netcdf-writer 0.9.0

//! Pure-Rust NetCDF writer.
//!
//! The current implementation writes CDF-1, CDF-2, and CDF-5 classic-family
//! files, plus a conservative NetCDF-4/HDF5 subset backed by `hdf5-writer`.
//! NetCDF-4 emission is intentionally strict about metadata it cannot yet
//! represent losslessly.

use std::io::Write;

#[cfg(feature = "netcdf4")]
use hdf5_writer::{
    AttributeBuilder as H5AttributeBuilder, ByteOrder as H5ByteOrder,
    CompoundField as H5CompoundField, DatasetBuilder as H5DatasetBuilder, Datatype as H5Datatype,
    EnumMember as H5EnumMember, FilterDescription as H5FilterDescription, Hdf5Builder,
    ReferenceType as H5ReferenceType, StringEncoding as H5StringEncoding,
    StringPadding as H5StringPadding, StringSize as H5StringSize, VarLenKind as H5VarLenKind,
    WriteOptions as H5WriteOptions, FILTER_DEFLATE as H5_FILTER_DEFLATE,
    FILTER_FLETCHER32 as H5_FILTER_FLETCHER32, FILTER_SHUFFLE as H5_FILTER_SHUFFLE,
    UNLIMITED as H5_UNLIMITED,
};

pub use netcdf_core::{
    NcAttrValue, NcAttribute, NcCompoundField, NcDimension, NcEnumMember, NcFormat, NcGroup,
    NcIntegerValue, NcSliceInfo, NcSliceInfoElem, NcType, NcVariable, NC_FILL_BYTE, NC_FILL_CHAR,
    NC_FILL_DOUBLE, NC_FILL_FLOAT, NC_FILL_INT, NC_FILL_INT64, NC_FILL_SHORT, NC_FILL_UBYTE,
    NC_FILL_UINT, NC_FILL_UINT64, NC_FILL_USHORT,
};

const ABSENT: u32 = 0x0000_0000;
const NC_DIMENSION: u32 = 0x0000_000A;
const NC_VARIABLE: u32 = 0x0000_000B;
const NC_ATTRIBUTE: u32 = 0x0000_000C;

/// NetCDF writer errors.
#[derive(Debug, thiserror::Error)]
pub enum Error {
    #[error("I/O error: {0}")]
    Io(#[from] std::io::Error),

    #[error("NetCDF core error: {0}")]
    Core(#[from] netcdf_core::Error),

    #[error("invalid definition: {0}")]
    InvalidDefinition(String),

    #[error("type mismatch: expected {expected}, got {actual}")]
    TypeMismatch { expected: String, actual: String },

    #[error("data length mismatch: expected {expected}, got {actual}")]
    DataLengthMismatch { expected: usize, actual: usize },

    /// The schema needs the NetCDF-4 data model (e.g. a type, name, or
    /// structure the requested classic format cannot represent). Callers can
    /// match on this and retry with [`NcWriteFormat::Nc4`].
    #[error("requires the NetCDF-4 data model: {reason}")]
    RequiresNetcdf4 { reason: String },

    /// A value exceeds the capacity of the requested classic format (for
    /// example a 32-bit offset or size field). Callers can match on this and
    /// retry with a larger classic format such as CDF-5.
    #[error("exceeds the capacity of the requested classic format: {reason}")]
    FormatCapacityExceeded { reason: String },

    #[error("unsupported write feature: {0}")]
    UnsupportedFeature(String),
}

pub type Result<T> = std::result::Result<T, Error>;

/// Requested NetCDF output format.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum NcWriteFormat {
    /// Pick CDF-1, CDF-2, or CDF-5 from the exact schema and layout.
    AutoClassic,
    Classic,
    Offset64,
    Cdf5,
    Nc4,
    Nc4Classic,
}

/// NetCDF write options.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct NcWriteOptions {
    pub format: NcWriteFormat,
}

impl Default for NcWriteOptions {
    fn default() -> Self {
        Self {
            format: NcWriteFormat::AutoClassic,
        }
    }
}

impl NcWriteOptions {
    pub fn classic() -> Self {
        Self {
            format: NcWriteFormat::Classic,
        }
    }

    pub fn offset64() -> Self {
        Self {
            format: NcWriteFormat::Offset64,
        }
    }

    pub fn cdf5() -> Self {
        Self {
            format: NcWriteFormat::Cdf5,
        }
    }
}

/// Dimension handle returned by [`NcFileBuilder::add_dimension`].
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct DimensionId(usize);

/// Variable handle returned by [`NcFileBuilder::add_variable`].
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct VariableId(usize);

/// Rust types that can be written as NetCDF classic-family variable data.
pub trait NcWriteType: Copy {
    fn nc_type() -> NcType;
    fn write_one_be(self, dst: &mut Vec<u8>);

    fn write_one_le(self, dst: &mut Vec<u8>) {
        let start = dst.len();
        self.write_one_be(dst);
        dst[start..].reverse();
    }
}

/// Rust primitive types that can be used as a NetCDF variable fill value.
///
/// Setting a fill value writes the canonical `_FillValue` variable attribute.
/// NetCDF-4 output also records the value as an HDF5 dataset fill value, which
/// allows fixed-size datasets to be represented without eagerly materializing
/// every element.
pub trait NcFillValueType: NcWriteType {
    fn fill_attr_value(self) -> NcAttrValue;
}

macro_rules! impl_write_type {
    ($ty:ty, $nc_type:expr, $write:expr) => {
        impl NcWriteType for $ty {
            fn nc_type() -> NcType {
                $nc_type
            }

            fn write_one_be(self, dst: &mut Vec<u8>) {
                $write(self, dst)
            }
        }
    };
}

macro_rules! impl_fill_value_type {
    ($ty:ty, $attr:ident) => {
        impl NcFillValueType for $ty {
            fn fill_attr_value(self) -> NcAttrValue {
                NcAttrValue::$attr(vec![self])
            }
        }
    };
}

impl_write_type!(i8, NcType::Byte, |value: i8, dst: &mut Vec<u8>| dst
    .push(value as u8));
impl_write_type!(u8, NcType::UByte, |value: u8, dst: &mut Vec<u8>| dst
    .push(value));
impl_write_type!(i16, NcType::Short, |value: i16, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(u16, NcType::UShort, |value: u16, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(i32, NcType::Int, |value: i32, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(u32, NcType::UInt, |value: u32, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(i64, NcType::Int64, |value: i64, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(u64, NcType::UInt64, |value: u64, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(f32, NcType::Float, |value: f32, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));
impl_write_type!(f64, NcType::Double, |value: f64, dst: &mut Vec<u8>| dst
    .extend_from_slice(&value.to_be_bytes()));

impl_fill_value_type!(i8, Bytes);
impl_fill_value_type!(u8, UBytes);
impl_fill_value_type!(i16, Shorts);
impl_fill_value_type!(u16, UShorts);
impl_fill_value_type!(i32, Ints);
impl_fill_value_type!(u32, UInts);
impl_fill_value_type!(i64, Int64s);
impl_fill_value_type!(u64, UInt64s);
impl_fill_value_type!(f32, Floats);
impl_fill_value_type!(f64, Doubles);

const FILL_VALUE_ATTR_NAME: &str = "_FillValue";

#[derive(Debug, Clone)]
struct DimensionDef {
    name: String,
    size: u64,
    is_unlimited: bool,
    current_size: u64,
}

#[derive(Debug, Clone)]
struct GroupAttributeDef {
    group_path: String,
    attribute: NcAttribute,
}

#[derive(Debug, Clone)]
struct VariableDef {
    name: String,
    dim_ids: Vec<DimensionId>,
    dtype: NcType,
    attributes: Vec<NcAttribute>,
    data: Vec<u8>,
    data_encoding: VariableDataEncoding,
    string_values: Option<Vec<String>>,
    vlen_values: Option<Vec<Vec<u8>>>,
    storage: VariableStorageDef,
    fill_value: Option<VariableFillValue>,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum VariableDataEncoding {
    ClassicBigEndian,
    Hdf5Native,
}

#[derive(Debug, Clone)]
struct VariableFillValue {
    classic_bytes: Vec<u8>,
    hdf5_bytes: Vec<u8>,
}

#[derive(Debug, Clone, Default)]
struct VariableStorageDef {
    chunk_shape: Option<Vec<u64>>,
    shuffle: bool,
    deflate_level: Option<u8>,
    fletcher32: bool,
}

impl VariableStorageDef {
    fn has_filters(&self) -> bool {
        self.shuffle || self.deflate_level.is_some() || self.fletcher32
    }

    fn has_nc4_options(&self) -> bool {
        self.chunk_shape.is_some() || self.has_filters()
    }
}

/// Builder for a single NetCDF file.
#[derive(Debug, Clone, Default)]
pub struct NcFileBuilder {
    dimensions: Vec<DimensionDef>,
    attributes: Vec<NcAttribute>,
    group_attributes: Vec<GroupAttributeDef>,
    variables: Vec<VariableDef>,
}

impl NcFileBuilder {
    pub fn new() -> Self {
        Self::default()
    }

    pub fn add_dimension(&mut self, name: impl Into<String>, size: u64) -> Result<DimensionId> {
        let name = name.into();
        validate_name(&name, "dimension")?;
        self.add_dimension_def(name, size, false)
    }

    pub fn add_dimension_path(
        &mut self,
        path: impl Into<String>,
        size: u64,
    ) -> Result<DimensionId> {
        let path = validate_path_name(path.into(), "dimension")?;
        self.add_dimension_def(path, size, false)
    }

    pub fn add_unlimited_dimension(&mut self, name: impl Into<String>) -> Result<DimensionId> {
        let name = name.into();
        validate_name(&name, "dimension")?;
        self.add_dimension_def(name, 0, true)
    }

    pub fn add_unlimited_dimension_path(&mut self, path: impl Into<String>) -> Result<DimensionId> {
        let path = validate_path_name(path.into(), "dimension")?;
        self.add_dimension_def(path, 0, true)
    }

    pub fn add_attribute(&mut self, name: impl Into<String>, value: NcAttrValue) -> Result<()> {
        let name = name.into();
        validate_name(&name, "attribute")?;
        ensure_unique_attr(&self.attributes, &name)?;
        self.attributes.push(NcAttribute { name, value });
        Ok(())
    }

    pub fn add_group_attribute(
        &mut self,
        group_path: impl Into<String>,
        name: impl Into<String>,
        value: NcAttrValue,
    ) -> Result<()> {
        let group_path = validate_path_name(group_path.into(), "group")?;
        let name = name.into();
        validate_name(&name, "attribute")?;
        if self
            .group_attributes
            .iter()
            .any(|attr| attr.group_path == group_path && attr.attribute.name == name)
        {
            return Err(Error::InvalidDefinition(format!(
                "duplicate attribute '{name}' on group '{group_path}'"
            )));
        }
        self.group_attributes.push(GroupAttributeDef {
            group_path,
            attribute: NcAttribute { name, value },
        });
        Ok(())
    }

    pub fn add_variable<T: NcWriteType>(
        &mut self,
        name: impl Into<String>,
        dimensions: &[DimensionId],
    ) -> Result<VariableId> {
        self.add_variable_with_type(name, dimensions, T::nc_type())
    }

    pub fn add_variable_path<T: NcWriteType>(
        &mut self,
        path: impl Into<String>,
        dimensions: &[DimensionId],
    ) -> Result<VariableId> {
        self.add_variable_path_with_type(path, dimensions, T::nc_type())
    }

    pub fn add_char_variable(
        &mut self,
        name: impl Into<String>,
        dimensions: &[DimensionId],
    ) -> Result<VariableId> {
        self.add_variable_with_type(name, dimensions, NcType::Char)
    }

    pub fn add_char_variable_path(
        &mut self,
        path: impl Into<String>,
        dimensions: &[DimensionId],
    ) -> Result<VariableId> {
        self.add_variable_path_with_type(path, dimensions, NcType::Char)
    }

    pub fn add_string_variable(
        &mut self,
        name: impl Into<String>,
        dimensions: &[DimensionId],
    ) -> Result<VariableId> {
        self.add_variable_with_type(name, dimensions, NcType::String)
    }

    pub fn add_string_variable_path(
        &mut self,
        path: impl Into<String>,
        dimensions: &[DimensionId],
    ) -> Result<VariableId> {
        self.add_variable_path_with_type(path, dimensions, NcType::String)
    }

    pub fn add_user_defined_variable(
        &mut self,
        name: impl Into<String>,
        dimensions: &[DimensionId],
        dtype: NcType,
    ) -> Result<VariableId> {
        validate_supported_user_defined_type(&dtype)?;
        self.add_variable_with_type(name, dimensions, dtype)
    }

    pub fn add_user_defined_variable_path(
        &mut self,
        path: impl Into<String>,
        dimensions: &[DimensionId],
        dtype: NcType,
    ) -> Result<VariableId> {
        validate_supported_user_defined_type(&dtype)?;
        self.add_variable_path_with_type(path, dimensions, dtype)
    }

    pub fn add_variable_attribute(
        &mut self,
        variable: VariableId,
        name: impl Into<String>,
        value: NcAttrValue,
    ) -> Result<()> {
        let name = name.into();
        validate_name(&name, "attribute")?;
        let variable = self.variable_mut(variable)?;
        ensure_unique_attr(&variable.attributes, &name)?;
        variable.attributes.push(NcAttribute { name, value });
        Ok(())
    }

    pub fn set_variable_fill_value<T: NcFillValueType>(
        &mut self,
        variable: VariableId,
        value: T,
    ) -> Result<()> {
        let variable = self.variable_mut(variable)?;
        let expected = T::nc_type();
        if variable.dtype != expected {
            return Err(Error::TypeMismatch {
                expected: format!("{:?}", variable.dtype),
                actual: format!("{expected:?}"),
            });
        }

        let mut classic_bytes = Vec::with_capacity(expected.size()?);
        value.write_one_be(&mut classic_bytes);
        let mut hdf5_bytes = Vec::with_capacity(expected.size()?);
        value.write_one_le(&mut hdf5_bytes);
        set_variable_fill_value_metadata(
            variable,
            value.fill_attr_value(),
            classic_bytes,
            hdf5_bytes,
        )
    }

    pub fn set_variable_chunking(
        &mut self,
        variable: VariableId,
        chunk_shape: impl Into<Vec<u64>>,
    ) -> Result<()> {
        let variable = self.variable_mut(variable)?;
        let chunk_shape = chunk_shape.into();
        validate_chunk_shape_for_variable(variable, &chunk_shape)?;
        variable.storage.chunk_shape = Some(chunk_shape);
        Ok(())
    }

    pub fn set_variable_shuffle(&mut self, variable: VariableId, enabled: bool) -> Result<()> {
        let variable = self.variable_mut(variable)?;
        reject_scalar_filtered_variable(variable)?;
        variable.storage.shuffle = enabled;
        Ok(())
    }

    pub fn set_variable_deflate(
        &mut self,
        variable: VariableId,
        level: Option<u8>,
        shuffle: bool,
    ) -> Result<()> {
        if let Some(level) = level {
            if level > 9 {
                return Err(Error::InvalidDefinition(
                    "deflate level must be in 0..=9".into(),
                ));
            }
        }
        let variable = self.variable_mut(variable)?;
        reject_scalar_filtered_variable(variable)?;
        variable.storage.deflate_level = level;
        variable.storage.shuffle = shuffle;
        Ok(())
    }

    pub fn set_variable_fletcher32(&mut self, variable: VariableId, enabled: bool) -> Result<()> {
        let variable = self.variable_mut(variable)?;
        reject_scalar_filtered_variable(variable)?;
        variable.storage.fletcher32 = enabled;
        Ok(())
    }

    pub fn write_variable<T: NcWriteType>(
        &mut self,
        variable: VariableId,
        values: &[T],
    ) -> Result<()> {
        {
            let variable = self.variable_mut(variable)?;
            let expected = T::nc_type();
            if variable.dtype != expected {
                return Err(Error::TypeMismatch {
                    expected: format!("{:?}", variable.dtype),
                    actual: format!("{:?}", expected),
                });
            }
            let mut data = Vec::with_capacity(std::mem::size_of_val(values));
            for &value in values {
                value.write_one_be(&mut data);
            }
            variable.data = data;
            variable.data_encoding = VariableDataEncoding::ClassicBigEndian;
            variable.string_values = None;
            variable.vlen_values = None;
        }
        self.update_unlimited_extents_from_element_count(variable, values.len() as u64)
    }

