u-insight 0.9.0

//! C FFI bindings for u-insight.
//!
//! Exposes profiling and analysis functionality via a C-compatible interface.
//!
//! # Design (AD-4)
//!
//! - **Opaque handles**: `*mut ProfileContext` / `*mut AnalysisContext`
//! - **`#[repr(C)]`**: All data transfer structs
//! - **Integer error codes**: 0 = success, negative = error
//! - **Thread-local error message**: `insight_last_error()`
//! - **`catch_unwind`**: All FFI entry points wrapped to prevent panic propagation
//!
//! # Safety
//!
//! All functions use `catch_unwind` to prevent panics from crossing the FFI boundary.
//! Null pointer arguments return error code -1.

use std::cell::RefCell;
use std::ffi::{CStr, CString};
use std::os::raw::c_char;
use std::panic;
use std::ptr;
use std::slice;

use crate::analysis::mutual_info_classif;
use crate::clustering::{
    dbscan, gap_statistic, hdbscan, hierarchical, kmeans, mini_batch_kmeans, DbscanConfig,
    HdbscanConfig, HierarchicalConfig, KMeansConfig, Linkage, MiniBatchKMeansConfig,
};
use crate::csv_parser::CsvParser;
use crate::distribution::{distribution_analysis, DistributionConfig};
use crate::feature_importance::{feature_analysis, permutation_importance, FeatureConfig};
use crate::json_parser::JsonParser;
use crate::pca::{pca, PcaConfig};
use crate::profiling::profile_dataframe;

// ── Error handling ────────────────────────────────────────────────────

/// Error codes returned by FFI functions.
pub const INSIGHT_OK: i32 = 0;
pub const INSIGHT_ERR_NULL_PTR: i32 = -1;
pub const INSIGHT_ERR_INVALID_INPUT: i32 = -2;
pub const INSIGHT_ERR_PARSE_FAILED: i32 = -3;
pub const INSIGHT_ERR_ANALYSIS_FAILED: i32 = -4;
pub const INSIGHT_ERR_PANIC: i32 = -99;
// New granular error codes — 1:1 with InsightError variants
pub const INSIGHT_ERR_INSUFFICIENT_DATA: i32 = -5;
pub const INSIGHT_ERR_INVALID_PARAM: i32 = -6;
pub const INSIGHT_ERR_DEGENERATE_DATA: i32 = -7;
pub const INSIGHT_ERR_COMPUTATION_FAILED: i32 = -8;

thread_local! {
    static LAST_ERROR: RefCell<Option<CString>> = const { RefCell::new(None) };
}

fn error_to_code(e: &crate::error::InsightError) -> i32 {
    use crate::error::InsightError;
    match e {
        InsightError::CsvParse { .. } | InsightError::JsonParse { .. } => INSIGHT_ERR_PARSE_FAILED,
        InsightError::MissingValues { .. }
        | InsightError::NonNumericColumn { .. }
        | InsightError::ColumnNotFound { .. }
        | InsightError::DimensionMismatch { .. } => INSIGHT_ERR_INVALID_INPUT,
        InsightError::InsufficientData { .. } => INSIGHT_ERR_INSUFFICIENT_DATA,
        InsightError::InvalidParameter { .. } => INSIGHT_ERR_INVALID_PARAM,
        InsightError::DegenerateData { .. } => INSIGHT_ERR_DEGENERATE_DATA,
        InsightError::ComputationFailed { .. } => INSIGHT_ERR_COMPUTATION_FAILED,
        InsightError::Io(_) => INSIGHT_ERR_ANALYSIS_FAILED,
    }
}

fn set_last_error(msg: &str) {
    LAST_ERROR.with(|cell| {
        *cell.borrow_mut() = CString::new(msg).ok();
    });
}

/// Returns the last error message, or null if no error.
/// The returned string is valid until the next FFI call on this thread.
///
/// # Safety
/// The caller must not free the returned pointer.
#[no_mangle]
pub extern "C" fn insight_last_error() -> *const c_char {
    LAST_ERROR.with(|cell| {
        let borrow = cell.borrow();
        match borrow.as_ref() {
            Some(cstr) => cstr.as_ptr(),
            None => ptr::null(),
        }
    })
}

/// Clears the last error message.
#[no_mangle]
pub extern "C" fn insight_clear_error() {
    LAST_ERROR.with(|cell| {
        *cell.borrow_mut() = None;
    });
}

// ── Profile Context (opaque handle) ──────────────────────────────────

/// Opaque handle for a profiling context.
/// Holds the parsed DataFrame and computed profiles.
pub struct ProfileContext {
    dataframe: crate::dataframe::DataFrame,
    column_profiles: Vec<crate::profiling::ColumnProfile>,
}

/// C-compatible column profile summary.
#[repr(C)]
pub struct CColumnSummary {
    /// Column index.
    pub index: u32,
    /// Number of valid (non-null) values.
    pub valid_count: u64,
    /// Number of null values.
    pub null_count: u64,
    /// Column data type: 0=Numeric, 1=Boolean, 2=Categorical, 3=Text.
    pub data_type: u32,
    /// For numeric columns: mean. NaN for non-numeric.
    pub mean: f64,
    /// For numeric columns: standard deviation. NaN for non-numeric.
    pub std_dev: f64,
    /// For numeric columns: minimum. NaN for non-numeric.
    pub min: f64,
    /// For numeric columns: maximum. NaN for non-numeric.
    pub max: f64,
}

/// Creates a profile context from a CSV string.
///
/// # Safety
/// - `csv_data` must be a valid null-terminated UTF-8 string.
/// - The returned handle must be freed with `insight_profile_free`.
#[no_mangle]
pub unsafe extern "C" fn insight_profile_csv(csv_data: *const c_char) -> *mut ProfileContext {
    let result = panic::catch_unwind(|| {
        if csv_data.is_null() {
            set_last_error("null csv_data pointer");
            return ptr::null_mut();
        }

        let c_str = unsafe { CStr::from_ptr(csv_data) };
        let csv = match c_str.to_str() {
            Ok(s) => s,
            Err(e) => {
                set_last_error(&format!("invalid UTF-8: {e}"));
                return ptr::null_mut();
            }
        };

        let df = match CsvParser::new().parse_str(csv) {
            Ok(df) => df,
            Err(e) => {
                set_last_error(&format!("CSV parse error: {e}"));
                return ptr::null_mut();
            }
        };

        let profiles = profile_dataframe(&df);

        let ctx = Box::new(ProfileContext {
            dataframe: df,
            column_profiles: profiles,
        });
        Box::into_raw(ctx)
    });

    match result {
        Ok(ptr) => ptr,
        Err(_) => {
            set_last_error("panic in insight_profile_csv");
            ptr::null_mut()
        }
    }
}

/// Creates a profile context from a column-major JSON string.
///
/// Expected format: `{"col1": [v1, v2, ...], "col2": [...]}`
///
/// Values can be numbers, booleans, strings, or null. Column types are
/// inferred automatically: number → Numeric, bool → Boolean,
/// string → Categorical/Text (based on cardinality), null → missing.
///
/// # Safety
/// - `json_data` must be a valid null-terminated UTF-8 string.
/// - The returned handle must be freed with `insight_profile_free`.
#[no_mangle]
pub unsafe extern "C" fn insight_profile_json(json_data: *const c_char) -> *mut ProfileContext {
    let result = panic::catch_unwind(|| {
        if json_data.is_null() {
            set_last_error("null json_data pointer");
            return ptr::null_mut();
        }

        let c_str = unsafe { CStr::from_ptr(json_data) };
        let json = match c_str.to_str() {
            Ok(s) => s,
            Err(e) => {
                set_last_error(&format!("invalid UTF-8: {e}"));
                return ptr::null_mut();
            }
        };

        let df = match JsonParser::new().parse_str(json) {
            Ok(df) => df,
            Err(e) => {
                set_last_error(&format!("JSON parse error: {e}"));
                return ptr::null_mut();
            }
        };

        let profiles = profile_dataframe(&df);

        let ctx = Box::new(ProfileContext {
            dataframe: df,
            column_profiles: profiles,
        });
        Box::into_raw(ctx)
    });

    match result {
        Ok(ptr) => ptr,
        Err(_) => {
            set_last_error("panic in insight_profile_json");
            ptr::null_mut()
        }
    }
}

/// Frees a profile context.
///
/// # Safety
/// `ctx` must be a valid pointer from `insight_profile_csv` or `insight_profile_json`, or null.
#[no_mangle]
pub unsafe extern "C" fn insight_profile_free(ctx: *mut ProfileContext) {
    if !ctx.is_null() {
        let _ = unsafe { Box::from_raw(ctx) };
    }
}

/// Returns the number of rows in the profiled dataset.
///
/// # Safety
/// `ctx` must be a valid, non-null profile context.
#[no_mangle]
pub unsafe extern "C" fn insight_profile_row_count(ctx: *const ProfileContext) -> i64 {
    if ctx.is_null() {
        set_last_error("null context");
        return -1;
    }
    let ctx = unsafe { &*ctx };
    ctx.dataframe.row_count() as i64
}

/// Returns the number of columns in the profiled dataset.
///
/// # Safety
/// `ctx` must be a valid, non-null profile context.
#[no_mangle]
pub unsafe extern "C" fn insight_profile_col_count(ctx: *const ProfileContext) -> i64 {
    if ctx.is_null() {
        set_last_error("null context");
        return -1;
    }
    let ctx = unsafe { &*ctx };
    ctx.dataframe.column_count() as i64
}

