u-analytics 0.17.0

//! WASM bindings for u-analytics.
//!
//! Exposes SPC (Statistical Process Control) and capability analysis functions
//! to JavaScript/TypeScript via `wasm-bindgen`.
//!
//! # Feature
//!
//! Only compiled when the `wasm` feature is enabled:
//! ```toml
//! [dependencies]
//! u-analytics = { version = "...", features = ["wasm"] }
//! ```

use serde::{Deserialize, Serialize};
use wasm_bindgen::prelude::*;

// The shapes below are the wire contract, shared with the C FFI so the two
// transports cannot drift apart again. See `crate::wire`.
use crate::wire::{
    at_least, c_chart_dto, capability_dto, count_pairs, count_rows, default_cost, default_min_seg,
    default_penalty, g_chart_dto, gage_rr_anova_dto, gage_rr_xbar_r_dto,
    hypothesis::{
        chi_squared_gof_dto, chi_squared_independence_dto, fisher_exact_dto, groups_test_dto,
        jarque_bera_dto, mann_kendall_dto, mann_whitney_dto, one_sample_t_dto, one_way_anova_dto,
        p_adjust_dto, paired_t_dto, shapiro_wilk_dto, two_sample_t_dto, wilcoxon_dto,
    },
    imr_dto, laney_p_dto, laney_u_dto, np_chart_dto, p_chart_dto, pelt_dto,
    percentile_capability_dto, rate_pairs, rules_from_json, run_rules_dto, sample_size_value,
    t_chart_dto, u_chart_dto, xbar_r_dto, xbar_s_dto, AttributeStandardDto, CapabilityInputDto,
    GageRRInputDto, LimitsInputDto, PeltInputDto, PeltPenaltyDto, PeltResultDto,
    PercentileCapabilityInputDto, SeasonalityInputDto, SpectralResidualInputDto, WireError,
};

// ---------------------------------------------------------------------------
// Serializable DTO types
// ---------------------------------------------------------------------------

#[derive(Serialize, tsify::Tsify)]
#[tsify(missing_as_null)]
struct AdNormalityDto {
    statistic: f64,
    statistic_modified: f64,
    p_value: f64,
}

// ---------------------------------------------------------------------------
// Helpers
// ---------------------------------------------------------------------------

/// Every refusal crosses into JavaScript as an `Error` carrying `code` (a
/// stable reason), `index` (the offending array position, or `null`) and
/// `parameter` (the option the refusal is about, or `null`), with
/// the human-readable text as its `message` -- the shape Node gives its own
/// errors, so `err.message` still reads as before while a program can branch on
/// `err.code` and point at `err.index`.
fn js_err(error: impl Into<WireError>) -> JsValue {
    let error = error.into();
    let js = js_sys::Error::new(&error.message);
    let index = error
        .index
        .map_or(JsValue::NULL, |i| JsValue::from_f64(i as f64));
    // `Reflect::set` on a freshly created ordinary object cannot fail.
    let _ = js_sys::Reflect::set(
        &js,
        &JsValue::from_str("code"),
        &JsValue::from_str(error.code),
    );
    let _ = js_sys::Reflect::set(&js, &JsValue::from_str("index"), &index);
    let parameter = error
        .parameter
        .as_deref()
        .map_or(JsValue::NULL, JsValue::from_str);
    let _ = js_sys::Reflect::set(&js, &JsValue::from_str("parameter"), &parameter);
    js.into()
}

impl From<serde_wasm_bindgen::Error> for WireError {
    fn from(e: serde_wasm_bindgen::Error) -> Self {
        WireError::invalid_input(e.to_string())
    }
}

/// Serializes a response. An absent value (`None`) crosses as `null`, not as a
/// missing key: the FFI renders the same wire value with `serde_json`, which
/// writes `null`, and this crate documents `null` -- serde-wasm-bindgen's
/// default of omitting the key made the two transports disagree on every
/// optional field.
fn to_js<T: Serialize>(val: &T) -> Result<JsValue, JsValue> {
    val.serialize(&serde_wasm_bindgen::Serializer::new().serialize_missing_as_null(true))
        .map_err(js_err)
}

/// A NaN or ±Infinity found in a JS argument, and where it sits.
///
/// JSON has no non-finite numbers, so on the way to the wire schema
/// `serde_json` turns one into `null` and the caller would be told a value has
/// the wrong type. [`find_non_finite`] looks before that happens, so the
/// refusal names the real reason and the place.
struct NonFinite {
    /// The argument's name, then `.key` and `[i]` steps down to the array or
    /// field that holds the number.
    parameter: String,
    /// The number's position, when it is an array element.
    index: Option<usize>,
    value: f64,
}

impl NonFinite {
    fn message(&self) -> String {
        let at = match self.index {
            Some(i) => format!("{}[{i}]", self.parameter),
            None => self.parameter.clone(),
        };
        let got = if self.value.is_nan() {
            "NaN"
        } else if self.value > 0.0 {
            "Infinity"
        } else {
            "-Infinity"
        };
        format!("{at}: expected a finite number, got {got}")
    }
}

/// The first NaN or ±Infinity in `value`, searching arrays, iterables and
/// plain objects. `allow_nan` lets NaN through for an input that reads it as a
/// missing value; it then arrives as `null`.
fn find_non_finite(value: &JsValue, parameter: &str, allow_nan: bool) -> Option<NonFinite> {
    let refused = |n: f64| !n.is_finite() && !(allow_nan && n.is_nan());
    let found = |index: Option<usize>, value: f64| NonFinite {
        parameter: parameter.to_string(),
        index,
        value,
    };
    if let Some(n) = value.as_f64() {
        return refused(n).then(|| found(None, n));
    }
    if !value.is_object() {
        return None;
    }
    if let Ok(Some(items)) = js_sys::try_iter(value) {
        for (i, item) in items.enumerate() {
            // An iterator that throws is left for serde to report.
            let item = item.ok()?;
            match item.as_f64() {
                Some(n) if refused(n) => return Some(found(Some(i), n)),
                Some(_) => {}
                None => {
                    let inner = find_non_finite(&item, &format!("{parameter}[{i}]"), allow_nan);
                    if inner.is_some() {
                        return inner;
                    }
                }
            }
        }
        return None;
    }
    let object: &js_sys::Object = wasm_bindgen::JsCast::unchecked_ref(value);
    for entry in js_sys::Object::entries(object).iter() {
        let pair: js_sys::Array = wasm_bindgen::JsCast::unchecked_into(entry);
        let key = pair.get(0).as_string().unwrap_or_default();
        let inner = find_non_finite(&pair.get(1), &format!("{parameter}.{key}"), allow_nan);
        if inner.is_some() {
            return inner;
        }
    }
    None
}

/// Deserialize a native JS value, rejecting JSON strings with an actionable
/// message and prefixing the offending parameter name to any serde error.
fn from_js<T: serde::de::DeserializeOwned>(value: JsValue, param: &str) -> Result<T, JsValue> {
    if value.as_string().is_some() {
        return Err(js_err(format!(
            "{param}: expected a native JS object/array, got a string — \
             pass the value directly, not JSON.stringify(...)"
        )));
    }
    if let Some(found) = find_non_finite(&value, param, false) {
        return Err(js_err(
            WireError::new(
                crate::wire::code::VALUE_NOT_FINITE,
                found.index,
                found.message(),
            )
            .about(found.parameter),
        ));
    }
    // serde-wasm-bindgen reads only a struct's declared fields from a JS
    // object, so `deny_unknown_fields` never sees extra keys. Round-trip
    // through serde_json::Value so the strict wire schema is enforced.
    let json: serde_json::Value = serde_wasm_bindgen::from_value(value).map_err(|e| {
        js_err(WireError::new(
            crate::wire::code::MALFORMED_INPUT,
            None,
            format!("{param}: {e}"),
        ))
    })?;
    from_json(json, param).map_err(js_err)
}

/// The half of [`from_js`] that enforces the wire schema, split out so it can
/// be exercised without a `JsValue` -- which cannot be constructed off
/// `wasm32`. Tests calling this walk the same deserialization path a JS caller
/// does, rather than a parallel one that could drift from it.
fn from_json<T: serde::de::DeserializeOwned>(
    json: serde_json::Value,
    param: &str,
) -> Result<T, WireError> {
    serde_json::from_value(json).map_err(|e| {
        WireError::new(
            crate::wire::code::MALFORMED_INPUT,
            None,
            format!("{param}: {e}"),
        )
    })
}

/// One element of a JS number array, as found.
#[derive(Debug, Clone, PartialEq)]
enum Element {
    Number(f64),
    /// Anything else, by its JS type name (`"null"`, `"string"`, …).
    Other(String),
}

/// The values of a number array, refusing the first element that is not a
/// finite number at its index. Kept apart from the `JsValue` walk so it runs
/// off `wasm32`.
fn numbers_from(
    param: &'static str,
    elements: impl IntoIterator<Item = Element>,
) -> Result<Vec<f64>, WireError> {
    let mut out = Vec::new();
    for (i, element) in elements.into_iter().enumerate() {
        match element {
            Element::Number(x) if x.is_finite() => out.push(x),
            Element::Number(x) => {
                return Err(WireError::new(
                    crate::wire::code::VALUE_NOT_FINITE,
                    Some(i),
                    format!("{param}[{i}]: expected a finite number, got {x}"),
                )
                .about(param))
            }
            Element::Other(kind) => {
                return Err(WireError::new(
                    crate::wire::code::MALFORMED_INPUT,
                    Some(i),
                    format!("{param}[{i}]: expected a number, got {kind}"),
                )
                .about(param))
            }
        }
    }
    Ok(out)
}

/// Reads a `number[]` or `Float64Array` argument as sent.
///
/// A `&[f64]` parameter would let the generated glue copy a plain array into a
/// typed array first, so `null` would arrive as 0 and a string as NaN.
fn read_numbers(value: &JsValue, param: &'static str) -> Result<Vec<f64>, WireError> {
    use wasm_bindgen::JsCast;
    if let Some(typed) = value.dyn_ref::<js_sys::Float64Array>() {
        return numbers_from(param, typed.to_vec().into_iter().map(Element::Number));
    }
    if !js_sys::Array::is_array(value) {
        return Err(WireError::new(
            crate::wire::code::MALFORMED_INPUT,
            None,
            format!("{param}: expected an array of numbers or a Float64Array"),
        )
        .about(param));
    }
    let array: &js_sys::Array = value.unchecked_ref();
    numbers_from(
        param,
        array.iter().map(|item| match item.as_f64() {
            Some(x) => Element::Number(x),
            None if item.is_null() => Element::Other("null".to_string()),
            None => Element::Other(item.js_typeof().as_string().unwrap_or_default()),
        }),
    )
}

/// Places a row reader's refusal in its row, as the crate's error contract
/// reads: `parameter` is the path to the array (`data[1]`) and `index` the
/// position in it -- the shape `find_non_finite` already gives a NaN there.
fn at_row(e: WireError, param: &'static str, row: usize) -> WireError {
    let path = format!("{param}[{row}]");
    let message = e.message.replacen(param, &path, 1);
    WireError::new(e.code, e.index, message).about(path)
}

/// Reads a `number[][]` argument as sent -- each row as [`read_numbers`] does
/// -- so `null` in a row is refused at that row rather than by the parser.
fn read_number_rows(value: &JsValue, param: &'static str) -> Result<Vec<Vec<f64>>, WireError> {
    use wasm_bindgen::JsCast;
    if !js_sys::Array::is_array(value) {
        return Err(WireError::new(
            crate::wire::code::MALFORMED_INPUT,
            None,
            format!("{param}: expected an array of number arrays"),
        )
        .about(param));
    }
    let array: &js_sys::Array = value.unchecked_ref();
    array
        .iter()
        .enumerate()
        .map(|(row, item)| read_numbers(&item, param).map_err(|e| at_row(e, param, row)))
        .collect()
}

// ---------------------------------------------------------------------------
// WASM exports
// ---------------------------------------------------------------------------

