runmat_runtime/builtins/strings/core/

string.rs

//! MATLAB-compatible `string` builtin with GPU-aware conversion semantics for RunMat.

use runmat_builtins::{
    BuiltinCompletionPolicy, BuiltinDescriptor, BuiltinErrorDescriptor, BuiltinExtensionDescriptor,
    BuiltinExtensionMode, BuiltinIntegerBackendRule, BuiltinIntegerCapabilityDescriptor,
    BuiltinIntegerComputationDomain, BuiltinIntegerInputAvailability,
    BuiltinIntegerInputCapability, BuiltinIntegerOutputClassRule, BuiltinIntegerOverflowRule,
    BuiltinIntegerOverloadKind, BuiltinIntegerScalarDoubleRule, BuiltinOutputMode,
    BuiltinParamArity, BuiltinParamDescriptor, BuiltinParamType, BuiltinSignatureDescriptor,
};
use runmat_macros::runtime_builtin;
use runmat_value::{
    CharArray, ComplexTensor, IntValue, LogicalArray, NumericScalar, SparseTensor, StringArray,
    Tensor, Value,
};

use crate::builtins::common::format::{complex_to_string, format_variadic, number_to_string};
use crate::builtins::common::map_control_flow_with_builtin;
use crate::builtins::common::spec::{
    BroadcastSemantics, BuiltinFusionSpec, BuiltinGpuSpec, ConstantStrategy, GpuOpKind,
    ReductionNaN, ResidencyPolicy, ShapeRequirements,
};
use crate::builtins::common::tensor;
use crate::builtins::strings::type_resolvers::string_array_type;
use crate::{build_runtime_error, gather_if_needed_async, BuiltinResult, RuntimeError};

const STRING_OUTPUT_S: [BuiltinParamDescriptor; 1] = [BuiltinParamDescriptor {
    name: "S",
    ty: BuiltinParamType::Any,
    arity: BuiltinParamArity::Required,
    default: None,
    description: "String scalar/array result.",
}];

const STRING_INPUTS_VALUE: [BuiltinParamDescriptor; 1] = [BuiltinParamDescriptor {
    name: "X",
    ty: BuiltinParamType::Any,
    arity: BuiltinParamArity::Required,
    default: None,
    description: "Input value to convert to string array.",
}];

const STRING_INPUTS_VALUE_ENCODING: [BuiltinParamDescriptor; 2] = [
    BuiltinParamDescriptor {
        name: "X",
        ty: BuiltinParamType::Any,
        arity: BuiltinParamArity::Required,
        default: None,
        description: "Input value to convert to string array.",
    },
    BuiltinParamDescriptor {
        name: "encoding",
        ty: BuiltinParamType::StringScalar,
        arity: BuiltinParamArity::Optional,
        default: Some("\"UTF-8\""),
        description: "Character encoding (UTF-8 aliases supported).",
    },
];

const STRING_INPUTS_FORMAT: [BuiltinParamDescriptor; 2] = [
    BuiltinParamDescriptor {
        name: "formatSpec",
        ty: BuiltinParamType::Any,
        arity: BuiltinParamArity::Required,
        default: None,
        description: "Format specification text/cell/string array.",
    },
    BuiltinParamDescriptor {
        name: "A",
        ty: BuiltinParamType::Any,
        arity: BuiltinParamArity::Variadic,
        default: None,
        description: "Formatting data arguments.",
    },
];

const STRING_SIGNATURES: [BuiltinSignatureDescriptor; 3] = [
    BuiltinSignatureDescriptor {
        label: "S = string(X)",
        inputs: &STRING_INPUTS_VALUE,
        outputs: &STRING_OUTPUT_S,
    },
    BuiltinSignatureDescriptor {
        label: "S = string(X, encoding)",
        inputs: &STRING_INPUTS_VALUE_ENCODING,
        outputs: &STRING_OUTPUT_S,
    },
    BuiltinSignatureDescriptor {
        label: "S = string(formatSpec, A...)",
        inputs: &STRING_INPUTS_FORMAT,
        outputs: &STRING_OUTPUT_S,
    },
];

const STRING_ERROR_INVALID_INPUT: BuiltinErrorDescriptor = BuiltinErrorDescriptor {
    code: "RM.STRING.INVALID_INPUT",
    identifier: Some("RunMat:string:InvalidInput"),
    when: "Input conversion/formatting/encoding constraints are violated.",
    message: "string: invalid input",
};

const STRING_ERRORS: [BuiltinErrorDescriptor; 1] = [STRING_ERROR_INVALID_INPUT];
const STRING_SPARSE_DENSE_ELEMENT_LIMIT: usize = 10_000_000;

pub const STRING_DESCRIPTOR: BuiltinDescriptor = BuiltinDescriptor {
    signatures: &STRING_SIGNATURES,
    output_mode: BuiltinOutputMode::Fixed,
    completion_policy: BuiltinCompletionPolicy::Public,
    errors: &STRING_ERRORS,
};

const STRING_FORMAT_EXTENSION: BuiltinExtensionDescriptor = BuiltinExtensionDescriptor {
    id: "string-format-spec",
    mode: BuiltinExtensionMode::RunMatOnly,
    description: "string(formatSpec, A...) printf-style formatting is a RunMat extension",
    error_identifier: Some("RunMat:compatibility:StringFormatSpecExtension"),
};

const STRING_ENCODING_EXTENSION: BuiltinExtensionDescriptor = BuiltinExtensionDescriptor {
    id: "string-encoding",
    mode: BuiltinExtensionMode::RunMatOnly,
    description: "string(X, encoding) character-encoding selection is a RunMat extension",
    error_identifier: Some("RunMat:compatibility:StringEncodingExtension"),
};

const STRING_EXPLICIT_GPU_EXTENSION: BuiltinExtensionDescriptor = BuiltinExtensionDescriptor {
    id: "string-explicit-gpu-input",
    mode: BuiltinExtensionMode::RunMatOnly,
    description: "string with explicit gpuArray numeric input is a RunMat extension",
    error_identifier: Some("RunMat:compatibility:StringExplicitGpuInputExtension"),
};

pub const STRING_EXTENSIONS: [BuiltinExtensionDescriptor; 3] = [
    STRING_FORMAT_EXTENSION,
    STRING_ENCODING_EXTENSION,
    STRING_EXPLICIT_GPU_EXTENSION,
];

const STRING_VALUE_INTEGER_INPUTS: [BuiltinIntegerInputCapability; 1] =
    [BuiltinIntegerInputCapability {
        name: "A",
        classes: &crate::builtins::common::integer_capability::ALL_INTEGER_CLASSES,
        availability: BuiltinIntegerInputAvailability::Documented,
        scalar_double: BuiltinIntegerScalarDoubleRule::NotApplicable,
        notes: "Every built-in integer class converts element by element to decimal text from authoritative storage, preserving input shape and exact wide values.",
    }];