/// Gets a summary for a specific column.
///
/// # Safety
/// `ctx` must be valid. `out` must point to a valid `CColumnSummary`.
#[no_mangle]
pub unsafe extern "C" fn insight_profile_column(
    ctx: *const ProfileContext,
    col_idx: u32,
    out: *mut CColumnSummary,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if ctx.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }
        let ctx = unsafe { &*ctx };
        let idx = col_idx as usize;

        if idx >= ctx.column_profiles.len() {
            set_last_error("column index out of range");
            return INSIGHT_ERR_INVALID_INPUT;
        }

        let profile = &ctx.column_profiles[idx];
        let col = ctx.dataframe.column(idx);

        let (data_type, valid_count, null_count) = match col {
            Some(c) => {
                let dt = match c.data_type() {
                    crate::dataframe::DataType::Numeric => 0u32,
                    crate::dataframe::DataType::Boolean => 1,
                    crate::dataframe::DataType::Categorical => 2,
                    crate::dataframe::DataType::Text => 3,
                };
                let vc = c.valid_count() as u64;
                let nc = c.null_count() as u64;
                (dt, vc, nc)
            }
            None => (3, 0, 0),
        };

        let (mean, std_dev, min, max) = match &profile.numeric {
            Some(np) => (np.mean, np.std_dev, np.min, np.max),
            None => (f64::NAN, f64::NAN, f64::NAN, f64::NAN),
        };

        unsafe {
            (*out) = CColumnSummary {
                index: col_idx,
                valid_count,
                null_count,
                data_type,
                mean,
                std_dev,
                min,
                max,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_profile_column");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── K-Means FFI ──────────────────────────────────────────────────────

/// C-compatible K-Means result.
#[repr(C)]
pub struct CKMeansResult {
    /// Number of clusters.
    pub k: u32,
    /// WCSS value.
    pub wcss: f64,
    /// Number of iterations.
    pub iterations: u32,
    /// Cluster labels (length = n_rows). Caller must free with `insight_free_labels`.
    pub labels: *mut u32,
    /// Number of labels.
    pub n_labels: u32,
}

/// Runs K-Means on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CKMeansResult`.
/// - The caller must free `out.labels` with `insight_free_labels`.
#[no_mangle]
pub unsafe extern "C" fn insight_kmeans(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    k: u32,
    out: *mut CKMeansResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let n = n_rows as usize;
        let d = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, n * d) };

        // Convert to Vec<Vec<f64>>
        let points: Vec<Vec<f64>> = (0..n).map(|i| raw[i * d..(i + 1) * d].to_vec()).collect();

        let config = KMeansConfig::new(k as usize);
        let km_result = match kmeans(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        // Allocate labels array
        let mut labels: Vec<u32> = km_result.labels.iter().map(|&l| l as u32).collect();
        let labels_ptr = labels.as_mut_ptr();
        let labels_len = labels.len() as u32;
        std::mem::forget(labels);

        unsafe {
            (*out) = CKMeansResult {
                k: km_result.k as u32,
                wcss: km_result.wcss,
                iterations: km_result.iterations as u32,
                labels: labels_ptr,
                n_labels: labels_len,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_kmeans");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees a labels array allocated by `insight_kmeans`.
///
/// # Safety
/// `labels` must have been allocated by an insight FFI function, or be null.
#[no_mangle]
pub unsafe extern "C" fn insight_free_labels(labels: *mut u32, count: u32) {
    if !labels.is_null() {
        let _ = unsafe { Vec::from_raw_parts(labels, count as usize, count as usize) };
    }
}

// ── PCA FFI ──────────────────────────────────────────────────────────

/// C-compatible PCA result.
#[repr(C)]
pub struct CPcaResult {
    /// Number of components retained.
    pub n_components: u32,
    /// Explained variance ratios (length = n_components).
    /// Caller must free with `insight_free_f64_array`.
    pub explained_variance: *mut f64,
    /// Number of variance values.
    pub n_variance: u32,
}

/// Runs PCA on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CPcaResult`.
/// - Caller must free `out.explained_variance` with `insight_free_f64_array`.
#[no_mangle]
pub unsafe extern "C" fn insight_pca(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    n_components: u32,
    auto_scale: i32,
    out: *mut CPcaResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let n = n_rows as usize;
        let d = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, n * d) };

        let points: Vec<Vec<f64>> = (0..n).map(|i| raw[i * d..(i + 1) * d].to_vec()).collect();

        let config = PcaConfig::new(n_components as usize).auto_scale(auto_scale != 0);
        let pca_result = match pca(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        let mut evr: Vec<f64> = pca_result.explained_variance_ratio;
        let evr_ptr = evr.as_mut_ptr();
        let evr_len = evr.len() as u32;
        std::mem::forget(evr);

        unsafe {
            (*out) = CPcaResult {
                n_components: pca_result.n_components as u32,
                explained_variance: evr_ptr,
                n_variance: evr_len,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_pca");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees an f64 array allocated by an insight FFI function.
///
/// # Safety
/// `ptr` must have been allocated by an insight FFI function, or be null.
#[no_mangle]
pub unsafe extern "C" fn insight_free_f64_array(ptr: *mut f64, count: u32) {
    if !ptr.is_null() {
        let _ = unsafe { Vec::from_raw_parts(ptr, count as usize, count as usize) };
    }
}

// ── DBSCAN FFI ──────────────────────────────────────────────────────

/// C-compatible DBSCAN result.
#[repr(C)]
pub struct CDbscanResult {
    /// Number of clusters discovered.
    pub n_clusters: u32,
    /// Number of noise points.
    pub noise_count: u32,
    /// Cluster labels (length = n_rows). -1 = noise, >= 0 = cluster id.
    /// Caller must free with `insight_free_i32_array`.
    pub labels: *mut i32,
    /// Number of labels.
    pub n_labels: u32,
}

/// Runs DBSCAN on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CDbscanResult`.
/// - Caller must free `out.labels` with `insight_free_i32_array`.
#[no_mangle]
pub unsafe extern "C" fn insight_dbscan(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    epsilon: f64,
    min_samples: u32,
    out: *mut CDbscanResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let n = n_rows as usize;
        let d = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, n * d) };

        let points: Vec<Vec<f64>> = (0..n).map(|i| raw[i * d..(i + 1) * d].to_vec()).collect();

        let config = DbscanConfig::new(epsilon, min_samples as usize);
        let db_result = match dbscan(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        // Convert labels: None → -1, Some(id) → id as i32
        let mut labels: Vec<i32> = db_result
            .labels
            .iter()
            .map(|l| match l {
                Some(id) => *id as i32,
                None => -1,
            })
            .collect();
        let labels_ptr = labels.as_mut_ptr();
        let labels_len = labels.len() as u32;
        std::mem::forget(labels);

        unsafe {
            (*out) = CDbscanResult {
                n_clusters: db_result.n_clusters as u32,
                noise_count: db_result.noise_count as u32,
                labels: labels_ptr,
                n_labels: labels_len,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_dbscan");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees an i32 array allocated by an insight FFI function.
///
/// # Safety
/// `ptr` must have been allocated by an insight FFI function, or be null.
#[no_mangle]
pub unsafe extern "C" fn insight_free_i32_array(ptr: *mut i32, count: u32) {
    if !ptr.is_null() {
        let _ = unsafe { Vec::from_raw_parts(ptr, count as usize, count as usize) };
    }
}

// ── Distribution Analysis FFI ───────────────────────────────────────

/// C-compatible distribution analysis result (normality assessment).
#[repr(C)]
pub struct CDistributionResult {
    /// Number of observations.
    pub n: u32,
    /// KS test statistic (NaN if unavailable).
    pub ks_statistic: f64,
    /// KS test p-value (NaN if unavailable).
    pub ks_p_value: f64,
    /// Jarque-Bera test statistic (NaN if unavailable).
    pub jb_statistic: f64,
    /// Jarque-Bera test p-value (NaN if unavailable).
    pub jb_p_value: f64,
    /// Shapiro-Wilk W statistic (NaN if unavailable).
    pub sw_statistic: f64,
    /// Shapiro-Wilk p-value (NaN if unavailable).
    pub sw_p_value: f64,
    /// Anderson-Darling A*² statistic (NaN if unavailable).
    pub ad_statistic: f64,
    /// Anderson-Darling p-value (NaN if unavailable).
    pub ad_p_value: f64,
    /// Whether data appears normally distributed. 1 = normal, 0 = not normal.
    pub is_normal: i32,
}