    pub fn write_variable_slice<T: NcWriteType>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[T],
    ) -> Result<()> {
        let variable_def = self.variable(variable)?;
        let expected = T::nc_type();
        if variable_def.dtype != expected {
            return Err(Error::TypeMismatch {
                expected: format!("{:?}", variable_def.dtype),
                actual: format!("{expected:?}"),
            });
        }
        let resolved =
            self.resolve_variable_write_selection(variable_def, selection, expected.size()?)?;
        if values.len() != resolved.elements {
            return Err(Error::DataLengthMismatch {
                expected: resolved.elements,
                actual: values.len(),
            });
        }

        let elem_size = expected.size()?;
        let mut encoded = Vec::with_capacity(checked_mul_usize(
            values.len(),
            elem_size,
            "variable slice byte size",
        )?);
        for &value in values {
            value.write_one_be(&mut encoded);
        }

        let strides = row_major_strides(&resolved.shape, "writer slice stride")?;
        let total_elements =
            checked_shape_elements(&resolved.shape, "writer variable element count")?;
        {
            let variable_def = self.variable_mut(variable)?;
            ensure_variable_slice_buffer(
                variable_def,
                &resolved.old_shape,
                &resolved.shape,
                total_elements,
                elem_size,
                resolved.can_grow,
            )?;
            scatter_slice_bytes(
                &mut variable_def.data,
                elem_size,
                &resolved.dims,
                &strides,
                &encoded,
            )?;
        }
        self.update_unlimited_extents_from_shape(variable, &resolved.shape)
    }

    pub fn write_char_variable(&mut self, variable: VariableId, bytes: &[u8]) -> Result<()> {
        {
            let variable = self.variable_mut(variable)?;
            if variable.dtype != NcType::Char {
                return Err(Error::TypeMismatch {
                    expected: "Char".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            }
            variable.data = bytes.to_vec();
            variable.data_encoding = VariableDataEncoding::ClassicBigEndian;
            variable.string_values = None;
            variable.vlen_values = None;
        }
        self.update_unlimited_extents_from_element_count(variable, bytes.len() as u64)
    }

    pub fn write_char_variable_slice(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        bytes: &[u8],
    ) -> Result<()> {
        let variable_def = self.variable(variable)?;
        if variable_def.dtype != NcType::Char {
            return Err(Error::TypeMismatch {
                expected: "Char".into(),
                actual: format!("{:?}", variable_def.dtype),
            });
        }
        let resolved = self.resolve_variable_write_selection(variable_def, selection, 1)?;
        if bytes.len() != resolved.elements {
            return Err(Error::DataLengthMismatch {
                expected: resolved.elements,
                actual: bytes.len(),
            });
        }

        let strides = row_major_strides(&resolved.shape, "writer char slice stride")?;
        let total_elements =
            checked_shape_elements(&resolved.shape, "writer char variable element count")?;
        {
            let variable_def = self.variable_mut(variable)?;
            ensure_variable_slice_buffer(
                variable_def,
                &resolved.old_shape,
                &resolved.shape,
                total_elements,
                1,
                resolved.can_grow,
            )?;
            scatter_slice_bytes(&mut variable_def.data, 1, &resolved.dims, &strides, bytes)?;
        }
        self.update_unlimited_extents_from_shape(variable, &resolved.shape)
    }

    pub fn write_char_variable_strings<S: AsRef<str>>(
        &mut self,
        variable: VariableId,
        values: &[S],
    ) -> Result<()> {
        let (name, width, inferred_unlimited) = {
            let variable_def = self.variable(variable)?;
            if variable_def.dtype != NcType::Char {
                return Err(Error::TypeMismatch {
                    expected: "Char".into(),
                    actual: format!("{:?}", variable_def.dtype),
                });
            }
            let inferred_unlimited =
                self.validate_char_string_value_count(variable_def, values.len())?;
            (
                variable_def.name.clone(),
                self.char_string_axis_width(variable_def)?,
                inferred_unlimited,
            )
        };
        let encoded = encode_char_string_values(&name, width, values)?;
        self.write_char_variable(variable, &encoded)?;
        if let Some((dimension, size)) = inferred_unlimited {
            if size > self.dimensions[dimension.0].current_size {
                self.dimensions[dimension.0].current_size = size;
            }
        }
        Ok(())
    }

    pub fn write_char_variable_strings_slice<S: AsRef<str>>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[S],
    ) -> Result<()> {
        let (name, width, resolved) = {
            let variable_def = self.variable(variable)?;
            if variable_def.dtype != NcType::Char {
                return Err(Error::TypeMismatch {
                    expected: "Char".into(),
                    actual: format!("{:?}", variable_def.dtype),
                });
            }
            (
                variable_def.name.clone(),
                self.char_string_axis_width(variable_def)?,
                self.resolve_char_string_write_selection(variable_def, selection)?,
            )
        };
        if values.len() != resolved.elements {
            return Err(Error::DataLengthMismatch {
                expected: resolved.elements,
                actual: values.len(),
            });
        }

        let encoded = encode_char_string_values(&name, width, values)?;
        let mut selections = selection.selections.clone();
        selections.push(NcSliceInfoElem::Slice {
            start: 0,
            end: u64::try_from(width).map_err(|_| {
                Error::InvalidDefinition("char string width exceeds u64 capacity".into())
            })?,
            step: 1,
        });
        self.write_char_variable_slice(variable, &NcSliceInfo { selections }, &encoded)
    }

    pub fn write_user_defined_variable_bytes(
        &mut self,
        variable: VariableId,
        bytes: &[u8],
    ) -> Result<()> {
        let elem_size = {
            let variable = self.variable_mut(variable)?;
            validate_fixed_width_user_defined_type(&variable.dtype)?;
            let elem_size = variable.dtype.size()?;
            variable.data = bytes.to_vec();
            variable.data_encoding = VariableDataEncoding::Hdf5Native;
            variable.string_values = None;
            variable.vlen_values = None;
            elem_size
        };
        if bytes.len() % elem_size != 0 {
            return Err(Error::DataLengthMismatch {
                expected: bytes.len() + (elem_size - bytes.len() % elem_size),
                actual: bytes.len(),
            });
        }
        self.update_unlimited_extents_from_element_count(variable, (bytes.len() / elem_size) as u64)
    }

    pub fn write_user_defined_variable_slice_bytes(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        bytes: &[u8],
    ) -> Result<()> {
        let elem_size = {
            let variable = self.variable(variable)?;
            validate_fixed_width_user_defined_type(&variable.dtype)?;
            variable.dtype.size()?
        };
        self.write_native_variable_slice_bytes(variable, selection, bytes, elem_size)
    }

    pub fn write_enum_variable(
        &mut self,
        variable: VariableId,
        values: &[NcIntegerValue],
    ) -> Result<()> {
        {
            let variable = self.variable_mut(variable)?;
            let NcType::Enum { base, .. } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "Enum".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            let mut data = Vec::with_capacity(values.len() * base.size()?);
            for &value in values {
                data.extend_from_slice(&nc_enum_value_to_le_bytes(base, value)?);
            }
            variable.data = data;
            variable.data_encoding = VariableDataEncoding::Hdf5Native;
            variable.string_values = None;
            variable.vlen_values = None;
        }
        self.update_unlimited_extents_from_element_count(variable, values.len() as u64)
    }

    pub fn write_enum_variable_slice(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[NcIntegerValue],
    ) -> Result<()> {
        let base = {
            let variable = self.variable(variable)?;
            let NcType::Enum { base, .. } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "Enum".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            base.clone()
        };
        let elem_size = base.size()?;
        let mut data = Vec::with_capacity(checked_mul_usize(
            values.len(),
            elem_size,
            "enum slice byte size",
        )?);
        for &value in values {
            data.extend_from_slice(&nc_enum_value_to_le_bytes(&base, value)?);
        }
        self.write_native_variable_slice_bytes(variable, selection, &data, elem_size)
    }

    pub fn write_opaque_variable<S: AsRef<[u8]>>(
        &mut self,
        variable: VariableId,
        values: &[S],
    ) -> Result<()> {
        {
            let variable = self.variable_mut(variable)?;
            let NcType::Opaque { size, .. } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "Opaque".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            let size = usize::try_from(*size).map_err(|_| {
                Error::InvalidDefinition("opaque element size exceeds platform usize".into())
            })?;
            let mut data = Vec::with_capacity(values.len() * size);
            for value in values {
                let value = value.as_ref();
                if value.len() != size {
                    return Err(Error::DataLengthMismatch {
                        expected: size,
                        actual: value.len(),
                    });
                }
                data.extend_from_slice(value);
            }
            variable.data = data;
            variable.data_encoding = VariableDataEncoding::Hdf5Native;
            variable.string_values = None;
            variable.vlen_values = None;
        }
        self.update_unlimited_extents_from_element_count(variable, values.len() as u64)
    }

    pub fn write_opaque_variable_slice<S: AsRef<[u8]>>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[S],
    ) -> Result<()> {
        let size = {
            let variable = self.variable(variable)?;
            let NcType::Opaque { size, .. } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "Opaque".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            usize::try_from(*size).map_err(|_| {
                Error::InvalidDefinition("opaque element size exceeds platform usize".into())
            })?
        };
        let mut data = Vec::with_capacity(checked_mul_usize(
            values.len(),
            size,
            "opaque slice byte size",
        )?);
        for value in values {
            let value = value.as_ref();
            if value.len() != size {
                return Err(Error::DataLengthMismatch {
                    expected: size,
                    actual: value.len(),
                });
            }
            data.extend_from_slice(value);
        }
        self.write_native_variable_slice_bytes(variable, selection, &data, size)
    }

    pub fn write_array_variable<T: NcWriteType>(
        &mut self,
        variable: VariableId,
        values: &[T],
    ) -> Result<()> {
        let variable_elements = {
            let variable = self.variable_mut(variable)?;
            let NcType::Array { base, .. } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "Array".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            let expected = T::nc_type();
            if base.as_ref() != &expected {
                return Err(Error::TypeMismatch {
                    expected: format!("{:?}", base),
                    actual: format!("{expected:?}"),
                });
            }
            let mut data = Vec::with_capacity(std::mem::size_of_val(values));
            for &value in values {
                value.write_one_le(&mut data);
            }
            let elem_size = variable.dtype.size()?;
            if data.len() % elem_size != 0 {
                return Err(Error::DataLengthMismatch {
                    expected: data.len() + (elem_size - data.len() % elem_size),
                    actual: data.len(),
                });
            }
            let variable_elements = (data.len() / elem_size) as u64;
            variable.data = data;
            variable.data_encoding = VariableDataEncoding::Hdf5Native;
            variable.string_values = None;
            variable.vlen_values = None;
            variable_elements
        };
        self.update_unlimited_extents_from_element_count(variable, variable_elements)
    }

    pub fn write_array_variable_slice<T: NcWriteType>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[T],
    ) -> Result<()> {
        let elem_size = {
            let variable = self.variable(variable)?;
            let NcType::Array { base, .. } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "Array".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            let expected = T::nc_type();
            if base.as_ref() != &expected {
                return Err(Error::TypeMismatch {
                    expected: format!("{:?}", base),
                    actual: format!("{expected:?}"),
                });
            }
            variable.dtype.size()?
        };
        let mut data = Vec::with_capacity(checked_mul_usize(
            values.len(),
            T::nc_type().size()?,
            "array slice byte size",
        )?);
        for &value in values {
            value.write_one_le(&mut data);
        }
        self.write_native_variable_slice_bytes(variable, selection, &data, elem_size)
    }

    pub fn write_vlen_variable_bytes<S: AsRef<[u8]>>(
        &mut self,
        variable: VariableId,
        values: &[S],
    ) -> Result<()> {
        {
            let variable = self.variable_mut(variable)?;
            let NcType::VLen { base } = &variable.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "VLen".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            };
            validate_vlen_base_nc4_type(base)?;
            let base_size = base.size()?;
            let mut sequences = Vec::with_capacity(values.len());
            for value in values {
                let value = value.as_ref();
                if value.len() % base_size != 0 {
                    return Err(Error::InvalidDefinition(format!(
                        "vlen sequence byte length {} is not a multiple of base element size {base_size}",
                        value.len()
                    )));
                }
                sequences.push(value.to_vec());
            }
            variable.data.clear();
            variable.data_encoding = VariableDataEncoding::Hdf5Native;
            variable.string_values = None;
            variable.vlen_values = Some(sequences);
        }
        self.update_unlimited_extents_from_element_count(variable, values.len() as u64)
    }

    pub fn write_vlen_variable<T: NcWriteType>(
        &mut self,
        variable: VariableId,
        values: &[Vec<T>],
    ) -> Result<()> {
        {
            let variable_def = self.variable_mut(variable)?;
            let NcType::VLen { base } = &variable_def.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "VLen".into(),
                    actual: format!("{:?}", variable_def.dtype),
                });
            };
            let expected = T::nc_type();
            if base.as_ref() != &expected {
                return Err(Error::TypeMismatch {
                    expected: format!("{:?}", base),
                    actual: format!("{expected:?}"),
                });
            }

            let mut sequences = Vec::with_capacity(values.len());
            for sequence in values {
                let mut bytes = Vec::with_capacity(std::mem::size_of_val(sequence.as_slice()));
                for &value in sequence {
                    value.write_one_le(&mut bytes);
                }
                sequences.push(bytes);
            }
            variable_def.data.clear();
            variable_def.data_encoding = VariableDataEncoding::Hdf5Native;
            variable_def.string_values = None;
            variable_def.vlen_values = Some(sequences);
        }
        self.update_unlimited_extents_from_element_count(variable, values.len() as u64)
    }

    pub fn write_vlen_variable_slice_bytes<S: AsRef<[u8]>>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[S],
    ) -> Result<()> {
        let (resolved, base_size) = {
            let variable_def = self.variable(variable)?;
            let NcType::VLen { base } = &variable_def.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "VLen".into(),
                    actual: format!("{:?}", variable_def.dtype),
                });
            };
            validate_vlen_base_nc4_type(base)?;
            (
                self.resolve_variable_write_selection(variable_def, selection, 1)?,
                base.size()?,
            )
        };
        if values.len() != resolved.elements {
            return Err(Error::DataLengthMismatch {
                expected: resolved.elements,
                actual: values.len(),
            });
        }

        let mut sequences = Vec::with_capacity(values.len());
        for value in values {
            let value = value.as_ref();
            if value.len() % base_size != 0 {
                return Err(Error::InvalidDefinition(format!(
                    "vlen sequence byte length {} is not a multiple of base element size {base_size}",
                    value.len()
                )));
            }
            sequences.push(value.to_vec());
        }

        let strides = row_major_strides(&resolved.shape, "writer vlen slice stride")?;
        let total_elements =
            checked_shape_elements(&resolved.shape, "writer vlen variable element count")?;
        {
            let variable_def = self.variable_mut(variable)?;
            ensure_variable_vlen_slice_buffer(
                variable_def,
                &resolved.old_shape,
                &resolved.shape,
                total_elements,
                resolved.can_grow,
            )?;
            let vlen_values = variable_def.vlen_values.as_mut().ok_or_else(|| {
                Error::InvalidDefinition(format!(
                    "vlen variable '{}' has no initialized sequence data",
                    variable_def.name
                ))
            })?;
            scatter_slice_values(vlen_values, &resolved.dims, &strides, &sequences)?;
        }
        self.update_unlimited_extents_from_shape(variable, &resolved.shape)
    }

    pub fn write_vlen_variable_slice<T: NcWriteType>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[Vec<T>],
    ) -> Result<()> {
        let resolved = {
            let variable_def = self.variable(variable)?;
            let NcType::VLen { base } = &variable_def.dtype else {
                return Err(Error::TypeMismatch {
                    expected: "VLen".into(),
                    actual: format!("{:?}", variable_def.dtype),
                });
            };
            let expected = T::nc_type();
            if base.as_ref() != &expected {
                return Err(Error::TypeMismatch {
                    expected: format!("{:?}", base),
                    actual: format!("{expected:?}"),
                });
            }
            self.resolve_variable_write_selection(variable_def, selection, 1)?
        };
        if values.len() != resolved.elements {
            return Err(Error::DataLengthMismatch {
                expected: resolved.elements,
                actual: values.len(),
            });
        }

        let mut sequences = Vec::with_capacity(values.len());
        for sequence in values {
            let mut bytes = Vec::with_capacity(std::mem::size_of_val(sequence.as_slice()));
            for &value in sequence {
                value.write_one_le(&mut bytes);
            }
            sequences.push(bytes);
        }

        let strides = row_major_strides(&resolved.shape, "writer vlen slice stride")?;
        let total_elements =
            checked_shape_elements(&resolved.shape, "writer vlen variable element count")?;
        {
            let variable_def = self.variable_mut(variable)?;
            ensure_variable_vlen_slice_buffer(
                variable_def,
                &resolved.old_shape,
                &resolved.shape,
                total_elements,
                resolved.can_grow,
            )?;
            let vlen_values = variable_def.vlen_values.as_mut().ok_or_else(|| {
                Error::InvalidDefinition(format!(
                    "vlen variable '{}' has no initialized sequence data",
                    variable_def.name
                ))
            })?;
            scatter_slice_values(vlen_values, &resolved.dims, &strides, &sequences)?;
        }
        self.update_unlimited_extents_from_shape(variable, &resolved.shape)
    }