const STRING_FORMAT_INTEGER_INPUTS: [BuiltinIntegerInputCapability; 1] =
    [BuiltinIntegerInputCapability {
        name: "A",
        classes: &crate::builtins::common::integer_capability::ALL_INTEGER_CLASSES,
        availability: BuiltinIntegerInputAvailability::RunMatOnly,
        scalar_double: BuiltinIntegerScalarDoubleRule::NotApplicable,
        notes: "RunMat printf-style format arguments accept every integer class and retain exact typed scalars until the selected formatter consumes them.",
    }];

pub const STRING_INTEGER_CAPABILITIES: [BuiltinIntegerCapabilityDescriptor; 2] = [
    BuiltinIntegerCapabilityDescriptor {
        form: "str = string(A)",
        inputs: &STRING_VALUE_INTEGER_INPUTS,
        computation_domain: BuiltinIntegerComputationDomain::FunctionSpecific,
        output_class: BuiltinIntegerOutputClassRule::FunctionSpecific,
        overflow: BuiltinIntegerOverflowRule::NotApplicable,
        backend: BuiltinIntegerBackendRule::GatherFallback,
        overload: BuiltinIntegerOverloadKind::ElementwiseShapePreserving,
        notes: "Host and automatically resident integer arrays convert exactly to host strings, including wide integer values. Explicit gpuArray input is independently mode-gated because the compatibility target does not document interactive GPU-array input for string.",
    },
    BuiltinIntegerCapabilityDescriptor {
        form: "str = string(formatSpec, A...)",
        inputs: &STRING_FORMAT_INTEGER_INPUTS,
        computation_domain: BuiltinIntegerComputationDomain::FunctionSpecific,
        output_class: BuiltinIntegerOutputClassRule::FunctionSpecific,
        overflow: BuiltinIntegerOverflowRule::NotApplicable,
        backend: BuiltinIntegerBackendRule::GatherFallback,
        overload: BuiltinIntegerOverloadKind::Multiple,
        notes: "This printf-style formatting form is a mode-gated RunMat extension and does not redefine the documented string(A) conversion surface.",
    },
];

#[runmat_macros::register_gpu_spec(builtin_path = "crate::builtins::strings::core::string")]
pub const GPU_SPEC: BuiltinGpuSpec = BuiltinGpuSpec {
    name: "string",
    op_kind: GpuOpKind::Custom("conversion"),
    supported_precisions: &[],
    broadcast: BroadcastSemantics::None,
    provider_hooks: &[],
    constant_strategy: ConstantStrategy::InlineLiteral,
    residency: ResidencyPolicy::GatherImmediately,
    nan_mode: ReductionNaN::Include,
    two_pass_threshold: None,
    workgroup_size: None,
    accepts_nan_mode: false,
    notes: "Always converts on the CPU; GPU tensors are gathered to host memory before conversion.",
};

#[runmat_macros::register_fusion_spec(builtin_path = "crate::builtins::strings::core::string")]
pub const FUSION_SPEC: BuiltinFusionSpec = BuiltinFusionSpec {
    name: "string",
    shape: ShapeRequirements::Any,
    constant_strategy: ConstantStrategy::InlineLiteral,
    elementwise: None,
    reduction: None,
    emits_nan: false,
    notes:
        "Conversion builtin; not eligible for fusion and always materialises host string arrays.",
};

#[runtime_builtin(
    name = "string",
    category = "strings/core",
    summary = "Convert numeric, logical, and text inputs into string arrays.",
    keywords = "string,convert,text,char,gpu",
    accel = "sink",
    type_resolver(string_array_type),
    descriptor(crate::builtins::strings::core::string::STRING_DESCRIPTOR),
    extensions(crate::builtins::strings::core::string::STRING_EXTENSIONS),
    integer_capabilities(crate::builtins::strings::core::string::STRING_INTEGER_CAPABILITIES),
    builtin_path = "crate::builtins::strings::core::string"
)]
async fn string_builtin(value: Value, rest: Vec<Value>) -> crate::BuiltinResult<Value> {
    if crate::builtins::common::validation::value_contains_explicit_gpu(&value)
        || rest
            .iter()
            .any(crate::builtins::common::validation::value_contains_explicit_gpu)
    {
        crate::compatibility::ensure_builtin_extension_enabled(
            &STRING_EXPLICIT_GPU_EXTENSION,
            "string",
        )?;
    }
    if !rest.is_empty() {
        let extension = if rest.len() == 1 && is_encoding_call_shape(&value, &rest[0]) {
            &STRING_ENCODING_EXTENSION
        } else {
            &STRING_FORMAT_EXTENSION
        };
        crate::compatibility::ensure_builtin_extension_enabled(extension, "string")?;
    }
    if rest.is_empty() {
        let gathered = gather_if_needed_async(&value)
            .await
            .map_err(|flow| remap_string_flow(flow))?;
        let array = convert_to_string_array(gathered, StringEncoding::Utf8).await?;
        return Ok(Value::StringArray(array));
    }

    let mut args = rest;
    let format_value = gather_if_needed_async(&value)
        .await
        .map_err(|flow| remap_string_flow(flow))?;

    if args.len() == 1 {
        let arg = args.pop().unwrap();
        let gathered_arg = gather_if_needed_async(&arg)
            .await
            .map_err(|flow| remap_string_flow(flow))?;
        if let Some(encoding) = try_encoding_argument(&format_value, &gathered_arg)? {
            let array = convert_to_string_array(format_value, encoding).await?;
            return Ok(Value::StringArray(array));
        }
        let formatted = format_from_spec(format_value, vec![gathered_arg]).await?;
        return Ok(Value::StringArray(formatted));
    }

    let mut gathered_args = Vec::with_capacity(args.len());
    for arg in args {
        gathered_args.push(
            gather_if_needed_async(&arg)
                .await
                .map_err(|flow| remap_string_flow(flow))?,
        );
    }
    let formatted = format_from_spec(format_value, gathered_args).await?;
    Ok(Value::StringArray(formatted))
}

fn is_encoding_call_shape(first: &Value, candidate: &Value) -> bool {
    if !matches!(
        first,
        Value::CharArray(_) | Value::String(_) | Value::StringArray(_) | Value::Cell(_)
    ) || has_format_placeholders(first)
    {
        return false;
    }
    if let Value::Cell(cell) = first {
        if !cell_contains_only_text_scalars(cell) {
            return false;
        }
    }
    value_to_scalar_text(candidate).is_some()
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum StringEncoding {
    Utf8,
}

fn try_encoding_argument(
    first: &Value,
    candidate: &Value,
) -> BuiltinResult<Option<StringEncoding>> {
    if !matches!(
        first,
        Value::CharArray(_) | Value::String(_) | Value::StringArray(_) | Value::Cell(_)
    ) {
        return Ok(None);
    }
    if has_format_placeholders(first) {
        return Ok(None);
    }
    if let Value::Cell(cell) = first {
        if !cell_contains_only_text_scalars(cell) {
            return Ok(None);
        }
    }
    let Some(text) = value_to_scalar_text(candidate) else {
        return Ok(None);
    };
    parse_encoding_text(&text).map(Some)
}