/// Runs distribution analysis (normality testing) on a data vector.
///
/// # Safety
/// - `data` must point to `n` contiguous f64 values.
/// - `out` must point to a valid `CDistributionResult`.
#[no_mangle]
pub unsafe extern "C" fn insight_distribution(
    data: *const f64,
    n: u32,
    significance_level: f64,
    out: *mut CDistributionResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let len = n as usize;
        let raw = unsafe { slice::from_raw_parts(data, len) };
        let values: Vec<f64> = raw.to_vec();

        let config = DistributionConfig {
            significance_level,
            compute_ecdf: false,
            compute_histogram: false,
            compute_qq_plot: false,
            ..Default::default()
        };

        let dist_result = match distribution_analysis(&values, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        let (ks_stat, ks_p) = dist_result
            .normality
            .ks_test
            .map_or((f64::NAN, f64::NAN), |t| (t.statistic, t.p_value));
        let (jb_stat, jb_p) = dist_result
            .normality
            .jarque_bera
            .map_or((f64::NAN, f64::NAN), |t| (t.statistic, t.p_value));
        let (sw_stat, sw_p) = dist_result
            .normality
            .shapiro_wilk
            .map_or((f64::NAN, f64::NAN), |t| (t.statistic, t.p_value));
        let (ad_stat, ad_p) = dist_result
            .normality
            .anderson_darling
            .map_or((f64::NAN, f64::NAN), |t| (t.statistic, t.p_value));

        unsafe {
            (*out) = CDistributionResult {
                n: dist_result.n as u32,
                ks_statistic: ks_stat,
                ks_p_value: ks_p,
                jb_statistic: jb_stat,
                jb_p_value: jb_p,
                sw_statistic: sw_stat,
                sw_p_value: sw_p,
                ad_statistic: ad_stat,
                ad_p_value: ad_p,
                is_normal: if dist_result.normality.is_normal {
                    1
                } else {
                    0
                },
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_distribution");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Feature Importance FFI ──────────────────────────────────────────

/// C-compatible feature importance result.
#[repr(C)]
pub struct CFeatureImportanceResult {
    /// Importance scores per feature (length = n_cols). Higher = more important.
    /// Caller must free with `insight_free_f64_array`.
    pub scores: *mut f64,
    /// Number of scores.
    pub n_scores: u32,
    /// Condition number of the feature correlation matrix.
    pub condition_number: f64,
    /// Number of low-variance features detected.
    pub n_low_variance: u32,
    /// Number of high-correlation pairs detected.
    pub n_high_corr_pairs: u32,
}

/// Runs feature importance analysis on column-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CFeatureImportanceResult`.
/// - Caller must free `out.scores` with `insight_free_f64_array`.
#[no_mangle]
pub unsafe extern "C" fn insight_feature_importance(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    out: *mut CFeatureImportanceResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let n = n_rows as usize;
        let d = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, n * d) };

        // Convert row-major to column-major vectors
        let columns: Vec<Vec<f64>> = (0..d)
            .map(|col| (0..n).map(|row| raw[row * d + col]).collect())
            .collect();

        let names: Vec<String> = (0..d).map(|i| format!("f{i}")).collect();
        let config = FeatureConfig::default();

        let fi_result = match feature_analysis(&columns, &names, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        // Extract scores in column order
        let mut scores: Vec<f64> = (0..d)
            .map(|i| {
                let name = &names[i];
                fi_result
                    .feature_scores
                    .iter()
                    .find(|s| s.name == *name)
                    .map_or(0.0, |s| s.importance)
            })
            .collect();
        let scores_ptr = scores.as_mut_ptr();
        let scores_len = scores.len() as u32;
        std::mem::forget(scores);

        unsafe {
            (*out) = CFeatureImportanceResult {
                scores: scores_ptr,
                n_scores: scores_len,
                condition_number: fi_result.condition_number,
                n_low_variance: fi_result.low_variance.len() as u32,
                n_high_corr_pairs: fi_result.high_correlations.len() as u32,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_feature_importance");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Isolation Forest FFI ─────────────────────────────────────────────

/// C-compatible anomaly detection result (shared by Isolation Forest and LOF).
#[repr(C)]
pub struct CAnomalyResult {
    /// Anomaly score for each point. Higher = more anomalous.
    /// Caller must free with `insight_free_f64_array`.
    pub scores: *mut f64,
    /// Binary anomaly labels (1 = anomaly, 0 = normal).
    /// Caller must free with `insight_free_i32_array`.
    pub anomalies: *mut i32,
    /// Number of data points.
    pub n: u32,
    /// Number of anomalies detected.
    pub anomaly_count: u32,
    /// Threshold used for classification.
    pub threshold: f64,
}

/// Runs Isolation Forest anomaly detection on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CAnomalyResult`.
/// - Caller must free `out.scores` with `insight_free_f64_array` and
///   `out.anomalies` with `insight_free_i32_array`.
#[no_mangle]
pub unsafe extern "C" fn insight_isolation_forest(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    n_estimators: u32,
    contamination: f64,
    seed: u64,
    out: *mut CAnomalyResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let rows = n_rows as usize;
        let cols = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, rows * cols) };

        let points: Vec<Vec<f64>> = (0..rows)
            .map(|i| raw[i * cols..(i + 1) * cols].to_vec())
            .collect();

        let config = crate::isolation_forest::IsolationForestConfig::default()
            .n_estimators(n_estimators as usize)
            .contamination(contamination)
            .seed(Some(seed));

        let iforest = match crate::isolation_forest::isolation_forest(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        let mut scores = iforest.scores.into_boxed_slice();
        let anomalies_i32: Vec<i32> = iforest.anomalies.iter().map(|&a| a as i32).collect();
        let mut anomalies = anomalies_i32.into_boxed_slice();

        unsafe {
            (*out) = CAnomalyResult {
                scores: scores.as_mut_ptr(),
                anomalies: anomalies.as_mut_ptr(),
                n: n_rows,
                anomaly_count: iforest.anomaly_count as u32,
                threshold: iforest.threshold,
            };
        }

        std::mem::forget(scores);
        std::mem::forget(anomalies);

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_isolation_forest");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── LOF FFI ─────────────────────────────────────────────────────────

/// Runs Local Outlier Factor anomaly detection on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CAnomalyResult`.
/// - Caller must free `out.scores` with `insight_free_f64_array` and
///   `out.anomalies` with `insight_free_i32_array`.
#[no_mangle]
pub unsafe extern "C" fn insight_lof(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    k: u32,
    threshold: f64,
    out: *mut CAnomalyResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let rows = n_rows as usize;
        let cols = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, rows * cols) };

        let points: Vec<Vec<f64>> = (0..rows)
            .map(|i| raw[i * cols..(i + 1) * cols].to_vec())
            .collect();

        let config = crate::lof::LofConfig::default()
            .k(k as usize)
            .threshold(threshold);

        let lof_result = match crate::lof::lof(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        let mut scores = lof_result.scores.into_boxed_slice();
        let anomalies_i32: Vec<i32> = lof_result.anomalies.iter().map(|&a| a as i32).collect();
        let mut anomalies = anomalies_i32.into_boxed_slice();

        unsafe {
            (*out) = CAnomalyResult {
                scores: scores.as_mut_ptr(),
                anomalies: anomalies.as_mut_ptr(),
                n: n_rows,
                anomaly_count: lof_result.anomaly_count as u32,
                threshold: lof_result.threshold,
            };
        }

        std::mem::forget(scores);
        std::mem::forget(anomalies);

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_lof");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Correlation FFI ─────────────────────────────────────────────────

/// C-compatible correlation result.
#[repr(C)]
pub struct CCorrelationResult {
    /// Number of variables (n).
    pub n_vars: u32,
    /// Flat n×n correlation matrix (row-major). Caller must free with `insight_free_f64_array`.
    pub matrix: *mut f64,
    /// Number of high-correlation pairs found.
    pub n_high_pairs: u32,
}

/// Method codes for [`insight_correlation`].
pub const INSIGHT_CORR_PEARSON: u32 = 0;
/// Spearman rank correlation.
pub const INSIGHT_CORR_SPEARMAN: u32 = 1;
/// Kendall tau-b rank correlation.
pub const INSIGHT_CORR_KENDALL: u32 = 2;

/// Computes a correlation matrix over row-major numeric data.
///
/// `data`: flat array of `n_rows × n_cols` f64 values, row-major.
/// `method`: one of `INSIGHT_CORR_PEARSON` (0) / `_SPEARMAN` (1) / `_KENDALL` (2).
/// `out`: pointer to a `CCorrelationResult`.
///
/// Returns 0 on success, negative on error. Unknown `method` values
/// return `INSIGHT_ERR_INVALID_PARAM`.
///
/// # Safety
/// `data` must point to `n_rows * n_cols` f64s. `out` must be valid.
#[no_mangle]
pub unsafe extern "C" fn insight_correlation(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    method: u32,
    out: *mut CCorrelationResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let nr = n_rows as usize;
        let nc = n_cols as usize;
        if nr < 2 || nc < 2 {
            set_last_error("need at least 2 rows and 2 columns");
            return INSIGHT_ERR_INVALID_INPUT;
        }

        let corr_method = match method {
            INSIGHT_CORR_PEARSON => crate::analysis::CorrelationMethod::Pearson,
            INSIGHT_CORR_SPEARMAN => crate::analysis::CorrelationMethod::Spearman,
            INSIGHT_CORR_KENDALL => crate::analysis::CorrelationMethod::Kendall,
            _ => {
                set_last_error("invalid correlation method (use 0=Pearson, 1=Spearman, 2=Kendall)");
                return INSIGHT_ERR_INVALID_PARAM;
            }
        };

        let raw = unsafe { slice::from_raw_parts(data, nr * nc) };

        // Convert row-major flat to column-major Vec<Vec<f64>>
        let mut columns: Vec<Vec<f64>> = vec![Vec::with_capacity(nr); nc];
        for row in 0..nr {
            for col in 0..nc {
                columns[col].push(raw[row * nc + col]);
            }
        }

        let names: Vec<String> = (0..nc).map(|i| format!("c{i}")).collect();
        let config = crate::analysis::CorrelationConfig {
            method: corr_method,
            high_threshold: 0.7,
        };

        match crate::analysis::correlation_analysis(&columns, &names, &config) {
            Ok(result) => {
                let out_ref = unsafe { &mut *out };
                out_ref.n_vars = nc as u32;
                out_ref.n_high_pairs = result.high_pairs.len() as u32;