    pub fn write_string_variable<S: AsRef<str>>(
        &mut self,
        variable: VariableId,
        values: &[S],
    ) -> Result<()> {
        {
            let variable = self.variable_mut(variable)?;
            if variable.dtype != NcType::String {
                return Err(Error::TypeMismatch {
                    expected: "String".into(),
                    actual: format!("{:?}", variable.dtype),
                });
            }

            let mut strings = Vec::with_capacity(values.len());
            for value in values {
                let value = value.as_ref();
                if value.as_bytes().contains(&0) {
                    return Err(Error::InvalidDefinition(
                        "NC_STRING variable values cannot contain NUL bytes".into(),
                    ));
                }
                strings.push(value.to_string());
            }
            variable.data.clear();
            variable.data_encoding = VariableDataEncoding::Hdf5Native;
            variable.string_values = Some(strings);
            variable.vlen_values = None;
        }
        self.update_unlimited_extents_from_element_count(variable, values.len() as u64)
    }

    pub fn write_string_variable_slice<S: AsRef<str>>(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        values: &[S],
    ) -> Result<()> {
        let variable_def = self.variable(variable)?;
        if variable_def.dtype != NcType::String {
            return Err(Error::TypeMismatch {
                expected: "String".into(),
                actual: format!("{:?}", variable_def.dtype),
            });
        }
        let resolved = self.resolve_variable_write_selection(variable_def, selection, 1)?;
        if values.len() != resolved.elements {
            return Err(Error::DataLengthMismatch {
                expected: resolved.elements,
                actual: values.len(),
            });
        }

        let mut strings = Vec::with_capacity(values.len());
        for value in values {
            let value = value.as_ref();
            if value.as_bytes().contains(&0) {
                return Err(Error::InvalidDefinition(
                    "NC_STRING variable values cannot contain NUL bytes".into(),
                ));
            }
            strings.push(value.to_string());
        }

        let strides = row_major_strides(&resolved.shape, "writer string slice stride")?;
        let total_elements =
            checked_shape_elements(&resolved.shape, "writer string variable element count")?;
        {
            let variable_def = self.variable_mut(variable)?;
            ensure_variable_string_slice_buffer(
                variable_def,
                &resolved.old_shape,
                &resolved.shape,
                total_elements,
                resolved.can_grow,
            )?;
            let string_values = variable_def.string_values.as_mut().ok_or_else(|| {
                Error::InvalidDefinition(format!(
                    "NC_STRING variable '{}' has no initialized string data",
                    variable_def.name
                ))
            })?;
            scatter_slice_values(string_values, &resolved.dims, &strides, &strings)?;
        }
        self.update_unlimited_extents_from_shape(variable, &resolved.shape)
    }

    pub fn write<W: Write>(&self, mut writer: W, options: NcWriteOptions) -> Result<NcFormat> {
        if options.format == NcWriteFormat::AutoClassic {
            return self.write_auto_classic(&mut writer);
        }

        let format = self.select_format(options)?;
        match format {
            NcFormat::Classic | NcFormat::Offset64 | NcFormat::Cdf5 => {
                let plan = ClassicWritePlan::build(self, format)?;
                plan.write(&mut writer)?;
                Ok(format)
            }
            NcFormat::Nc4 | NcFormat::Nc4Classic => self.write_nc4(&mut writer, format),
        }
    }

    pub fn to_vec(&self, options: NcWriteOptions) -> Result<(NcFormat, Vec<u8>)> {
        let mut data = Vec::new();
        let format = self.write(&mut data, options)?;
        Ok((format, data))
    }

    fn write_auto_classic(&self, writer: &mut impl Write) -> Result<NcFormat> {
        if self.requires_cdf5() {
            self.validate_for_format(NcFormat::Cdf5)?;
            let plan = ClassicWritePlan::build(self, NcFormat::Cdf5)?;
            plan.write(writer)?;
            return Ok(NcFormat::Cdf5);
        }

        match ClassicWritePlan::build(self, NcFormat::Classic) {
            Ok(plan) => {
                plan.write(writer)?;
                Ok(NcFormat::Classic)
            }
            Err(classic_err) => match ClassicWritePlan::build(self, NcFormat::Offset64) {
                Ok(plan) => {
                    plan.write(writer)?;
                    Ok(NcFormat::Offset64)
                }
                Err(_) => Err(classic_err),
            },
        }
    }

    #[cfg(feature = "netcdf4")]
    fn write_nc4(&self, writer: &mut impl Write, format: NcFormat) -> Result<NcFormat> {
        self.validate_for_nc4_bridge(format)?;
        let hdf5_plan = self.build_hdf5_plan(format)?;
        // Encode straight to bytes and write them to the caller's sink, rather
        // than routing through an intermediate seekable in-memory writer.
        let bytes = hdf5_plan
            .encode(H5WriteOptions::default())
            .map_err(hdf5_error_to_unsupported)?;
        writer.write_all(&bytes)?;
        Ok(format)
    }

    #[cfg(not(feature = "netcdf4"))]
    fn write_nc4(&self, _writer: &mut impl Write, _format: NcFormat) -> Result<NcFormat> {
        Err(Error::UnsupportedFeature(
            "NetCDF-4 writing requires the netcdf4 feature".into(),
        ))
    }

    #[cfg(feature = "netcdf4")]
    fn validate_for_nc4_bridge(&self, format: NcFormat) -> Result<()> {
        if format == NcFormat::Nc4Classic {
            validate_nc4_classic_model(self)?;
            for variable in &self.variables {
                validate_classic_type(&variable.dtype)?;
                for attr in &variable.attributes {
                    validate_nc4_classic_attr_value(&attr.value)?;
                }
            }
            for attr in &self.attributes {
                validate_nc4_classic_attr_value(&attr.value)?;
            }
        }
        let dimension_sizes = self.nc4_dimension_sizes()?;
        for variable in &self.variables {
            validate_nc4_variable_data(variable, &dimension_sizes)?;
        }
        for (dim_id, dimension) in self.dimensions.iter().enumerate() {
            let dimension_id = DimensionId(dim_id);
            if self
                .coordinate_variable_for_dimension(dimension_id)
                .is_none()
                && self
                    .variables
                    .iter()
                    .any(|variable| variable.name == dimension.name)
            {
                return Err(Error::UnsupportedFeature(format!(
                    "NetCDF-4 dimension '{}' needs a hidden dimension-scale dataset, but a scalar variable already uses that name",
                    dimension.name
                )));
            }
        }
        Ok(())
    }

    #[cfg(feature = "netcdf4")]
    fn build_hdf5_plan(&self, format: NcFormat) -> Result<hdf5_writer::Hdf5WritePlan> {
        let mut builder = Hdf5Builder::new();
        let dimension_sizes = self.nc4_dimension_sizes()?;
        if format == NcFormat::Nc4Classic {
            builder = builder.attribute(
                H5AttributeBuilder::scalar("_nc3_strict", 1_i32)
                    .map_err(hdf5_error_to_unsupported)?,
            );
        }
        for attr in &self.attributes {
            builder = builder.attribute(nc_attr_to_hdf5(attr)?);
        }
        for group_attr in &self.group_attributes {
            builder = builder.group_attribute(
                group_attr.group_path.as_str(),
                nc_attr_to_hdf5(&group_attr.attribute)?,
            );
        }

        for (dim_id, dimension) in self.dimensions.iter().enumerate() {
            if self
                .coordinate_variable_for_dimension(DimensionId(dim_id))
                .is_some()
            {
                continue;
            }
            let size = dimension_sizes[dim_id];
            let zeros = vec![0_i32; checked_usize(size, "dimension scale size")?];
            let mut dataset = H5DatasetBuilder::typed_data(&dimension.name, vec![size], &zeros)
                .map_err(hdf5_error_to_unsupported)?
                .attribute(H5AttributeBuilder::fixed_string("CLASS", "DIMENSION_SCALE"))
                .attribute(H5AttributeBuilder::fixed_string(
                    "NAME",
                    format!(
                        "This is a netCDF dimension but not a netCDF variable. {}",
                        path_leaf_name(&dimension.name)
                    ),
                ))
                .attribute(
                    H5AttributeBuilder::scalar("_Netcdf4Dimid", dim_id as i32)
                        .map_err(hdf5_error_to_unsupported)?,
                );
            if let Some(reference_list) = self.reference_list_attribute(DimensionId(dim_id))? {
                dataset = dataset.attribute(reference_list);
            }
            if dimension.is_unlimited {
                dataset = dataset
                    .max_shape(vec![H5_UNLIMITED])
                    .chunked(nc4_default_chunk_shape(
                        &[size],
                        std::mem::size_of::<i32>(),
                    )?);
            }
            builder = builder.dataset(dataset);
        }

        for variable in &self.variables {
            let shape = variable
                .dim_ids
                .iter()
                .map(|id| dimension_sizes[id.0])
                .collect::<Vec<_>>();
            let mut dataset = if variable.dtype == NcType::String {
                let values = variable.string_values.as_ref().ok_or_else(|| {
                    Error::InvalidDefinition(format!(
                        "NC_STRING variable '{}' has no string data",
                        variable.name
                    ))
                })?;
                H5DatasetBuilder::vlen_string_data(&variable.name, shape.clone(), values)
                    .map_err(hdf5_error_to_unsupported)?
            } else if let NcType::VLen { base } = &variable.dtype {
                let values = variable.vlen_values.as_ref().ok_or_else(|| {
                    Error::InvalidDefinition(format!(
                        "NC_VLEN variable '{}' has no sequence data",
                        variable.name
                    ))
                })?;
                H5DatasetBuilder::vlen_sequence_data(
                    &variable.name,
                    nc_type_to_hdf5(base)?,
                    shape.clone(),
                    values.clone(),
                )
                .map_err(hdf5_error_to_unsupported)?
            } else {
                let data = match variable.data_encoding {
                    VariableDataEncoding::ClassicBigEndian => {
                        convert_classic_be_data_to_hdf5_le(&variable.dtype, &variable.data)?
                    }
                    VariableDataEncoding::Hdf5Native => variable.data.clone(),
                };
                let dataset = H5DatasetBuilder::new(
                    &variable.name,
                    nc_type_to_hdf5(&variable.dtype)?,
                    shape.clone(),
                );
                if data.is_empty() && variable.fill_value.is_some() {
                    dataset
                } else {
                    dataset.raw_data(data)
                }
            };
            if let Some(fill_value) = &variable.fill_value {
                dataset = dataset.fill_value(fill_value.hdf5_bytes.clone());
            }
            let chunk_shape = nc4_variable_chunk_shape(
                variable,
                &shape,
                self.nc4_variable_max_shape(variable).is_some(),
            )?;
            if let Some(max_shape) = self.nc4_variable_max_shape(variable) {
                dataset = dataset.max_shape(max_shape);
            }
            if let Some(chunk_shape) = chunk_shape {
                dataset = dataset.chunked(chunk_shape);
            }
            for filter in nc4_variable_filters(variable) {
                dataset = dataset.filter(filter);
            }

            if self.variable_is_coordinate_scale(variable)? {
                let dim_id = variable.dim_ids[0];
                dataset = dataset
                    .attribute(H5AttributeBuilder::fixed_string("CLASS", "DIMENSION_SCALE"))
                    .attribute(H5AttributeBuilder::fixed_string(
                        "NAME",
                        path_leaf_name(&self.dimensions[dim_id.0].name),
                    ))
                    .attribute(
                        H5AttributeBuilder::scalar("_Netcdf4Dimid", dim_id.0 as i32)
                            .map_err(hdf5_error_to_unsupported)?,
                    );
                if let Some(reference_list) = self.reference_list_attribute(dim_id)? {
                    dataset = dataset.attribute(reference_list);
                }
            } else if variable.dim_ids.is_empty() {
                dataset = dataset.attribute(empty_dimension_list_attribute());
            } else {
                dataset = dataset.attribute(self.dimension_list_attribute(variable)?);
            }

            for attr in &variable.attributes {
                dataset = dataset.attribute(nc_attr_to_hdf5(attr)?);
            }

            builder = builder.dataset(dataset);
        }

        builder.into_plan().map_err(hdf5_error_to_unsupported)
    }

    #[cfg(feature = "netcdf4")]
    fn nc4_dimension_sizes(&self) -> Result<Vec<u64>> {
        infer_nc4_dimension_sizes(self)
    }

    #[cfg(feature = "netcdf4")]
    fn nc4_variable_max_shape(&self, variable: &VariableDef) -> Option<Vec<u64>> {
        variable
            .dim_ids
            .iter()
            .any(|id| self.dimensions[id.0].is_unlimited)
            .then(|| {
                variable
                    .dim_ids
                    .iter()
                    .map(|id| {
                        let dimension = &self.dimensions[id.0];
                        if dimension.is_unlimited {
                            H5_UNLIMITED
                        } else {
                            dimension.size
                        }
                    })
                    .collect()
            })
    }

    #[cfg(feature = "netcdf4")]
    fn variable_is_coordinate_scale(&self, variable: &VariableDef) -> Result<bool> {
        Ok(match variable.dim_ids.as_slice() {
            [dim_id] => self.dimension(*dim_id)?.name == variable.name,
            _ => false,
        })
    }

    #[cfg(feature = "netcdf4")]
    fn coordinate_variable_for_dimension(&self, dimension: DimensionId) -> Option<&VariableDef> {
        let dim = self.dimensions.get(dimension.0)?;
        self.variables.iter().find(|variable| {
            variable.name == dim.name && variable.dim_ids.as_slice() == [dimension]
        })
    }

    /// Build the `REFERENCE_LIST` back-reference attribute for a dimension
    /// scale: one entry per (variable, dimension position) that references the
    /// dimension, excluding the scale's own coordinate variable. Returns `None`
    /// when no data variable references the dimension (netcdf-c omits the
    /// attribute in that case).
    #[cfg(feature = "netcdf4")]
    fn reference_list_attribute(&self, dim_id: DimensionId) -> Result<Option<H5AttributeBuilder>> {
        let dim_name = &self.dimensions[dim_id.0].name;
        let mut entries = Vec::new();
        for variable in &self.variables {
            let is_scale_itself =
                variable.name == *dim_name && variable.dim_ids.as_slice() == [dim_id];
            if is_scale_itself {
                continue;
            }
            for (position, vid) in variable.dim_ids.iter().enumerate() {
                if *vid == dim_id {
                    entries.push((variable.name.clone(), position as u32));
                }
            }
        }
        Ok((!entries.is_empty())
            .then(|| H5AttributeBuilder::object_reference_list("REFERENCE_LIST", entries)))
    }