fn parse_encoding_text(raw: &str) -> BuiltinResult<StringEncoding> {
    let trimmed = raw.trim();
    let lowered = trimmed.to_ascii_lowercase();
    match lowered.as_str() {
        "utf-8" | "utf8" | "unicode" | "system" => Ok(StringEncoding::Utf8),
        _ => Err(string_flow(format!(
            "string: unsupported character encoding '{trimmed}'; only UTF-8 is available"
        ))),
    }
}

fn cell_contains_only_text_scalars(cell: &runmat_value::CellArray) -> bool {
    cell.data.iter().all(|ptr| match &ptr {
        Value::String(_) => true,
        Value::StringArray(sa) => sa.data.len() <= 1,
        Value::CharArray(ca) => ca.rows <= 1,
        _ => false,
    })
}

fn text_has_format_placeholder(text: &str) -> bool {
    let mut chars = text.chars().peekable();
    while let Some(ch) = chars.next() {
        if ch != '%' {
            continue;
        }
        if let Some('%') = chars.peek() {
            chars.next();
            continue;
        }
        while matches!(chars.peek(), Some(flag) if matches!(flag, '+' | '-' | '0' | '#')) {
            chars.next();
        }
        while matches!(chars.peek(), Some(digit) if digit.is_ascii_digit()) {
            chars.next();
        }
        if let Some('.') = chars.peek() {
            chars.next();
            while matches!(chars.peek(), Some(digit) if digit.is_ascii_digit()) {
                chars.next();
            }
        }
        if let Some(conv) = chars.peek() {
            if conv.is_ascii_alphabetic() {
                return true;
            }
        }
    }
    false
}

fn has_format_placeholders(value: &Value) -> bool {
    match value {
        Value::String(s) => text_has_format_placeholder(s),
        Value::StringArray(sa) => sa.data.iter().any(|s| text_has_format_placeholder(s)),
        Value::CharArray(ca) => {
            for row in 0..ca.rows {
                let mut row_str = String::with_capacity(ca.cols);
                for col in 0..ca.cols {
                    row_str.push(ca.data[row * ca.cols + col]);
                }
                if text_has_format_placeholder(&row_str) {
                    return true;
                }
            }
            false
        }
        Value::Cell(cell) => {
            for ptr in &cell.data {
                let element = (ptr).clone();
                if has_format_placeholders(&element) {
                    return true;
                }
            }
            false
        }
        _ => false,
    }
}

pub(crate) struct FormatSpecData {
    pub(crate) specs: Vec<String>,
    pub(crate) shape: Vec<usize>,
}

struct ArgumentData {
    values: Vec<Value>,
    shape: Vec<usize>,
}

fn string_flow(message: impl Into<String>) -> RuntimeError {
    string_error_with_detail(&STRING_ERROR_INVALID_INPUT, message)
}

fn string_error_with_detail(
    error: &'static BuiltinErrorDescriptor,
    detail: impl Into<String>,
) -> RuntimeError {
    let detail = detail.into();
    let message = if detail.starts_with("string:") {
        detail
    } else {
        format!("{}: {detail}", error.message)
    };
    let mut builder = build_runtime_error(message).with_builtin("string");
    if let Some(identifier) = error.identifier {
        builder = builder.with_identifier(identifier);
    }
    builder.build()
}

fn remap_string_flow(err: RuntimeError) -> RuntimeError {
    map_control_flow_with_builtin(err, "string")
}

pub(crate) async fn format_from_spec(
    format_value: Value,
    args: Vec<Value>,
) -> crate::BuiltinResult<StringArray> {
    let spec = extract_format_spec(format_value).await?;
    let mut arguments = Vec::with_capacity(args.len());
    for arg in args {
        arguments.push(extract_argument_data(arg).await?);
    }

    let (target_len, mut target_shape) = resolve_target_shape(&spec, &arguments)?;

    if target_len == 0 {
        let shape = if target_shape.is_empty() {
            if spec.shape.is_empty() {
                vec![0, 0]
            } else {
                spec.shape.clone()
            }
        } else {
            target_shape
        };
        return StringArray::new(Vec::new(), shape)
            .map_err(|e| string_flow(format!("string: {e}")));
    }

    let spec_len = spec.specs.len();
    if spec_len == 0 {
        return Err(string_flow(
            "string: formatSpec must contain at least one element when formatting with data",
        ));
    }

    for arg in &arguments {
        if target_len > 0 && arg.values.is_empty() {
            return Err(string_flow(
                "string: format data arguments must be scalars or match formatSpec size",
            ));
        }
    }

    let mut output = Vec::with_capacity(target_len);
    for idx in 0..target_len {
        let spec_idx = if spec_len == 1 { 0 } else { idx };
        let spec_str = &spec.specs[spec_idx];
        let mut per_call = Vec::with_capacity(arguments.len());
        for arg in &arguments {
            let value =
                match arg.values.len() {
                    0 => continue,
                    1 => arg.values[0].clone(),
                    len if len == target_len => arg.values[idx].clone(),
                    _ => return Err(string_flow(
                        "string: format data arguments must be scalars or match formatSpec size",
                    )),
                };
            per_call.push(value);
        }
        let formatted =
            format_variadic(spec_str, &per_call).map_err(|flow| remap_string_flow(flow))?;
        output.push(formatted);
    }

    if target_shape.is_empty() {
        target_shape = if spec_len > 1 {
            spec.shape.clone()
        } else {
            vec![target_len, 1]
        };
    }

    if tensor::element_count(&target_shape) != target_len {
        target_shape = vec![target_len, 1];
    }

    StringArray::new(output, target_shape).map_err(|e| string_flow(format!("string: {e}")))
}

fn resolve_target_shape(
    spec: &FormatSpecData,
    args: &[ArgumentData],
) -> BuiltinResult<(usize, Vec<usize>)> {
    let mut target_len = spec.specs.len();
    let mut target_shape = if target_len > 1 || (target_len == 1 && !spec.shape.is_empty()) {
        spec.shape.clone()
    } else {
        Vec::new()
    };

    for arg in args {
        let len = arg.values.len();
        if len == 0 {
            continue;
        }
        if target_len == 0 {
            target_len = len;
            target_shape = arg.shape.clone();
            continue;
        }
        if len == 1 {
            continue;
        }
        if target_len == 1 {
            target_len = len;
            target_shape = arg.shape.clone();
            continue;
        }
        if len != target_len {
            return Err(string_flow(
                "string: format data arguments must be scalars or match formatSpec size",
            ));
        }
        if target_shape.is_empty() && len > 1 {
            target_shape = arg.shape.clone();
        }
    }

    if target_len == 0 {
        let shape = if spec.shape.is_empty() {
            vec![0, 0]
        } else {
            spec.shape.clone()
        };
        return Ok((0, shape));
    }

    if target_shape.is_empty() {
        target_shape = if spec.shape.is_empty() {
            vec![target_len, 1]
        } else {
            spec.shape.clone()
        };
        if spec.specs.len() == 1 && tensor::element_count(&target_shape) != target_len {
            target_shape = vec![target_len, 1];
        }
    }

    if tensor::element_count(&target_shape) != target_len {
        target_shape = vec![target_len, 1];
    }