                // Flatten the correlation matrix to row-major
                let mut flat = Vec::with_capacity(nc * nc);
                for r in 0..nc {
                    for c in 0..nc {
                        flat.push(result.matrix.get(r, c));
                    }
                }
                let mut boxed = flat.into_boxed_slice();
                out_ref.matrix = boxed.as_mut_ptr();
                std::mem::forget(boxed);

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_correlation");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Regression FFI ──────────────────────────────────────────────────

/// C-compatible simple regression result.
#[repr(C)]
pub struct CRegressionResult {
    /// Intercept (β₀).
    pub intercept: f64,
    /// Slope (β₁).
    pub slope: f64,
    /// R² (coefficient of determination).
    pub r_squared: f64,
    /// Adjusted R².
    pub adj_r_squared: f64,
    /// P-value for the F-test.
    pub f_p_value: f64,
}

/// Computes simple linear regression (one predictor).
///
/// `x`: predictor array of length `n`.
/// `y`: target array of length `n`.
/// `out`: pointer to a `CRegressionResult`.
///
/// Returns 0 on success, negative on error.
///
/// # Safety
/// `x` and `y` must point to `n` f64s. `out` must be valid.
#[no_mangle]
pub unsafe extern "C" fn insight_regression(
    x: *const f64,
    y: *const f64,
    n: u32,
    out: *mut CRegressionResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if x.is_null() || y.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let len = n as usize;
        if len < 3 {
            set_last_error("need at least 3 data points");
            return INSIGHT_ERR_INVALID_INPUT;
        }

        let x_slice = unsafe { slice::from_raw_parts(x, len) };
        let y_slice = unsafe { slice::from_raw_parts(y, len) };

        let x_vecs = vec![x_slice.to_vec()];
        let names = vec!["x".to_string()];

        match crate::analysis::regression_analysis(&x_vecs, &names, y_slice, "y") {
            Ok(reg) => {
                let out_ref = unsafe { &mut *out };
                out_ref.intercept = reg.coefficients[0];
                out_ref.slope = reg.coefficients[1];
                out_ref.r_squared = reg.r_squared;
                out_ref.adj_r_squared = reg.adj_r_squared;
                out_ref.f_p_value = reg.f_p_value;

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_regression");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Mahalanobis FFI ──────────────────────────────────────────────────

/// C-compatible Mahalanobis distance result.
#[repr(C)]
pub struct CMahalanobisResult {
    /// Mahalanobis distances (length = n_rows).
    /// Caller must free with `insight_free_f64_array`.
    pub distances: *mut f64,
    /// Anomaly flags (1 = outlier, 0 = normal, length = n_rows).
    /// Caller must free with `insight_free_i32_array`.
    pub anomalies: *mut i32,
    /// Number of data points.
    pub n: u32,
    /// Chi-squared threshold used.
    pub threshold: f64,
    /// Number of outliers detected.
    pub outlier_count: u32,
}

/// Runs Mahalanobis distance multivariate outlier detection on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values (row-major).
/// - `out` must point to a valid `CMahalanobisResult`.
/// - Caller must free output arrays.
#[no_mangle]
pub unsafe extern "C" fn insight_mahalanobis(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    chi2_quantile: f64,
    out: *mut CMahalanobisResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let nr = n_rows as usize;
        let nc = n_cols as usize;
        if nr < nc + 1 {
            set_last_error("need n > p for Mahalanobis distance");
            return INSIGHT_ERR_INVALID_INPUT;
        }

        let raw = unsafe { slice::from_raw_parts(data, nr * nc) };
        let points: Vec<Vec<f64>> = (0..nr)
            .map(|i| raw[i * nc..(i + 1) * nc].to_vec())
            .collect();

        let config = crate::mahalanobis::MahalanobisConfig {
            chi2_quantile: if chi2_quantile > 0.0 && chi2_quantile < 1.0 {
                chi2_quantile
            } else {
                0.975
            },
        };

        match crate::mahalanobis::mahalanobis(&points, &config) {
            Ok(r) => {
                let out_ref = unsafe { &mut *out };
                out_ref.n = nr as u32;
                out_ref.threshold = r.threshold;
                out_ref.outlier_count = r.outlier_count as u32;

                let mut dists = r.distances.into_boxed_slice();
                out_ref.distances = dists.as_mut_ptr();
                std::mem::forget(dists);

                let anoms: Vec<i32> = r.anomalies.iter().map(|&a| a as i32).collect();
                let mut anoms_box = anoms.into_boxed_slice();
                out_ref.anomalies = anoms_box.as_mut_ptr();
                std::mem::forget(anoms_box);

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_mahalanobis");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Cramér's V FFI ──────────────────────────────────────────────────

/// C-compatible Cramér's V result.
#[repr(C)]
pub struct CCramersVResult {
    /// Cramér's V (0 to 1).
    pub v: f64,
    /// Chi-squared statistic.
    pub chi_squared: f64,
    /// P-value.
    pub p_value: f64,
}

/// Computes Cramér's V for a contingency table.
///
/// `table`: flat row-major contingency table (observed frequencies).
///
/// # Safety
/// `table` must point to `n_rows * n_cols` f64s. `out` must be valid.
#[no_mangle]
pub unsafe extern "C" fn insight_cramers_v(
    table: *const f64,
    n_rows: u32,
    n_cols: u32,
    out: *mut CCramersVResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if table.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let nr = n_rows as usize;
        let nc = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(table, nr * nc) };

        match crate::analysis::cramers_v(raw, nr, nc) {
            Some(r) => {
                let out_ref = unsafe { &mut *out };
                out_ref.v = r.v;
                out_ref.chi_squared = r.chi_squared;
                out_ref.p_value = r.p_value;
                INSIGHT_OK
            }
            None => {
                set_last_error("Cramér's V computation failed");
                INSIGHT_ERR_ANALYSIS_FAILED
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_cramers_v");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── ANOVA Feature Selection FFI ─────────────────────────────────────

/// C-compatible ANOVA feature result.
#[repr(C)]
pub struct CAnovaFeature {
    /// Feature index.
    pub index: u32,
    /// F-statistic.
    pub f_statistic: f64,
    /// P-value.
    pub p_value: f64,
}

/// C-compatible ANOVA selection result.
#[repr(C)]
pub struct CAnovaSelectionResult {
    /// Per-feature results sorted by p-value ascending.
    /// Caller must free with `insight_free_anova_features`.
    pub features: *mut CAnovaFeature,
    /// Number of features.
    pub n_features: u32,
    /// Number of significant features.
    pub n_selected: u32,
}

/// Runs ANOVA F-test feature selection.
///
/// `data`: row-major n_rows × n_features.
/// `target`: class labels (u32), length n_rows.
///
/// # Safety
/// All pointers must be valid. Caller frees output.
#[no_mangle]
pub unsafe extern "C" fn insight_anova_select(
    data: *const f64,
    n_rows: u32,
    n_features: u32,
    target: *const u32,
    significance_level: f64,
    out: *mut CAnovaSelectionResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || target.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let nr = n_rows as usize;
        let nf = n_features as usize;
        let raw = unsafe { slice::from_raw_parts(data, nr * nf) };
        let target_raw = unsafe { slice::from_raw_parts(target, nr) };

        // Convert row-major to column-major
        let columns: Vec<Vec<f64>> = (0..nf)
            .map(|col| (0..nr).map(|row| raw[row * nf + col]).collect())
            .collect();

        let names: Vec<String> = (0..nf).map(|i| format!("f{i}")).collect();
        let target_usize: Vec<usize> = target_raw.iter().map(|&t| t as usize).collect();

        match crate::analysis::anova_feature_selection(
            &columns,
            &names,
            &target_usize,
            significance_level,
        ) {
            Ok(r) => {
                let out_ref = unsafe { &mut *out };
                out_ref.n_features = r.features.len() as u32;
                out_ref.n_selected = r.selected_indices.len() as u32;

                let c_features: Vec<CAnovaFeature> = r
                    .features
                    .iter()
                    .enumerate()
                    .map(|(i, f)| CAnovaFeature {
                        index: i as u32,
                        f_statistic: f.f_statistic,
                        p_value: f.p_value,
                    })
                    .collect();
                let mut boxed = c_features.into_boxed_slice();
                out_ref.features = boxed.as_mut_ptr();
                std::mem::forget(boxed);

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_anova_select");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees ANOVA feature result array.
///
/// # Safety
/// `ptr` must have been allocated by `insight_anova_select`, or be null.
#[no_mangle]
pub unsafe extern "C" fn insight_free_anova_features(ptr: *mut CAnovaFeature, count: u32) {
    if !ptr.is_null() {
        let _ = unsafe { Vec::from_raw_parts(ptr, count as usize, count as usize) };
    }
}

// ── Version ──────────────────────────────────────────────────────────

/// Returns the version string of u-insight.
///
/// # Safety
/// The returned string is a static string literal. Do not free it.
#[no_mangle]
pub extern "C" fn insight_version() -> *const c_char {
    // "0.1.0\0"
    c"0.1.0".as_ptr()
}