/// Compute an X-bar R chart from subgroups.
///
/// # Input JSON
///
/// Array of arrays: `[[x1, x2, ...], [x1, x2, ...], ...]`
/// All subgroups must have the same length, within the range the crate's
/// factor tables cover (see `MIN_SUBGROUP_SIZE`/`MAX_SUBGROUP_SIZE`).
///
/// # Output JSON
///
/// Object with fields: `xbar_cl`, `xbar_ucl`, `xbar_lcl`, `r_cl`, `r_ucl`,
/// `r_lcl`, `sigma_hat`, `xbar_points`, `r_points`, `in_control`.
///
/// `sigma_hat` is `R-bar / d2` -- the short-term sigma a capability study
/// needs. Pass it to [`process_capability`] as `sigma_within`; it is `null`
/// when there is not enough data for control limits.
///
/// # Options (optional second argument)
///
/// `{ rules?: string[] }` -- which run tests to apply. Names are the same
/// values that appear in each point's `violations`, so the set is written in
/// the vocabulary the output already uses:
/// `BeyondLimits`, `NineOneSide`, `SixTrend`, `FourteenAlternating`,
/// `TwoOfThreeBeyond2Sigma`, `FourOfFiveBeyond1Sigma`, `FifteenWithin1Sigma`,
/// `EightBeyond1Sigma`.
///
/// Omitted, `undefined`, `null`, or an object without `rules` all mean all
/// eight (Nelson), which is what this binding did before the option existed.
/// `{ rules: [] }` applies none, leaving control limits only.
#[wasm_bindgen(unchecked_return_type = "XbarRChartDto")]
pub fn xbar_r_chart(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] data: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "RuleOptions | null")] options: Option<JsValue>,
) -> Result<JsValue, JsValue> {
    let subgroups = read_number_rows(&data, "data").map_err(js_err)?;
    let rules = rules_option(options)?;
    to_js(&xbar_r_dto(subgroups, rules).map_err(js_err)?)
}

/// Compute an X-bar S chart from subgroups.
///
/// # Input JSON
///
/// Array of arrays, as for [`xbar_r_chart`]. The standard deviation uses
/// every value in a subgroup rather than only the extremes, which is why this
/// chart is the usual choice once subgroups exceed about ten.
///
/// # Output JSON
///
/// Object with fields: `xbar_cl`, `xbar_ucl`, `xbar_lcl`, `s_cl`, `s_ucl`,
/// `s_lcl`, `sigma_hat` (`S-bar / c4`), `xbar_points`, `s_points`,
/// `in_control`.
///
/// # Options (optional second argument)
///
/// `{ rules?: string[] }`, exactly as for [`xbar_r_chart`].
#[wasm_bindgen(unchecked_return_type = "XbarSChartDto")]
pub fn xbar_s_chart(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] data: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "RuleOptions | null")] options: Option<JsValue>,
) -> Result<JsValue, JsValue> {
    let subgroups = read_number_rows(&data, "data").map_err(js_err)?;
    let rules = rules_option(options)?;
    to_js(&xbar_s_dto(subgroups, rules).map_err(js_err)?)
}

/// Compute an Individual and Moving Range (I-MR) chart.
///
/// # Input JSON
///
/// Array of individual observations in time order: `[x1, x2, ...]` (need >= 2).
///
/// # Output JSON
///
/// Object with fields: `i_cl`, `i_ucl`, `i_lcl`, `mr_cl`, `mr_ucl`, `mr_lcl`,
/// `sigma_hat` (`MR-bar / d2(2)`), `i_points`, `mr_points`, `in_control`.
/// `mr_points` starts at index 1: the first observation has no moving range.
///
/// # Options (optional second argument)
///
/// `{ rules?: string[] }`, exactly as for [`xbar_r_chart`].
#[wasm_bindgen(unchecked_return_type = "ImrChartDto")]
pub fn imr_chart(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] values: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "RuleOptions | null")] options: Option<JsValue>,
) -> Result<JsValue, JsValue> {
    let values = read_numbers(&values, "values").map_err(js_err)?;
    let rules = rules_option(options)?;
    to_js(&imr_dto(values, rules).map_err(js_err)?)
}

/// Apply run tests to a series against control limits the caller supplies.
///
/// The same engine the charts use, reachable on its own: for a statistic the
/// crate does not chart, or limits fixed from an earlier phase-I study.
///
/// # Input JSON
///
/// - `values`: `[x1, x2, ...]`
/// - `limits`: `{ ucl, cl, lcl }` with `lcl <= cl <= ucl`
/// - `options` (optional): `{ rules?: string[] }`, as for [`xbar_r_chart`]
///
/// # Output JSON
///
/// One `{ index, value, violations }` per input value, in input order.
#[wasm_bindgen(unchecked_return_type = "ChartPointDto[]")]
pub fn run_rules(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] values: JsValue,
    #[wasm_bindgen(unchecked_param_type = "LimitsInputDto")] limits: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "RuleOptions | null")] options: Option<JsValue>,
) -> Result<JsValue, JsValue> {
    let values = read_numbers(&values, "values").map_err(js_err)?;
    let limits: LimitsInputDto = from_js(limits, "limits")?;
    let rules = rules_option(options)?;
    to_js(&run_rules_dto(values, limits, rules).map_err(js_err)?)
}

// ---------------------------------------------------------------------------
// TypeScript declarations for the option arguments
// ---------------------------------------------------------------------------

/// The optional last argument of the charts and `run_rules`, declared for
/// TypeScript. The variables charts read a rule set; each attributes chart
/// reads only its own Phase I standard, so each is declared with only that.
#[wasm_bindgen(typescript_custom_section)]
const OPTIONS_TS: &'static str = r#"
/** A run test, by the name a point's `violations` reports it under. */
export type RunRule = "BeyondLimits" | "NineOneSide" | "SixTrend" | "FourteenAlternating" | "TwoOfThreeBeyond2Sigma" | "FourOfFiveBeyond1Sigma" | "FifteenWithin1Sigma" | "EightBeyond1Sigma";
/** Which run tests to apply. Omitted or `null`: all eight. `[]`: none. */
export interface RuleOptions {
    rules?: RunRule[] | null;
}
/** A known centre line from a Phase I study. */
export interface PChartOptions {
    p_bar?: number | null;
}
/** A known centre line and dispersion factor from a Phase I study. */
export interface LaneyPChartOptions {
    p_bar?: number | null;
    phi?: number | null;
}
/** A known centre line from a Phase I study. */
export interface UChartOptions {
    u_bar?: number | null;
}
/** A known centre line and dispersion factor from a Phase I study. */
export interface LaneyUChartOptions {
    u_bar?: number | null;
    phi?: number | null;
}
"#;

// ---------------------------------------------------------------------------
// Pure cores of the variables-chart bindings
// ---------------------------------------------------------------------------
//
// Each binding above is a thin adapter over one of these, so the contract a
// JavaScript caller observes can be tested on a host with no WebAssembly
// runner.

/// The optional `{ rules?: [...] }` argument the variables charts and
/// `run_rules` share.
fn rules_option(options: Option<JsValue>) -> Result<crate::spc::RuleSet, JsValue> {
    match options {
        Some(o) if !o.is_undefined() && !o.is_null() => {
            rules_from_json(Some(from_js(o, "options")?)).map_err(js_err)
        }
        _ => Ok(crate::spc::RuleSet::default()),
    }
}

/// Reads the optional `{ p_bar?, u_bar?, phi? }` standard of an attributes chart.
fn standard_option(options: Option<JsValue>) -> Result<AttributeStandardDto, JsValue> {
    match options {
        Some(o) if !o.is_undefined() && !o.is_null() => from_js(o, "options"),
        _ => Ok(AttributeStandardDto::default()),
    }
}

/// Compute a P chart from (defectives, sample_size) pairs.
///
/// # Input JSON
///
/// Array of `[defectives, sample_size]` pairs (as integers):
/// `[[3, 100], [5, 100], ...]`.
///
/// # Errors
///
/// Throws an `Error` with `code` and `index` (the row): `count_not_whole`
/// for a defective count that is not a whole number >= 0,
/// `sample_size_not_whole` for a sample size that is not a whole number >= 1
/// (zero, negative, fractional or not a number alike),
/// `defectives_exceed_sample`, `malformed_input` for a row that is not a pair,
/// `insufficient_data` for an empty array, `standard_out_of_range` for a
/// `p_bar` outside (0, 1).
///
/// # Options
///
/// `{ p_bar?: number }` -- a known centre line from a Phase I study. Every
/// limit then uses it with the sample's own size (Phase II); without it p-bar
/// is estimated from `samples`.
///
/// # Output JSON
///
/// Object with fields: `p_bar` (the centre line used), `points` (array, each
/// with `z` -- the standardized value, on which every limit is +/-3),
/// `in_control`.
#[wasm_bindgen(unchecked_return_type = "PChartDto")]
pub fn p_chart(
    #[wasm_bindgen(unchecked_param_type = "[number, number][]")] samples: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "PChartOptions | null")] options: Option<
        JsValue,
    >,
) -> Result<JsValue, JsValue> {
    let json: serde_json::Value = from_js(samples, "samples")?;
    let samples = count_pairs(&json, "samples").map_err(js_err)?;
    let standard = standard_option(options)?;
    to_js(&p_chart_dto(&samples, &standard).map_err(js_err)?)
}
/// Compute process capability indices (Cp, Cpk, Pp, Ppk, Cpm).
///
/// # Input JSON
///
/// ```text
/// { data: number[], usl?: number, lsl?: number,
///   sigma_within?: number, target?: number }
/// ```
///
/// At least one of `usl`/`lsl` is required; supplying one gives a one-sided
/// specification, which is routine for characteristics such as flatness,
/// contamination or runout.
///
/// `sigma_within` is the short-term (within-subgroup) standard deviation,
/// normally estimated from a control chart as R-bar/d2 or S-bar/c4. It is not
/// derivable from `data`: a flat measurement vector no longer carries the
/// subgroup structure. Omit it and the short-term indices are reported as
/// `null` rather than being computed from the long-term sigma.
///
/// `target` sets the Cpm target. Omit it and `cpm` is `null`: Cpm measures
/// clustering about a declared target, and one substituted on the caller's
/// behalf could not be told apart from it in the result.
///
/// # Output JSON
///
/// Object with fields: `mean`, `sigma_source`, `std_dev_within`,
/// `std_dev_overall`, `cp`, `cpk`, `cpu`, `cpl`, `pp`, `ppk`, `ppu`, `ppl`,
/// `cpm`. Indices that the specification does not support are `null`
/// (a one-sided specification has no `cp`, `pp` or `cpm`).
///
/// `sigma_source` is `"within"` when `sigma_within` was supplied and
/// `"overall"` otherwise. In the `"overall"` case `std_dev_within`, `cp`,
/// `cpk`, `cpu` and `cpl` are all `null`: the short-term indices are
/// undefined without a short-term sigma, and reporting the long-term one in
/// their place would make `cp` equal `pp` for every input. `cpm` uses neither
/// sigma -- it is the spread of `data` about `target` -- so it is reported in
/// both cases.
#[wasm_bindgen(unchecked_return_type = "CapabilityDto")]
pub fn process_capability(
    #[wasm_bindgen(unchecked_param_type = "CapabilityInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: CapabilityInputDto = from_js(input, "input")?;
    to_js(&capability_dto(input).map_err(js_err)?)
}

/// Anderson-Darling normality test (Stephens 1974).
///
/// H₀: data is normally distributed.
///
/// # Parameters
///
/// - `data`: `number[]` or `Float64Array` of observations (need >= 3, not all equal)
///
/// # Output JSON
///
/// Object with fields: `statistic` (A²), `statistic_modified` (A²*), `p_value`.
#[wasm_bindgen(unchecked_return_type = "AdNormalityDto")]
pub fn anderson_darling_normality(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] data: JsValue,
) -> Result<JsValue, JsValue> {
    let data = at_least(read_numbers(&data, "data").map_err(js_err)?, 3, "data").map_err(js_err)?;
    let result = crate::testing::anderson_darling_normality(&data).ok_or_else(|| {
        js_err(
            WireError::invalid_input(
                "data: every value is the same, so normality cannot be tested",
            )
            .about("data"),
        )
    })?;

    let dto = AdNormalityDto {
        statistic: result.statistic,
        statistic_modified: result.statistic_modified,
        p_value: result.p_value,
    };
    to_js(&dto)
}

/// Compute the Laney P' chart from (defectives, sample_size) pairs.
///
/// Adjusts control limits for overdispersion via a φ correction factor.
///
/// # Input JSON
///
/// Array of `[defectives, sample_size]` pairs:
/// `[[3, 100], [5, 100], ...]` (need >= 3 subgroups).
///
/// # Errors
///
/// As [`p_chart`], with `insufficient_data` below 3 samples (1 with a standard).
///
/// # Options
///
/// `{ p_bar?: number, phi?: number }` -- both from a Phase I study, or neither:
/// phi scales the standard error about that centre, so the two are one
/// standard.
///
/// # Output JSON
///
/// Object with fields: `p_bar`, `phi` (the values used), `points` (array, each
/// with `z`).
#[wasm_bindgen(unchecked_return_type = "LaneyPChartDto")]
pub fn laney_p_chart(
    #[wasm_bindgen(unchecked_param_type = "[number, number][]")] samples: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "LaneyPChartOptions | null")] options: Option<
        JsValue,
    >,
) -> Result<JsValue, JsValue> {
    let json: serde_json::Value = from_js(samples, "samples")?;
    let samples = count_pairs(&json, "samples").map_err(js_err)?;
    let standard = standard_option(options)?;
    to_js(&laney_p_dto(&samples, &standard).map_err(js_err)?)
}