    Ok((target_len, target_shape))
}

pub(crate) async fn extract_format_spec(value: Value) -> BuiltinResult<FormatSpecData> {
    match value {
        Value::String(s) => Ok(FormatSpecData {
            specs: vec![s],
            shape: vec![1, 1],
        }),
        Value::StringArray(sa) => Ok(FormatSpecData {
            specs: sa.data.clone(),
            shape: sa.shape.clone(),
        }),
        Value::CharArray(ca) => {
            let array = char_array_to_string_array(ca, StringEncoding::Utf8)?;
            Ok(FormatSpecData {
                specs: array.data,
                shape: array.shape,
            })
        }
        Value::Cell(cell) => {
            let mut specs = Vec::with_capacity(cell.data.len());
            for col in 0..cell.cols {
                for row in 0..cell.rows {
                    let idx = row * cell.cols + col;
                    let element = &cell.data[idx];
                    let value = (element).clone();
                    let gathered = gather_if_needed_async(&value)
                        .await
                        .map_err(|flow| remap_string_flow(flow))?;
                    let text = value_to_scalar_text(&gathered).ok_or_else(|| {
                        string_flow("string: formatSpec cell elements must be text scalars")
                    })?;
                    specs.push(text);
                }
            }
            Ok(FormatSpecData {
                specs,
                shape: vec![cell.rows, cell.cols],
            })
        }
        _ => Err(string_flow(
            "string: formatSpec must be text (string, char, or cellstr)",
        )),
    }
}

#[async_recursion::async_recursion(?Send)]
async fn extract_argument_data(value: Value) -> BuiltinResult<ArgumentData> {
    match value {
        Value::String(s) => Ok(ArgumentData {
            values: vec![Value::String(s)],
            shape: vec![1, 1],
        }),
        Value::StringArray(sa) => Ok(ArgumentData {
            values: sa.data.into_iter().map(Value::String).collect(),
            shape: sa.shape,
        }),
        Value::CharArray(ca) => {
            let array = char_array_to_string_array(ca, StringEncoding::Utf8)?;
            Ok(ArgumentData {
                values: array.data.into_iter().map(Value::String).collect(),
                shape: array.shape,
            })
        }
        Value::Symbolic(expr) => Ok(ArgumentData {
            values: vec![Value::String(expr.to_string())],
            shape: vec![1, 1],
        }),
        Value::SymbolicArray(array) => Ok(ArgumentData {
            values: array
                .data
                .into_iter()
                .map(|expr| Value::String(expr.to_string()))
                .collect(),
            shape: array.shape,
        }),
        Value::Num(n) => Ok(ArgumentData {
            values: vec![Value::Num(n)],
            shape: vec![1, 1],
        }),
        Value::Int(i) => Ok(ArgumentData {
            values: vec![Value::Int(i)],
            shape: vec![1, 1],
        }),
        Value::Bool(b) => Ok(ArgumentData {
            values: vec![Value::Num(if b { 1.0 } else { 0.0 })],
            shape: vec![1, 1],
        }),
        Value::Tensor(tensor) => tensor_into_argument_data(tensor),
        Value::SparseTensor(s) => {
            ensure_sparse_dense_conversion(&s, "format argument")?;
            if s.is_logical() {
                let logical = s.to_dense_logical().map_err(string_flow)?;
                Ok(ArgumentData {
                    values: logical
                        .data
                        .into_iter()
                        .map(|value| Value::Num(f64::from(value)))
                        .collect(),
                    shape: logical.shape,
                })
            } else {
                tensor_into_argument_data(s.to_dense().map_err(string_flow)?)
            }
        }
        Value::Complex(re, im) => Ok(ArgumentData {
            values: vec![Value::String(complex_to_string(re, im))],
            shape: vec![1, 1],
        }),
        Value::ComplexTensor(t) => Ok(ArgumentData {
            values: t
                .materialize_f64()
                .into_iter()
                .map(|(re, im)| Value::String(complex_to_string(re, im)))
                .collect(),
            shape: t.shape,
        }),
        Value::LogicalArray(la) => Ok(ArgumentData {
            values: la
                .data
                .into_iter()
                .map(|byte| Value::Num(if byte != 0 { 1.0 } else { 0.0 }))
                .collect(),
            shape: la.shape,
        }),
        Value::Cell(cell) => {
            let mut values = Vec::with_capacity(cell.data.len());
            for col in 0..cell.cols {
                for row in 0..cell.rows {
                    let idx = row * cell.cols + col;
                    let element = &cell.data[idx];
                    let value = (element).clone();
                    let gathered = gather_if_needed_async(&value)
                        .await
                        .map_err(|flow| remap_string_flow(flow))?;
                    let value = match gathered {
                        Value::String(s) => Value::String(s),
                        Value::StringArray(sa) if sa.data.len() == 1 => {
                            Value::String(sa.data[0].clone())
                        }
                        Value::CharArray(ca) => {
                            if ca.rows != 1 {
                                return Err(string_flow(
                                    "string: cell format arguments must contain char row vectors",
                                ));
                            }
                            let mut row_str = String::with_capacity(ca.cols);
                            for ch in ca.data {
                                row_str.push(ch);
                            }
                            Value::String(row_str)
                        }
                        Value::Num(n) => Value::Num(n),
                        Value::Int(i) => Value::Int(i),
                        Value::Bool(b) => Value::Num(if b { 1.0 } else { 0.0 }),
                        Value::Tensor(t) => {
                            if !tensor::is_scalar_tensor(&t) {
                                return Err(string_flow(
                                    "string: cell format arguments must contain scalar values",
                                ));
                            }
                            Value::Num(tensor::tensor_value_f64(&t, 0))
                        }
                        Value::LogicalArray(la) => {
                            if la.data.len() != 1 {
                                return Err(string_flow(
                                    "string: cell format arguments must contain scalar values",
                                ));
                            }
                            Value::Num(if la.data[0] != 0 { 1.0 } else { 0.0 })
                        }
                        Value::Complex(re, im) => Value::String(complex_to_string(re, im)),
                        Value::Symbolic(expr) => Value::String(expr.to_string()),
                        Value::ComplexTensor(t) => {
                            if !complex_tensor_is_scalar(&t) {
                                return Err(string_flow(
                                    "string: cell format arguments must contain scalar values",
                                ));
                            }
                            Value::String(t.format_element(0))
                        }
                        other => {
                            return Err(string_flow(format!(
                                "string: unsupported cell format argument {other:?}; expected scalar text or numeric values"
                            )))
                        }
                    };
                    values.push(value);
                }
            }
            Ok(ArgumentData {
                values,
                shape: vec![cell.rows, cell.cols],
            })
        }
        Value::GpuTensor(handle) => {
            let gathered = gather_if_needed_async(&Value::GpuTensor(handle))
                .await
                .map_err(|flow| remap_string_flow(flow))?;
            extract_argument_data(gathered).await
        }
        Value::MException(_)
        | Value::HandleObject(_)
        | Value::ObjectArray(_)
        | Value::Object(_)
        | Value::Listener(_)
        | Value::Struct(_)
        | Value::OutputList(_)
        | Value::Future(_)
        | Value::Task(_)
        | Value::Pool(_)
        | Value::Job(_)
        | Value::Foreign(_) => Err(string_flow("string: unsupported format argument type")),
        Value::FunctionHandle(_)
        | Value::ExternalFunctionHandle(_)
        | Value::MethodFunctionHandle(_)
        | Value::BoundFunctionHandle { .. }
        | Value::Closure(_)
        | Value::ClassRef(_) => Err(string_flow("string: unsupported format argument type")),
    }
}