// ── Hierarchical Clustering FFI ──────────────────────────────────────

/// C-compatible hierarchical clustering result.
#[repr(C)]
pub struct CHierarchicalResult {
    /// Number of flat clusters (0 if no cut).
    pub n_clusters: u32,
    /// Flat cluster labels (length = n_rows). -1 if no labels.
    /// Caller must free with `insight_free_i32_array`.
    pub labels: *mut i32,
    /// Number of labels.
    pub n_labels: u32,
    /// Number of merges in the dendrogram.
    pub n_merges: u32,
    /// Merge distances (length = n_merges).
    /// Caller must free with `insight_free_f64_array`.
    pub merge_distances: *mut f64,
    /// Merge sizes (length = n_merges).
    /// Caller must free with `insight_free_i32_array`.
    pub merge_sizes: *mut i32,
}

/// Runs hierarchical agglomerative clustering on row-major data.
///
/// # Parameters
///
/// - `linkage`: 0 = Single, 1 = Complete, 2 = Average, 3 = Ward.
/// - `n_clusters`: Desired number of flat clusters (0 = no cut).
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values.
/// - `out` must point to a valid `CHierarchicalResult`.
/// - Caller must free output arrays with appropriate free functions.
#[no_mangle]
pub unsafe extern "C" fn insight_hierarchical(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    linkage: u32,
    n_clusters: u32,
    out: *mut CHierarchicalResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let n = n_rows as usize;
        let d = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, n * d) };

        let points: Vec<Vec<f64>> = (0..n).map(|i| raw[i * d..(i + 1) * d].to_vec()).collect();

        let linkage_method = match linkage {
            0 => Linkage::Single,
            1 => Linkage::Complete,
            2 => Linkage::Average,
            _ => Linkage::Ward,
        };

        let config = if n_clusters > 0 {
            HierarchicalConfig::with_k(n_clusters as usize).linkage(linkage_method)
        } else {
            HierarchicalConfig {
                linkage: linkage_method,
                n_clusters: None,
                distance_threshold: None,
            }
        };

        let hc_result = match hierarchical(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        // Labels
        let (labels_ptr, labels_len, nc) = match &hc_result.labels {
            Some(labels) => {
                let mut l: Vec<i32> = labels.iter().map(|&v| v as i32).collect();
                let ptr = l.as_mut_ptr();
                let len = l.len() as u32;
                std::mem::forget(l);
                (ptr, len, hc_result.n_clusters.unwrap_or(0) as u32)
            }
            None => (ptr::null_mut(), 0, 0),
        };

        // Merge distances and sizes
        let nm = hc_result.merges.len();
        let mut distances: Vec<f64> = hc_result.merges.iter().map(|m| m.distance).collect();
        let mut sizes: Vec<i32> = hc_result.merges.iter().map(|m| m.size as i32).collect();
        let dist_ptr = distances.as_mut_ptr();
        let size_ptr = sizes.as_mut_ptr();
        std::mem::forget(distances);
        std::mem::forget(sizes);

        unsafe {
            (*out) = CHierarchicalResult {
                n_clusters: nc,
                labels: labels_ptr,
                n_labels: labels_len,
                n_merges: nm as u32,
                merge_distances: dist_ptr,
                merge_sizes: size_ptr,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_hierarchical");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── HDBSCAN FFI ─────────────────────────────────────────────────────

/// C-compatible HDBSCAN result.
#[repr(C)]
pub struct CHdbscanResult {
    /// Number of clusters found (excluding noise).
    pub n_clusters: u32,
    /// Number of noise points.
    pub noise_count: u32,
    /// Cluster labels (length = n_rows). -1 = noise.
    /// Caller must free with `insight_free_i32_array`.
    pub labels: *mut i32,
    /// Membership probabilities (length = n_rows).
    /// Caller must free with `insight_free_f64_array`.
    pub probabilities: *mut f64,
    /// Number of data points.
    pub n_labels: u32,
}

/// Runs HDBSCAN on row-major data.
///
/// # Safety
/// - `data` must point to `n_rows * n_cols` contiguous f64 values.
/// - `out` must point to a valid `CHdbscanResult`.
/// - Caller must free output arrays with appropriate free functions.
#[no_mangle]
pub unsafe extern "C" fn insight_hdbscan(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    min_cluster_size: u32,
    min_samples: u32,
    out: *mut CHdbscanResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let n = n_rows as usize;
        let d = n_cols as usize;
        let raw = unsafe { slice::from_raw_parts(data, n * d) };

        let points: Vec<Vec<f64>> = (0..n).map(|i| raw[i * d..(i + 1) * d].to_vec()).collect();

        let mut config = HdbscanConfig::new(min_cluster_size as usize);
        if min_samples > 0 {
            config = config.min_samples(min_samples as usize);
        }

        let hdb_result = match hdbscan(&points, &config) {
            Ok(r) => r,
            Err(e) => {
                set_last_error(&e.to_string());
                return error_to_code(&e);
            }
        };

        // Convert labels: None → -1, Some(id) → id as i32
        let mut labels: Vec<i32> = hdb_result
            .labels
            .iter()
            .map(|l| match l {
                Some(id) => *id as i32,
                None => -1,
            })
            .collect();
        let labels_ptr = labels.as_mut_ptr();
        let labels_len = labels.len() as u32;
        std::mem::forget(labels);

        // Probabilities
        let mut probs = hdb_result.probabilities;
        let probs_ptr = probs.as_mut_ptr();
        std::mem::forget(probs);

        unsafe {
            (*out) = CHdbscanResult {
                n_clusters: hdb_result.n_clusters as u32,
                noise_count: hdb_result.noise_count as u32,
                labels: labels_ptr,
                probabilities: probs_ptr,
                n_labels: labels_len,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_hdbscan");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Mutual Information FFI ──────────────────────────────────────────

/// Result of mutual information feature selection via FFI.
#[repr(C)]
pub struct CMutualInfoFeature {
    pub index: u32,
    pub mi: f64,
}

/// Aggregate result for mutual information.
#[repr(C)]
pub struct CMutualInfoResult {
    pub features: *mut CMutualInfoFeature,
    pub n_features: u32,
}

/// Computes mutual information between continuous features and a categorical target.
///
/// # Parameters
/// - `data`: row-major feature matrix (n_rows × n_features)
/// - `target`: categorical target array (u32, length n_rows)
/// - `n_bins`: number of bins (0 = auto via Sturges' rule)
///
/// # Safety
/// Caller must free `out.features` via `insight_free_mi_features`.
#[no_mangle]
pub unsafe extern "C" fn insight_mutual_info(
    data: *const f64,
    n_rows: u32,
    n_features: u32,
    target: *const u32,
    n_bins: u32,
    out: *mut CMutualInfoResult,
) -> i32 {
    if data.is_null() || target.is_null() || out.is_null() {
        set_last_error("null pointer argument");
        return INSIGHT_ERR_NULL_PTR;
    }

    let result = panic::catch_unwind(|| {
        let nr = n_rows as usize;
        let nf = n_features as usize;
        let raw = slice::from_raw_parts(data, nr * nf);
        let target_raw = slice::from_raw_parts(target, nr);

        let mut features: Vec<Vec<f64>> = Vec::with_capacity(nf);
        for col in 0..nf {
            let mut v = Vec::with_capacity(nr);
            for row in 0..nr {
                v.push(raw[row * nf + col]);
            }
            features.push(v);
        }

        let names: Vec<String> = (0..nf).map(|i| format!("f{}", i)).collect();
        let target_usize: Vec<usize> = target_raw.iter().map(|&t| t as usize).collect();
        let bins = if n_bins == 0 {
            None
        } else {
            Some(n_bins as usize)
        };

        match mutual_info_classif(&features, &names, &target_usize, bins) {
            Ok(mi_result) => {
                let n = mi_result.features.len();
                let mut c_features: Vec<CMutualInfoFeature> = mi_result
                    .features
                    .iter()
                    .map(|f| CMutualInfoFeature {
                        index: f.index as u32,
                        mi: f.mi,
                    })
                    .collect();

                let ptr = c_features.as_mut_ptr();
                std::mem::forget(c_features);

                (*out) = CMutualInfoResult {
                    features: ptr,
                    n_features: n as u32,
                };

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_mutual_info");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees a CMutualInfoFeature array allocated by `insight_mutual_info`.
///
/// # Safety
/// `ptr` must have been returned by `insight_mutual_info` with matching `count`.
#[no_mangle]
pub unsafe extern "C" fn insight_free_mi_features(ptr: *mut CMutualInfoFeature, count: u32) {
    if !ptr.is_null() && count > 0 {
        let _ = Vec::from_raw_parts(ptr, count as usize, count as usize);
    }
}