    #[cfg(feature = "netcdf4")]
    fn dimension_list_attribute(&self, variable: &VariableDef) -> Result<H5AttributeBuilder> {
        let target_sequences = variable
            .dim_ids
            .iter()
            .map(|dimension| Ok(vec![self.dimension(*dimension)?.name.clone()]))
            .collect::<Result<Vec<_>>>()?;
        Ok(H5AttributeBuilder::vlen_object_references(
            "DIMENSION_LIST",
            target_sequences,
        ))
    }

    fn add_dimension_def(
        &mut self,
        name: String,
        size: u64,
        is_unlimited: bool,
    ) -> Result<DimensionId> {
        if !is_unlimited && size == 0 {
            return Err(Error::InvalidDefinition(
                "fixed dimensions must have non-zero size; use add_unlimited_dimension".into(),
            ));
        }
        self.ensure_unique_dimension(&name)?;
        let id = DimensionId(self.dimensions.len());
        self.dimensions.push(DimensionDef {
            name,
            size,
            is_unlimited,
            current_size: if is_unlimited { 0 } else { size },
        });
        Ok(id)
    }

    fn add_variable_with_type(
        &mut self,
        name: impl Into<String>,
        dimensions: &[DimensionId],
        dtype: NcType,
    ) -> Result<VariableId> {
        let name = name.into();
        validate_name(&name, "variable")?;
        self.add_variable_def(name, dimensions, dtype)
    }

    fn add_variable_path_with_type(
        &mut self,
        path: impl Into<String>,
        dimensions: &[DimensionId],
        dtype: NcType,
    ) -> Result<VariableId> {
        let path = validate_path_name(path.into(), "variable")?;
        self.add_variable_def(path, dimensions, dtype)
    }

    fn add_variable_def(
        &mut self,
        name: String,
        dimensions: &[DimensionId],
        dtype: NcType,
    ) -> Result<VariableId> {
        if self.variables.iter().any(|v| v.name == name) {
            return Err(Error::InvalidDefinition(format!(
                "duplicate variable '{name}'"
            )));
        }
        for dim in dimensions {
            self.dimension(*dim)?;
        }
        let id = VariableId(self.variables.len());
        self.variables.push(VariableDef {
            name,
            dim_ids: dimensions.to_vec(),
            dtype,
            attributes: Vec::new(),
            data: Vec::new(),
            data_encoding: VariableDataEncoding::ClassicBigEndian,
            string_values: None,
            vlen_values: None,
            storage: VariableStorageDef::default(),
            fill_value: None,
        });
        Ok(id)
    }

    fn dimension(&self, id: DimensionId) -> Result<&DimensionDef> {
        self.dimensions
            .get(id.0)
            .ok_or_else(|| Error::InvalidDefinition("invalid dimension handle".into()))
    }

    fn variable(&self, id: VariableId) -> Result<&VariableDef> {
        self.variables
            .get(id.0)
            .ok_or_else(|| Error::InvalidDefinition("invalid variable handle".into()))
    }

    fn variable_mut(&mut self, id: VariableId) -> Result<&mut VariableDef> {
        self.variables
            .get_mut(id.0)
            .ok_or_else(|| Error::InvalidDefinition("invalid variable handle".into()))
    }

    fn resolve_variable_write_selection(
        &self,
        variable: &VariableDef,
        selection: &NcSliceInfo,
        elem_size: usize,
    ) -> Result<ResolvedWriteSelection> {
        let extents = self.variable_write_extents(variable, elem_size)?;
        resolve_write_selection(&variable.name, &extents, selection)
    }

    fn resolve_char_string_write_selection(
        &self,
        variable: &VariableDef,
        selection: &NcSliceInfo,
    ) -> Result<ResolvedWriteSelection> {
        if variable.dim_ids.is_empty() {
            return Err(Error::InvalidDefinition(format!(
                "char variable '{}' must have a string-width dimension",
                variable.name
            )));
        }
        let extents = self.variable_write_extents(variable, 1)?;
        resolve_write_selection(&variable.name, &extents[..extents.len() - 1], selection)
    }

    fn variable_write_extents(
        &self,
        variable: &VariableDef,
        _elem_size: usize,
    ) -> Result<Vec<WriteDimensionExtent>> {
        variable
            .dim_ids
            .iter()
            .enumerate()
            .map(|(position, id)| {
                let dimension = &self.dimensions[id.0];
                Ok(WriteDimensionExtent {
                    current: if dimension.is_unlimited {
                        dimension.current_size
                    } else {
                        dimension.size
                    },
                    is_unlimited: dimension.is_unlimited,
                    position,
                })
            })
            .collect()
    }

    fn char_string_axis_width(&self, variable: &VariableDef) -> Result<usize> {
        let width_dim_id = variable.dim_ids.last().ok_or_else(|| {
            Error::InvalidDefinition(format!(
                "char variable '{}' must have a string-width dimension",
                variable.name
            ))
        })?;
        let dimension = &self.dimensions[width_dim_id.0];
        let width = if dimension.is_unlimited {
            dimension.current_size
        } else {
            dimension.size
        };
        if dimension.is_unlimited && width == 0 {
            return Err(Error::InvalidDefinition(format!(
                "char variable '{}' string-width dimension cannot be inferred from string values",
                variable.name
            )));
        }
        require_usize(width, "char string width")
    }

    fn validate_char_string_value_count(
        &self,
        variable: &VariableDef,
        value_count: usize,
    ) -> Result<Option<(DimensionId, u64)>> {
        if variable.dim_ids.is_empty() {
            return Err(Error::InvalidDefinition(format!(
                "char variable '{}' must have a string-width dimension",
                variable.name
            )));
        }

        let value_count = u64::try_from(value_count).map_err(|_| {
            Error::InvalidDefinition("char string value count exceeds u64 capacity".into())
        })?;
        let mut known_product = 1u64;
        let mut unknown = Vec::new();
        for dim_id in &variable.dim_ids[..variable.dim_ids.len() - 1] {
            let dimension = &self.dimensions[dim_id.0];
            let size = if dimension.is_unlimited {
                dimension.current_size
            } else {
                dimension.size
            };
            if dimension.is_unlimited && size == 0 {
                if !unknown.contains(dim_id) {
                    unknown.push(*dim_id);
                }
                continue;
            }
            known_product = checked_mul_u64(known_product, size, "char string value count shape")?;
        }

        let mut inferred_unlimited = None;
        if unknown.is_empty() {
            if value_count != known_product {
                return Err(Error::DataLengthMismatch {
                    expected: require_usize(known_product, "char string value count")?,
                    actual: require_usize(value_count, "char string value count")?,
                });
            }
        } else if unknown.len() == 1 {
            if known_product == 0 {
                if value_count != 0 {
                    return Err(Error::DataLengthMismatch {
                        expected: 0,
                        actual: require_usize(value_count, "char string value count")?,
                    });
                }
            } else if value_count % known_product != 0 {
                return Err(Error::InvalidDefinition(format!(
                    "char string value count {value_count} is not divisible by known leading dimension product {known_product}"
                )));
            } else {
                inferred_unlimited = Some((unknown[0], value_count / known_product));
            }
        } else {
            let names = unknown
                .iter()
                .map(|id| self.dimensions[id.0].name.as_str())
                .collect::<Vec<_>>()
                .join(", ");
            return Err(Error::InvalidDefinition(format!(
                "cannot infer multiple unlimited char string dimensions {names} for variable '{}'",
                variable.name
            )));
        }

        Ok(inferred_unlimited)
    }

    fn update_unlimited_extents_from_element_count(
        &mut self,
        variable: VariableId,
        elements: u64,
    ) -> Result<()> {
        let variable = self.variable(variable)?;
        let mut known_product = 1u64;
        let mut unknown = Vec::new();
        for dim_id in &variable.dim_ids {
            let dimension = &self.dimensions[dim_id.0];
            let size = if dimension.is_unlimited {
                dimension.current_size
            } else {
                dimension.size
            };
            if dimension.is_unlimited && size == 0 {
                if !unknown.contains(dim_id) {
                    unknown.push(*dim_id);
                }
                continue;
            }
            known_product = match known_product.checked_mul(size) {
                Some(product) => product,
                None => return Ok(()),
            };
        }

        if unknown.len() != 1 || known_product == 0 {
            return Ok(());
        }
        if elements % known_product != 0 {
            return Ok(());
        }
        self.dimensions[unknown[0].0].current_size = elements / known_product;
        Ok(())
    }

    fn update_unlimited_extents_from_shape(
        &mut self,
        variable: VariableId,
        shape: &[u64],
    ) -> Result<()> {
        let dim_ids = self.variable(variable)?.dim_ids.clone();
        if dim_ids.len() != shape.len() {
            return Err(Error::InvalidDefinition(format!(
                "resolved shape rank {} does not match variable '{}' rank {}",
                shape.len(),
                self.variable(variable)?.name,
                dim_ids.len()
            )));
        }
        for (dim_id, &size) in dim_ids.iter().zip(shape) {
            let dimension = &mut self.dimensions[dim_id.0];
            if dimension.is_unlimited && size > dimension.current_size {
                dimension.current_size = size;
            }
        }
        Ok(())
    }

    fn write_native_variable_slice_bytes(
        &mut self,
        variable: VariableId,
        selection: &NcSliceInfo,
        bytes: &[u8],
        elem_size: usize,
    ) -> Result<()> {
        let variable_def = self.variable(variable)?;
        let resolved = self.resolve_variable_write_selection(variable_def, selection, elem_size)?;
        let expected = checked_mul_usize(
            resolved.elements,
            elem_size,
            "native variable slice byte size",
        )?;
        if bytes.len() != expected {
            return Err(Error::DataLengthMismatch {
                expected,
                actual: bytes.len(),
            });
        }

        let strides = row_major_strides(&resolved.shape, "writer native slice stride")?;
        let total_elements =
            checked_shape_elements(&resolved.shape, "writer native variable element count")?;
        {
            let variable_def = self.variable_mut(variable)?;
            ensure_variable_native_slice_buffer(
                variable_def,
                &resolved.old_shape,
                &resolved.shape,
                total_elements,
                elem_size,
                resolved.can_grow,
            )?;
            scatter_slice_bytes(
                &mut variable_def.data,
                elem_size,
                &resolved.dims,
                &strides,
                bytes,
            )?;
        }
        self.update_unlimited_extents_from_shape(variable, &resolved.shape)
    }

    fn ensure_unique_dimension(&self, name: &str) -> Result<()> {
        if self.dimensions.iter().any(|d| d.name == name) {
            return Err(Error::InvalidDefinition(format!(
                "duplicate dimension '{name}'"
            )));
        }
        Ok(())
    }

    fn select_format(&self, options: NcWriteOptions) -> Result<NcFormat> {
        match options.format {
            NcWriteFormat::Classic => {
                self.validate_for_format(NcFormat::Classic)?;
                Ok(NcFormat::Classic)
            }
            NcWriteFormat::Offset64 => {
                self.validate_for_format(NcFormat::Offset64)?;
                Ok(NcFormat::Offset64)
            }
            NcWriteFormat::Cdf5 => {
                self.validate_for_format(NcFormat::Cdf5)?;
                Ok(NcFormat::Cdf5)
            }
            NcWriteFormat::Nc4 => Ok(NcFormat::Nc4),
            NcWriteFormat::Nc4Classic => Ok(NcFormat::Nc4Classic),
            NcWriteFormat::AutoClassic => {
                let preferred = if self.requires_cdf5() {
                    NcFormat::Cdf5
                } else {
                    NcFormat::Classic
                };
                match self.validate_for_format(preferred) {
                    Ok(()) => Ok(preferred),
                    Err(_) if preferred == NcFormat::Classic => {
                        self.validate_for_format(NcFormat::Offset64)?;
                        Ok(NcFormat::Offset64)
                    }
                    Err(err) => Err(err),
                }
            }
        }
    }

    fn requires_cdf5(&self) -> bool {
        self.dimensions.iter().any(|d| d.size > u32::MAX as u64)
            || self.variables.iter().any(|v| {
                matches!(
                    v.dtype,
                    NcType::UByte | NcType::UShort | NcType::UInt | NcType::Int64 | NcType::UInt64
                )
            })
            || self.attributes.iter().any(|a| attr_requires_cdf5(&a.value))
            || self
                .group_attributes
                .iter()
                .any(|a| attr_requires_cdf5(&a.attribute.value))
            || self
                .variables
                .iter()
                .flat_map(|v| &v.attributes)
                .any(|a| attr_requires_cdf5(&a.value))
    }

    fn validate_for_format(&self, format: NcFormat) -> Result<()> {
        if !matches!(
            format,
            NcFormat::Classic | NcFormat::Offset64 | NcFormat::Cdf5
        ) {
            return Err(Error::UnsupportedFeature(
                "only classic-family NetCDF formats are implemented".into(),
            ));
        }
        let unlimited_count = self.dimensions.iter().filter(|d| d.is_unlimited).count();
        if unlimited_count > 1 {
            return Err(Error::RequiresNetcdf4 {
                reason: "classic NetCDF supports at most one unlimited dimension".into(),
            });
        }
        if format != NcFormat::Cdf5 && self.requires_cdf5() {
            return Err(Error::FormatCapacityExceeded {
                reason: "CDF-5 is required for unsigned integer, 64-bit integer, or 64-bit count \
                         data"
                    .into(),
            });
        }
        validate_root_only_names(self)?;
        for variable in &self.variables {
            if variable.storage.has_nc4_options() {
                return Err(Error::UnsupportedFeature(format!(
                    "variable '{}' uses NetCDF-4 storage options",
                    variable.name
                )));
            }
            validate_classic_type(&variable.dtype)?;
            let is_record = variable_is_record(self, variable)?;
            if variable
                .dim_ids
                .iter()
                .skip(1)
                .any(|id| self.dimensions[id.0].is_unlimited)
            {
                return Err(Error::InvalidDefinition(format!(
                    "record dimension must be first for variable '{}'",
                    variable.name
                )));
            }
            let expected = expected_variable_bytes(self, variable, is_record)?;
            if variable.data.len() != expected {
                return Err(Error::DataLengthMismatch {
                    expected,
                    actual: variable.data.len(),
                });
            }
        }
        Ok(())
    }
}

#[derive(Debug, Clone)]
struct PlannedVar {
    def: VariableDef,
    is_record: bool,
    /// Header `vsize` field: the per-record (or fixed) byte size, always
    /// rounded up to a multiple of 4 as the classic spec requires.
    header_vsize: u64,
    /// Bytes this variable actually occupies per record (or in total for
    /// fixed variables) in the data layout. Equal to `header_vsize` except
    /// for the spec's lone-record-variable case, where records are packed
    /// without padding.
    layout_vsize: u64,
    begin: u64,
    fixed_data_len: u64,
    record_slab_len: u64,
}

#[derive(Debug, Clone)]
struct ClassicWritePlan {
    num_records: u64,
    header: Vec<u8>,
    fixed_vars: Vec<PlannedVar>,
    record_vars: Vec<PlannedVar>,
}

impl ClassicWritePlan {
    fn build(builder: &NcFileBuilder, format: NcFormat) -> Result<Self> {
        builder.validate_for_format(format)?;
        let num_records = infer_num_records(builder)?;
        let mut planned = plan_variables(builder, 0)?;
        let header_without_offsets = encode_header(builder, format, num_records, &planned)?;
        planned = plan_variables(builder, header_without_offsets.len() as u64)?;
        let header = encode_header(builder, format, num_records, &planned)?;
        let planned = plan_variables(builder, header.len() as u64)?;
        let header = encode_header(builder, format, num_records, &planned)?;

        if format == NcFormat::Classic {
            for var in &planned {
                require_u32(var.begin, "CDF-1 variable offset")?;
            }
        }
        if matches!(format, NcFormat::Classic | NcFormat::Offset64) {
            require_u32(num_records, "record count")?;
        }

        let fixed_vars = planned.iter().filter(|v| !v.is_record).cloned().collect();
        let record_vars = planned.iter().filter(|v| v.is_record).cloned().collect();
        Ok(Self {
            num_records,
            header,
            fixed_vars,
            record_vars,
        })
    }

    fn write(&self, writer: &mut impl Write) -> Result<()> {
        writer.write_all(&self.header)?;
        for var in &self.fixed_vars {
            writer.write_all(&var.def.data)?;
            write_fill_padding(
                writer,
                var.layout_vsize - var.fixed_data_len,
                &variable_fill_pattern(&var.def)?,
            )?;
        }
        for record in 0..self.num_records {
            for var in &self.record_vars {
                let start = usize::try_from(record * var.record_slab_len).map_err(|_| {
                    Error::InvalidDefinition("record byte offset exceeds platform usize".into())
                })?;
                let slab_len = usize::try_from(var.record_slab_len).map_err(|_| {
                    Error::InvalidDefinition("record slab length exceeds platform usize".into())
                })?;
                writer.write_all(&var.def.data[start..start + slab_len])?;
                write_fill_padding(
                    writer,
                    var.layout_vsize - var.record_slab_len,
                    &variable_fill_pattern(&var.def)?,
                )?;
            }
        }
        Ok(())
    }
}

/// The classic-encoded fill pattern used to pad a variable's data out to its
/// padded size, matching netcdf-c (which fills the slack with fill values).
/// The HDF5-native (little-endian) fill pattern used to initialize gaps when a
/// native-encoded variable buffer is created or grown, mirroring netcdf-c. An
/// explicit fill value wins; otherwise primitive types use their default fill
/// and user-defined types (which have no primitive default) use zeros.
fn native_fill_pattern(def: &VariableDef, elem_size: usize) -> Result<Vec<u8>> {
    if let Some(fill_value) = &def.fill_value {
        return Ok(fill_value.hdf5_bytes.clone());
    }
    match default_classic_fill_bytes(&def.dtype) {
        Ok(classic) => convert_classic_be_data_to_hdf5_le(&def.dtype, &classic),
        Err(_) => Ok(vec![0; elem_size]),
    }
}

fn variable_fill_pattern(def: &VariableDef) -> Result<Vec<u8>> {
    match &def.fill_value {
        Some(fill_value) => Ok(fill_value.classic_bytes.clone()),
        None => default_classic_fill_bytes(&def.dtype),
    }
}

fn plan_variables(builder: &NcFileBuilder, data_start: u64) -> Result<Vec<PlannedVar>> {
    let mut fixed_offset = data_start;
    let mut planned = Vec::with_capacity(builder.variables.len());
    for def in &builder.variables {
        let is_record = variable_is_record(builder, def)?;
        let elem_size = def.dtype.size()? as u64;
        let record_slab_len = if is_record {
            variable_shape_elements(builder, def, 1)? * elem_size
        } else {
            0
        };
        let fixed_data_len = if is_record {
            0
        } else {
            variable_shape_elements(builder, def, 0)? * elem_size
        };
        let vsize = if is_record {
            record_slab_len
        } else {
            fixed_data_len
        };
        let header_vsize = pad4_u64(vsize)?;
        let begin = if is_record { 0 } else { fixed_offset };
        if !is_record {
            fixed_offset = checked_add(fixed_offset, header_vsize, "fixed data offset")?;
        }
        planned.push(PlannedVar {
            def: def.clone(),
            is_record,
            header_vsize,
            layout_vsize: header_vsize,
            begin,
            fixed_data_len,
            record_slab_len,
        });
    }