#[async_recursion::async_recursion(?Send)]
async fn convert_to_string_array(
    value: Value,
    encoding: StringEncoding,
) -> BuiltinResult<StringArray> {
    if let Some(array) = crate::builtins::datetime::datetime_string_array(&value)
        .map_err(|err| string_flow(err.message().to_string()))?
    {
        return Ok(array);
    }
    if let Some(array) = crate::builtins::duration::duration_string_array(&value)
        .map_err(|err| string_flow(err.message().to_string()))?
    {
        return Ok(array);
    }
    match value {
        Value::String(s) => string_scalar(s),
        Value::StringArray(sa) => Ok(sa),
        Value::CharArray(ca) => char_array_to_string_array(ca, encoding),
        Value::Symbolic(expr) => string_scalar(expr.to_string()),
        Value::SymbolicArray(array) => StringArray::new(
            array.data.into_iter().map(|expr| expr.to_string()).collect(),
            array.shape,
        )
        .map_err(|e| string_flow(format!("string: {e}"))),
        Value::Tensor(tensor) => tensor_to_string_array(tensor),
        Value::SparseTensor(sparse) => {
            ensure_sparse_dense_conversion(&sparse, "dense string array")?;
            if sparse.is_logical() {
                logical_array_to_string_array(sparse.to_dense_logical().map_err(string_flow)?)
            } else {
                tensor_to_string_array(sparse.to_dense().map_err(string_flow)?)
            }
        }
        Value::ComplexTensor(tensor) => complex_tensor_to_string_array(tensor),
        Value::LogicalArray(logical) => logical_array_to_string_array(logical),
        Value::Cell(cell) => cell_array_to_string_array(cell, encoding).await,
        Value::Num(n) => string_scalar(number_to_string(n)),
        Value::Int(i) => string_scalar(int_value_to_string(&i)),
        Value::Bool(b) => string_scalar(bool_to_string(b).to_string()),
        Value::Complex(re, im) => string_scalar(complex_to_string(re, im)),
        Value::GpuTensor(handle) => {
            // Defensive fallback: gather and retry.
            let gathered = gather_if_needed_async(&Value::GpuTensor(handle))
                .await
                .map_err(|flow| remap_string_flow(flow))?;
            convert_to_string_array(gathered, encoding).await
        }
        Value::ObjectArray(_) | Value::Object(_) | Value::HandleObject(_) | Value::Listener(_) => Err(string_flow(
            "string: unsupported conversion from handle-based objects. Use class-specific formatters.",
        )),
        Value::Struct(_) => Err(string_flow(
            "string: structs are not supported for automatic conversion",
        )),
        Value::FunctionHandle(_) | Value::ExternalFunctionHandle(_) | Value::MethodFunctionHandle(_) | Value::BoundFunctionHandle { .. }
        | Value::Closure(_)
        | Value::ClassRef(_)
        | Value::MException(_)
            | Value::Future(_)
            | Value::Task(_)
            | Value::Pool(_)
            | Value::Job(_)
            | Value::Foreign(_)
        | Value::OutputList(_) => Err(
            string_flow("string: unsupported conversion for function or exception handles"),
        ),
    }
}

fn string_scalar<S: Into<String>>(text: S) -> BuiltinResult<StringArray> {
    StringArray::new(vec![text.into()], vec![1, 1]).map_err(|e| string_flow(format!("string: {e}")))
}

fn value_to_scalar_text(value: &Value) -> Option<String> {
    match value {
        Value::String(s) => Some(s.clone()),
        Value::StringArray(sa) if sa.data.len() == 1 => Some(sa.data[0].clone()),
        Value::CharArray(ca) if ca.rows == 1 => Some(ca.data.iter().collect()),
        _ => None,
    }
}

fn char_array_to_string_array(
    array: CharArray,
    _encoding: StringEncoding,
) -> BuiltinResult<StringArray> {
    let mut rows: Vec<String> = Vec::with_capacity(array.rows);
    for r in 0..array.rows {
        let mut row = String::with_capacity(array.cols);
        for c in 0..array.cols {
            row.push(array.data[r * array.cols + c]);
        }
        rows.push(row);
    }
    let shape = if array.rows == 0 {
        vec![0, 1]
    } else {
        vec![array.rows, 1]
    };
    StringArray::new(rows, shape).map_err(|e| string_flow(format!("string: {e}")))
}

fn tensor_to_string_array(tensor: Tensor) -> BuiltinResult<StringArray> {
    let shape = tensor.shape.clone();
    let storage = tensor.into_numeric_storage().map_err(string_flow)?;
    let mut strings = Vec::with_capacity(storage.len());
    for idx in 0..storage.len() {
        let value = storage
            .value_at(idx)
            .ok_or_else(|| string_flow("string: numeric tensor storage is inconsistent"))?;
        strings.push(numeric_scalar_to_string(value));
    }
    StringArray::new(strings, shape).map_err(|e| string_flow(format!("string: {e}")))
}

fn tensor_into_argument_data(tensor: Tensor) -> BuiltinResult<ArgumentData> {
    let shape = tensor.shape.clone();
    let storage = tensor.into_numeric_storage().map_err(string_flow)?;
    let mut values = Vec::with_capacity(storage.len());
    for idx in 0..storage.len() {
        let value = storage
            .value_at(idx)
            .ok_or_else(|| string_flow("string: numeric tensor storage is inconsistent"))?;
        values.push(numeric_scalar_to_value(value));
    }
    Ok(ArgumentData { values, shape })
}

fn numeric_scalar_to_value(value: NumericScalar) -> Value {
    match value {
        NumericScalar::F64(value) => Value::Num(value),
        NumericScalar::F32(value) => Value::Num(f64::from(value)),
        integer => Value::Int(
            integer
                .into_int_value()
                .expect("non-floating numeric scalar is integer"),
        ),
    }
}

fn numeric_scalar_to_string(value: NumericScalar) -> String {
    match value {
        NumericScalar::F64(value) => number_to_string(value),
        NumericScalar::F32(value) => number_to_string(f64::from(value)),
        integer => int_value_to_string(
            &integer
                .into_int_value()
                .expect("non-floating numeric scalar is integer"),
        ),
    }
}