// ── Mini-Batch K-Means FFI ─────────────────────────────────────────

/// Runs mini-batch K-Means clustering.
///
/// # Parameters
/// - `data`: row-major matrix (n_rows × n_cols)
/// - `k`: number of clusters
/// - `batch_size`: mini-batch size (0 = default 100)
/// - `max_iter`: max iterations (0 = default 100)
/// - `seed`: random seed
///
/// # Safety
/// Caller must free `out.labels` via `insight_free_labels` and
/// `out.centroids` via `insight_free_f64_array(out.centroids, k * n_cols)`.
#[no_mangle]
pub unsafe extern "C" fn insight_mini_batch_kmeans(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    k: u32,
    batch_size: u32,
    max_iter: u32,
    seed: u64,
    out: *mut CKMeansResult,
) -> i32 {
    if data.is_null() || out.is_null() {
        set_last_error("null pointer argument");
        return INSIGHT_ERR_NULL_PTR;
    }

    let result = panic::catch_unwind(|| {
        let nr = n_rows as usize;
        let nc = n_cols as usize;
        let raw = slice::from_raw_parts(data, nr * nc);

        let mut rows: Vec<Vec<f64>> = Vec::with_capacity(nr);
        for r in 0..nr {
            rows.push(raw[r * nc..(r + 1) * nc].to_vec());
        }

        let config = MiniBatchKMeansConfig {
            k: k as usize,
            batch_size: if batch_size == 0 {
                100
            } else {
                batch_size as usize
            },
            max_iter: if max_iter == 0 {
                100
            } else {
                max_iter as usize
            },
            tol: 1e-4,
            seed: Some(seed),
        };

        match mini_batch_kmeans(&rows, &config) {
            Ok(km_result) => {
                let mut labels: Vec<u32> = km_result.labels.iter().map(|&l| l as u32).collect();
                let labels_len = labels.len() as u32;
                let labels_ptr = labels.as_mut_ptr();
                std::mem::forget(labels);

                (*out) = CKMeansResult {
                    k,
                    wcss: km_result.wcss,
                    iterations: km_result.iterations as u32,
                    labels: labels_ptr,
                    n_labels: labels_len,
                };

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_mini_batch_kmeans");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Gap Statistic FFI ──────────────────────────────────────────────

/// Result of gap statistic analysis via FFI.
#[repr(C)]
pub struct CGapStatResult {
    /// Optimal K selected by gap criterion.
    pub best_k: u32,
    /// Number of K values tested.
    pub n_values: u32,
    /// Array of (k, gap) pairs flattened: [k0, gap0, k1, gap1, ...].
    pub gap_values: *mut f64,
    /// Array of (k, stderr) pairs flattened: [k0, se0, k1, se1, ...].
    pub std_errors: *mut f64,
}

/// Computes gap statistic to find optimal K for clustering.
///
/// # Parameters
/// - `data`: row-major matrix (n_rows × n_cols)
/// - `k_min`, `k_max`: K range to test
/// - `n_refs`: number of reference datasets
/// - `seed`: random seed
///
/// # Safety
/// Caller must free `out.gap_values` and `out.std_errors` via
/// `insight_free_f64_array(ptr, n_values * 2)`.
#[no_mangle]
pub unsafe extern "C" fn insight_gap_statistic(
    data: *const f64,
    n_rows: u32,
    n_cols: u32,
    k_min: u32,
    k_max: u32,
    n_refs: u32,
    seed: u64,
    out: *mut CGapStatResult,
) -> i32 {
    if data.is_null() || out.is_null() {
        set_last_error("null pointer argument");
        return INSIGHT_ERR_NULL_PTR;
    }

    let result = panic::catch_unwind(|| {
        let nr = n_rows as usize;
        let nc = n_cols as usize;
        let raw = slice::from_raw_parts(data, nr * nc);

        let mut rows: Vec<Vec<f64>> = Vec::with_capacity(nr);
        for r in 0..nr {
            rows.push(raw[r * nc..(r + 1) * nc].to_vec());
        }

        match gap_statistic(&rows, k_min as usize, k_max as usize, n_refs as usize, seed) {
            Ok(gap_result) => {
                let n = gap_result.gap_values.len();

                // Flatten (k, gap) pairs
                let mut gaps: Vec<f64> = Vec::with_capacity(n * 2);
                for &(k, g) in &gap_result.gap_values {
                    gaps.push(k as f64);
                    gaps.push(g);
                }
                let gaps_ptr = gaps.as_mut_ptr();
                std::mem::forget(gaps);

                // Flatten (k, se) pairs
                let mut ses: Vec<f64> = Vec::with_capacity(n * 2);
                for &(k, s) in &gap_result.std_errors {
                    ses.push(k as f64);
                    ses.push(s);
                }
                let ses_ptr = ses.as_mut_ptr();
                std::mem::forget(ses);

                (*out) = CGapStatResult {
                    best_k: gap_result.best_k as u32,
                    n_values: n as u32,
                    gap_values: gaps_ptr,
                    std_errors: ses_ptr,
                };

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_gap_statistic");
            INSIGHT_ERR_PANIC
        }
    }
}

// ── Permutation Importance FFI ─────────────────────────────────────

/// Per-feature permutation importance result via FFI.
#[repr(C)]
pub struct CPermImportanceFeature {
    pub index: u32,
    pub importance: f64,
    pub std_dev: f64,
}

/// Aggregate permutation importance result via FFI.
#[repr(C)]
pub struct CPermImportanceResult {
    pub baseline_score: f64,
    pub features: *mut CPermImportanceFeature,
    pub n_features: u32,
}

/// Computes permutation importance for regression features.
///
/// # Parameters
/// - `data`: row-major feature matrix (n_rows × n_features)
/// - `target`: continuous target array (length n_rows)
/// - `n_repeats`: number of permutation repetitions
/// - `seed`: random seed
///
/// # Safety
/// Caller must free `out.features` via `insight_free_perm_features`.
#[no_mangle]
pub unsafe extern "C" fn insight_permutation_importance(
    data: *const f64,
    n_rows: u32,
    n_features: u32,
    target: *const f64,
    n_repeats: u32,
    seed: u64,
    out: *mut CPermImportanceResult,
) -> i32 {
    if data.is_null() || target.is_null() || out.is_null() {
        set_last_error("null pointer argument");
        return INSIGHT_ERR_NULL_PTR;
    }

    let result = panic::catch_unwind(|| {
        let nr = n_rows as usize;
        let nf = n_features as usize;
        let raw = slice::from_raw_parts(data, nr * nf);
        let target_raw = slice::from_raw_parts(target, nr);

        let mut features: Vec<Vec<f64>> = Vec::with_capacity(nf);
        for col in 0..nf {
            let mut v = Vec::with_capacity(nr);
            for row in 0..nr {
                v.push(raw[row * nf + col]);
            }
            features.push(v);
        }

        let names: Vec<String> = (0..nf).map(|i| format!("f{}", i)).collect();

        match permutation_importance(&features, &names, target_raw, n_repeats as usize, seed) {
            Ok(pi_result) => {
                let n = pi_result.features.len();
                let mut c_features: Vec<CPermImportanceFeature> = pi_result
                    .features
                    .iter()
                    .map(|f| CPermImportanceFeature {
                        index: f.index as u32,
                        importance: f.importance,
                        std_dev: f.std_dev,
                    })
                    .collect();

                let ptr = c_features.as_mut_ptr();
                std::mem::forget(c_features);

                (*out) = CPermImportanceResult {
                    baseline_score: pi_result.baseline_score,
                    features: ptr,
                    n_features: n as u32,
                };

                INSIGHT_OK
            }
            Err(e) => {
                set_last_error(&e.to_string());
                error_to_code(&e)
            }
        }
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_permutation_importance");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees a CPermImportanceFeature array.
///
/// # Safety
/// `ptr` must have been returned by `insight_permutation_importance` with matching `count`.
#[no_mangle]
pub unsafe extern "C" fn insight_free_perm_features(ptr: *mut CPermImportanceFeature, count: u32) {
    if !ptr.is_null() && count > 0 {
        let _ = Vec::from_raw_parts(ptr, count as usize, count as usize);
    }
}

// ── PELT changepoint detection ───────────────────────────────────────

/// Result of PELT changepoint detection.
#[repr(C)]
pub struct CPeltResult {
    /// Detected changepoint indices (0-based). Caller must free with
    /// `insight_free_pelt_result`.
    pub changepoints: *mut u32,
    /// Number of changepoints detected.
    pub n_changepoints: u32,
}

/// Runs PELT changepoint detection on a univariate time series.
///
/// # Parameters
///
/// - `data`: pointer to `n` contiguous f64 values
/// - `n`: number of data points
/// - `cost`: cost function (0 = L2 mean change, 1 = Normal mean+variance)
/// - `penalty`: penalty value. Pass 0.0 to use BIC (automatic).
/// - `min_segment_len`: minimum segment length (must be >= 2)
/// - `out`: pointer to `CPeltResult` (filled on success)
///
/// # Safety
///
/// - `data` must point to `n` contiguous f64 values.
/// - `out` must point to a valid `CPeltResult`.
/// - Caller must free `out` with `insight_free_pelt_result`.
#[no_mangle]
pub unsafe extern "C" fn insight_pelt(
    data: *const f64,
    n: u32,
    cost: u32,
    penalty: f64,
    min_segment_len: u32,
    out: *mut CPeltResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let len = n as usize;
        let raw = unsafe { slice::from_raw_parts(data, len) };

        let cost_fn = match cost {
            0 => u_analytics::detection::CostFunction::L2,
            1 => u_analytics::detection::CostFunction::Normal,
            _ => {
                set_last_error("cost must be 0 (L2) or 1 (Normal)");
                return INSIGHT_ERR_INVALID_PARAM;
            }
        };

        let pen = if penalty == 0.0 {
            u_analytics::detection::Penalty::Bic
        } else if penalty > 0.0 && penalty.is_finite() {
            u_analytics::detection::Penalty::Custom(penalty)
        } else {
            set_last_error("penalty must be 0.0 (BIC) or a positive finite number");
            return INSIGHT_ERR_INVALID_PARAM;
        };

        let min_seg = min_segment_len as usize;
        let pelt = match u_analytics::detection::Pelt::with_min_segment_len(cost_fn, pen, min_seg) {
            Some(p) => p,
            None => {
                set_last_error("invalid parameters (min_segment_len must be >= 2)");
                return INSIGHT_ERR_INVALID_PARAM;
            }
        };