    // The classic spec packs records without padding when the file has
    // exactly one record variable; the header vsize stays padded.
    let record_var_count = planned.iter().filter(|var| var.is_record).count();
    if record_var_count == 1 {
        for var in &mut planned {
            if var.is_record {
                var.layout_vsize = var.record_slab_len;
            }
        }
    }

    let record_data_start = fixed_offset;
    let mut record_offset = record_data_start;
    for var in &mut planned {
        if var.is_record {
            var.begin = record_offset;
            record_offset = checked_add(record_offset, var.layout_vsize, "record offset")?;
        }
    }
    Ok(planned)
}

fn encode_header(
    builder: &NcFileBuilder,
    format: NcFormat,
    num_records: u64,
    planned: &[PlannedVar],
) -> Result<Vec<u8>> {
    let mut out = Vec::new();
    out.extend_from_slice(match format {
        NcFormat::Classic => b"CDF\x01",
        NcFormat::Offset64 => b"CDF\x02",
        NcFormat::Cdf5 => b"CDF\x05",
        _ => unreachable!("classic header only"),
    });
    write_count(&mut out, format, num_records)?;
    encode_dimensions(&mut out, builder, format)?;
    encode_attributes(&mut out, format, &builder.attributes)?;
    encode_variables(&mut out, builder, format, planned)?;
    Ok(out)
}

fn encode_dimensions(out: &mut Vec<u8>, builder: &NcFileBuilder, format: NcFormat) -> Result<()> {
    if builder.dimensions.is_empty() {
        out.extend_from_slice(&ABSENT.to_be_bytes());
        write_count(out, format, 0)?;
        return Ok(());
    }
    out.extend_from_slice(&NC_DIMENSION.to_be_bytes());
    write_count(out, format, builder.dimensions.len() as u64)?;
    for dim in &builder.dimensions {
        write_name(out, format, &dim.name)?;
        write_count(out, format, if dim.is_unlimited { 0 } else { dim.size })?;
    }
    Ok(())
}

fn encode_attributes(out: &mut Vec<u8>, format: NcFormat, attrs: &[NcAttribute]) -> Result<()> {
    if attrs.is_empty() {
        out.extend_from_slice(&ABSENT.to_be_bytes());
        write_count(out, format, 0)?;
        return Ok(());
    }
    out.extend_from_slice(&NC_ATTRIBUTE.to_be_bytes());
    write_count(out, format, attrs.len() as u64)?;
    for attr in attrs {
        write_name(out, format, &attr.name)?;
        let (dtype, count, bytes) = encode_attr_value(&attr.value)?;
        write_u32(
            out,
            dtype.classic_type_code().ok_or_else(|| {
                Error::InvalidDefinition(format!("{dtype:?} is not valid in classic attributes"))
            })?,
        );
        write_count(out, format, count)?;
        out.extend_from_slice(&bytes);
        pad_vec_to_4(out);
    }
    Ok(())
}

fn encode_variables(
    out: &mut Vec<u8>,
    builder: &NcFileBuilder,
    format: NcFormat,
    planned: &[PlannedVar],
) -> Result<()> {
    if planned.is_empty() {
        out.extend_from_slice(&ABSENT.to_be_bytes());
        write_count(out, format, 0)?;
        return Ok(());
    }
    out.extend_from_slice(&NC_VARIABLE.to_be_bytes());
    write_count(out, format, planned.len() as u64)?;
    for var in planned {
        write_name(out, format, &var.def.name)?;
        write_count(out, format, var.def.dim_ids.len() as u64)?;
        for dim_id in &var.def.dim_ids {
            write_count(out, format, dim_id.0 as u64)?;
        }
        encode_attributes(out, format, &var.def.attributes)?;
        write_u32(
            out,
            var.def.dtype.classic_type_code().ok_or_else(|| {
                Error::InvalidDefinition(format!(
                    "{:?} is not valid in classic variables",
                    var.def.dtype
                ))
            })?,
        );
        // CDF-1/2 store vsize as a 32-bit field; the spec reserves 2^32 - 1
        // as the marker for sizes that exceed it (readers recompute from the
        // dimensions anyway).
        let header_vsize = match format {
            NcFormat::Classic | NcFormat::Offset64 if var.header_vsize > u64::from(u32::MAX) => {
                u64::from(u32::MAX)
            }
            _ => var.header_vsize,
        };
        write_count(out, format, header_vsize)?;
        match format {
            NcFormat::Classic => write_u32(out, require_u32(var.begin, "CDF-1 begin")?),
            NcFormat::Offset64 | NcFormat::Cdf5 => write_u64(out, var.begin),
            _ => unreachable!("classic header only"),
        }
        let _ = builder;
    }
    Ok(())
}

fn encode_attr_value(value: &NcAttrValue) -> Result<(NcType, u64, Vec<u8>)> {
    let mut out = Vec::new();
    let (dtype, count) = match value {
        NcAttrValue::Bytes(values) => {
            out.extend(values.iter().map(|v| *v as u8));
            (NcType::Byte, values.len() as u64)
        }
        NcAttrValue::Chars(value) => {
            out.extend_from_slice(value.as_bytes());
            (NcType::Char, value.len() as u64)
        }
        NcAttrValue::Shorts(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::Short, values.len() as u64)
        }
        NcAttrValue::Ints(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::Int, values.len() as u64)
        }
        NcAttrValue::Floats(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::Float, values.len() as u64)
        }
        NcAttrValue::Doubles(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::Double, values.len() as u64)
        }
        NcAttrValue::UBytes(values) => {
            out.extend_from_slice(values);
            (NcType::UByte, values.len() as u64)
        }
        NcAttrValue::UShorts(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::UShort, values.len() as u64)
        }
        NcAttrValue::UInts(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::UInt, values.len() as u64)
        }
        NcAttrValue::Int64s(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::Int64, values.len() as u64)
        }
        NcAttrValue::UInt64s(values) => {
            for value in values {
                out.extend_from_slice(&value.to_be_bytes());
            }
            (NcType::UInt64, values.len() as u64)
        }
        NcAttrValue::Strings(_) => {
            return Err(Error::UnsupportedFeature(
                "NC_STRING attributes require NetCDF-4".into(),
            ));
        }
    };
    Ok((dtype, count, out))
}

#[cfg(feature = "netcdf4")]
fn nc_type_to_hdf5(dtype: &NcType) -> Result<H5Datatype> {
    let byte_order = H5ByteOrder::LittleEndian;
    match dtype {
        NcType::Byte => Ok(H5Datatype::FixedPoint {
            size: 1,
            signed: true,
            byte_order,
        }),
        NcType::UByte => Ok(H5Datatype::FixedPoint {
            size: 1,
            signed: false,
            byte_order,
        }),
        NcType::Short => Ok(H5Datatype::FixedPoint {
            size: 2,
            signed: true,
            byte_order,
        }),
        NcType::UShort => Ok(H5Datatype::FixedPoint {
            size: 2,
            signed: false,
            byte_order,
        }),
        NcType::Int => Ok(H5Datatype::FixedPoint {
            size: 4,
            signed: true,
            byte_order,
        }),
        NcType::UInt => Ok(H5Datatype::FixedPoint {
            size: 4,
            signed: false,
            byte_order,
        }),
        NcType::Int64 => Ok(H5Datatype::FixedPoint {
            size: 8,
            signed: true,
            byte_order,
        }),
        NcType::UInt64 => Ok(H5Datatype::FixedPoint {
            size: 8,
            signed: false,
            byte_order,
        }),
        NcType::Float => Ok(H5Datatype::FloatingPoint {
            size: 4,
            byte_order,
        }),
        NcType::Double => Ok(H5Datatype::FloatingPoint {
            size: 8,
            byte_order,
        }),
        NcType::Char => Ok(H5Datatype::String {
            size: H5StringSize::Fixed(1),
            encoding: H5StringEncoding::Ascii,
            padding: H5StringPadding::NullPad,
        }),
        NcType::String => Ok(H5Datatype::String {
            size: H5StringSize::Variable,
            encoding: H5StringEncoding::Utf8,
            padding: H5StringPadding::NullTerminate,
        }),
        NcType::Enum { base, members } => {
            let h5_base = nc_type_to_hdf5(base)?;
            let members = members
                .iter()
                .map(|member| {
                    Ok(H5EnumMember {
                        name: member.name.clone(),
                        value: nc_enum_value_to_hdf5(base, member.value)?,
                    })
                })
                .collect::<Result<Vec<_>>>()?;
            Ok(H5Datatype::Enum {
                base: Box::new(h5_base),
                members,
            })
        }
        NcType::Compound { size, fields } => {
            let fields = fields
                .iter()
                .map(|field| {
                    let byte_offset = u32::try_from(field.offset).map_err(|_| {
                        Error::UnsupportedFeature(format!(
                            "compound field '{}' offset exceeds HDF5 u32 capacity",
                            field.name
                        ))
                    })?;
                    Ok(H5CompoundField {
                        name: field.name.clone(),
                        byte_offset,
                        datatype: nc_type_to_hdf5(&field.dtype)?,
                    })
                })
                .collect::<Result<Vec<_>>>()?;
            Ok(H5Datatype::Compound {
                size: *size,
                fields,
            })
        }
        NcType::Opaque { size, tag } => Ok(H5Datatype::Opaque {
            size: *size,
            tag: tag.clone(),
        }),
        NcType::Array { base, dims } => Ok(H5Datatype::Array {
            base: Box::new(nc_type_to_hdf5(base)?),
            dims: dims.clone(),
        }),
        NcType::VLen { base } => {
            validate_vlen_base_nc4_type(base)?;
            Ok(H5Datatype::VarLen {
                base: Box::new(nc_type_to_hdf5(base)?),
                kind: H5VarLenKind::Sequence,
                encoding: H5StringEncoding::Ascii,
                padding: H5StringPadding::NullTerminate,
            })
        }
    }
}

#[cfg(feature = "netcdf4")]
fn nc_enum_value_to_hdf5(base: &NcType, value: NcIntegerValue) -> Result<Vec<u8>> {
    nc_enum_value_to_le_bytes(base, value)
}

fn nc_enum_value_to_le_bytes(base: &NcType, value: NcIntegerValue) -> Result<Vec<u8>> {
    match (base, value) {
        (NcType::Byte, NcIntegerValue::I8(value)) => Ok(vec![value as u8]),
        (NcType::UByte, NcIntegerValue::U8(value)) => Ok(vec![value]),
        (NcType::Short, NcIntegerValue::I16(value)) => Ok(value.to_le_bytes().to_vec()),
        (NcType::UShort, NcIntegerValue::U16(value)) => Ok(value.to_le_bytes().to_vec()),
        (NcType::Int, NcIntegerValue::I32(value)) => Ok(value.to_le_bytes().to_vec()),
        (NcType::UInt, NcIntegerValue::U32(value)) => Ok(value.to_le_bytes().to_vec()),
        (NcType::Int64, NcIntegerValue::I64(value)) => Ok(value.to_le_bytes().to_vec()),
        (NcType::UInt64, NcIntegerValue::U64(value)) => Ok(value.to_le_bytes().to_vec()),
        (base, value) => Err(Error::InvalidDefinition(format!(
            "enum value {value:?} is incompatible with base type {base:?}"
        ))),
    }
}

#[cfg(feature = "netcdf4")]
fn nc4_hdf5_element_size(dtype: &NcType) -> Result<usize> {
    match dtype {
        NcType::String | NcType::VLen { .. } => Ok(16),
        other => other.size().map_err(Error::Core),
    }
}

fn convert_classic_be_data_to_hdf5_le(dtype: &NcType, data: &[u8]) -> Result<Vec<u8>> {
    let width = dtype.size()?;
    if width == 1 {
        return Ok(data.to_vec());
    }
    if data.len() % width != 0 {
        return Err(Error::InvalidDefinition(format!(
            "variable data length {} is not a multiple of element size {width}",
            data.len()
        )));
    }

    match dtype {
        NcType::Short | NcType::UShort => Ok(data
            .chunks_exact(2)
            .flat_map(|chunk| [chunk[1], chunk[0]])
            .collect()),
        NcType::Int | NcType::UInt | NcType::Float => Ok(data
            .chunks_exact(4)
            .flat_map(|chunk| [chunk[3], chunk[2], chunk[1], chunk[0]])
            .collect()),
        NcType::Int64 | NcType::UInt64 | NcType::Double => Ok(data
            .chunks_exact(8)
            .flat_map(|chunk| {
                [
                    chunk[7], chunk[6], chunk[5], chunk[4], chunk[3], chunk[2], chunk[1], chunk[0],
                ]
            })
            .collect()),
        other => Err(Error::UnsupportedFeature(format!(
            "NetCDF-4 data conversion is not implemented for {other:?}"
        ))),
    }
}

#[cfg(feature = "netcdf4")]
fn nc_attr_to_hdf5(attribute: &NcAttribute) -> Result<H5AttributeBuilder> {
    match &attribute.value {
        NcAttrValue::Chars(value) => Ok(H5AttributeBuilder::fixed_string(&attribute.name, value)),
        NcAttrValue::Bytes(values) => hdf5_numeric_attr(&attribute.name, NcType::Byte, values),
        NcAttrValue::UBytes(values) => hdf5_numeric_attr(&attribute.name, NcType::UByte, values),
        NcAttrValue::Shorts(values) => hdf5_numeric_attr(&attribute.name, NcType::Short, values),
        NcAttrValue::UShorts(values) => hdf5_numeric_attr(&attribute.name, NcType::UShort, values),
        NcAttrValue::Ints(values) => hdf5_numeric_attr(&attribute.name, NcType::Int, values),
        NcAttrValue::UInts(values) => hdf5_numeric_attr(&attribute.name, NcType::UInt, values),
        NcAttrValue::Int64s(values) => hdf5_numeric_attr(&attribute.name, NcType::Int64, values),
        NcAttrValue::UInt64s(values) => hdf5_numeric_attr(&attribute.name, NcType::UInt64, values),
        NcAttrValue::Floats(values) => hdf5_numeric_attr(&attribute.name, NcType::Float, values),
        NcAttrValue::Doubles(values) => hdf5_numeric_attr(&attribute.name, NcType::Double, values),
        NcAttrValue::Strings(values) => H5AttributeBuilder::vlen_strings(&attribute.name, values)
            .map_err(hdf5_error_to_unsupported),
    }
}

#[cfg(feature = "netcdf4")]
trait NcAttrElement {
    fn write_le(&self, dst: &mut Vec<u8>);
}

#[cfg(feature = "netcdf4")]
macro_rules! impl_attr_element_bytes {
    ($ty:ty) => {
        impl NcAttrElement for $ty {
            fn write_le(&self, dst: &mut Vec<u8>) {
                dst.push(*self as u8);
            }
        }
    };
}

#[cfg(feature = "netcdf4")]
macro_rules! impl_attr_element_le {
    ($ty:ty) => {
        impl NcAttrElement for $ty {
            fn write_le(&self, dst: &mut Vec<u8>) {
                dst.extend_from_slice(&self.to_le_bytes());
            }
        }
    };
}

#[cfg(feature = "netcdf4")]
impl_attr_element_bytes!(i8);
#[cfg(feature = "netcdf4")]
impl_attr_element_bytes!(u8);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(i16);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(u16);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(i32);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(u32);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(i64);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(u64);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(f32);
#[cfg(feature = "netcdf4")]
impl_attr_element_le!(f64);

#[cfg(feature = "netcdf4")]
fn hdf5_numeric_attr<T: NcAttrElement>(
    name: &str,
    dtype: NcType,
    values: &[T],
) -> Result<H5AttributeBuilder> {
    let mut raw = Vec::with_capacity(values.len() * dtype.size()?);
    for value in values {
        value.write_le(&mut raw);
    }
    Ok(H5AttributeBuilder::new(
        name,
        nc_type_to_hdf5(&dtype)?,
        vec![values.len() as u64],
        raw,
    ))
}