/// Compute an NP chart: the count of defectives in samples of one size.
///
/// # Input
///
/// - `defectives`: `[d1, d2, ...]`, each at most `sample_size`
/// - `sample_size`: the constant sample size, above 0
///
/// # Errors
///
/// Throws an `Error` with `code` and `index`: `count_not_whole` with the row
/// for a defective count, `sample_size_not_whole` with `index: null` for the
/// `sample_size` argument, `defectives_exceed_sample`, `insufficient_data` for
/// an empty array.
///
/// # Output JSON
///
/// Object with fields: `cl`, `ucl`, `lcl`, `points` (array), `in_control`.
#[wasm_bindgen(unchecked_return_type = "FixedLimitChartDto")]
pub fn np_chart(
    #[wasm_bindgen(unchecked_param_type = "number[]")] defectives: JsValue,
    #[wasm_bindgen(unchecked_param_type = "number")] sample_size: JsValue,
) -> Result<JsValue, JsValue> {
    let defectives: serde_json::Value = from_js(defectives, "defectives")?;
    let defectives = count_rows(&defectives, "defectives").map_err(js_err)?;
    let sample_size: serde_json::Value = from_js(sample_size, "sample_size")?;
    let sample_size = sample_size_value(&sample_size, "sample_size").map_err(js_err)?;
    to_js(&np_chart_dto(&defectives, sample_size).map_err(js_err)?)
}

/// Compute a C chart: the count of defects per inspection unit of one size.
///
/// # Input JSON
///
/// `[c1, c2, ...]` (need >= 1).
///
/// # Errors
///
/// Throws an `Error` with `code` and `index`: `count_not_whole` with the row,
/// `insufficient_data` for an empty array.
///
/// # Output JSON
///
/// Object with fields: `cl`, `ucl`, `lcl`, `points` (array), `in_control`.
#[wasm_bindgen(unchecked_return_type = "FixedLimitChartDto")]
pub fn c_chart(
    #[wasm_bindgen(unchecked_param_type = "number[]")] defects: JsValue,
) -> Result<JsValue, JsValue> {
    let defects: serde_json::Value = from_js(defects, "defects")?;
    let defects = count_rows(&defects, "defects").map_err(js_err)?;
    to_js(&c_chart_dto(&defects).map_err(js_err)?)
}

/// Compute a U chart: defects per unit when the quantity inspected varies.
///
/// # Input JSON
///
/// Array of `[defects, units]` pairs. `units` may be fractional (an area, a
/// length) and must be positive.
///
/// # Errors
///
/// Throws an `Error` with `code` and `index` (the row): `count_not_whole`,
/// `units_not_positive`, `malformed_input`, `insufficient_data`,
/// `standard_out_of_range`.
///
/// # Options
///
/// `{ u_bar?: number }` -- a known centre line (Phase II), as for [`p_chart`].
///
/// # Output JSON
///
/// Object with fields: `u_bar`, `points` (array, each with `z`), `in_control`.
#[wasm_bindgen(unchecked_return_type = "UChartDto")]
pub fn u_chart(
    #[wasm_bindgen(unchecked_param_type = "[number, number][]")] samples: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "UChartOptions | null")] options: Option<
        JsValue,
    >,
) -> Result<JsValue, JsValue> {
    let json: serde_json::Value = from_js(samples, "samples")?;
    let samples = rate_pairs(&json, "samples").map_err(js_err)?;
    let standard = standard_option(options)?;
    to_js(&u_chart_dto(&samples, &standard).map_err(js_err)?)
}

/// Compute the Laney U' chart from `[defects, units]` pairs (need >= 3).
///
/// Adjusts the U chart's limits for overdispersion via a φ correction factor.
///
/// # Errors
///
/// As [`u_chart`], with `insufficient_data` below 3 samples (1 with a standard).
///
/// # Options
///
/// `{ u_bar?: number, phi?: number }` -- both or neither, as for
/// [`laney_p_chart`].
///
/// # Output JSON
///
/// Object with fields: `u_bar`, `phi`, `points` (array, each with `z`).
#[wasm_bindgen(unchecked_return_type = "LaneyUChartDto")]
pub fn laney_u_chart(
    #[wasm_bindgen(unchecked_param_type = "[number, number][]")] samples: JsValue,
    #[wasm_bindgen(unchecked_optional_param_type = "LaneyUChartOptions | null")] options: Option<
        JsValue,
    >,
) -> Result<JsValue, JsValue> {
    let json: serde_json::Value = from_js(samples, "samples")?;
    let samples = rate_pairs(&json, "samples").map_err(js_err)?;
    let standard = standard_option(options)?;
    to_js(&laney_u_dto(&samples, &standard).map_err(js_err)?)
}

// ---------------------------------------------------------------------------
// Pure cores of the attributes-chart bindings
// ---------------------------------------------------------------------------
//
// Every attributes chart drops a row it cannot use, and every later point then
// carries the index of the wrong row. The cores refuse such a row by its index
// before the chart sees it.

/// Compute the G chart for rare-event monitoring (inter-event conforming counts).
///
/// Suitable when defect rates are very low (< 1%).
///
/// # Parameters
///
/// - `gaps`: `number[]` or `Float64Array` of inter-event conforming counts
///   (need >= 3, each >= 0)
///
/// # Output JSON
///
/// Object with fields: `g_bar`, `points` (array with ucl/cl/lcl/out_of_control).
#[wasm_bindgen(unchecked_return_type = "GChartDto")]
pub fn g_chart(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] gaps: JsValue,
) -> Result<JsValue, JsValue> {
    let gaps = read_numbers(&gaps, "gaps").map_err(js_err)?;
    to_js(&g_chart_dto(&gaps).map_err(js_err)?)
}

/// Compute the T chart for rare-event monitoring (inter-event times).
///
/// Control limits are derived from exponential distribution percentiles.
///
/// # Parameters
///
/// - `times`: `number[]` or `Float64Array` of inter-event times (need >= 3, each > 0)
///
/// # Output JSON
///
/// Object with fields: `t_bar`, `points` (array with ucl/cl/lcl/out_of_control).
#[wasm_bindgen(unchecked_return_type = "TChartDto")]
pub fn t_chart(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] times: JsValue,
) -> Result<JsValue, JsValue> {
    let times = read_numbers(&times, "times").map_err(js_err)?;
    to_js(&t_chart_dto(&times).map_err(js_err)?)
}

// ---------------------------------------------------------------------------
// Hypothesis tests
// ---------------------------------------------------------------------------
//
// Every sample is a `number[]` or `Float64Array` read as sent (`read_numbers`):
// `null` or a string is `malformed_input` and a NaN `value_not_finite`, each at
// its index. Too few values is `insufficient_data`; data with no variation,
// for which the statistic is undefined, is `invalid_input`.

/// One-sample t test of `H0: mean = mu0` (need >= 2 values).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn one_sample_t_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] data: JsValue,
    mu0: f64,
) -> Result<JsValue, JsValue> {
    let data = read_numbers(&data, "data").map_err(js_err)?;
    to_js(&one_sample_t_dto(&data, mu0).map_err(js_err)?)
}

/// Welch's two-sample t test (unequal variances; need >= 2 values each).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn two_sample_t_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] a: JsValue,
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] b: JsValue,
) -> Result<JsValue, JsValue> {
    let a = read_numbers(&a, "a").map_err(js_err)?;
    let b = read_numbers(&b, "b").map_err(js_err)?;
    to_js(&two_sample_t_dto(&a, &b).map_err(js_err)?)
}

/// Paired t test on `y - x` (same length, need >= 2 pairs).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn paired_t_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] x: JsValue,
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] y: JsValue,
) -> Result<JsValue, JsValue> {
    let x = read_numbers(&x, "x").map_err(js_err)?;
    let y = read_numbers(&y, "y").map_err(js_err)?;
    to_js(&paired_t_dto(&x, &y).map_err(js_err)?)
}

/// Mann-Whitney U test, normal approximation with tie correction (need >= 2 values each).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn mann_whitney_u_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] a: JsValue,
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] b: JsValue,
) -> Result<JsValue, JsValue> {
    let a = read_numbers(&a, "a").map_err(js_err)?;
    let b = read_numbers(&b, "b").map_err(js_err)?;
    to_js(&mann_whitney_dto(&a, &b).map_err(js_err)?)
}

/// Wilcoxon signed-rank test on paired samples (same length; pairs with `x = y` are dropped, >= 2 must remain).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn wilcoxon_signed_rank_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] x: JsValue,
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] y: JsValue,
) -> Result<JsValue, JsValue> {
    let x = read_numbers(&x, "x").map_err(js_err)?;
    let y = read_numbers(&y, "y").map_err(js_err)?;
    to_js(&wilcoxon_dto(&x, &y).map_err(js_err)?)
}

/// Chi-squared goodness of fit: `observed` counts (>= 0) against `expected` frequencies (> 0), same length >= 2.
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn chi_squared_goodness_of_fit(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] observed: JsValue,
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] expected: JsValue,
) -> Result<JsValue, JsValue> {
    let observed = read_numbers(&observed, "observed").map_err(js_err)?;
    let expected = read_numbers(&expected, "expected").map_err(js_err)?;
    to_js(&chi_squared_gof_dto(&observed, &expected).map_err(js_err)?)
}

/// Jarque-Bera normality test (need >= 8 values).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn jarque_bera_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] data: JsValue,
) -> Result<JsValue, JsValue> {
    let data = read_numbers(&data, "data").map_err(js_err)?;
    to_js(&jarque_bera_dto(&data).map_err(js_err)?)
}

/// Shapiro-Wilk normality test, Royston (1995) (3 to 5000 values). Returns `{ w, p_value }`.
#[wasm_bindgen(unchecked_return_type = "ShapiroWilkDto")]
pub fn shapiro_wilk_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] data: JsValue,
) -> Result<JsValue, JsValue> {
    let data = read_numbers(&data, "data").map_err(js_err)?;
    to_js(&shapiro_wilk_dto(&data).map_err(js_err)?)
}

/// Mann-Kendall trend test with Kendall's tau and Sen's slope (need >= 4 values).
#[wasm_bindgen(unchecked_return_type = "MannKendallDto")]
pub fn mann_kendall_test(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] data: JsValue,
) -> Result<JsValue, JsValue> {
    let data = read_numbers(&data, "data").map_err(js_err)?;
    to_js(&mann_kendall_dto(&data).map_err(js_err)?)
}

/// One-way ANOVA across `groups` (>= 2 groups of >= 2 values).
#[wasm_bindgen(unchecked_return_type = "AnovaDto")]
pub fn one_way_anova(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] groups: JsValue,
) -> Result<JsValue, JsValue> {
    let groups = read_number_rows(&groups, "groups").map_err(js_err)?;
    to_js(&one_way_anova_dto(&groups).map_err(js_err)?)
}

/// Kruskal-Wallis H test across `groups` (>= 2 groups of >= 2 values).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn kruskal_wallis_test(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] groups: JsValue,
) -> Result<JsValue, JsValue> {
    let groups = read_number_rows(&groups, "groups").map_err(js_err)?;
    to_js(&groups_test_dto("kruskal_wallis_test", &groups).map_err(js_err)?)
}

/// Levene's test (Brown-Forsythe, median-centred) for equal variances across `groups`.
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn levene_test(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] groups: JsValue,
) -> Result<JsValue, JsValue> {
    let groups = read_number_rows(&groups, "groups").map_err(js_err)?;
    to_js(&groups_test_dto("levene_test", &groups).map_err(js_err)?)
}

/// Bartlett's test for equal variances across `groups` (each group must vary).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn bartlett_test(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] groups: JsValue,
) -> Result<JsValue, JsValue> {
    let groups = read_number_rows(&groups, "groups").map_err(js_err)?;
    to_js(&groups_test_dto("bartlett_test", &groups).map_err(js_err)?)
}

/// Chi-squared test of independence on a contingency `table` (>= 2 x 2, counts >= 0).
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn chi_squared_independence(
    #[wasm_bindgen(unchecked_param_type = "number[][]")] table: JsValue,
) -> Result<JsValue, JsValue> {
    let table = read_number_rows(&table, "table").map_err(js_err)?;
    to_js(&chi_squared_independence_dto(&table).map_err(js_err)?)
}

/// Fisher's exact test on a 2 x 2 `table` of whole counts `[[a, b], [c, d]]`.
#[wasm_bindgen(unchecked_return_type = "TestResultDto")]
pub fn fisher_exact_test(
    #[wasm_bindgen(unchecked_param_type = "[[number, number], [number, number]]")] table: JsValue,
) -> Result<JsValue, JsValue> {
    let table: serde_json::Value = from_js(table, "table")?;
    to_js(&fisher_exact_dto(&table).map_err(js_err)?)
}