        let pelt_result = pelt.detect(raw);

        let mut changepoints: Vec<u32> = pelt_result
            .changepoints
            .iter()
            .map(|&cp| cp as u32)
            .collect();
        let n_cp = changepoints.len() as u32;
        let cp_ptr = if changepoints.is_empty() {
            ptr::null_mut()
        } else {
            let p = changepoints.as_mut_ptr();
            std::mem::forget(changepoints);
            p
        };

        unsafe {
            (*out) = CPeltResult {
                changepoints: cp_ptr,
                n_changepoints: n_cp,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_pelt");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Runs PELT on multi-signal (multivariate) data.
///
/// # Parameters
///
/// - `data`: row-major array of shape `[n_samples, n_channels]`.
///   Each row is one observation across all channels — same convention
///   as every other multi-dimensional FFI in this crate
///   (PCA, KMeans, DBSCAN, IsolationForest, etc.).
/// - `n_samples`: number of time-series observations (rows)
/// - `n_channels`: number of signal channels (columns)
/// - Other params same as `insight_pelt`.
///
/// # Safety
///
/// - `data` must point to `n_samples * n_channels` contiguous f64 values.
/// - `out` must point to a valid `CPeltResult`.
/// - Caller must free `out` with `insight_free_pelt_result`.
#[no_mangle]
pub unsafe extern "C" fn insight_pelt_multi(
    data: *const f64,
    n_samples: u32,
    n_channels: u32,
    cost: u32,
    penalty: f64,
    min_segment_len: u32,
    out: *mut CPeltResult,
) -> i32 {
    let result = panic::catch_unwind(|| {
        if data.is_null() || out.is_null() {
            set_last_error("null pointer");
            return INSIGHT_ERR_NULL_PTR;
        }

        let ns = n_samples as usize;
        let ch = n_channels as usize;
        let raw = unsafe { slice::from_raw_parts(data, ns * ch) };

        // Transpose row-major [n_samples, n_channels] into per-channel slices
        // expected by `Pelt::detect_multi` (one slice per channel).
        let signals: Vec<Vec<f64>> = (0..ch)
            .map(|c| (0..ns).map(|s| raw[s * ch + c]).collect())
            .collect();
        let refs: Vec<&[f64]> = signals.iter().map(|s| s.as_slice()).collect();

        let cost_fn = match cost {
            0 => u_analytics::detection::CostFunction::L2,
            1 => u_analytics::detection::CostFunction::Normal,
            _ => {
                set_last_error("cost must be 0 (L2) or 1 (Normal)");
                return INSIGHT_ERR_INVALID_PARAM;
            }
        };

        let pen = if penalty == 0.0 {
            u_analytics::detection::Penalty::Bic
        } else if penalty > 0.0 && penalty.is_finite() {
            u_analytics::detection::Penalty::Custom(penalty)
        } else {
            set_last_error("penalty must be 0.0 (BIC) or positive finite");
            return INSIGHT_ERR_INVALID_PARAM;
        };

        let min_seg = min_segment_len as usize;
        let pelt = match u_analytics::detection::Pelt::with_min_segment_len(cost_fn, pen, min_seg) {
            Some(p) => p,
            None => {
                set_last_error("invalid parameters");
                return INSIGHT_ERR_INVALID_PARAM;
            }
        };

        let pelt_result = match pelt.detect_multi(&refs) {
            Some(r) => r,
            None => {
                set_last_error("all signals must have the same length");
                return INSIGHT_ERR_INVALID_INPUT;
            }
        };

        let mut changepoints: Vec<u32> = pelt_result
            .changepoints
            .iter()
            .map(|&cp| cp as u32)
            .collect();
        let n_cp = changepoints.len() as u32;
        let cp_ptr = if changepoints.is_empty() {
            ptr::null_mut()
        } else {
            let p = changepoints.as_mut_ptr();
            std::mem::forget(changepoints);
            p
        };

        unsafe {
            (*out) = CPeltResult {
                changepoints: cp_ptr,
                n_changepoints: n_cp,
            };
        }

        INSIGHT_OK
    });

    match result {
        Ok(code) => code,
        Err(_) => {
            set_last_error("panic in insight_pelt_multi");
            INSIGHT_ERR_PANIC
        }
    }
}

/// Frees a `CPeltResult` allocated by `insight_pelt` or `insight_pelt_multi`.
///
/// # Safety
///
/// The result must have been allocated by `insight_pelt` or `insight_pelt_multi`
/// and not yet freed.
#[no_mangle]
pub unsafe extern "C" fn insight_free_pelt_result(result: *mut CPeltResult) {
    if !result.is_null() {
        let r = unsafe { &*result };
        if !r.changepoints.is_null() && r.n_changepoints > 0 {
            let _ = unsafe {
                Vec::from_raw_parts(
                    r.changepoints,
                    r.n_changepoints as usize,
                    r.n_changepoints as usize,
                )
            };
        }
    }
}

// ── Tests ─────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    use std::ffi::CString;

    #[test]
    fn ffi_version() {
        let v = insight_version();
        let s = unsafe { CStr::from_ptr(v) }.to_str().unwrap();
        assert_eq!(s, "0.1.0");
    }

    #[test]
    fn ffi_error_lifecycle() {
        insight_clear_error();
        assert!(insight_last_error().is_null());

        set_last_error("test error");
        let msg = unsafe { CStr::from_ptr(insight_last_error()) }
            .to_str()
            .unwrap();
        assert_eq!(msg, "test error");

        insight_clear_error();
        assert!(insight_last_error().is_null());
    }

    #[test]
    fn ffi_profile_csv_roundtrip() {
        let csv = CString::new("name,value\nAlice,1.5\nBob,2.3\n").unwrap();
        let ctx = unsafe { insight_profile_csv(csv.as_ptr()) };
        assert!(!ctx.is_null());

        let rows = unsafe { insight_profile_row_count(ctx) };
        assert_eq!(rows, 2);

        let cols = unsafe { insight_profile_col_count(ctx) };
        assert_eq!(cols, 2);

        // Get numeric column profile (column 1 = "value")
        let mut summary = CColumnSummary {
            index: 0,
            valid_count: 0,
            null_count: 0,
            data_type: 0,
            mean: 0.0,
            std_dev: 0.0,
            min: 0.0,
            max: 0.0,
        };
        let rc = unsafe { insight_profile_column(ctx, 1, &mut summary) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(summary.data_type, 0); // Numeric
        assert!((summary.mean - 1.9).abs() < 0.01);

        unsafe { insight_profile_free(ctx) };
    }

    #[test]
    fn ffi_profile_null_ptr() {
        let ctx = unsafe { insight_profile_csv(ptr::null()) };
        assert!(ctx.is_null());
    }

    #[test]
    fn ffi_kmeans_basic() {
        // 4 points, 2D, 2 clusters
        let data: Vec<f64> = vec![0.0, 0.0, 0.5, 0.5, 10.0, 10.0, 10.5, 10.5];

        let mut result = CKMeansResult {
            k: 0,
            wcss: 0.0,
            iterations: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
        };

        let rc = unsafe { insight_kmeans(data.as_ptr(), 4, 2, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.k, 2);
        assert_eq!(result.n_labels, 4);

        // Read labels
        let labels = unsafe { slice::from_raw_parts(result.labels, result.n_labels as usize) };
        assert_eq!(labels[0], labels[1]);
        assert_eq!(labels[2], labels[3]);
        assert_ne!(labels[0], labels[2]);

        // Clean up
        unsafe { insight_free_labels(result.labels, result.n_labels) };
    }

    #[test]
    fn ffi_pca_basic() {
        // 4 points, 2D
        let data: Vec<f64> = vec![1.0, 0.0, 2.0, 0.0, 3.0, 0.0, 4.0, 0.0];

        let mut result = CPcaResult {
            n_components: 0,
            explained_variance: ptr::null_mut(),
            n_variance: 0,
        };

        let rc = unsafe { insight_pca(data.as_ptr(), 4, 2, 1, 0, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_components, 1);

        let evr =
            unsafe { slice::from_raw_parts(result.explained_variance, result.n_variance as usize) };
        assert!(
            (evr[0] - 1.0).abs() < 1e-10,
            "PC1 should explain all variance"
        );

        unsafe { insight_free_f64_array(result.explained_variance, result.n_variance) };
    }

    #[test]
    fn ffi_kmeans_null_ptr() {
        let mut result = CKMeansResult {
            k: 0,
            wcss: 0.0,
            iterations: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
        };
        let rc = unsafe { insight_kmeans(ptr::null(), 4, 2, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── DBSCAN FFI tests ─────────────────────────────────────────