fn complex_tensor_to_string_array(tensor: ComplexTensor) -> BuiltinResult<StringArray> {
    let len = tensor::complex_tensor_element_len(&tensor);
    let mut strings = Vec::with_capacity(len);
    for idx in 0..len {
        strings.push(tensor.format_element(idx));
    }
    StringArray::new(strings, tensor.shape).map_err(|e| string_flow(format!("string: {e}")))
}

fn logical_array_to_string_array(logical: LogicalArray) -> BuiltinResult<StringArray> {
    let mut strings = Vec::with_capacity(logical.data.len());
    for &byte in &logical.data {
        strings.push(bool_to_string(byte != 0).to_string());
    }
    StringArray::new(strings, logical.shape).map_err(|e| string_flow(format!("string: {e}")))
}

async fn cell_array_to_string_array(
    cell: runmat_value::CellArray,
    _encoding: StringEncoding,
) -> BuiltinResult<StringArray> {
    let mut strings = Vec::with_capacity(cell.data.len());
    for col in 0..cell.cols {
        for row in 0..cell.rows {
            let idx = row * cell.cols + col;
            let element = &cell.data[idx];
            let value = (element).clone();
            let gathered = gather_if_needed_async(&value)
                .await
                .map_err(|flow| remap_string_flow(flow))?;
            strings.push(cell_element_to_string(&gathered)?);
        }
    }
    StringArray::new(strings, vec![cell.rows, cell.cols])
        .map_err(|e| string_flow(format!("string: {e}")))
}

fn cell_element_to_string(value: &Value) -> BuiltinResult<String> {
    if let Some(array) = crate::builtins::datetime::datetime_string_array(value)
        .map_err(|err| string_flow(err.message().to_string()))?
    {
        if array.data.len() == 1 {
            return Ok(array.data[0].clone());
        }
        return Err(string_flow("string: cell datetime values must be scalar"));
    }
    if let Some(array) = crate::builtins::duration::duration_string_array(value)
        .map_err(|err| string_flow(err.message().to_string()))?
    {
        if array.data.len() == 1 {
            return Ok(array.data[0].clone());
        }
        return Err(string_flow("string: cell duration values must be scalar"));
    }
    match value {
        Value::String(s) => Ok(s.clone()),
        Value::StringArray(sa) => {
            if sa.data.len() == 1 {
                Ok(sa.data[0].clone())
            } else {
                Err(string_flow(
                    "string: cell elements must contain string scalars, not string arrays",
                ))
            }
        }
        Value::CharArray(ca) => {
            if ca.rows == 1 {
                Ok(ca.data.iter().collect())
            } else {
                Err(string_flow(
                    "string: cell character arrays must be row vectors",
                ))
            }
        }
        Value::Num(n) => Ok(number_to_string(*n)),
        Value::Int(i) => Ok(int_value_to_string(i)),
        Value::Bool(b) => Ok(bool_to_string(*b).to_string()),
        Value::LogicalArray(array) => {
            if array.data.len() == 1 {
                Ok(bool_to_string(array.data[0] != 0).to_string())
            } else {
                Err(string_flow("string: cell logical values must be scalar"))
            }
        }
        Value::Tensor(t) => {
            if tensor::is_scalar_tensor(t) {
                if let Some(value) = t.integer_storage().and_then(|storage| storage.value_at(0)) {
                    Ok(int_value_to_string(&value))
                } else {
                    Ok(number_to_string(tensor::tensor_value_f64(t, 0)))
                }
            } else {
                Err(string_flow("string: cell numeric values must be scalar"))
            }
        }
        Value::Complex(re, im) => Ok(complex_to_string(*re, *im)),
        Value::ComplexTensor(t) => {
            if complex_tensor_is_scalar(t) {
                Ok(t.format_element(0))
            } else {
                Err(string_flow("string: cell complex values must be scalar"))
            }
        }
        other => Err(string_flow(format!(
            "string: unsupported cell element type {:?}; expected text or scalar values",
            other
        ))),
    }
}

fn complex_tensor_is_scalar(tensor: &ComplexTensor) -> bool {
    tensor.shape.iter().product::<usize>() == 1
}

fn ensure_sparse_dense_conversion(sparse: &SparseTensor, target: &str) -> BuiltinResult<()> {
    let total_elements = sparse
        .rows
        .checked_mul(sparse.cols)
        .ok_or_else(|| string_flow("string: sparse matrix dimensions overflow"))?;
    if total_elements > STRING_SPARSE_DENSE_ELEMENT_LIMIT {
        return Err(string_flow(format!(
            "string: cannot convert sparse tensor {}x{} with {} stored entries to {target} ({} elements exceeds safe threshold)",
            sparse.rows,
            sparse.cols,
            sparse.nnz(),
            total_elements
        )));
    }
    Ok(())
}

fn bool_to_string(value: bool) -> &'static str {
    if value {
        "true"
    } else {
        "false"
    }
}

fn int_value_to_string(value: &IntValue) -> String {
    match value {
        IntValue::I8(v) => v.to_string(),
        IntValue::I16(v) => v.to_string(),
        IntValue::I32(v) => v.to_string(),
        IntValue::I64(v) => v.to_string(),
        IntValue::U8(v) => v.to_string(),
        IntValue::U16(v) => v.to_string(),
        IntValue::U32(v) => v.to_string(),
        IntValue::U64(v) => v.to_string(),
    }
}

#[cfg(test)]
pub(crate) mod tests {
    use super::*;
    use crate::builtins::common::test_support;
    use runmat_builtins::{ResolveContext, Type};
    use runmat_value::{
        CellArray, IntValue, IntegerComplexStorage, IntegerStorage, StringArray, StructValue,
    };

    fn string_builtin(value: Value, rest: Vec<Value>) -> BuiltinResult<Value> {
        let _runmat = crate::compatibility::push_runmat_extensions_enabled(true);
        futures::executor::block_on(super::string_builtin(value, rest))
    }

    fn error_message(err: crate::RuntimeError) -> String {
        err.message().to_string()
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_numeric_scalar() {
        let out = string_builtin(Value::Num(42.0), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["42".to_string()]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_numeric_tensor_preserves_shape() {
        let tensor = Tensor::new(vec![1.0, 2.0, 3.0, 4.0], vec![2, 2]).unwrap();
        let out = string_builtin(Value::Tensor(tensor), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![2, 2]);
                assert_eq!(sa.data, vec!["1", "2", "3", "4"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[test]
    fn string_from_logical_sparse_uses_boolean_text() {
        let sparse =
            SparseTensor::new_logical(2, 2, vec![0, 1, 2], vec![1, 0]).expect("logical sparse");
        let out = string_builtin(Value::SparseTensor(sparse), Vec::new()).expect("string");
        let Value::StringArray(out) = out else {
            panic!("expected string array");
        };
        assert_eq!(out.shape, vec![2, 2]);
        assert_eq!(out.data, vec!["false", "true", "true", "false"]);
    }

