chunk-write a pre-compressed array"
        );
        writer.close_file().ok();
        std::fs::remove_file(&path).ok();
    }

    #[test]
    fn test_unsupported_codec_rejected_on_open() {
        // A Zlib-compressed array has no C direct-chunk-write analog; dataset
        // creation rejects it rather than writing an unreadable file.
        let path = temp_path("hdf5_dcw_zlib");
        let mut writer = Hdf5Writer::new();
        let comp = crate::codec::compress_zlib(&ramp_u16(4, 5));
        assert_eq!(comp.codec.as_ref().unwrap().name, CodecName::Zlib);
        writer.open_file(&path, NDFileMode::Single, &comp).unwrap();
        assert!(
            writer.write_file(&comp).is_err(),
            "an unsupported (zlib) codec must be rejected, not silently mis-written"
        );
        writer.close_file().ok();
        std::fs::remove_file(&path).ok();
    }
}