/// Bonferroni-adjusted p-values, `min(p * m, 1)`, in input order.
#[wasm_bindgen]
pub fn bonferroni_correction(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] p_values: JsValue,
) -> Result<Vec<f64>, JsValue> {
    let p = read_numbers(&p_values, "p_values").map_err(js_err)?;
    p_adjust_dto("bonferroni_correction", &p).map_err(js_err)
}

/// Benjamini-Hochberg (1995) FDR-adjusted p-values, in input order.
#[wasm_bindgen]
pub fn benjamini_hochberg(
    #[wasm_bindgen(unchecked_param_type = "number[] | Float64Array")] p_values: JsValue,
) -> Result<Vec<f64>, JsValue> {
    let p = read_numbers(&p_values, "p_values").map_err(js_err)?;
    p_adjust_dto("benjamini_hochberg", &p).map_err(js_err)
}

// ---------------------------------------------------------------------------
// PELT changepoint detection
// ---------------------------------------------------------------------------

/// Detect changepoints using the PELT algorithm (Killick et al., 2012).
///
/// # Input JSON
///
/// ```json
/// {
///   "data": [1.0, 1.1, 0.9, 5.0, 5.1, 4.9],
///   "cost": "l2",
///   "penalty": "bic",
///   "min_segment_len": 2
/// }
/// ```
///
/// - `data` (required): Array of f64 values.
/// - `cost` (optional): `"l2"` (mean change, default) or `"normal"` (mean+variance).
/// - `penalty` (optional): `"bic"` (default) or a positive number.
/// - `min_segment_len` (optional): Minimum segment length (default 2, must be >= 2).
///
/// # Output JSON
///
/// ```json
/// { "changepoints": [3], "n_segments": 2 }
/// ```
#[wasm_bindgen(unchecked_return_type = "PeltResultDto")]
pub fn detect_changepoints(
    #[wasm_bindgen(unchecked_param_type = "PeltInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: PeltInputDto = from_js(input, "input")?;
    to_js(&pelt_dto(input).map_err(js_err)?)
}

// ---------------------------------------------------------------------------
// Multi-signal PELT
// ---------------------------------------------------------------------------

#[derive(Deserialize, tsify::Tsify)]
#[tsify(missing_as_null)]
#[serde(deny_unknown_fields)]
struct MultiPeltInputDto {
    signals: Vec<Vec<f64>>,
    #[serde(default = "default_cost")]
    #[tsify(optional)]
    #[tsify(type = "\"l2\" | \"normal\"")]
    cost: String,
    #[serde(default = "default_penalty")]
    #[tsify(optional)]
    #[tsify(type = "\"bic\" | number")]
    penalty: PeltPenaltyDto,
    #[serde(default = "default_min_seg")]
    #[tsify(optional)]
    min_segment_len: usize,
}

/// Detect changepoints in multi-signal data using PELT.
///
/// # Input JSON
///
/// ```json
/// {
///   "signals": [[0.0, 0.0, 5.0, 5.0], [0.0, 0.0, 3.0, 3.0]],
///   "cost": "l2",
///   "penalty": "bic",
///   "min_segment_len": 2
/// }
/// ```
///
/// - `signals` (required): Array of signal channels (each same length).
/// - `cost`, `penalty`, `min_segment_len`: Same as `detect_changepoints`.
///
/// # Output JSON
///
/// ```json
/// { "changepoints": [2], "n_segments": 2 }
/// ```
#[wasm_bindgen(unchecked_return_type = "PeltResultDto")]
pub fn detect_changepoints_multi(
    #[wasm_bindgen(unchecked_param_type = "MultiPeltInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: MultiPeltInputDto = from_js(input, "input")?;

    if input.signals.is_empty() {
        return Err(js_err("signals must not be empty"));
    }

    let cost = match input.cost.as_str() {
        "l2" => crate::detection::CostFunction::L2,
        "normal" => crate::detection::CostFunction::Normal,
        other => return Err(js_err(format!("unknown cost function: {other}"))),
    };

    let penalty = match input.penalty {
        PeltPenaltyDto::Named(ref s) if s == "bic" => crate::detection::Penalty::Bic,
        PeltPenaltyDto::Named(ref s) => return Err(js_err(format!("unknown penalty: {s}"))),
        PeltPenaltyDto::Value(v) => crate::detection::Penalty::Custom(v),
    };

    let pelt = crate::detection::Pelt::with_min_segment_len(cost, penalty, input.min_segment_len)
        .ok_or_else(|| js_err("invalid parameters"))?;

    let refs: Vec<&[f64]> = input.signals.iter().map(|s| s.as_slice()).collect();
    let result = pelt
        .detect_multi(&refs)
        .ok_or_else(|| js_err("all signals must have the same length"))?;

    let dto = PeltResultDto {
        n_segments: result.changepoints.len() + 1,
        changepoints: result.changepoints,
    };
    to_js(&dto)
}

// ---------------------------------------------------------------------------
// Gage R&R (MSA)
// ---------------------------------------------------------------------------

/// Compute Gage R&R using the X̄-R (Average & Range) method.
///
/// # Input JSON
///
/// ```json
/// {
///   "measurements": [[[0.29, 0.41], [0.08, 0.25]], [[1.34, 1.17], [1.19, 0.94]]],
///   "tolerance": 4.0
/// }
/// ```
///
/// `measurements[part][operator][trial]`. Tolerance is optional.
///
/// # Output JSON
///
/// Object with fields: `ev`, `av`, `grr`, `pv`, `tv`, `percent_ev`, `percent_av`,
/// `percent_grr`, `percent_pv`, `percent_tolerance`, `ndc`, `status`.
#[wasm_bindgen(unchecked_return_type = "GageRRResultDto")]
pub fn gage_rr_xbar_r(
    #[wasm_bindgen(unchecked_param_type = "GageRRInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: GageRRInputDto = from_js(input, "input")?;
    to_js(&gage_rr_xbar_r_dto(input).map_err(js_err)?)
}

/// Compute Gage R&R using the two-factor crossed ANOVA method.
///
/// # Input JSON
///
/// Same format as `gage_rr_xbar_r`.
///
/// # Output JSON
///
/// Object with fields: `anova_table`, `variance_components`, `ev`, `av`, `grr`,
/// `pv`, `tv`, `percent_grr`, `percent_tolerance`, `ndc`, `status`,
/// `interaction_significant`, `interaction_pooled`.
#[wasm_bindgen(unchecked_return_type = "GageRRAnovaResultDto")]
pub fn gage_rr_anova(
    #[wasm_bindgen(unchecked_param_type = "GageRRInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: GageRRInputDto = from_js(input, "input")?;
    to_js(&gage_rr_anova_dto(input).map_err(js_err)?)
}

/// Compute percentile-based capability indices (ISO 22514-2).
///
/// # Input JSON
///
/// ```json
/// {
///   "data": [1.0, 2.0, ...],
///   "lsl": 0.0,
///   "usl": 10.0
/// }
/// ```
///
/// At least one of `lsl` or `usl` must be provided. Requires >= 20 data points.
///
/// # Output JSON
///
/// Object with fields: `cp_star`, `cpk_star`, `cpu_star`, `cpl_star`,
/// `median`, `percentile_lower`, `percentile_upper`.
#[wasm_bindgen(unchecked_return_type = "PercentileCapabilityDto")]
pub fn percentile_capability(
    #[wasm_bindgen(unchecked_param_type = "PercentileCapabilityInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: PercentileCapabilityInputDto = from_js(input, "input")?;
    to_js(&percentile_capability_dto(input).map_err(js_err)?)
}

// ---------------------------------------------------------------------------
// CUSUM / EWMA -- online (sequential) shift detection
// ---------------------------------------------------------------------------

#[derive(Deserialize, tsify::Tsify)]
#[tsify(missing_as_null)]
#[serde(deny_unknown_fields)]
struct CusumInputDto {
    data: Vec<f64>,
    target: f64,
    sigma: f64,
    #[serde(default = "default_cusum_k")]
    #[tsify(optional)]
    k: f64,
    #[serde(default = "default_cusum_h")]
    #[tsify(optional)]
    h: f64,
}

fn default_cusum_k() -> f64 {
    0.5
}

fn default_cusum_h() -> f64 {
    5.0
}

#[derive(Serialize, Debug, tsify::Tsify)]
#[tsify(missing_as_null)]
struct CusumPointDto {
    index: usize,
    s_upper: f64,
    s_lower: f64,
    signal: bool,
}

#[derive(Serialize, Debug, tsify::Tsify)]
#[tsify(missing_as_null)]
struct CusumDto {
    /// Decision interval actually used, echoed so a caller can draw the
    /// boundary without restating its own input. Unlike EWMA's widening
    /// limits this one is constant, so it belongs to the chart, not the point.
    h: f64,
    points: Vec<CusumPointDto>,
    signal_indices: Vec<usize>,
    in_control: bool,
}

#[derive(Deserialize, tsify::Tsify)]
#[tsify(missing_as_null)]
#[serde(deny_unknown_fields)]
struct EwmaInputDto {
    data: Vec<f64>,
    target: f64,
    sigma: f64,
    #[serde(default = "default_ewma_lambda")]
    #[tsify(optional)]
    lambda: f64,
    #[serde(default = "default_ewma_l_factor")]
    #[tsify(optional)]
    l_factor: f64,
}

fn default_ewma_lambda() -> f64 {
    0.2
}

fn default_ewma_l_factor() -> f64 {
    3.0
}

#[derive(Serialize, Debug, tsify::Tsify)]
#[tsify(missing_as_null)]
struct EwmaPointDto {
    index: usize,
    ewma: f64,
    ucl: f64,
    lcl: f64,
    signal: bool,
}

#[derive(Serialize, Debug, tsify::Tsify)]
#[tsify(missing_as_null)]
struct EwmaDto {
    points: Vec<EwmaPointDto>,
    signal_indices: Vec<usize>,
    in_control: bool,
}

/// CUSUM chart (Page, 1954) -- detects small persistent shifts in the mean.
///
/// # Input JSON
///
/// ```json
/// { "data": [10.1, 9.8, 12.0], "target": 10.0, "sigma": 1.0, "k": 0.5, "h": 5.0 }
/// ```
///
/// - `data` (required): observations, in order. Must be non-empty.
/// - `target` (required): process target mean (mu_0).
/// - `sigma` (required): known process standard deviation. Must be positive.
/// - `k` (optional): reference value / allowance, default `0.5` (optimal for a
///   1-sigma shift). Must be non-negative.
/// - `h` (optional): decision interval, default `5.0` (ARL_0 ~ 465). Must be positive.
///
/// # Output JSON
///
/// ```json
/// {
///   "h": 5.0,
///   "points": [{ "index": 0, "s_upper": 0.0, "s_lower": 0.0, "signal": false }],
///   "signal_indices": [],
///   "in_control": true
/// }
/// ```
///
/// Both cumulative sums are on the **standardized** scale (`z = (x - target) / sigma`)
/// and start at zero, so they compare against `h` directly.
#[wasm_bindgen(unchecked_return_type = "CusumDto")]
pub fn cusum(
    #[wasm_bindgen(unchecked_param_type = "CusumInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: CusumInputDto = from_js(input, "input")?;
    let dto = cusum_dto(input).map_err(js_err)?;
    to_js(&dto)
}

/// The half of [`cusum`] below the `JsValue` boundary, so the contract is
/// testable off `wasm32`.
fn cusum_dto(input: CusumInputDto) -> Result<CusumDto, String> {
    if input.data.is_empty() {
        return Err("data must not be empty".to_owned());
    }

    let chart = crate::detection::Cusum::with_params(input.target, input.sigma, input.k, input.h)
        .ok_or_else(|| {
        format!(
            "invalid parameters (target must be finite, sigma > 0, k >= 0, h > 0); \
                 got target={}, sigma={}, k={}, h={}",
            input.target, input.sigma, input.k, input.h
        )
    })?;

    let results = chart.analyze(&input.data);
    let signal_indices: Vec<usize> = results
        .iter()
        .filter(|r| r.signal)
        .map(|r| r.index)
        .collect();

    Ok(CusumDto {
        h: input.h,
        in_control: signal_indices.is_empty(),
        signal_indices,
        points: results
            .into_iter()
            .map(|r| CusumPointDto {
                index: r.index,
                s_upper: r.s_upper,
                s_lower: r.s_lower,
                signal: r.signal,
            })
            .collect(),
    })
}