#[cfg(feature = "netcdf4")]
fn empty_dimension_list_attribute() -> H5AttributeBuilder {
    H5AttributeBuilder::new(
        "DIMENSION_LIST",
        H5Datatype::VarLen {
            base: Box::new(H5Datatype::Reference {
                ref_type: H5ReferenceType::Object,
                size: 8,
            }),
            kind: H5VarLenKind::Sequence,
            encoding: H5StringEncoding::Ascii,
            padding: H5StringPadding::NullTerminate,
        },
        vec![0],
        Vec::new(),
    )
}

#[cfg(feature = "netcdf4")]
fn hdf5_error_to_unsupported(err: hdf5_writer::Error) -> Error {
    Error::UnsupportedFeature(format!("HDF5 writer error: {err}"))
}

#[cfg(feature = "netcdf4")]
fn checked_usize(value: u64, context: &str) -> Result<usize> {
    usize::try_from(value).map_err(|_| {
        Error::InvalidDefinition(format!(
            "{context} value {value} exceeds platform usize capacity"
        ))
    })
}

fn expected_variable_bytes(
    builder: &NcFileBuilder,
    variable: &VariableDef,
    is_record: bool,
) -> Result<usize> {
    let elements = if is_record {
        let records = infer_num_records_for_var(builder, variable)?;
        records
            .checked_mul(variable_shape_elements(builder, variable, 1)?)
            .ok_or_else(|| Error::InvalidDefinition("variable element count overflow".into()))?
    } else {
        variable_shape_elements(builder, variable, 0)?
    };
    let bytes = elements
        .checked_mul(variable.dtype.size()? as u64)
        .ok_or_else(|| Error::InvalidDefinition("variable byte size overflow".into()))?;
    usize::try_from(bytes)
        .map_err(|_| Error::InvalidDefinition("variable byte size exceeds platform usize".into()))
}

#[derive(Debug, Clone)]
struct ResolvedWriteSelection {
    dims: Vec<ResolvedWriteSelectionDim>,
    elements: usize,
    old_shape: Vec<u64>,
    shape: Vec<u64>,
    can_grow: bool,
}

#[derive(Debug, Clone)]
enum ResolvedWriteSelectionDim {
    Index(u64),
    Slice { start: u64, step: u64, count: usize },
}

#[derive(Debug, Clone, Copy)]
struct WriteDimensionExtent {
    current: u64,
    is_unlimited: bool,
    position: usize,
}

fn resolve_write_selection(
    variable_name: &str,
    extents: &[WriteDimensionExtent],
    selection: &NcSliceInfo,
) -> Result<ResolvedWriteSelection> {
    if selection.selections.len() != extents.len() {
        return Err(Error::InvalidDefinition(format!(
            "selection has {} dimensions but variable '{}' has {}",
            selection.selections.len(),
            variable_name,
            extents.len()
        )));
    }

    let mut dims = Vec::with_capacity(extents.len());
    let mut old_shape = Vec::with_capacity(extents.len());
    let mut shape = Vec::with_capacity(extents.len());
    let mut elements = 1usize;
    let mut can_grow = false;
    for (selection, extent) in selection.selections.iter().zip(extents) {
        let dim_size = extent.current;
        old_shape.push(dim_size);
        let mut target_dim_size = dim_size;
        match selection {
            NcSliceInfoElem::Index(index) => {
                if extent.is_unlimited {
                    target_dim_size = checked_add(*index, 1, "unlimited slice extent")?;
                } else if *index >= dim_size {
                    return Err(Error::InvalidDefinition(format!(
                        "index {index} out of bounds for dimension {} (size {dim_size})",
                        extent.position
                    )));
                }
                dims.push(ResolvedWriteSelectionDim::Index(*index));
            }
            NcSliceInfoElem::Slice { start, end, step } => {
                if *step == 0 {
                    return Err(Error::InvalidDefinition("slice step cannot be 0".into()));
                }
                if !extent.is_unlimited && *start > dim_size {
                    return Err(Error::InvalidDefinition(format!(
                        "slice start {start} out of bounds for dimension {} (size {dim_size})",
                        extent.position
                    )));
                }
                let actual_end = if extent.is_unlimited {
                    if *end == u64::MAX {
                        dim_size
                    } else {
                        *end
                    }
                } else if *end == u64::MAX {
                    dim_size
                } else {
                    (*end).min(dim_size)
                };
                if extent.is_unlimited && *start > actual_end {
                    return Err(Error::InvalidDefinition(format!(
                        "slice start {start} exceeds end {actual_end} for unlimited dimension {}",
                        extent.position
                    )));
                }
                let count_u64 = if *start >= actual_end {
                    0
                } else {
                    (actual_end - *start).div_ceil(*step)
                };
                let count = require_usize(count_u64, "slice result dimension")?;
                if extent.is_unlimited && count > 0 {
                    let last_index = checked_add(
                        *start,
                        checked_mul_u64((count - 1) as u64, *step, "slice coordinate step")?,
                        "slice coordinate",
                    )?;
                    target_dim_size = checked_add(last_index, 1, "unlimited slice extent")?;
                }
                elements = elements.checked_mul(count).ok_or_else(|| {
                    Error::InvalidDefinition(
                        "slice result element count exceeds platform usize".into(),
                    )
                })?;
                dims.push(ResolvedWriteSelectionDim::Slice {
                    start: *start,
                    step: *step,
                    count,
                });
            }
        }
        if extent.is_unlimited && target_dim_size > dim_size {
            can_grow = true;
        }
        shape.push(target_dim_size.max(dim_size));
    }

    Ok(ResolvedWriteSelection {
        dims,
        elements,
        old_shape,
        shape,
        can_grow,
    })
}

fn ensure_variable_slice_buffer(
    variable: &mut VariableDef,
    old_shape: &[u64],
    new_shape: &[u64],
    total_elements: u64,
    elem_size: usize,
    can_grow: bool,
) -> Result<()> {
    ensure_variable_byte_slice_buffer(
        variable,
        old_shape,
        new_shape,
        total_elements,
        elem_size,
        can_grow,
        VariableDataEncoding::ClassicBigEndian,
    )
}

fn ensure_variable_native_slice_buffer(
    variable: &mut VariableDef,
    old_shape: &[u64],
    new_shape: &[u64],
    total_elements: u64,
    elem_size: usize,
    can_grow: bool,
) -> Result<()> {
    ensure_variable_byte_slice_buffer(
        variable,
        old_shape,
        new_shape,
        total_elements,
        elem_size,
        can_grow,
        VariableDataEncoding::Hdf5Native,
    )
}

/// Create or grow a variable's byte payload for a slice write, in either the
/// classic big-endian or HDF5 native little-endian encoding. Gaps are filled
/// with the encoding-appropriate fill pattern (the variable's fill value, or
/// the type default).
fn ensure_variable_byte_slice_buffer(
    variable: &mut VariableDef,
    old_shape: &[u64],
    new_shape: &[u64],
    total_elements: u64,
    elem_size: usize,
    can_grow: bool,
    encoding: VariableDataEncoding,
) -> Result<()> {
    if variable.data_encoding != encoding && !variable.data.is_empty() {
        return Err(Error::UnsupportedFeature(format!(
            "slice writes cannot change the payload encoding of variable '{}'",
            variable.name
        )));
    }

    let expected = checked_mul_usize(
        require_usize(total_elements, "variable element count")?,
        elem_size,
        "variable byte size",
    )?;
    let fill_pattern = match encoding {
        VariableDataEncoding::Hdf5Native => native_fill_pattern(variable, elem_size)?,
        VariableDataEncoding::ClassicBigEndian => match &variable.fill_value {
            Some(fill_value) => fill_value.classic_bytes.clone(),
            None => default_classic_fill_bytes(&variable.dtype)?,
        },
    };
    if variable.data.is_empty() {
        variable.data = filled_data_buffer(expected, elem_size, &fill_pattern)?;
    } else if variable.data.len() < expected && can_grow {
        resize_variable_data(
            &mut variable.data,
            old_shape,
            new_shape,
            elem_size,
            &fill_pattern,
        )?;
    } else if variable.data.len() != expected {
        return Err(Error::DataLengthMismatch {
            expected,
            actual: variable.data.len(),
        });
    }

    variable.data_encoding = encoding;
    variable.string_values = None;
    variable.vlen_values = None;
    Ok(())
}

fn encode_char_string_values<S: AsRef<str>>(
    variable_name: &str,
    width: usize,
    values: &[S],
) -> Result<Vec<u8>> {
    let mut encoded = Vec::with_capacity(checked_mul_usize(
        values.len(),
        width,
        "char string byte count",
    )?);
    for value in values {
        let value = value.as_ref();
        let bytes = value.as_bytes();
        if bytes.contains(&0) {
            return Err(Error::InvalidDefinition(format!(
                "char variable '{variable_name}' string values cannot contain NUL bytes"
            )));
        }
        if bytes.len() > width {
            return Err(Error::InvalidDefinition(format!(
                "char variable '{variable_name}' string value is {} bytes, exceeding fixed width {width}",
                bytes.len()
            )));
        }
        encoded.extend_from_slice(bytes);
        encoded.resize(encoded.len() + (width - bytes.len()), 0);
    }
    Ok(encoded)
}

fn ensure_variable_string_slice_buffer(
    variable: &mut VariableDef,
    old_shape: &[u64],
    new_shape: &[u64],
    total_elements: u64,
    can_grow: bool,
) -> Result<()> {
    if variable.dtype != NcType::String {
        return Err(Error::TypeMismatch {
            expected: "String".into(),
            actual: format!("{:?}", variable.dtype),
        });
    }
    let expected = require_usize(total_elements, "string variable element count")?;
    let values = variable.string_values.get_or_insert_with(Vec::new);
    if values.is_empty() {
        values.resize(expected, String::new());
    } else if values.len() < expected && can_grow {
        resize_variable_values(values, old_shape, new_shape, String::new())?;
    } else if values.len() != expected {
        return Err(Error::DataLengthMismatch {
            expected,
            actual: values.len(),
        });
    }

    variable.data.clear();
    variable.data_encoding = VariableDataEncoding::Hdf5Native;
    variable.vlen_values = None;
    Ok(())
}

fn ensure_variable_vlen_slice_buffer(
    variable: &mut VariableDef,
    old_shape: &[u64],
    new_shape: &[u64],
    total_elements: u64,
    can_grow: bool,
) -> Result<()> {
    if !matches!(&variable.dtype, NcType::VLen { .. }) {
        return Err(Error::TypeMismatch {
            expected: "VLen".into(),
            actual: format!("{:?}", variable.dtype),
        });
    }
    let expected = require_usize(total_elements, "vlen variable element count")?;
    let values = variable.vlen_values.get_or_insert_with(Vec::new);
    if values.is_empty() {
        values.resize(expected, Vec::new());
    } else if values.len() < expected && can_grow {
        resize_variable_values(values, old_shape, new_shape, Vec::new())?;
    } else if values.len() != expected {
        return Err(Error::DataLengthMismatch {
            expected,
            actual: values.len(),
        });
    }

    variable.data.clear();
    variable.data_encoding = VariableDataEncoding::Hdf5Native;
    variable.string_values = None;
    Ok(())
}

fn resize_variable_data(
    data: &mut Vec<u8>,
    old_shape: &[u64],
    new_shape: &[u64],
    elem_size: usize,
    fill_pattern: &[u8],
) -> Result<()> {
    if old_shape.len() != new_shape.len() {
        return Err(Error::InvalidDefinition(format!(
            "cannot resize variable data from rank {} to rank {}",
            old_shape.len(),
            new_shape.len()
        )));
    }
    let old_elements = checked_shape_elements(old_shape, "old variable shape element count")?;
    let old_expected = checked_mul_usize(
        require_usize(old_elements, "old variable element count")?,
        elem_size,
        "old variable byte size",
    )?;
    if data.len() != old_expected {
        return Err(Error::DataLengthMismatch {
            expected: old_expected,
            actual: data.len(),
        });
    }

    let new_elements = checked_shape_elements(new_shape, "new variable shape element count")?;
    let new_len = checked_mul_usize(
        require_usize(new_elements, "new variable element count")?,
        elem_size,
        "new variable byte size",
    )?;
    if old_shape.get(1..) == new_shape.get(1..) {
        let additional = new_len.checked_sub(data.len()).ok_or_else(|| {
            Error::InvalidDefinition("new variable byte size is smaller than old size".into())
        })?;
        data.extend(filled_data_buffer(additional, elem_size, fill_pattern)?);
        return Ok(());
    }

    let old = std::mem::take(data);
    let mut resized = filled_data_buffer(new_len, elem_size, fill_pattern)?;
    copy_reshaped_variable_data(&old, &mut resized, old_shape, new_shape, elem_size)?;
    *data = resized;
    Ok(())
}

fn resize_variable_values<T: Clone>(
    values: &mut Vec<T>,
    old_shape: &[u64],
    new_shape: &[u64],
    fill_value: T,
) -> Result<()> {
    if old_shape.len() != new_shape.len() {
        return Err(Error::InvalidDefinition(format!(
            "cannot resize variable values from rank {} to rank {}",
            old_shape.len(),
            new_shape.len()
        )));
    }
    let old_elements = require_usize(
        checked_shape_elements(old_shape, "old variable value shape element count")?,
        "old variable value element count",
    )?;
    if values.len() != old_elements {
        return Err(Error::DataLengthMismatch {
            expected: old_elements,
            actual: values.len(),
        });
    }

    let new_len = require_usize(
        checked_shape_elements(new_shape, "new variable value shape element count")?,
        "new variable value element count",
    )?;
    if old_shape.get(1..) == new_shape.get(1..) {
        values.resize(new_len, fill_value);
        return Ok(());
    }

    let old = std::mem::take(values);
    let mut resized = vec![fill_value; new_len];
    copy_reshaped_values(&old, &mut resized, old_shape, new_shape)?;
    *values = resized;
    Ok(())
}

fn copy_reshaped_variable_data(
    old: &[u8],
    new: &mut [u8],
    old_shape: &[u64],
    new_shape: &[u64],
    elem_size: usize,
) -> Result<()> {
    let old_elements = checked_shape_elements(old_shape, "old variable copy element count")?;
    if old_elements == 0 {
        return Ok(());
    }
    if old_shape.is_empty() {
        if old.len() != elem_size || new.len() != elem_size {
            return Err(Error::DataLengthMismatch {
                expected: elem_size,
                actual: old.len().max(new.len()),
            });
        }
        new.copy_from_slice(old);
        return Ok(());
    }

    let old_strides = row_major_strides(old_shape, "old variable copy stride")?;
    let new_strides = row_major_strides(new_shape, "new variable copy stride")?;
    let rank = old_shape.len();
    let mut coords = vec![0u64; rank];
    for old_index in 0..old_elements {
        let mut remainder = old_index;
        for dim in 0..rank {
            let stride = old_strides[dim];
            coords[dim] = remainder / stride;
            remainder %= stride;
        }
        if coords
            .iter()
            .zip(new_shape)
            .any(|(&coord, &extent)| coord >= extent)
        {
            continue;
        }
        let new_index =
            coords
                .iter()
                .zip(&new_strides)
                .try_fold(0u64, |acc, (&coord, &stride)| {
                    checked_add(
                        acc,
                        checked_mul_u64(coord, stride, "new variable copy coordinate")?,
                        "new variable copy element index",
                    )
                })?;
        let old_start = checked_mul_usize(
            require_usize(old_index, "old variable copy element")?,
            elem_size,
            "old variable copy byte offset",
        )?;
        let new_start = checked_mul_usize(
            require_usize(new_index, "new variable copy element")?,
            elem_size,
            "new variable copy byte offset",
        )?;
        let old_end = old_start.checked_add(elem_size).ok_or_else(|| {
            Error::InvalidDefinition("old variable copy byte range exceeds usize".into())
        })?;
        let new_end = new_start.checked_add(elem_size).ok_or_else(|| {
            Error::InvalidDefinition("new variable copy byte range exceeds usize".into())
        })?;
        new[new_start..new_end].copy_from_slice(&old[old_start..old_end]);
    }
    Ok(())
}

fn copy_reshaped_values<T: Clone>(
    old: &[T],
    new: &mut [T],
    old_shape: &[u64],
    new_shape: &[u64],
) -> Result<()> {
    let old_elements = checked_shape_elements(old_shape, "old variable value copy element count")?;
    if old_elements == 0 {
        return Ok(());
    }
    if old_shape.is_empty() {
        if old.len() != 1 || new.len() != 1 {
            return Err(Error::DataLengthMismatch {
                expected: 1,
                actual: old.len().max(new.len()),
            });
        }
        new[0] = old[0].clone();
        return Ok(());
    }