    #[test]
    fn string_from_native_single_tensor_uses_authoritative_values() {
        let tensor = Tensor::from_f32(vec![1.25, -2.5], vec![1, 2]).expect("native single tensor");
        let out = string_builtin(Value::Tensor(tensor), Vec::new()).expect("string");
        match out {
            Value::StringArray(array) => {
                assert_eq!(array.shape, vec![1, 2]);
                assert_eq!(array.data, vec!["1.25", "-2.5"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_integer_tensor_preserves_exact_storage_values() {
        let tensor = Tensor::new_integer(
            IntegerStorage::U64(vec![u64::MAX, 9_007_199_254_740_993]),
            vec![1, 2],
        )
        .expect("integer tensor");
        let out = string_builtin(Value::Tensor(tensor), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 2]);
                assert_eq!(sa.data, vec!["18446744073709551615", "9007199254740993"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_signed_integer_tensor_preserves_exact_storage_values() {
        let tensor = Tensor::new_integer(
            IntegerStorage::I64(vec![i64::MIN, -9_007_199_254_740_993]),
            vec![1, 2],
        )
        .expect("integer tensor");
        let out = string_builtin(Value::Tensor(tensor), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 2]);
                assert_eq!(sa.data, vec!["-9223372036854775808", "-9007199254740993"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_complex_integer_tensor_preserves_exact_storage_values() {
        let storage = IntegerComplexStorage::new(
            IntegerStorage::U64(vec![u64::MAX, 9_007_199_254_740_993]),
            IntegerStorage::U64(vec![7, 0]),
        )
        .expect("matching integer complex storage");
        let tensor = ComplexTensor::new_integer(storage, vec![1, 2]).expect("complex integer");
        let out = string_builtin(Value::ComplexTensor(tensor), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 2]);
                assert_eq!(sa.data, vec!["18446744073709551615+7i", "9007199254740993"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_logical_array_uses_boolean_text() {
        let logical = LogicalArray::new(vec![1, 0, 1], vec![1, 3]).unwrap();
        let out = string_builtin(Value::LogicalArray(logical), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 3]);
                assert_eq!(sa.data, vec!["true", "false", "true"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_char_array_produces_column_vector() {
        let chars = CharArray::new("abc".chars().collect(), 1, 3).unwrap();
        let out = string_builtin(Value::CharArray(chars), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["abc"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_cell_array() {
        let cell = CellArray::new(vec![Value::Bool(true), Value::Int(IntValue::I32(7))], 1, 2)
            .expect("cell array");
        let out = string_builtin(Value::Cell(cell), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 2]);
                assert_eq!(sa.data, vec!["true", "7"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_cell_array_column_major() {
        let cell = CellArray::new(
            vec![
                Value::Int(IntValue::I32(1)),
                Value::Int(IntValue::I32(2)),
                Value::Int(IntValue::I32(3)),
                Value::Int(IntValue::I32(4)),
            ],
            2,
            2,
        )
        .expect("cell array");
        let out = string_builtin(Value::Cell(cell), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![2, 2]);
                assert_eq!(sa.data, vec!["1", "3", "2", "4"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_cell_scalar_integer_tensor_uses_exact_storage() {
        let tensor =
            Tensor::new_integer(IntegerStorage::U64(vec![9_007_199_254_740_993]), vec![1, 1])
                .expect("integer tensor");
        let cell = CellArray::new(vec![Value::Tensor(tensor)], 1, 1)
            .expect("cell with scalar integer tensor");
        let out = string_builtin(Value::Cell(cell), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["9007199254740993"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_from_cell_scalar_complex_integer_tensor_uses_exact_storage() {
        let storage = IntegerComplexStorage::new(
            IntegerStorage::I64(vec![1]),
            IntegerStorage::I64(vec![i64::MIN]),
        )
        .expect("matching integer complex storage");
        let tensor = ComplexTensor::new_integer(storage, vec![1, 1]).expect("complex integer");
        let cell = CellArray::new(vec![Value::ComplexTensor(tensor)], 1, 1)
            .expect("cell with scalar complex integer tensor");
        let out = string_builtin(Value::Cell(cell), Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec![format!("1-{}i", i64::MIN.unsigned_abs())]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_cell_element_requires_scalar_numeric() {
        let tensor = Tensor::new(vec![1.0, 2.0], vec![2, 1]).unwrap();
        let cell =
            CellArray::new(vec![Value::Tensor(tensor)], 1, 1).expect("cell with numeric tensor");
        let err = error_message(string_builtin(Value::Cell(cell), Vec::new()).unwrap_err());
        assert!(err.contains("cell numeric values must be scalar"));
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_rejects_struct_input() {
        let err = error_message(
            string_builtin(Value::Struct(StructValue::new()), Vec::new()).expect_err("string"),
        );
        assert!(err.contains("structs are not supported"));
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_errors_on_unsupported_encoding() {
        let err = error_message(
            string_builtin(
                Value::CharArray(CharArray::new_row("abc")),
                vec![Value::from("UTF-16")],
            )
            .unwrap_err(),
        );
        assert!(
            err.contains("unsupported character encoding"),
            "unexpected error message: {err}"
        );
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_accepts_system_encoding_alias() {
        let out = string_builtin(
            Value::CharArray(CharArray::new_row("hello")),
            vec![Value::from("system")],
        )
        .expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["hello"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_encoding_allows_percent_literal() {
        let out = string_builtin(
            Value::CharArray(CharArray::new_row("100% Done")),
            vec![Value::from("utf8")],
        )
        .expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["100% Done"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_spec_cell_requires_text_scalars() {
        let cell = CellArray::new(vec![Value::Num(1.0)], 1, 1).expect("cell");
        let err = error_message(
            string_builtin(Value::Cell(cell), vec![Value::from("data")]).expect_err("string"),
        );
        assert!(
            err.contains("formatSpec cell elements must be text scalars"),
            "unexpected error: {err}"
        );
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_cell_argument_requires_scalar_values() {
        let tensor = Tensor::new(vec![1.0, 2.0], vec![2, 1]).unwrap();
        let cell = CellArray::new(vec![Value::Tensor(tensor)], 1, 1).expect("cell argument values");
        let err = error_message(
            string_builtin(Value::from("%d"), vec![Value::Cell(cell)]).expect_err("string"),
        );
        assert!(err.contains("cell format arguments must contain scalar values"));
    }

    #[test]
    fn string_rejects_oversized_sparse_tensor_before_densifying() {
        let sparse = SparseTensor::zeros(STRING_SPARSE_DENSE_ELEMENT_LIMIT + 1, 1);
        let err = string_builtin(Value::SparseTensor(sparse), Vec::new()).unwrap_err();

        assert_eq!(err.identifier(), Some("RunMat:string:InvalidInput"));
        assert!(err.message().contains("exceeds safe threshold"));
    }