/// EWMA chart (Roberts, 1959) -- exponentially weighted moving average of the mean.
///
/// # Input JSON
///
/// ```json
/// { "data": [10.1, 9.8, 12.0], "target": 10.0, "sigma": 1.0, "lambda": 0.2, "l_factor": 3.0 }
/// ```
///
/// - `data` (required): observations, in order. Must be non-empty.
/// - `target` (required): process target mean (mu_0).
/// - `sigma` (required): known process standard deviation. Must be positive.
/// - `lambda` (optional): smoothing constant, default `0.2`. Must be in `(0, 1]`.
/// - `l_factor` (optional): control-limit width factor, default `3.0`. Must be positive.
///
/// # Output JSON
///
/// ```json
/// {
///   "points": [{ "index": 0, "ewma": 10.02, "ucl": 10.6, "lcl": 9.4, "signal": false }],
///   "signal_indices": [],
///   "in_control": true
/// }
/// ```
///
/// The limits **widen with the observation index** (they are exact, not asymptotic),
/// so they are returned per point rather than once for the chart.
#[wasm_bindgen(unchecked_return_type = "EwmaDto")]
pub fn ewma(
    #[wasm_bindgen(unchecked_param_type = "EwmaInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: EwmaInputDto = from_js(input, "input")?;
    let dto = ewma_dto(input).map_err(js_err)?;
    to_js(&dto)
}

/// The half of [`ewma`] below the `JsValue` boundary, so the contract is
/// testable off `wasm32`.
fn ewma_dto(input: EwmaInputDto) -> Result<EwmaDto, String> {
    if input.data.is_empty() {
        return Err("data must not be empty".to_owned());
    }

    let chart = crate::detection::Ewma::with_params(
        input.target,
        input.sigma,
        input.lambda,
        input.l_factor,
    )
    .ok_or_else(|| {
        format!(
            "invalid parameters (target must be finite, sigma > 0, 0 < lambda <= 1, \
             l_factor > 0); got target={}, sigma={}, lambda={}, l_factor={}",
            input.target, input.sigma, input.lambda, input.l_factor
        )
    })?;

    let results = chart.analyze(&input.data);
    let signal_indices: Vec<usize> = results
        .iter()
        .filter(|r| r.signal)
        .map(|r| r.index)
        .collect();

    Ok(EwmaDto {
        in_control: signal_indices.is_empty(),
        signal_indices,
        points: results
            .into_iter()
            .map(|r| EwmaPointDto {
                index: r.index,
                ewma: r.ewma,
                ucl: r.ucl,
                lcl: r.lcl,
                signal: r.signal,
            })
            .collect(),
    })
}

// ---------------------------------------------------------------------------
// Non-normal capability (Box-Cox) and sigma level <-> PPM
// ---------------------------------------------------------------------------

#[derive(Deserialize, tsify::Tsify)]
#[tsify(missing_as_null)]
#[serde(deny_unknown_fields)]
struct BoxcoxCapabilityInputDto {
    data: Vec<f64>,
    #[serde(default)]
    #[tsify(optional)]
    #[tsify(type = "number | null")]
    usl: Option<f64>,
    #[serde(default)]
    #[tsify(optional)]
    #[tsify(type = "number | null")]
    lsl: Option<f64>,
    /// `[min, max]` lambda search range; defaults to `[-5, 5]`.
    #[serde(default)]
    #[tsify(optional)]
    #[tsify(type = "[number, number] | null")]
    lambda_range: Option<[f64; 2]>,
}

#[derive(Serialize, Debug, tsify::Tsify)]
#[tsify(missing_as_null)]
struct BoxcoxCapabilityDto {
    /// Estimated optimal Box-Cox parameter. `0` is a log transform, `1` the
    /// identity, `0.5` approximately a square root.
    lambda: f64,
    /// `true` when the likelihood maximum is on an end of `lambda_range`, so
    /// `lambda` is that limit rather than an interior estimate.
    lambda_at_bound: bool,
    /// Every index below is on the **transformed** scale, which is where the
    /// normal-theory formulas are valid -- they are not comparable to indices
    /// computed on the raw non-normal data.
    cp: Option<f64>,
    cpk: Option<f64>,
    cpu: Option<f64>,
    cpl: Option<f64>,
    pp: Option<f64>,
    ppk: Option<f64>,
    ppu: Option<f64>,
    ppl: Option<f64>,
    cpm: Option<f64>,
}

/// Process capability for non-normal data, via a Box-Cox transformation.
///
/// # Input JSON
///
/// ```json
/// { "data": [1.2, 3.4, 9.1, 22.0], "usl": 100.0, "lsl": 1.0, "lambda_range": [-5, 5] }
/// ```
///
/// - `data` (required): at least 4 observations, **all strictly positive**.
/// - `usl` / `lsl` (optional): each must be positive. With neither, only the
///   lambda estimate is returned and every index is `null`.
/// - `lambda_range` (optional): `[min, max]` search range, default `[-5, 5]`
///   (the range Minitab searches).
///
/// # Output JSON
///
/// ```json
/// { "lambda": 0.13, "lambda_at_bound": false,
///   "cp": null, "cpk": null, "cpu": null, "cpl": null,
///   "pp": 1.42, "ppk": 1.19, "ppu": 1.19, "ppl": 1.65, "cpm": null }
/// ```
///
/// The optimal lambda is estimated by maximum likelihood over `lambda_range`;
/// `lambda_at_bound: true` means the likelihood was still rising at an end of
/// the range, so `lambda` is that limit, not an interior optimum. The
/// specification limits are transformed with that same lambda, and the indices
/// are computed on the transformed scale. `cp`/`cpk`/`cpu`/`cpl` need a
/// short-term sigma, which a flat vector cannot carry, so they come back
/// `null` here -- exactly as they do from `process_capability` without
/// `sigma_within`.
#[wasm_bindgen(unchecked_return_type = "BoxcoxCapabilityDto")]
pub fn boxcox_capability(
    #[wasm_bindgen(unchecked_param_type = "BoxcoxCapabilityInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: BoxcoxCapabilityInputDto = from_js(input, "input")?;
    let dto = boxcox_capability_dto(input).map_err(js_err)?;
    to_js(&dto)
}

/// The half of [`boxcox_capability`] below the `JsValue` boundary, so the
/// contract is testable off `wasm32`.
fn boxcox_capability_dto(input: BoxcoxCapabilityInputDto) -> Result<BoxcoxCapabilityDto, String> {
    let range = input
        .lambda_range
        .map_or(crate::capability::DEFAULT_LAMBDA_RANGE, |[lo, hi]| (lo, hi));
    let result = crate::capability::boxcox_capability(&input.data, input.usl, input.lsl, range)
        .map_err(|e| e.to_string())?;
    let i = result.indices.as_ref();
    Ok(BoxcoxCapabilityDto {
        lambda: result.lambda,
        lambda_at_bound: result.lambda_at_bound,
        cp: i.and_then(|i| i.cp),
        cpk: i.and_then(|i| i.cpk),
        cpu: i.and_then(|i| i.cpu),
        cpl: i.and_then(|i| i.cpl),
        pp: i.and_then(|i| i.pp),
        ppk: i.and_then(|i| i.ppk),
        ppu: i.and_then(|i| i.ppu),
        ppl: i.and_then(|i| i.ppl),
        cpm: i.and_then(|i| i.cpm),
    })
}

// ---------------------------------------------------------------------------
// Seasonality
// ---------------------------------------------------------------------------

/// Estimate the dominant period of a univariate series.
///
/// Two stages (AutoPeriod, Vlachos et al. 2005): the linearly detrended,
/// zero-padded periodogram is searched for peaks above a permutation
/// threshold (99th percentile of the largest ordinate over 100 seeded
/// shuffles), and each peak's frequency band is refined on the
/// autocorrelation function to an integer lag that is a local maximum above
/// the `1.96/√n` white-noise bound. Deterministic for a given series.
///
/// # Input JSON
///
/// ```json
/// { "data": [0, 1, 2, 3, 4, 5, 6, 0, 1, 2, 3, 4, 5, 6, 0, 1] }
/// ```
///
/// - `data` (required): at least 8 finite values.
///
/// # Output JSON
///
/// ```json
/// { "period": 7,
///   "candidates": [{ "period": 7, "acf": 0.71, "bin": 18, "power": 21.3, "power_share": 0.62 }],
///   "n": 16, "acf_threshold": 0.49, "power_threshold": 6.8 }
/// ```
///
/// `period` is `null` — explicitly, not an error — when no periodicity passes
/// both stages: a constant, a pure trend, white noise. Only periods from 2 to
/// `n / 2` are admissible. Each candidate carries the ACF at its lag (its
/// strength), the periodogram bin that produced it (1-based, of the padded
/// transform), that bin's power and share of the total.
#[wasm_bindgen(unchecked_return_type = "SeasonalityDto")]
pub fn estimate_period(
    #[wasm_bindgen(unchecked_param_type = "SeasonalityInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: SeasonalityInputDto = from_js(input, "input")?;
    to_js(&crate::wire::seasonality_dto(input).map_err(js_err)?)
}

/// Score every point of a series for anomalies by spectral residual saliency.
///
/// Ren et al. (2019): the log amplitude spectrum minus its moving average,
/// transformed back with the original phase, is large where the series has a
/// feature its regular structure does not explain — a spike, a step, a
/// dropout. No model is trained and no period is assumed.
///
/// # Input JSON
///
/// ```json
/// { "data": [1.0, 1.1, 0.9, 1.0, 6.0, 1.0, 1.1, 0.9, 1.0, 1.0, 1.1, 0.9],
///   "averaging_window": 3, "judgement_window": 40, "threshold": 3.0,
///   "min_zscore": 1.5, "sensitivity": 70, "batch_size": null }
/// ```
///
/// - `data` (required): at least 12 finite values.
/// - `averaging_window` (optional, default 3): moving average width on the log spectrum (`q`).
/// - `judgement_window` (optional, default 40): preceding saliencies a point is scored against (`z`).
/// - `threshold` (optional, default 3): score above which a point is an anomaly (`τ`).
/// - `min_zscore` (optional, default 1.5): the point must also stand this many
///   standard deviations from the level of the window before it; `0` disables.
/// - `sensitivity` (optional, default 70): coverage in percent of the band around
///   the expected value.
/// - `batch_size` (optional): score in consecutive batches of this size (>= 12).
///
/// # Output JSON
///
/// ```json
/// { "points": [{ "index": 4, "value": 6.0, "saliency": 2.1, "score": 5.3,
///                "expected": 1.0, "lower": 0.9, "upper": 1.1, "is_anomaly": true,
///                "near_edge": false }],
///   "anomalies": [4] }
/// ```
///
/// `expected` is the low-frequency reconstruction of the series with its
/// anomalies replaced by their neighbours; `lower`/`upper` is the band
/// around it. The band is chart information — the anomaly decision is the
/// `score`.
#[wasm_bindgen(unchecked_return_type = "SpectralResidualDto")]
pub fn spectral_residual(
    #[wasm_bindgen(unchecked_param_type = "SpectralResidualInputDto")] input: JsValue,
) -> Result<JsValue, JsValue> {
    let input: SpectralResidualInputDto = from_js(input, "input")?;
    to_js(&crate::wire::spectral_residual_dto(input).map_err(js_err)?)
}

/// Converts a sigma quality level to a defect rate in parts per million.
///
/// ```js
/// sigma_to_ppm(6.0);  // -> ~3.4
/// sigma_to_ppm(3.0);  // -> ~66807
/// ```
///
/// **This is the Motorola convention, which includes the 1.5-sigma shift**
/// (`PPM = 10^6 * (1 - Phi(sigma - 1.5))`). A "six sigma" process therefore
/// reports ~3.4 PPM rather than the ~0.002 PPM an unshifted normal tail gives.
/// A consumer comparing against an unshifted table will see a different number
/// for the same input, so the convention is named here rather than inferred.
///
/// Rejects a non-finite `sigma` instead of returning `NaN`.
#[wasm_bindgen]
pub fn sigma_to_ppm(sigma: f64) -> Result<f64, JsValue> {
    if !sigma.is_finite() {
        return Err(js_err(format!("sigma must be finite, got {sigma}")));
    }
    Ok(crate::capability::sigma_to_ppm(sigma))
}

/// Converts a defect rate in parts per million to a sigma quality level.
///
/// The inverse of [`sigma_to_ppm`], on the same (1.5-shifted) convention.
///
/// ```js
/// ppm_to_sigma(3.4);      // -> ~6.0
/// ppm_to_sigma(66807.0);  // -> ~3.0
/// ```
///
/// `ppm` must lie strictly inside `(0, 1_000_000)`: both ends are limits the
/// sigma scale does not reach, so they are rejected rather than mapped to an
/// infinity.
#[wasm_bindgen]
pub fn ppm_to_sigma(ppm: f64) -> Result<f64, JsValue> {
    crate::capability::ppm_to_sigma(ppm).ok_or_else(|| {
        js_err(format!(
            "ppm must be finite and strictly inside (0, 1000000), got {ppm}"
        ))
    })
}