    let old_strides = row_major_strides(old_shape, "old variable value copy stride")?;
    let new_strides = row_major_strides(new_shape, "new variable value copy stride")?;
    let rank = old_shape.len();
    let mut coords = vec![0u64; rank];
    for old_index in 0..old_elements {
        let mut remainder = old_index;
        for dim in 0..rank {
            let stride = old_strides[dim];
            coords[dim] = remainder / stride;
            remainder %= stride;
        }
        if coords
            .iter()
            .zip(new_shape)
            .any(|(&coord, &extent)| coord >= extent)
        {
            continue;
        }
        let new_index =
            coords
                .iter()
                .zip(&new_strides)
                .try_fold(0u64, |acc, (&coord, &stride)| {
                    checked_add(
                        acc,
                        checked_mul_u64(coord, stride, "new variable value copy coordinate")?,
                        "new variable value copy element index",
                    )
                })?;
        let old_index = require_usize(old_index, "old variable value copy element")?;
        let new_index = require_usize(new_index, "new variable value copy element")?;
        new[new_index] = old[old_index].clone();
    }
    Ok(())
}

fn filled_data_buffer(len: usize, elem_size: usize, fill_pattern: &[u8]) -> Result<Vec<u8>> {
    if elem_size == 0 {
        return if len == 0 {
            Ok(Vec::new())
        } else {
            Err(Error::InvalidDefinition(
                "non-empty variable data requires a non-zero element size".into(),
            ))
        };
    }
    if len % elem_size != 0 {
        return Err(Error::InvalidDefinition(format!(
            "variable byte length {len} is not a multiple of element size {elem_size}"
        )));
    }
    if fill_pattern.len() != elem_size {
        return Err(Error::InvalidDefinition(format!(
            "fill value byte length must match datatype element size: expected {elem_size}, got {}",
            fill_pattern.len()
        )));
    }
    let mut data = Vec::with_capacity(len);
    for _ in 0..len / elem_size {
        data.extend_from_slice(fill_pattern);
    }
    Ok(data)
}

fn default_classic_fill_bytes(dtype: &NcType) -> Result<Vec<u8>> {
    match dtype {
        NcType::Byte => Ok(vec![NC_FILL_BYTE as u8]),
        NcType::Char => Ok(vec![NC_FILL_CHAR]),
        NcType::Short => Ok(NC_FILL_SHORT.to_be_bytes().to_vec()),
        NcType::Int => Ok(NC_FILL_INT.to_be_bytes().to_vec()),
        NcType::Float => Ok(NC_FILL_FLOAT.to_be_bytes().to_vec()),
        NcType::Double => Ok(NC_FILL_DOUBLE.to_be_bytes().to_vec()),
        NcType::UByte => Ok(vec![NC_FILL_UBYTE]),
        NcType::UShort => Ok(NC_FILL_USHORT.to_be_bytes().to_vec()),
        NcType::UInt => Ok(NC_FILL_UINT.to_be_bytes().to_vec()),
        NcType::Int64 => Ok(NC_FILL_INT64.to_be_bytes().to_vec()),
        NcType::UInt64 => Ok(NC_FILL_UINT64.to_be_bytes().to_vec()),
        other => Err(Error::UnsupportedFeature(format!(
            "{other:?} does not have a primitive NetCDF default fill value"
        ))),
    }
}

fn scatter_slice_bytes(
    data: &mut [u8],
    elem_size: usize,
    dims: &[ResolvedWriteSelectionDim],
    strides: &[u64],
    encoded: &[u8],
) -> Result<()> {
    let mut src_elem = 0usize;
    scatter_slice_bytes_recursive(data, elem_size, dims, strides, encoded, 0, 0, &mut src_elem)?;
    let expected = encoded.len() / elem_size.max(1);
    if src_elem != expected {
        return Err(Error::InvalidDefinition(format!(
            "slice scatter wrote {src_elem} elements but expected {expected}"
        )));
    }
    Ok(())
}

fn scatter_slice_values<T: Clone>(
    data: &mut [T],
    dims: &[ResolvedWriteSelectionDim],
    strides: &[u64],
    values: &[T],
) -> Result<()> {
    let mut src_elem = 0usize;
    scatter_slice_values_recursive(data, dims, strides, values, 0, 0, &mut src_elem)?;
    if src_elem != values.len() {
        return Err(Error::InvalidDefinition(format!(
            "slice scatter wrote {src_elem} elements but expected {}",
            values.len()
        )));
    }
    Ok(())
}

fn scatter_slice_values_recursive<T: Clone>(
    data: &mut [T],
    dims: &[ResolvedWriteSelectionDim],
    strides: &[u64],
    values: &[T],
    dim: usize,
    dst_elem: u64,
    src_elem: &mut usize,
) -> Result<()> {
    if dim == dims.len() {
        let dst = require_usize(dst_elem, "slice destination element")?;
        if dst >= data.len() || *src_elem >= values.len() {
            return Err(Error::InvalidDefinition(
                "slice write exceeded planned value bounds".into(),
            ));
        }
        data[dst] = values[*src_elem].clone();
        *src_elem += 1;
        return Ok(());
    }

    match dims[dim] {
        ResolvedWriteSelectionDim::Index(index) => {
            let next = checked_add(
                dst_elem,
                checked_mul_u64(index, strides[dim], "slice index stride")?,
                "slice destination element",
            )?;
            scatter_slice_values_recursive(data, dims, strides, values, dim + 1, next, src_elem)
        }
        ResolvedWriteSelectionDim::Slice { start, step, count } => {
            for offset in 0..count {
                let coord = checked_add(
                    start,
                    checked_mul_u64(offset as u64, step, "slice coordinate step")?,
                    "slice coordinate",
                )?;
                let next = checked_add(
                    dst_elem,
                    checked_mul_u64(coord, strides[dim], "slice coordinate stride")?,
                    "slice destination element",
                )?;
                scatter_slice_values_recursive(
                    data,
                    dims,
                    strides,
                    values,
                    dim + 1,
                    next,
                    src_elem,
                )?;
            }
            Ok(())
        }
    }
}

#[allow(clippy::too_many_arguments)]
fn scatter_slice_bytes_recursive(
    data: &mut [u8],
    elem_size: usize,
    dims: &[ResolvedWriteSelectionDim],
    strides: &[u64],
    encoded: &[u8],
    dim: usize,
    dst_elem: u64,
    src_elem: &mut usize,
) -> Result<()> {
    if dim == dims.len() {
        let dst_start = checked_mul_usize(
            require_usize(dst_elem, "slice destination element")?,
            elem_size,
            "slice destination byte offset",
        )?;
        let dst_end = dst_start.checked_add(elem_size).ok_or_else(|| {
            Error::InvalidDefinition("slice destination byte range exceeds usize".into())
        })?;
        let src_start = checked_mul_usize(*src_elem, elem_size, "slice source byte offset")?;
        let src_end = src_start.checked_add(elem_size).ok_or_else(|| {
            Error::InvalidDefinition("slice source byte range exceeds usize".into())
        })?;
        if dst_end > data.len() || src_end > encoded.len() {
            return Err(Error::InvalidDefinition(
                "slice write exceeded planned data bounds".into(),
            ));
        }
        data[dst_start..dst_end].copy_from_slice(&encoded[src_start..src_end]);
        *src_elem += 1;
        return Ok(());
    }

    match dims[dim] {
        ResolvedWriteSelectionDim::Index(index) => {
            let offset = checked_add(
                dst_elem,
                checked_mul_u64(index, strides[dim], "slice index stride")?,
                "slice destination element",
            )?;
            scatter_slice_bytes_recursive(
                data,
                elem_size,
                dims,
                strides,
                encoded,
                dim + 1,
                offset,
                src_elem,
            )
        }
        ResolvedWriteSelectionDim::Slice { start, step, count } => {
            for i in 0..count {
                let coord = checked_add(
                    start,
                    checked_mul_u64(i as u64, step, "slice coordinate step")?,
                    "slice coordinate",
                )?;
                let offset = checked_add(
                    dst_elem,
                    checked_mul_u64(coord, strides[dim], "slice coordinate stride")?,
                    "slice destination element",
                )?;
                scatter_slice_bytes_recursive(
                    data,
                    elem_size,
                    dims,
                    strides,
                    encoded,
                    dim + 1,
                    offset,
                    src_elem,
                )?;
            }
            Ok(())
        }
    }
}

fn row_major_strides(shape: &[u64], context: &str) -> Result<Vec<u64>> {
    if shape.is_empty() {
        return Ok(Vec::new());
    }
    let mut strides = vec![1u64; shape.len()];
    for dim in (0..shape.len() - 1).rev() {
        strides[dim] = checked_mul_u64(strides[dim + 1], shape[dim + 1], context)?;
    }
    Ok(strides)
}

fn checked_shape_elements(shape: &[u64], context: &str) -> Result<u64> {
    shape
        .iter()
        .try_fold(1u64, |acc, &dim| checked_mul_u64(acc, dim, context))
}

#[cfg(feature = "netcdf4")]
fn validate_nc4_variable_data(variable: &VariableDef, dimension_sizes: &[u64]) -> Result<()> {
    let elements = nc4_variable_elements(variable, dimension_sizes)?;
    validate_nc4_variable_storage(variable)?;
    validate_nc4_variable_fill_value(variable)?;
    if variable.dtype == NcType::String {
        let values = variable.string_values.as_ref().ok_or_else(|| {
            Error::InvalidDefinition(format!(
                "NC_STRING variable '{}' has no string data",
                variable.name
            ))
        })?;
        let expected = checked_usize(elements, "NetCDF-4 string variable element count")?;
        if values.len() != expected {
            return Err(Error::DataLengthMismatch {
                expected,
                actual: values.len(),
            });
        }
        return Ok(());
    }
    if let NcType::VLen { base } = &variable.dtype {
        let values = variable.vlen_values.as_ref().ok_or_else(|| {
            Error::InvalidDefinition(format!(
                "NC_VLEN variable '{}' has no sequence data",
                variable.name
            ))
        })?;
        validate_vlen_base_nc4_type(base)?;
        let expected = checked_usize(elements, "NetCDF-4 vlen variable element count")?;
        if values.len() != expected {
            return Err(Error::DataLengthMismatch {
                expected,
                actual: values.len(),
            });
        }
        let base_size = base.size()?;
        for value in values {
            if value.len() % base_size != 0 {
                return Err(Error::InvalidDefinition(format!(
                    "NC_VLEN variable '{}' sequence byte length {} is not a multiple of base element size {base_size}",
                    variable.name,
                    value.len()
                )));
            }
        }
        return Ok(());
    }

    let bytes = checked_mul_u64(
        elements,
        variable.dtype.size()? as u64,
        "NetCDF-4 variable byte size",
    )?;
    let expected = usize::try_from(bytes).map_err(|_| {
        Error::InvalidDefinition("NetCDF-4 variable byte size exceeds platform usize".into())
    })?;
    if variable.data.is_empty() && variable.fill_value.is_some() {
        return Ok(());
    }
    if variable.data.len() != expected {
        return Err(Error::DataLengthMismatch {
            expected,
            actual: variable.data.len(),
        });
    }
    Ok(())
}

#[cfg(feature = "netcdf4")]
fn validate_nc4_variable_fill_value(variable: &VariableDef) -> Result<()> {
    let Some(fill_value) = &variable.fill_value else {
        return Ok(());
    };
    if matches!(&variable.dtype, NcType::String | NcType::VLen { .. }) {
        return Err(Error::UnsupportedFeature(format!(
            "NetCDF-4 variable-length fill values are not supported yet: '{}'",
            variable.name
        )));
    }
    let expected = nc4_hdf5_element_size(&variable.dtype)?;
    if fill_value.hdf5_bytes.len() != expected {
        return Err(Error::InvalidDefinition(format!(
            "variable '{}' fill value byte length must match datatype element size: expected {expected}, got {}",
            variable.name,
            fill_value.hdf5_bytes.len()
        )));
    }
    Ok(())
}

#[cfg(feature = "netcdf4")]
fn nc4_variable_elements(variable: &VariableDef, dimension_sizes: &[u64]) -> Result<u64> {
    let elements = variable.dim_ids.iter().try_fold(1u64, |acc, id| {
        checked_mul_u64(
            acc,
            dimension_sizes[id.0],
            "NetCDF-4 variable element count",
        )
    })?;
    Ok(elements)
}

#[cfg(feature = "netcdf4")]
fn infer_nc4_dimension_sizes(builder: &NcFileBuilder) -> Result<Vec<u64>> {
    let mut sizes = builder
        .dimensions
        .iter()
        .map(|dimension| {
            if dimension.is_unlimited {
                (dimension.current_size > 0).then_some(dimension.current_size)
            } else {
                Some(dimension.size)
            }
        })
        .collect::<Vec<_>>();

    loop {
        let mut changed = false;
        for variable in &builder.variables {
            changed |= infer_nc4_dimension_size_from_variable(builder, variable, &mut sizes)?;
        }
        if !changed {
            break;
        }
    }

    for variable in &builder.variables {
        let unresolved = variable
            .dim_ids
            .iter()
            .copied()
            .filter(|id| sizes[id.0].is_none())
            .collect::<Vec<_>>();
        if !unresolved.is_empty() && variable_has_data(variable) {
            let names = unresolved
                .iter()
                .map(|id| builder.dimensions[id.0].name.as_str())
                .collect::<Vec<_>>()
                .join(", ");
            return Err(Error::InvalidDefinition(format!(
                "cannot infer current size for NetCDF-4 unlimited dimension(s) {names} from variable '{}'",
                variable.name
            )));
        }
    }

    Ok(sizes.into_iter().map(|size| size.unwrap_or(0)).collect())
}

#[cfg(feature = "netcdf4")]
fn infer_nc4_dimension_size_from_variable(
    builder: &NcFileBuilder,
    variable: &VariableDef,
    sizes: &mut [Option<u64>],
) -> Result<bool> {
    let (elements, elem_size, actual_bytes) = variable_payload_element_count(variable)?;
    let mut known_product = 1u64;
    let mut unknown = Vec::new();
    for dim_id in &variable.dim_ids {
        match sizes[dim_id.0] {
            Some(size) => {
                known_product =
                    checked_mul_u64(known_product, size, "NetCDF-4 known dimension product")?;
            }
            None if !unknown.contains(dim_id) => unknown.push(*dim_id),
            None => {
                return Err(Error::UnsupportedFeature(format!(
                    "NetCDF-4 variable '{}' repeats unlimited dimension '{}'",
                    variable.name, builder.dimensions[dim_id.0].name
                )));
            }
        }
    }

    if unknown.len() != 1 || known_product == 0 {
        return Ok(false);
    }
    if elements % known_product != 0 {
        return Err(Error::DataLengthMismatch {
            expected: usize::try_from((elements / known_product + 1) * known_product * elem_size)
                .unwrap_or(usize::MAX),
            actual: actual_bytes,
        });
    }
    let inferred = elements / known_product;
    sizes[unknown[0].0] = Some(inferred);
    Ok(true)
}

#[cfg(feature = "netcdf4")]
fn variable_has_data(variable: &VariableDef) -> bool {
    match &variable.dtype {
        NcType::String => variable
            .string_values
            .as_ref()
            .is_some_and(|values| !values.is_empty()),
        NcType::VLen { .. } => variable
            .vlen_values
            .as_ref()
            .is_some_and(|values| !values.is_empty()),
        _ => !variable.data.is_empty(),
    }
}