    #[test]
    fn string_format_rejects_oversized_sparse_argument_before_densifying() {
        let sparse = SparseTensor::zeros(STRING_SPARSE_DENSE_ELEMENT_LIMIT + 1, 1);
        let err = string_builtin(Value::from("%g"), vec![Value::SparseTensor(sparse)]).unwrap_err();

        assert_eq!(err.identifier(), Some("RunMat:string:InvalidInput"));
        assert!(err.message().contains("format argument"));
        assert!(err.message().contains("exceeds safe threshold"));
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_handles_large_unsigned_int() {
        let value = Value::Int(IntValue::U64(u64::MAX));
        let out = string_builtin(value, Vec::new()).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec![u64::MAX.to_string()]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[test]
    fn string_descriptor_signatures_cover_core_forms() {
        let labels: Vec<&str> = STRING_DESCRIPTOR
            .signatures
            .iter()
            .map(|signature| signature.label)
            .collect();
        assert_eq!(
            labels,
            vec![
                "S = string(X)",
                "S = string(X, encoding)",
                "S = string(formatSpec, A...)",
            ]
        );

        let codes: Vec<&str> = STRING_DESCRIPTOR
            .errors
            .iter()
            .map(|error| error.code)
            .collect();
        assert_eq!(codes, vec!["RM.STRING.INVALID_INPUT"]);
    }

    #[test]
    fn string_struct_input_uses_stable_identifier() {
        let err = string_builtin(Value::Struct(StructValue::new()), Vec::new()).unwrap_err();
        assert_eq!(err.identifier(), Some("RunMat:string:InvalidInput"));
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_numeric_scalar() {
        let out = string_builtin(Value::from("%d"), vec![Value::Num(7.0)]).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["7"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_broadcast_over_tensor() {
        let tensor = Tensor::new(vec![1.0, 2.0, 3.0], vec![1, 3]).unwrap();
        let out =
            string_builtin(Value::from("Trial %d"), vec![Value::Tensor(tensor)]).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 3]);
                assert_eq!(sa.data, vec!["Trial 1", "Trial 2", "Trial 3"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[test]
    fn string_format_integer_tensor_preserves_wide_values() {
        let tensor = Tensor::new_integer(
            IntegerStorage::U64(vec![u64::MAX, 9_007_199_254_740_993]),
            vec![1, 2],
        )
        .expect("integer tensor");
        let out = string_builtin(Value::from("%u"), vec![Value::Tensor(tensor)]).expect("string");
        match out {
            Value::StringArray(array) => {
                assert_eq!(array.shape, vec![1, 2]);
                assert_eq!(
                    array.data,
                    vec![u64::MAX.to_string(), "9007199254740993".to_string()]
                );
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_string_array_spec_alignment() {
        let spec = StringArray::new(vec!["[%d]".into(), "Value %d".into()], vec![1, 2]).unwrap();
        let tensor = Tensor::new(vec![5.0, 6.0], vec![1, 2]).unwrap();
        let out =
            string_builtin(Value::StringArray(spec), vec![Value::Tensor(tensor)]).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 2]);
                assert_eq!(sa.data, vec!["[5]", "Value 6"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_prefers_placeholders_over_encoding_hint() {
        let out = string_builtin(Value::from("%s"), vec![Value::from("UTF-8")]).expect("string");
        match out {
            Value::StringArray(sa) => {
                assert_eq!(sa.shape, vec![1, 1]);
                assert_eq!(sa.data, vec!["UTF-8"]);
            }
            other => panic!("expected string array, got {other:?}"),
        }
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_format_mismatched_lengths_errors() {
        let spec = StringArray::new(vec!["%d".into(), "%d".into()], vec![2, 1]).unwrap();
        let tensor = Tensor::new(vec![1.0, 2.0, 3.0], vec![3, 1]).unwrap();
        let err = error_message(
            string_builtin(Value::StringArray(spec), vec![Value::Tensor(tensor)]).unwrap_err(),
        );
        assert!(err.contains("must be scalars or match formatSpec size"));
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_gpu_numeric_tensor() {
        test_support::with_test_provider(|provider| {
            let tensor = Tensor::new(vec![10.0, 20.0], vec![1, 2]).unwrap();
            let data = tensor.as_f64_slice().expect("double tensor");
            let view = runmat_accelerate_api::HostTensorView {
                data,
                shape: &tensor.shape,
            };
            let handle = provider.upload(&view).expect("upload");
            let result = string_builtin(Value::GpuTensor(handle), Vec::new())
                .expect("gpu string conversion");
            match result {
                Value::StringArray(sa) => {
                    assert_eq!(sa.shape, vec![1, 2]);
                    assert_eq!(sa.data, vec!["10", "20"]);
                }
                other => panic!("expected string array, got {other:?}"),
            }
        });
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    fn string_strict_mode_gates_explicit_gpu_before_provider_access() {
        let handle = runmat_accelerate_api::GpuTensorHandle {
            shape: vec![1, 1],
            device_id: u32::MAX,
            buffer_id: u64::MAX - 452,
            descriptor: Default::default(),
        };
        let handle = handle.with_provenance(runmat_accelerate_api::GpuHandleProvenance::Explicit);
        let _strict = crate::compatibility::push_runmat_extensions_enabled(false);
        let error = futures::executor::block_on(super::string_builtin(
            Value::GpuTensor(handle),
            Vec::new(),
        ))
        .expect_err("strict mode rejects explicit GPU input before gather");
        assert_eq!(
            error.identifier(),
            STRING_EXPLICIT_GPU_EXTENSION.error_identifier
        );
    }

    #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
    #[test]
    #[cfg(feature = "wgpu")]
    fn string_wgpu_numeric_tensor_matches_cpu() {
        let _ = runmat_accelerate::backend::wgpu::provider::register_wgpu_provider(
            runmat_accelerate::backend::wgpu::provider::WgpuProviderOptions::default(),
        );
        let tensor = Tensor::new(vec![4.0, 5.0, 6.0], vec![1, 3]).unwrap();
        let cpu = string_builtin(Value::Tensor(tensor.clone()), Vec::new())
            .expect("cpu string conversion");
        let data = tensor.as_f64_slice().expect("double tensor");
        let view = runmat_accelerate_api::HostTensorView {
            data,
            shape: &tensor.shape,
        };
        let handle = runmat_accelerate_api::provider()
            .unwrap()
            .upload(&view)
            .expect("gpu upload");
        let gpu =
            string_builtin(Value::GpuTensor(handle), Vec::new()).expect("gpu string conversion");
        match (cpu, gpu) {
            (Value::StringArray(expect), Value::StringArray(actual)) => {
                assert_eq!(actual.shape, expect.shape);
                assert_eq!(actual.data, expect.data);
            }
            other => panic!("unexpected results {other:?}"),
        }
    }

    #[test]
    fn string_type_is_string_array() {
        assert_eq!(
            string_array_type(&[Type::Num], &ResolveContext::new(Vec::new())),
            Type::cell_of(Type::String)
        );
    }
}