// ── Wire-schema strictness tests ─────────────────────────────────────

#[cfg(test)]
mod dto_strictness_tests {
    use serde_json::json;

    fn assert_rejects_unknown<T: serde::de::DeserializeOwned>(v: serde_json::Value) {
        match super::from_json::<T>(v, "input") {
            Ok(_) => panic!("unknown key must be rejected"),
            Err(e) => {
                assert!(e.message.contains("unknown field"), "{e}");
                assert_eq!(e.code, crate::wire::code::MALFORMED_INPUT);
            }
        }
    }

    #[test]
    fn capability_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::CapabilityInputDto>(
            json!({ "data": [1.0, 2.0], "usl": 3.0, "sigmaWithin": 0.5 }),
        );
    }

    #[test]
    fn pelt_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::PeltInputDto>(
            json!({ "data": [1.0, 2.0], "minSegmentLen": 3 }),
        );
    }

    #[test]
    fn multi_pelt_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::MultiPeltInputDto>(
            json!({ "signals": [[1.0, 2.0]], "penalty_value": 1.0 }),
        );
    }

    #[test]
    fn gage_rr_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::GageRRInputDto>(
            json!({ "measurements": [[[1.0]]], "tol": 0.5 }),
        );
    }

    #[test]
    fn cusum_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::CusumInputDto>(
            json!({ "data": [1.0, 2.0], "target": 1.0, "sigma": 1.0, "H": 5.0 }),
        );
    }

    #[test]
    fn ewma_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::EwmaInputDto>(
            json!({ "data": [1.0, 2.0], "target": 1.0, "sigma": 1.0, "lFactor": 3.0 }),
        );
    }

    #[test]
    fn boxcox_capability_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::BoxcoxCapabilityInputDto>(
            json!({ "data": [1.0, 2.0, 3.0, 4.0], "USL": 9.0 }),
        );
    }

    #[test]
    fn percentile_capability_input_rejects_unknown_keys() {
        assert_rejects_unknown::<super::PercentileCapabilityInputDto>(
            json!({ "data": [1.0, 2.0], "target": 1.5 }),
        );
    }
}

// ── Binding-contract tests ───────────────────────────────────────────
//
// These pin what a JavaScript caller observes, not what the underlying
// statistics compute -- the crate's own modules already cover the latter. The
// distinction matters: the defect these exist to prevent was never in
// `ProcessCapability`, which was correct throughout. It was in the binding,
// which called the long-term entry point and then labelled the result with
// short-term field names, so `cp == pp` held for every input while every unit
// test in the crate stayed green.
//
// They run under `cargo test --features wasm` on any host. `JsValue` is absent
// from every assertion below by design; the adapter that wraps it is two lines
// and has nothing left to get wrong.

#[cfg(test)]
mod binding_contract_tests {
    use super::*;
    use crate::wire::each_within;

    // --- hypothesis tests ---

    #[test]
    fn a_t_test_reports_what_the_crate_computes() {
        let data = [5.1, 4.9, 5.3, 5.0, 5.2];
        let dto = one_sample_t_dto(&data, 5.0).unwrap();
        let direct = crate::testing::one_sample_t_test(&data, 5.0).unwrap();
        assert_eq!(
            (dto.statistic, dto.df, dto.p_value),
            (direct.statistic, direct.df, direct.p_value)
        );
    }

    #[test]
    fn hypothesis_refusals_name_the_argument_and_row() {
        let e = paired_t_dto(&[1.0, 2.0, 3.0], &[1.0, 2.0]).unwrap_err();
        assert_eq!(
            (e.code, e.parameter.as_deref()),
            ("dimension_mismatch", Some("y"))
        );

        let e = one_way_anova_dto(&[vec![1.0, 2.0], vec![3.0]]).unwrap_err();
        assert_eq!(
            (e.code, e.index),
            (crate::wire::code::INSUFFICIENT_DATA, Some(1))
        );

        let e = chi_squared_independence_dto(&[vec![1.0, 2.0], vec![3.0]]).unwrap_err();
        assert_eq!((e.code, e.index), ("dimension_mismatch", Some(1)));

        let e = one_sample_t_dto(&[2.0, 2.0, 2.0], 1.0).unwrap_err();
        assert_eq!(
            (e.code, e.parameter.as_deref()),
            (crate::wire::code::INVALID_INPUT, Some("data"))
        );

        let e = chi_squared_gof_dto(&[3.0, 4.0], &[3.5, 0.0]).unwrap_err();
        assert_eq!(
            (e.code, e.index),
            (crate::wire::code::PARAMETER_OUT_OF_RANGE, Some(1))
        );
    }

    #[test]
    fn fisher_reads_a_two_by_two_table_of_whole_counts() {
        let ok = fisher_exact_dto(&serde_json::json!([[3, 1], [1, 3]])).unwrap();
        assert!(ok.p_value > 0.0 && ok.p_value <= 1.0);
        let e = fisher_exact_dto(&serde_json::json!([[3, 1.5], [1, 3]])).unwrap_err();
        assert_eq!(
            (e.code, e.index),
            (crate::wire::code::COUNT_NOT_WHOLE, Some(1))
        );
        let e = fisher_exact_dto(&serde_json::json!([[3, 1, 2], [1, 3, 0]])).unwrap_err();
        assert_eq!(e.code, "dimension_mismatch");
    }

    // --- number arrays are read as sent ---

    #[test]
    fn null_in_a_number_array_is_refused_at_its_index() {
        let e = numbers_from(
            "data",
            [
                Element::Number(1.0),
                Element::Other("null".into()),
                Element::Number(3.0),
            ],
        )
        .unwrap_err();
        assert_eq!(e.code, crate::wire::code::MALFORMED_INPUT);
        assert_eq!(e.index, Some(1));
        assert_eq!(e.parameter.as_deref(), Some("data"));
    }

    #[test]
    fn a_non_finite_element_is_value_not_finite() {
        let e = numbers_from("gaps", [2.0, f64::NAN].map(Element::Number)).unwrap_err();
        assert_eq!(e.code, crate::wire::code::VALUE_NOT_FINITE);
        assert_eq!(e.index, Some(1));
    }

    #[test]
    fn rare_event_charts_need_three_values_and_say_so() {
        let e = at_least(vec![1.0, 2.0], 3, "times").unwrap_err();
        assert_eq!(e.code, crate::wire::code::INSUFFICIENT_DATA);
        assert!(e.message.contains("at least 3"), "{}", e.message);
        // The binding's minimum is the chart's: two values make no chart.
        assert!(crate::spc::t_chart(&[1.0, 2.0]).is_none());
        assert!(crate::spc::t_chart(&[1.0, 2.0, 3.0]).is_some());
        assert!(crate::spc::g_chart(&[1.0, 2.0]).is_none());
    }

    #[test]
    fn a_value_outside_the_domain_is_named_where_it_sits() {
        let e = each_within(&[1.0, 0.0, 2.0], "times", "> 0", |v| v > 0.0).unwrap_err();
        assert_eq!(e.code, crate::wire::code::PARAMETER_OUT_OF_RANGE);
        assert_eq!(e.index, Some(1));
        assert_eq!(e.parameter.as_deref(), Some("times"));
        assert!(each_within(&[0.0, 2.0], "gaps", ">= 0", |v| v >= 0.0).is_ok());
    }

    // --- rules option on xbar_r_chart ---

    /// Every way of not asking for a rule set must mean the set the binding
    /// applied before the option existed. A caller that never passes options is
    /// the common case, and it must not have changed.
    #[test]
    fn rules_option_absent_means_the_default_set() {
        use crate::spc::RuleSet;
        for absent in [
            None,
            Some(serde_json::Value::Null),
            Some(serde_json::json!({})),
            Some(serde_json::json!({ "rules": null })),
        ] {
            assert_eq!(rules_from_json(absent).unwrap(), RuleSet::default());
        }
    }

    #[test]
    fn rules_option_selects_exactly_what_it_names() {
        use crate::spc::{RuleSet, ViolationType};
        let set = rules_from_json(Some(serde_json::json!({
            "rules": ["BeyondLimits", "SixTrend"]
        })))
        .unwrap();
        assert_eq!(
            set,
            RuleSet::from_iter([ViolationType::BeyondLimits, ViolationType::SixTrend])
        );
        assert!(!set.contains(ViolationType::NineOneSide));
    }

    #[test]
    fn rules_option_empty_array_applies_no_test() {
        use crate::spc::RuleSet;
        assert_eq!(
            rules_from_json(Some(serde_json::json!({ "rules": [] }))).unwrap(),
            RuleSet::none()
        );
    }

    #[test]
    fn rules_option_rejects_a_name_that_is_not_a_rule() {
        let err = rules_from_json(Some(serde_json::json!({ "rules": ["Rule1"] }))).unwrap_err();
        assert!(err.contains("Rule1"), "{err}");
        let err =
            rules_from_json(Some(serde_json::json!({ "rules": "BeyondLimits" }))).unwrap_err();
        assert!(err.contains("array"), "{err}");
    }

    use super::{capability_dto, from_json, CapabilityInputDto};
    use serde_json::json;

    /// The reporter's fixture (docket #219): six subgroups of five, with real
    /// between-subgroup drift, so the short-term and long-term families are
    /// genuinely different numbers.
    const SUBGROUPS: [[f64; 5]; 6] = [
        [9.9, 10.1, 10.0, 9.8, 10.2],
        [10.3, 9.7, 10.0, 10.1, 9.9],
        [9.8, 10.2, 10.1, 9.9, 10.0],
        [10.5, 9.5, 10.0, 10.2, 9.8],
        [9.6, 10.4, 10.0, 9.9, 10.1],
        [10.1, 9.9, 10.0, 10.3, 9.7],
    ];

    fn flat() -> Vec<f64> {
        SUBGROUPS.iter().flatten().copied().collect()
    }

    fn dto(v: serde_json::Value) -> Result<crate::wire::CapabilityDto, String> {
        capability_dto(from_json::<CapabilityInputDto>(v, "input").map_err(|e| e.message)?)
    }

    /// `sigma_hat` from the chart the measurements actually came from -- the
    /// quantity the flat vector cannot carry.
    fn sigma_within_from_chart() -> f64 {
        use crate::spc::{ControlChart, XBarRChart};
        let mut chart = XBarRChart::new(5).expect("5 is within range");
        for g in SUBGROUPS {
            chart.add_sample(&g).unwrap();
        }
        chart.sigma_hat().expect("limits available for 6 subgroups")
    }

    #[test]
    fn omitting_sigma_within_yields_nulls_not_long_term_numbers_under_short_term_names() {
        let d = dto(json!({ "data": flat(), "usl": 11.0, "lsl": 9.0 })).expect("valid input");

        assert_eq!(d.sigma_source, "overall");
        // The whole short-term family is absent rather than borrowed. A reader
        // can ignore a label; it cannot ignore a null.
        assert!(d.std_dev_within.is_none());
        assert!(d.cp.is_none());
        assert!(d.cpk.is_none());
        assert!(d.cpu.is_none());
        assert!(d.cpl.is_none());
        // The long-term family is what this input can support, so it is present.
        assert!(d.pp.is_some());
        assert!(d.ppk.is_some());
    }

    #[test]
    fn supplying_sigma_within_separates_the_two_index_families() {
        let sigma_within = sigma_within_from_chart();
        let d = dto(json!({
            "data": flat(), "usl": 11.0, "lsl": 9.0, "sigma_within": sigma_within,
        }))
        .expect("valid input");

        assert_eq!(d.sigma_source, "within");
        assert_eq!(d.std_dev_within, Some(sigma_within));

        let (cp, pp) = (d.cp.expect("cp"), d.pp.expect("pp"));
        let (cpk, ppk) = (d.cpk.expect("cpk"), d.ppk.expect("ppk"));

        // This is the shipped regression, stated as an assertion: the binding
        // used to make these pairs equal for every input by construction.
        assert!(
            (cp - pp).abs() > 1e-9,
            "cp and pp must come from different sigmas: cp={cp} pp={pp}"
        );
        assert!(
            (cpk - ppk).abs() > 1e-9,
            "cpk and ppk must come from different sigmas: cpk={cpk} ppk={ppk}"
        );
        // Stronger than "they differ": each family must be tied to *its own*
        // sigma. Cp = (USL-LSL)/(6*sigma), so the ratio of the two indices is
        // the inverse ratio of the two sigmas -- an identity that only holds
        // if the binding fed the short-term sigma to the short-term family and
        // the long-term sigma to the long-term one. The old binding satisfied
        // "they differ" trivially by never differing; it could not satisfy
        // this.
        let sigma_overall = d.std_dev_overall;
        assert!(
            (cp / pp - sigma_overall / sigma_within).abs() < 1e-12,
            "cp/pp must equal sigma_overall/sigma_within: \
             cp={cp} pp={pp} sigma_within={sigma_within} sigma_overall={sigma_overall}"
        );

        // Which of the two is larger is a property of the data, not of the
        // binding: this fixture's subgroup means sit close together, so the
        // range-based within estimate exceeds the pooled overall one.
        assert!(sigma_within > sigma_overall);
    }