#[cfg(feature = "netcdf4")]
fn variable_payload_element_count(variable: &VariableDef) -> Result<(u64, u64, usize)> {
    if variable.dtype == NcType::String {
        let values = variable.string_values.as_ref().map_or(0usize, Vec::len);
        return Ok((
            u64::try_from(values).map_err(|_| {
                Error::InvalidDefinition(
                    "NC_STRING variable element count exceeds u64 capacity".into(),
                )
            })?,
            1,
            values,
        ));
    }
    if matches!(&variable.dtype, NcType::VLen { .. }) {
        let values = variable.vlen_values.as_ref().map_or(0usize, Vec::len);
        return Ok((
            u64::try_from(values).map_err(|_| {
                Error::InvalidDefinition(
                    "NC_VLEN variable element count exceeds u64 capacity".into(),
                )
            })?,
            1,
            values,
        ));
    }

    let elem_size = variable.dtype.size()? as u64;
    let data_len = variable.data.len() as u64;
    if data_len % elem_size != 0 {
        return Err(Error::DataLengthMismatch {
            expected: usize::try_from((data_len / elem_size + 1) * elem_size).unwrap_or(usize::MAX),
            actual: variable.data.len(),
        });
    }
    Ok((data_len / elem_size, elem_size, variable.data.len()))
}

fn variable_shape_elements(
    builder: &NcFileBuilder,
    variable: &VariableDef,
    skip_dims: usize,
) -> Result<u64> {
    variable
        .dim_ids
        .iter()
        .skip(skip_dims)
        .try_fold(1u64, |acc, id| {
            let dim = &builder.dimensions[id.0];
            acc.checked_mul(dim.size).ok_or_else(|| {
                Error::InvalidDefinition(format!(
                    "variable '{}' element count overflows u64",
                    variable.name
                ))
            })
        })
}

fn variable_is_record(builder: &NcFileBuilder, variable: &VariableDef) -> Result<bool> {
    Ok(variable
        .dim_ids
        .first()
        .is_some_and(|id| builder.dimensions[id.0].is_unlimited))
}

fn infer_num_records(builder: &NcFileBuilder) -> Result<u64> {
    let mut records: Option<u64> = None;
    for variable in &builder.variables {
        if !variable_is_record(builder, variable)? {
            continue;
        }
        let current = infer_num_records_for_var(builder, variable)?;
        match records {
            Some(existing) if existing != current => {
                return Err(Error::InvalidDefinition(
                    "all record variables must contain the same record count".into(),
                ));
            }
            Some(_) => {}
            None => records = Some(current),
        }
    }
    Ok(records.unwrap_or(0))
}

fn infer_num_records_for_var(builder: &NcFileBuilder, variable: &VariableDef) -> Result<u64> {
    let record_elems = variable_shape_elements(builder, variable, 1)?;
    let elem_size = variable.dtype.size()? as u64;
    let record_bytes = record_elems
        .checked_mul(elem_size)
        .ok_or_else(|| Error::InvalidDefinition("record byte size overflow".into()))?;
    if record_bytes == 0 {
        return Ok(0);
    }
    let data_len = variable.data.len() as u64;
    if data_len % record_bytes != 0 {
        return Err(Error::DataLengthMismatch {
            expected: usize::try_from((data_len / record_bytes + 1) * record_bytes)
                .unwrap_or(usize::MAX),
            actual: variable.data.len(),
        });
    }
    Ok(data_len / record_bytes)
}

fn validate_classic_type(dtype: &NcType) -> Result<()> {
    if dtype.classic_type_code().is_none() {
        return Err(Error::RequiresNetcdf4 {
            reason: format!("the {dtype:?} datatype is not representable in classic NetCDF"),
        });
    }
    Ok(())
}

fn validate_chunk_shape_for_variable(variable: &VariableDef, chunk_shape: &[u64]) -> Result<()> {
    if variable.dim_ids.is_empty() {
        return Err(Error::InvalidDefinition(format!(
            "chunked NetCDF-4 scalar variables are not supported: '{}'",
            variable.name
        )));
    }
    if chunk_shape.len() != variable.dim_ids.len() {
        return Err(Error::InvalidDefinition(format!(
            "variable '{}' chunk rank must match variable rank: expected {}, got {}",
            variable.name,
            variable.dim_ids.len(),
            chunk_shape.len()
        )));
    }
    if chunk_shape.contains(&0) {
        return Err(Error::InvalidDefinition(format!(
            "variable '{}' chunk dimensions must be non-zero",
            variable.name
        )));
    }
    Ok(())
}

fn reject_scalar_filtered_variable(variable: &VariableDef) -> Result<()> {
    if variable.dim_ids.is_empty() {
        return Err(Error::InvalidDefinition(format!(
            "filtered NetCDF-4 scalar variables are not supported: '{}'",
            variable.name
        )));
    }
    Ok(())
}

#[cfg(feature = "netcdf4")]
fn validate_nc4_variable_storage(variable: &VariableDef) -> Result<()> {
    if let Some(chunk_shape) = &variable.storage.chunk_shape {
        validate_chunk_shape_for_variable(variable, chunk_shape)?;
    }
    if variable.storage.has_filters() {
        reject_scalar_filtered_variable(variable)?;
    }
    Ok(())
}

#[cfg(feature = "netcdf4")]
fn nc4_variable_chunk_shape(
    variable: &VariableDef,
    shape: &[u64],
    requires_chunking: bool,
) -> Result<Option<Vec<u64>>> {
    if let Some(chunk_shape) = &variable.storage.chunk_shape {
        validate_chunk_shape_for_variable(variable, chunk_shape)?;
        return Ok(Some(chunk_shape.clone()));
    }
    if requires_chunking || variable.storage.has_filters() {
        return Ok(Some(nc4_default_chunk_shape(
            shape,
            nc4_hdf5_element_size(&variable.dtype)?,
        )?));
    }
    Ok(None)
}

#[cfg(feature = "netcdf4")]
fn nc4_variable_filters(variable: &VariableDef) -> Vec<H5FilterDescription> {
    let mut filters = Vec::new();
    if variable.storage.shuffle {
        filters.push(H5FilterDescription {
            id: H5_FILTER_SHUFFLE,
            name: None,
            client_data: Vec::new(),
        });
    }
    if let Some(level) = variable.storage.deflate_level {
        filters.push(H5FilterDescription {
            id: H5_FILTER_DEFLATE,
            name: None,
            client_data: vec![u32::from(level)],
        });
    }
    if variable.storage.fletcher32 {
        filters.push(H5FilterDescription {
            id: H5_FILTER_FLETCHER32,
            name: None,
            client_data: Vec::new(),
        });
    }
    filters
}

fn validate_supported_user_defined_type(dtype: &NcType) -> Result<()> {
    if let NcType::VLen { base } = dtype {
        return validate_vlen_base_nc4_type(base);
    }
    if !matches!(
        dtype,
        NcType::Enum { .. }
            | NcType::Compound { .. }
            | NcType::Opaque { .. }
            | NcType::Array { .. }
    ) {
        return Err(Error::InvalidDefinition(format!(
            "user-defined variables require enum, compound, opaque, fixed-size array, or vlen type, got {dtype:?}"
        )));
    }
    validate_fixed_width_nc4_type(dtype)
}

fn validate_fixed_width_user_defined_type(dtype: &NcType) -> Result<()> {
    if !matches!(
        dtype,
        NcType::Enum { .. }
            | NcType::Compound { .. }
            | NcType::Opaque { .. }
            | NcType::Array { .. }
    ) {
        return Err(Error::InvalidDefinition(format!(
            "raw user-defined variable bytes require enum, compound, opaque, or fixed-size array type, got {dtype:?}"
        )));
    }
    validate_fixed_width_nc4_type(dtype)
}

fn validate_fixed_width_nc4_type(dtype: &NcType) -> Result<()> {
    match dtype {
        NcType::Byte
        | NcType::Char
        | NcType::Short
        | NcType::Int
        | NcType::Float
        | NcType::Double
        | NcType::UByte
        | NcType::UShort
        | NcType::UInt
        | NcType::Int64
        | NcType::UInt64
        | NcType::Opaque { .. } => Ok(()),
        NcType::Enum { base, .. } => match base.as_ref() {
            NcType::Byte
            | NcType::UByte
            | NcType::Short
            | NcType::UShort
            | NcType::Int
            | NcType::UInt
            | NcType::Int64
            | NcType::UInt64 => Ok(()),
            other => Err(Error::InvalidDefinition(format!(
                "enum base type must be fixed-width integer, got {other:?}"
            ))),
        },
        NcType::Compound { fields, .. } => {
            for field in fields {
                validate_fixed_width_nc4_type(&field.dtype)?;
            }
            Ok(())
        }
        NcType::Array { base, .. } => validate_fixed_width_nc4_type(base),
        NcType::String | NcType::VLen { .. } => Err(Error::UnsupportedFeature(format!(
            "{dtype:?} user-defined variable payloads require heap-backed sequence writing"
        ))),
    }
}

fn validate_vlen_base_nc4_type(dtype: &NcType) -> Result<()> {
    if dtype.size()? == 0 {
        return Err(Error::InvalidDefinition(
            "vlen base type must have non-zero byte size".into(),
        ));
    }
    match dtype {
        NcType::Byte
        | NcType::Char
        | NcType::Short
        | NcType::Int
        | NcType::Float
        | NcType::Double
        | NcType::UByte
        | NcType::UShort
        | NcType::UInt
        | NcType::Int64
        | NcType::UInt64
        | NcType::Opaque { .. } => Ok(()),
        NcType::Enum { .. } => validate_fixed_width_nc4_type(dtype),
        NcType::Compound { fields, .. } => {
            for field in fields {
                validate_vlen_base_nc4_type(&field.dtype)?;
            }
            Ok(())
        }
        NcType::Array { base, .. } => validate_vlen_base_nc4_type(base),
        NcType::String | NcType::VLen { .. } => Err(Error::UnsupportedFeature(format!(
            "{dtype:?} cannot be used as a vlen base until recursive heap-backed payload writing is implemented"
        ))),
    }
}

#[cfg(feature = "netcdf4")]
fn validate_nc4_classic_attr_value(value: &NcAttrValue) -> Result<()> {
    if matches!(value, NcAttrValue::Strings(_)) {
        return Err(Error::UnsupportedFeature(
            "NC_STRING attributes require full NetCDF-4".into(),
        ));
    }
    Ok(())
}

#[cfg(feature = "netcdf4")]
fn validate_nc4_classic_model(builder: &NcFileBuilder) -> Result<()> {
    validate_root_only_names(builder)?;
    let unlimited_count = builder
        .dimensions
        .iter()
        .filter(|dimension| dimension.is_unlimited)
        .count();
    if unlimited_count > 1 {
        return Err(Error::InvalidDefinition(
            "classic NetCDF supports at most one unlimited dimension".into(),
        ));
    }
    for variable in &builder.variables {
        if variable
            .dim_ids
            .iter()
            .skip(1)
            .any(|id| builder.dimensions[id.0].is_unlimited)
        {
            return Err(Error::InvalidDefinition(format!(
                "record dimension must be first for variable '{}'",
                variable.name
            )));
        }
    }
    Ok(())
}

#[cfg(feature = "netcdf4")]
fn nc4_default_chunk_shape(shape: &[u64], element_size: usize) -> Result<Vec<u64>> {
    if shape.is_empty() {
        return Err(Error::InvalidDefinition(
            "chunked NetCDF-4 scalar variables are not supported".into(),
        ));
    }
    let element_size = element_size.max(1);
    let target_elements = (1024usize * 1024usize / element_size).max(1) as u64;
    let mut remaining = target_elements;
    let mut chunk_shape = vec![1u64; shape.len()];
    for dim in (0..shape.len()).rev() {
        let extent = shape[dim].max(1);
        let chunk = extent.min(remaining.max(1));
        chunk_shape[dim] = chunk;
        remaining = (remaining / chunk).max(1);
    }
    Ok(chunk_shape)
}

fn attr_requires_cdf5(value: &NcAttrValue) -> bool {
    matches!(
        value,
        NcAttrValue::UBytes(_)
            | NcAttrValue::UShorts(_)
            | NcAttrValue::UInts(_)
            | NcAttrValue::Int64s(_)
            | NcAttrValue::UInt64s(_)
    )
}

fn validate_name(name: &str, kind: &str) -> Result<()> {
    if name.is_empty() {
        return Err(Error::InvalidDefinition(format!(
            "{kind} name must not be empty"
        )));
    }
    if name.contains('/') || name.bytes().any(|b| b == 0) {
        return Err(Error::InvalidDefinition(format!(
            "{kind} name '{name}' contains invalid characters"
        )));
    }
    Ok(())
}

fn validate_path_name(path: String, kind: &str) -> Result<String> {
    let trimmed = path.trim_matches('/').to_string();
    if trimmed.is_empty() {
        return Err(Error::InvalidDefinition(format!(
            "{kind} path must not be empty"
        )));
    }
    for component in trimmed.split('/') {
        validate_name(component, kind)?;
    }
    Ok(trimmed)
}

#[cfg(feature = "netcdf4")]
fn path_leaf_name(path: &str) -> &str {
    path.rsplit('/').next().unwrap_or(path)
}

fn validate_root_only_names(builder: &NcFileBuilder) -> Result<()> {
    for dimension in &builder.dimensions {
        if dimension.name.contains('/') {
            return Err(Error::RequiresNetcdf4 {
                reason: format!("grouped dimension '{}' needs NetCDF-4", dimension.name),
            });
        }
    }
    for variable in &builder.variables {
        if variable.name.contains('/') {
            return Err(Error::RequiresNetcdf4 {
                reason: format!("grouped variable '{}' needs NetCDF-4", variable.name),
            });
        }
    }
    if let Some(group_attr) = builder.group_attributes.first() {
        return Err(Error::RequiresNetcdf4 {
            reason: format!(
                "group attribute on '{}' needs NetCDF-4",
                group_attr.group_path
            ),
        });
    }
    Ok(())
}

fn ensure_unique_attr(attrs: &[NcAttribute], name: &str) -> Result<()> {
    if attrs.iter().any(|a| a.name == name) {
        return Err(Error::InvalidDefinition(format!(
            "duplicate attribute '{name}'"
        )));
    }
    Ok(())
}

fn set_variable_fill_value_metadata(
    variable: &mut VariableDef,
    value: NcAttrValue,
    classic_bytes: Vec<u8>,
    hdf5_bytes: Vec<u8>,
) -> Result<()> {
    if let Some(attribute) = variable
        .attributes
        .iter_mut()
        .find(|attribute| attribute.name == FILL_VALUE_ATTR_NAME)
    {
        if variable.fill_value.is_none() {
            return Err(Error::InvalidDefinition(format!(
                "variable '{}' already has a manual _FillValue attribute",
                variable.name
            )));
        }
        attribute.value = value;
    } else {
        variable.attributes.push(NcAttribute {
            name: FILL_VALUE_ATTR_NAME.to_string(),
            value,
        });
    }

    variable.fill_value = Some(VariableFillValue {
        classic_bytes,
        hdf5_bytes,
    });
    Ok(())
}

fn write_name(out: &mut Vec<u8>, format: NcFormat, name: &str) -> Result<()> {
    write_count(out, format, name.len() as u64)?;
    out.extend_from_slice(name.as_bytes());
    pad_vec_to_4(out);
    Ok(())
}

fn write_count(out: &mut Vec<u8>, format: NcFormat, value: u64) -> Result<()> {
    match format {
        NcFormat::Cdf5 => write_u64(out, value),
        NcFormat::Classic | NcFormat::Offset64 => write_u32(out, require_u32(value, "count")?),
        _ => unreachable!("classic count only"),
    }
    Ok(())
}

fn write_u32(out: &mut Vec<u8>, value: u32) {
    out.extend_from_slice(&value.to_be_bytes());
}

fn write_u64(out: &mut Vec<u8>, value: u64) {
    out.extend_from_slice(&value.to_be_bytes());
}

fn pad_vec_to_4(out: &mut Vec<u8>) {
    let pad = netcdf_core::padding_to_4(out.len());
    out.resize(out.len() + pad, 0);
}

/// Write up to 3 padding bytes, cycling the variable's fill pattern the way
/// netcdf-c fills the slack between a value slab and its 4-byte boundary.
fn write_fill_padding(writer: &mut impl Write, len: u64, fill_pattern: &[u8]) -> Result<()> {
    if len == 0 {
        return Ok(());
    }
    let len = usize::try_from(len)
        .map_err(|_| Error::InvalidDefinition("padding exceeds platform usize".into()))?;
    let mut pad = [0u8; 4];
    if !fill_pattern.is_empty() {
        for (i, byte) in pad.iter_mut().enumerate().take(len.min(4)) {
            *byte = fill_pattern[i % fill_pattern.len()];
        }
    }
    writer.write_all(&pad[..len.min(4)])?;
    Ok(())
}

fn pad4_u64(value: u64) -> Result<u64> {
    let rem = value % 4;
    if rem == 0 {
        Ok(value)
    } else {
        checked_add(value, 4 - rem, "4-byte padding")
    }
}

fn checked_add(lhs: u64, rhs: u64, context: &str) -> Result<u64> {
    lhs.checked_add(rhs)
        .ok_or_else(|| Error::InvalidDefinition(format!("{context} overflow")))
}

fn checked_mul_u64(lhs: u64, rhs: u64, context: &str) -> Result<u64> {
    lhs.checked_mul(rhs)
        .ok_or_else(|| Error::InvalidDefinition(format!("{context} overflow")))
}

fn checked_mul_usize(lhs: usize, rhs: usize, context: &str) -> Result<usize> {
    lhs.checked_mul(rhs)
        .ok_or_else(|| Error::InvalidDefinition(format!("{context} overflow")))
}

fn require_u32(value: u64, context: &str) -> Result<u32> {
    u32::try_from(value).map_err(|_| Error::FormatCapacityExceeded {
        reason: format!("{context} exceeds the 32-bit classic field capacity"),
    })
}

fn require_usize(value: u64, context: &str) -> Result<usize> {
    usize::try_from(value)
        .map_err(|_| Error::InvalidDefinition(format!("{context} exceeds platform usize capacity")))
}