    #[test]
    fn one_sided_specification_drops_only_the_indices_it_cannot_support() {
        let upper = dto(json!({ "data": flat(), "usl": 11.0 })).expect("usl-only is legal");
        assert!(upper.ppu.is_some(), "an upper limit supports Ppu");
        assert!(upper.ppl.is_none(), "no lower limit, no Ppl");
        assert!(upper.pp.is_none(), "Pp needs both limits");
        assert!(upper.cpm.is_none(), "Cpm needs both limits");

        let lower = dto(json!({ "data": flat(), "lsl": 9.0 })).expect("lsl-only is legal");
        assert!(lower.ppl.is_some());
        assert!(lower.ppu.is_none());
        assert!(lower.pp.is_none());
    }

    #[test]
    fn a_specification_with_no_limit_at_all_is_rejected() {
        let e = dto(json!({ "data": flat() })).expect_err("no limit is not a specification");
        assert!(e.contains("specification"), "{e}");
    }

    #[test]
    fn cpm_needs_a_declared_target_and_no_sigma() {
        // This test used to assert that omitting `target` reproduced the
        // midpoint exactly -- it pinned the substitution as the contract. Cpm
        // is defined only against a declared target, so there is nothing to
        // reproduce.
        let within = json!({ "data": flat(), "usl": 11.0, "lsl": 9.0, "sigma_within": 0.2 });
        let overall = json!({ "data": flat(), "usl": 11.0, "lsl": 9.0 });
        assert!(dto(within.clone()).expect("valid").cpm.is_none());
        assert!(dto(overall.clone()).expect("valid").cpm.is_none());

        // With a target, Cpm is the spread about it: no sigma enters, so the
        // two sigma sources give one value.
        let (mut within, mut overall) = (within, overall);
        within["target"] = json!(10.4);
        overall["target"] = json!(10.4);
        let a = dto(within).expect("valid").cpm.expect("cpm with a target");
        let b = dto(overall)
            .expect("valid")
            .cpm
            .expect("cpm without sigma_within");
        assert_eq!(a, b);
    }

    #[test]
    fn sigma_within_must_be_a_usable_standard_deviation() {
        // Representable on the wire: JSON carries these, so a JS caller can
        // actually send them.
        for bad in [0.0, -1.0] {
            let v = json!({ "data": flat(), "usl": 11.0, "lsl": 9.0, "sigma_within": bad });
            let e = dto(v).expect_err("a non-positive sigma is not a standard deviation");
            assert!(e.contains("sigma_within"), "{bad} -> {e}");
        }

        // NaN and infinity are *not* representable in JSON, so the wire cannot
        // deliver them however hard a caller tries -- `json!(f64::NAN)` is
        // `null`, which reads back as "omitted". The guard in the core is
        // therefore about direct Rust callers, and this is the only way to
        // reach it. Asserting it here documents which layer stops what.
        for bad in [f64::NAN, f64::INFINITY] {
            let input = CapabilityInputDto {
                data: flat(),
                usl: Some(11.0),
                lsl: Some(9.0),
                sigma_within: Some(bad),
                target: None,
            };
            let e = capability_dto(input).expect_err("non-finite sigma must be refused");
            assert!(e.contains("sigma_within"), "{bad} -> {e}");
        }
    }

    #[test]
    fn too_little_data_is_an_error_not_a_nan() {
        let e = dto(json!({ "data": [1.0], "usl": 11.0, "lsl": 9.0 }))
            .expect_err("one point has no dispersion");
        assert!(e.contains("insufficient"), "{e}");
    }

    // --- #220: X-bar-S, I-MR and the standalone rule engine ---

    /// Twenty subgroups whose mean steps up halfway, so the run tests have
    /// something to find. Every subgroup holds the same five offsets, so its
    /// mean is exactly 10.0 or 10.6.
    fn shifted_subgroups() -> Vec<Vec<f64>> {
        (0..20)
            .map(|g| {
                let shift = if g >= 10 { 0.6 } else { 0.0 };
                (0..5)
                    .map(|i| 10.0 + shift + 0.1 * (((g * 3 + i * 7) % 5) as f64 - 2.0))
                    .collect()
            })
            .collect()
    }

    #[test]
    fn xbar_s_chart_reports_the_crate_chart() {
        use crate::spc::{ControlChart, RuleSet, XBarSChart};
        let subgroups = shifted_subgroups();
        let d = xbar_s_dto(subgroups.clone(), RuleSet::default()).expect("valid input");

        let mut chart = XBarSChart::new(5).expect("5 is in range");
        for g in &subgroups {
            chart.add_sample(g).unwrap();
        }
        let x = chart.control_limits().expect("limits");
        let s = chart.s_limits().expect("limits");
        assert_eq!((d.xbar_ucl, d.xbar_cl, d.xbar_lcl), (x.ucl, x.cl, x.lcl));
        assert_eq!((d.s_ucl, d.s_cl, d.s_lcl), (s.ucl, s.cl, s.lcl));
        assert_eq!(d.sigma_hat, chart.sigma_hat());
        assert_eq!(d.xbar_points.len(), 20);
        assert_eq!(d.s_points.len(), 20);
        assert_eq!(d.in_control, chart.is_in_control());
    }

    #[test]
    fn a_ragged_subgroup_is_rejected_by_its_row_not_skipped() {
        // The chart would skip it and number every later point one short.
        use crate::spc::RuleSet;
        let mut subgroups = shifted_subgroups();
        subgroups[7].pop();
        for result in [
            xbar_s_dto(subgroups.clone(), RuleSet::default()).map(|_| ()),
            xbar_r_dto(subgroups, RuleSet::default()).map(|_| ()),
        ] {
            let e = result.expect_err("ragged subgroup");
            assert!(e.message.contains("subgroups[7]"), "{e}");
            assert_eq!(
                (e.code, e.index),
                (crate::wire::code::SUBGROUP_LENGTH_MISMATCH, Some(7))
            );
        }
    }

    #[test]
    fn imr_chart_reports_the_crate_chart() {
        use crate::spc::{ControlChart, IndividualMRChart, RuleSet};
        let values: Vec<f64> = shifted_subgroups().iter().map(|g| g[0]).collect();
        let d = imr_dto(values.clone(), RuleSet::default()).expect("valid input");

        let mut chart = IndividualMRChart::new();
        for &x in &values {
            chart.add_sample(&[x]).unwrap();
        }
        let i = chart.control_limits().expect("limits");
        let mr = chart.mr_limits().expect("limits");
        assert_eq!((d.i_ucl, d.i_cl, d.i_lcl), (i.ucl, i.cl, i.lcl));
        assert_eq!((d.mr_ucl, d.mr_cl, d.mr_lcl), (mr.ucl, mr.cl, mr.lcl));
        assert_eq!(d.sigma_hat, chart.sigma_hat());
        assert_eq!(d.i_points.len(), values.len());
        // MR_0 is undefined: the moving-range series starts at the second value.
        assert_eq!(d.mr_points.len(), values.len() - 1);
        assert_eq!(d.mr_points[0].index, 1);
    }

    #[test]
    fn imr_chart_needs_two_values() {
        use crate::spc::RuleSet;
        let e = imr_dto(vec![1.0], RuleSet::default()).expect_err("one value has no range");
        assert!(e.message.contains("two"), "{e}");
    }

    /// The standalone engine must be the engine the charts use: fed a chart's
    /// own points and limits, it has to find exactly what the chart found.
    #[test]
    fn run_rules_finds_what_the_chart_found() {
        use crate::spc::RuleSet;
        let chart = xbar_r_dto(shifted_subgroups(), RuleSet::default()).expect("valid");
        let flagged = chart
            .xbar_points
            .iter()
            .filter(|p| !p.violations.is_empty())
            .count();
        assert!(
            flagged > 0,
            "the fixture must give the rules something to find"
        );

        let values = chart.xbar_points.iter().map(|p| p.value).collect();
        let limits = LimitsInputDto {
            ucl: chart.xbar_ucl,
            cl: chart.xbar_cl,
            lcl: chart.xbar_lcl,
        };
        let standalone = run_rules_dto(values, limits, RuleSet::default()).expect("valid");

        assert_eq!(standalone.len(), chart.xbar_points.len());
        for (a, b) in standalone.iter().zip(&chart.xbar_points) {
            assert_eq!(a.index, b.index);
            assert_eq!(a.violations, b.violations, "point {}", a.index);
        }
    }

    #[test]
    fn run_rules_applies_only_the_rules_it_is_given() {
        use crate::spc::{RuleSet, ViolationType};
        // One point far outside, the rest on the centre line.
        let mut values = vec![0.0; 10];
        values[4] = 5.0;
        let limits = || LimitsInputDto {
            ucl: 3.0,
            cl: 0.0,
            lcl: -3.0,
        };

        let all = run_rules_dto(values.clone(), limits(), RuleSet::default()).expect("valid");
        assert!(all[4].violations.contains(&"BeyondLimits".to_owned()));

        let none = run_rules_dto(values.clone(), limits(), RuleSet::none()).expect("valid");
        assert!(none.iter().all(|p| p.violations.is_empty()));

        let only_trend = RuleSet::none().with(ViolationType::SixTrend);
        let trend = run_rules_dto(values, limits(), only_trend).expect("valid");
        assert!(trend[4].violations.is_empty());
    }

    #[test]
    fn run_rules_reports_every_value_in_order() {
        use crate::spc::RuleSet;
        let limits = LimitsInputDto {
            ucl: 5.0,
            cl: 2.0,
            lcl: -1.0,
        };
        let out = run_rules_dto(vec![1.0, 2.0, 3.0], limits, RuleSet::default()).expect("valid");
        let indices: Vec<usize> = out.iter().map(|p| p.index).collect();
        assert_eq!(indices, vec![0, 1, 2]);
        assert_eq!(out[2].value, 3.0);
    }

    #[test]
    fn run_rules_rejects_limits_out_of_order() {
        use crate::spc::RuleSet;
        let limits = LimitsInputDto {
            ucl: 1.0,
            cl: 2.0,
            lcl: 0.0,
        };
        let e = run_rules_dto(vec![1.0], limits, RuleSet::default()).expect_err("cl above ucl");
        assert!(e.contains("lcl <= cl <= ucl"), "{e}");
    }

    #[test]
    fn limits_input_rejects_unknown_keys() {
        let e = from_json::<LimitsInputDto>(
            json!({ "ucl": 1.0, "cl": 0.0, "lcl": -1.0, "usl": 2.0 }),
            "limits",
        )
        .err()
        .expect("unknown key");
        assert!(e.message.contains("unknown field"), "{e}");
    }

    // --- #220: attributes charts ---

    #[test]
    fn np_chart_reports_the_crate_chart() {
        use crate::spc::NPChart;
        let defectives = [5, 8, 3, 6, 4, 7, 2, 9, 5, 6];
        let d = np_chart_dto(&defectives, 100).expect("valid");
        let mut chart = NPChart::new(100).expect("valid size");
        for &x in &defectives {
            chart.add_sample(x).unwrap();
        }
        let (ucl, cl, lcl) = chart.control_limits().expect("limits");
        assert_eq!((d.ucl, d.cl, d.lcl), (ucl, cl, lcl));
        assert_eq!(d.points.len(), defectives.len());
        assert_eq!(d.in_control, chart.is_in_control());
    }

    #[test]
    fn np_chart_refuses_rows_it_cannot_chart() {
        use crate::wire::code;
        let e = np_chart_dto(&[5, 101, 3], 100).expect_err("more defectives than inspected");
        assert!(e.message.contains("defectives[1]"), "{e}");
        assert_eq!((e.code, e.index), (code::DEFECTIVES_EXCEED_SAMPLE, Some(1)));
        // A zero sample size is a value, not a panic that would trap the module.
        let e = np_chart_dto(&[0, 0], 0).expect_err("nothing inspected");
        assert!(e.message.contains("sample size"), "{e}");
        assert_eq!((e.code, e.index), (code::SAMPLE_SIZE_NOT_WHOLE, None));
    }

    #[test]
    fn c_chart_reports_the_crate_chart() {
        use crate::spc::CChart;
        let defects = [3, 5, 2, 4, 6, 3, 1, 4];
        let d = c_chart_dto(&defects).expect("valid");
        let mut chart = CChart::new();
        for &x in &defects {
            chart.add_sample(x);
        }
        let (ucl, cl, lcl) = chart.control_limits().expect("limits");
        assert_eq!((d.ucl, d.cl, d.lcl), (ucl, cl, lcl));
        assert!(c_chart_dto(&[]).is_err());
    }

    #[test]
    fn u_chart_reports_the_crate_chart_and_refuses_empty_units() {
        use crate::spc::UChart;
        let samples = [(3, 1.0), (5, 1.5), (2, 0.8), (4, 1.2)];
        let d = u_chart_dto(&samples, &Default::default()).expect("valid");
        let mut chart = UChart::new();
        for &(x, u) in &samples {
            chart.add_sample(x, u).unwrap();
        }
        assert_eq!(Some(d.u_bar), chart.u_bar());
        assert_eq!(d.points.len(), samples.len());

        let e = u_chart_dto(&[(3, 1.0), (5, 0.0), (2, 0.8)], &Default::default())
            .expect_err("zero units");
        assert!(e.message.contains("samples[1]"), "{e}");
        assert_eq!(
            (e.code, e.index),
            (crate::wire::code::UNITS_NOT_POSITIVE, Some(1))
        );
    }

    #[test]
    fn laney_u_chart_reports_the_crate_chart() {
        let samples = [(3, 1.0), (5, 1.5), (2, 0.8), (4, 1.2), (6, 1.1)];
        let d = laney_u_dto(&samples, &Default::default()).expect("valid");
        let expected = crate::spc::laney_u_chart(&samples, None).expect("valid");
        assert_eq!((d.u_bar, d.phi), (expected.u_bar, expected.phi));
        let e = laney_u_dto(&samples[..2], &Default::default()).expect_err("fewer than 3 samples");
        assert_eq!(
            (e.code, e.index),
            (crate::wire::code::INSUFFICIENT_DATA, None)
        );
        let e = laney_u_dto(&[(3, 1.0), (5, -1.0), (2, 0.8)], &Default::default())
            .expect_err("negative units");
        assert!(e.message.contains("samples[1]"), "{e}");
        assert_eq!(
            (e.code, e.index),
            (crate::wire::code::UNITS_NOT_POSITIVE, Some(1))
        );
    }

    /// `p_chart` used to hand every pair to the chart, which dropped the ones
    /// it could not use -- shifting the index of every later point.
    #[test]
    fn p_chart_refuses_a_sample_with_no_proportion() {
        use crate::wire::code;
        let e = p_chart_dto(&[(3, 100), (12, 10), (4, 100)], &Default::default())
            .expect_err("12 of 10");
        assert!(e.message.contains("samples[1]"), "{e}");
        assert_eq!((e.code, e.index), (code::DEFECTIVES_EXCEED_SAMPLE, Some(1)));
        let e = p_chart_dto(&[(3, 100), (0, 0)], &Default::default()).expect_err("0 of 0");
        assert!(e.message.contains("samples[1]"), "{e}");
        assert_eq!((e.code, e.index), (code::SAMPLE_SIZE_NOT_WHOLE, Some(1)));
        let ok = p_chart_dto(&[(3, 100), (5, 120), (2, 80)], &Default::default()).expect("valid");
        assert_eq!(ok.points.len(), 3);
    }

    // Box-Cox capability and sigma level <-> PPM

    #[test]
    fn boxcox_reports_lambda_and_long_term_indices_matching_the_crate() {
        let data: Vec<f64> = (1..=20).map(|i| (i as f64 * 0.3_f64).exp()).collect();
        let input: super::BoxcoxCapabilityInputDto =
            super::from_json(json!({ "data": data, "usl": 100.0, "lsl": 1.0 }), "input")
                .expect("valid input");
        let dto = super::boxcox_capability_dto(input).expect("skewed data is analysable");

        let native = crate::capability::boxcox_capability(
            &data,
            Some(100.0),
            Some(1.0),
            crate::capability::DEFAULT_LAMBDA_RANGE,
        )
        .expect("same call");
        let native_i = native.indices.expect("limits given");
        assert!((dto.lambda - native.lambda).abs() < 1e-12);
        assert_eq!(dto.lambda_at_bound, native.lambda_at_bound);
        assert_eq!(dto.pp, native_i.pp);
        assert_eq!(dto.ppk, native_i.ppk);

        // A flat vector carries no subgroup structure, so the short-term
        // indices must stay absent rather than borrow the long-term sigma.
        assert!(dto.cp.is_none() && dto.cpk.is_none());
        assert!(
            dto.ppk.is_some(),
            "long-term indices are the available ones"
        );
    }

    #[test]
    fn boxcox_refuses_the_inputs_the_transform_cannot_take() {
        let cases = [
            json!({ "data": [1.0, 2.0, 0.0, 4.0], "usl": 9.0 }), // non-positive value
            json!({ "data": [1.0, 2.0, 3.0], "usl": 9.0 }),      // fewer than four points
            json!({ "data": [1.0, 2.0, 3.0, 4.0], "lambda_range": [2.0, -2.0] }), // empty range
        ];
        for case in cases {
            let parsed: super::BoxcoxCapabilityInputDto =
                super::from_json(case.clone(), "input").expect("parses");
            assert!(
                super::boxcox_capability_dto(parsed).is_err(),
                "should be refused: {case}"
            );
        }
    }

    #[test]
    fn boxcox_without_limits_and_with_a_narrow_range_reports_the_bound() {
        // Normal quantiles through the inverse transform at lambda = 4: the
        // likelihood peaks near 4, so a [-2, 2] range stops at its upper end.
        let n = 100;
        let data: Vec<f64> = (1..=n)
            .map(|i| {
                let p = (i as f64 - 0.5) / n as f64;
                let z = 10.0 + 2.0 * u_numflow::special::inverse_normal_cdf(p);
                (4.0 * z + 1.0).powf(0.25)
            })
            .collect();
        let narrow: super::BoxcoxCapabilityInputDto = super::from_json(
            json!({ "data": data, "lambda_range": [-2.0, 2.0] }),
            "input",
        )
        .expect("valid input");
        let dto = super::boxcox_capability_dto(narrow).expect("no limits is allowed");
        assert!(dto.lambda_at_bound);
        assert_eq!(dto.lambda, 2.0);
        assert!(
            dto.pp.is_none() && dto.ppk.is_none(),
            "no limits, no indices"
        );

        let default: super::BoxcoxCapabilityInputDto =
            super::from_json(json!({ "data": data }), "input").expect("valid input");
        let dto = super::boxcox_capability_dto(default).expect("default range");
        assert!(!dto.lambda_at_bound, "lambda={}", dto.lambda);
    }

    #[test]
    fn sigma_and_ppm_round_trip_on_the_shifted_convention() {
        // The 1.5-sigma shift is the whole contract here: an unshifted table
        // would put six sigma near 0.002 PPM, not 3.4.
        for sigma in [3.0_f64, 4.5, 6.0] {
            let ppm = crate::capability::sigma_to_ppm(sigma);
            let back = crate::capability::ppm_to_sigma(ppm).expect("in range");
            // The inverse normal CDF here is a rational approximation, so the
            // round trip closes to ~3e-4, not to machine precision. Pinning the
            // measured accuracy rather than an aspirational one.
            assert!((back - sigma).abs() < 1e-3, "{sigma} -> {ppm} -> {back}");
        }
        assert!((crate::capability::sigma_to_ppm(6.0) - 3.4).abs() < 1.0);
        assert!((crate::capability::sigma_to_ppm(3.0) - 66_807.0).abs() < 500.0);
    }

    #[test]
    fn ppm_to_sigma_rejects_both_ends_of_the_range() {
        assert!(crate::capability::ppm_to_sigma(0.0).is_none());
        assert!(crate::capability::ppm_to_sigma(1_000_000.0).is_none());
        assert!(crate::capability::ppm_to_sigma(-1.0).is_none());
    }

    // ── CUSUM / EWMA ────────────────────────────────────────────────

    fn cusum_in(data: &[f64], target: f64, sigma: f64) -> super::CusumInputDto {
        super::from_json(
            json!({ "data": data, "target": target, "sigma": sigma }),
            "input",
        )
        .expect("valid cusum input")
    }

    fn ewma_in(data: &[f64], target: f64, sigma: f64) -> super::EwmaInputDto {
        super::from_json(
            json!({ "data": data, "target": target, "sigma": sigma }),
            "input",
        )
        .expect("valid ewma input")
    }

    #[test]
    fn cusum_defaults_match_the_crate_and_stay_in_control_on_stable_data() {
        let data = [10.1, 9.8, 10.2, 9.9, 10.0, 10.1, 9.7, 10.3];
        let dto = super::cusum_dto(cusum_in(&data, 10.0, 1.0)).expect("in-control run");

        // The binding must not invent its own defaults: k=0.5, h=5.0 (Page 1954).
        assert_eq!(dto.h, 5.0);
        assert_eq!(dto.points.len(), data.len());
        assert!(dto.in_control);
        assert!(dto.signal_indices.is_empty());

        // Same numbers the crate produces, point for point.
        let native = crate::detection::Cusum::new(10.0, 1.0)
            .expect("valid chart")
            .analyze(&data);
        for (p, n) in dto.points.iter().zip(native.iter()) {
            assert_eq!(p.index, n.index);
            assert!((p.s_upper - n.s_upper).abs() < 1e-12);
            assert!((p.s_lower - n.s_lower).abs() < 1e-12);
            assert_eq!(p.signal, n.signal);
        }
    }

    #[test]
    fn cusum_signals_a_sustained_upward_shift_and_reports_where() {
        let mut data = vec![10.0; 10];
        data.extend(vec![12.0; 10]);
        let dto = super::cusum_dto(cusum_in(&data, 10.0, 1.0)).expect("shifted run");

        assert!(!dto.in_control);
        assert!(!dto.signal_indices.is_empty());
        // Every flagged index is inside the shifted half, and signal_indices
        // agrees with the per-point flags (the redundancy must not drift).
        assert!(
            dto.signal_indices.iter().all(|&i| i >= 10),
            "{:?}",
            dto.signal_indices
        );
        let from_points: Vec<usize> = dto
            .points
            .iter()
            .filter(|p| p.signal)
            .map(|p| p.index)
            .collect();
        assert_eq!(from_points, dto.signal_indices);
    }

    #[test]
    fn cusum_refuses_empty_data_and_invalid_parameters() {
        let empty: super::CusumInputDto =
            super::from_json(json!({ "data": [], "target": 1.0, "sigma": 1.0 }), "input")
                .expect("parses");
        assert!(super::cusum_dto(empty)
            .expect_err("empty")
            .contains("empty"));

        let bad: super::CusumInputDto = super::from_json(
            json!({ "data": [1.0], "target": 1.0, "sigma": 0.0 }),
            "input",
        )
        .expect("parses");
        let e = super::cusum_dto(bad).expect_err("sigma must be > 0");
        assert!(e.contains("sigma"), "{e}");
    }

    #[test]
    fn ewma_limits_widen_with_the_index_and_match_the_crate() {
        let data = [10.1, 9.8, 10.2, 9.9, 10.0, 10.1];
        let dto = super::ewma_dto(ewma_in(&data, 10.0, 1.0)).expect("in-control run");

        assert_eq!(dto.points.len(), data.len());
        assert!(dto.in_control);

        // Exact (not asymptotic) limits: the half-width grows monotonically.
        let widths: Vec<f64> = dto.points.iter().map(|p| p.ucl - p.lcl).collect();
        for w in widths.windows(2) {
            assert!(w[1] > w[0], "limits must widen: {widths:?}");
        }

        let native = crate::detection::Ewma::new(10.0, 1.0)
            .expect("valid chart")
            .analyze(&data);
        for (p, n) in dto.points.iter().zip(native.iter()) {
            assert_eq!(p.index, n.index);
            assert!((p.ewma - n.ewma).abs() < 1e-12);
            assert!((p.ucl - n.ucl).abs() < 1e-12);
            assert!((p.lcl - n.lcl).abs() < 1e-12);
            assert_eq!(p.signal, n.signal);
        }
    }

    #[test]
    fn ewma_refuses_empty_data_and_a_lambda_outside_zero_to_one() {
        let empty: super::EwmaInputDto =
            super::from_json(json!({ "data": [], "target": 1.0, "sigma": 1.0 }), "input")
                .expect("parses");
        assert!(super::ewma_dto(empty).expect_err("empty").contains("empty"));

        for lambda in [0.0, 1.5] {
            let bad: super::EwmaInputDto = super::from_json(
                json!({ "data": [1.0, 2.0], "target": 1.0, "sigma": 1.0, "lambda": lambda }),
                "input",
            )
            .expect("parses");
            let e = super::ewma_dto(bad).expect_err("lambda out of range");
            assert!(e.contains("lambda"), "{e}");
        }
    }

    #[test]
    fn laney_p_chart_refuses_a_sample_with_no_proportion() {
        let e = laney_p_dto(&[(3, 100), (0, 0), (4, 100), (2, 100)], &Default::default())
            .expect_err("0 of 0");
        assert!(e.message.contains("samples[1]"), "{e}");
        assert_eq!(e.index, Some(1));
        let e = laney_p_dto(&[(3, 100), (5, 120)], &Default::default())
            .expect_err("fewer than 3 samples");
        assert_eq!(e.code, crate::wire::code::INSUFFICIENT_DATA);
    }
}