    #[test]
    fn ffi_dbscan_basic() {
        // 2 clusters + 1 noise point
        let data: Vec<f64> = vec![
            0.0, 0.0, 0.5, 0.0, 0.0, 0.5, // cluster A
            10.0, 10.0, 10.5, 10.0, 10.0, 10.5, // cluster B
            50.0, 50.0, // noise
        ];

        let mut result = CDbscanResult {
            n_clusters: 0,
            noise_count: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
        };

        let rc = unsafe { insight_dbscan(data.as_ptr(), 7, 2, 1.5, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_clusters, 2);
        assert_eq!(result.noise_count, 1);
        assert_eq!(result.n_labels, 7);

        let labels = unsafe { slice::from_raw_parts(result.labels, result.n_labels as usize) };
        // Noise point (last) should be -1
        assert_eq!(labels[6], -1);
        // First 3 in same cluster
        assert_eq!(labels[0], labels[1]);
        assert_eq!(labels[0], labels[2]);
        // Next 3 in same cluster
        assert_eq!(labels[3], labels[4]);
        assert_eq!(labels[3], labels[5]);
        // Different clusters
        assert_ne!(labels[0], labels[3]);

        unsafe { insight_free_i32_array(result.labels, result.n_labels) };
    }

    #[test]
    fn ffi_dbscan_null_ptr() {
        let mut result = CDbscanResult {
            n_clusters: 0,
            noise_count: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
        };
        let rc = unsafe { insight_dbscan(ptr::null(), 4, 2, 1.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Distribution FFI tests ───────────────────────────────────

    #[test]
    fn ffi_distribution_basic() {
        // Roughly normal data
        let data: Vec<f64> = vec![
            -2.5, -2.0, -1.8, -1.5, -1.2, -1.0, -0.8, -0.5, -0.3, -0.1, 0.1, 0.3, 0.5, 0.8, 1.0,
            1.2, 1.5, 1.8, 2.0, 2.5,
        ];

        let mut result = CDistributionResult {
            n: 0,
            ks_statistic: 0.0,
            ks_p_value: 0.0,
            jb_statistic: 0.0,
            jb_p_value: 0.0,
            sw_statistic: 0.0,
            sw_p_value: 0.0,
            ad_statistic: 0.0,
            ad_p_value: 0.0,
            is_normal: 0,
        };

        let rc = unsafe { insight_distribution(data.as_ptr(), 20, 0.05, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n, 20);
        assert!(result.ks_statistic > 0.0);
        assert!(result.ks_p_value > 0.0);
        assert!(!result.jb_statistic.is_nan());
        assert!(!result.sw_statistic.is_nan());
        assert!(result.sw_p_value > 0.0);
        assert!(!result.ad_statistic.is_nan());
        assert!(result.ad_p_value > 0.0);
        assert_eq!(result.is_normal, 1); // should be normal
    }

    #[test]
    fn ffi_distribution_null_ptr() {
        let mut result = CDistributionResult {
            n: 0,
            ks_statistic: 0.0,
            ks_p_value: 0.0,
            jb_statistic: 0.0,
            jb_p_value: 0.0,
            sw_statistic: 0.0,
            sw_p_value: 0.0,
            ad_statistic: 0.0,
            ad_p_value: 0.0,
            is_normal: 0,
        };
        let rc = unsafe { insight_distribution(ptr::null(), 10, 0.05, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Feature Importance FFI tests ─────────────────────────────

    #[test]
    fn ffi_feature_importance_basic() {
        // 3 features, 10 rows (row-major)
        let data: Vec<f64> = vec![
            1.0, 2.0, 5.0, 2.0, 4.0, 4.0, 3.0, 6.0, 3.0, 4.0, 8.0, 2.0, 5.0, 10.0, 1.0, 6.0, 12.0,
            6.0, 7.0, 14.0, 5.0, 8.0, 16.0, 4.0, 9.0, 18.0, 3.0, 10.0, 20.0, 2.0,
        ];

        let mut result = CFeatureImportanceResult {
            scores: ptr::null_mut(),
            n_scores: 0,
            condition_number: 0.0,
            n_low_variance: 0,
            n_high_corr_pairs: 0,
        };

        let rc = unsafe { insight_feature_importance(data.as_ptr(), 10, 3, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_scores, 3);

        let scores = unsafe { slice::from_raw_parts(result.scores, result.n_scores as usize) };
        // All scores should be in [0, 1]
        for &s in scores {
            assert!((0.0..=1.0).contains(&s), "score {s} out of range");
        }

        // f0 and f1 are perfectly correlated → should have high_corr_pairs
        assert!(result.n_high_corr_pairs > 0);

        unsafe { insight_free_f64_array(result.scores, result.n_scores) };
    }

    #[test]
    fn ffi_feature_importance_null_ptr() {
        let mut result = CFeatureImportanceResult {
            scores: ptr::null_mut(),
            n_scores: 0,
            condition_number: 0.0,
            n_low_variance: 0,
            n_high_corr_pairs: 0,
        };
        let rc = unsafe { insight_feature_importance(ptr::null(), 10, 3, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Isolation Forest FFI tests ──

    #[test]
    fn ffi_isolation_forest_roundtrip() {
        // Dense cluster + outlier (row-major: 2 cols)
        let mut data: Vec<f64> = Vec::new();
        for i in 0..20 {
            data.push(i as f64 * 0.1);
            data.push(0.0);
        }
        data.push(100.0);
        data.push(100.0);

        let mut result = CAnomalyResult {
            scores: ptr::null_mut(),
            anomalies: ptr::null_mut(),
            n: 0,
            anomaly_count: 0,
            threshold: 0.0,
        };

        let rc =
            unsafe { insight_isolation_forest(data.as_ptr(), 21, 2, 50, 0.1, 42, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n, 21);

        let scores = unsafe { slice::from_raw_parts(result.scores, 21) };
        // Outlier (last point) should have highest score
        let max_idx = scores
            .iter()
            .enumerate()
            .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
            .unwrap()
            .0;
        assert_eq!(max_idx, 20);

        unsafe {
            insight_free_f64_array(result.scores, result.n);
            insight_free_i32_array(result.anomalies, result.n);
        }
    }

    #[test]
    fn ffi_isolation_forest_null() {
        let mut result = CAnomalyResult {
            scores: ptr::null_mut(),
            anomalies: ptr::null_mut(),
            n: 0,
            anomaly_count: 0,
            threshold: 0.0,
        };
        let rc = unsafe { insight_isolation_forest(ptr::null(), 10, 2, 50, 0.1, 42, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── LOF FFI tests ──

    #[test]
    fn ffi_lof_roundtrip() {
        // Dense cluster + outlier (row-major: 2 cols)
        let mut data: Vec<f64> = Vec::new();
        for i in 0..20 {
            data.push(i as f64 * 0.1);
            data.push(0.0);
        }
        data.push(100.0);
        data.push(100.0);

        let mut result = CAnomalyResult {
            scores: ptr::null_mut(),
            anomalies: ptr::null_mut(),
            n: 0,
            anomaly_count: 0,
            threshold: 0.0,
        };

        let rc = unsafe { insight_lof(data.as_ptr(), 21, 2, 5, 1.5, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n, 21);

        let scores = unsafe { slice::from_raw_parts(result.scores, 21) };
        // Outlier should have highest LOF
        let max_idx = scores
            .iter()
            .enumerate()
            .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
            .unwrap()
            .0;
        assert_eq!(max_idx, 20);
        assert!(scores[20] > 1.5, "outlier LOF = {}", scores[20]);

        unsafe {
            insight_free_f64_array(result.scores, result.n);
            insight_free_i32_array(result.anomalies, result.n);
        }
    }

    #[test]
    fn ffi_lof_null() {
        let mut result = CAnomalyResult {
            scores: ptr::null_mut(),
            anomalies: ptr::null_mut(),
            n: 0,
            anomaly_count: 0,
            threshold: 0.0,
        };
        let rc = unsafe { insight_lof(ptr::null(), 10, 2, 5, 1.5, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Hierarchical FFI tests ──────────────────────────────────────

    #[test]
    fn ffi_hierarchical_basic() {
        // 6 points in 2D, 2 obvious clusters
        let data: Vec<f64> = vec![
            0.0, 0.0, 0.1, 0.1, 0.05, 0.05, // cluster 1
            10.0, 10.0, 10.1, 10.1, 10.05, 10.05, // cluster 2
        ];
        let mut result = CHierarchicalResult {
            n_clusters: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
            n_merges: 0,
            merge_distances: ptr::null_mut(),
            merge_sizes: ptr::null_mut(),
        };

        let rc = unsafe { insight_hierarchical(data.as_ptr(), 6, 2, 3, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_clusters, 2);
        assert_eq!(result.n_labels, 6);
        assert_eq!(result.n_merges, 5); // n-1 merges

        // Verify labels
        let labels = unsafe { slice::from_raw_parts(result.labels, result.n_labels as usize) };
        assert_eq!(labels[0], labels[1]);
        assert_eq!(labels[3], labels[4]);
        assert_ne!(labels[0], labels[3]);

        // Clean up
        unsafe {
            insight_free_i32_array(result.labels, result.n_labels);
            insight_free_f64_array(result.merge_distances, result.n_merges);
            insight_free_i32_array(result.merge_sizes, result.n_merges);
        }
    }

    #[test]
    fn ffi_hierarchical_null_ptr() {
        let mut result = CHierarchicalResult {
            n_clusters: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
            n_merges: 0,
            merge_distances: ptr::null_mut(),
            merge_sizes: ptr::null_mut(),
        };
        let rc = unsafe { insight_hierarchical(ptr::null(), 6, 2, 3, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── HDBSCAN FFI tests ───────────────────────────────────────────

    #[test]
    fn ffi_hdbscan_basic() {
        let data: Vec<f64> = vec![
            0.0, 0.0, 0.1, 0.0, 0.0, 0.1, 0.1, 0.1, 0.05, 0.05, 10.0, 10.0, 10.1, 10.0, 10.0, 10.1,
            10.1, 10.1, 10.05, 10.05, 50.0, 50.0, // noise
        ];
        let mut result = CHdbscanResult {
            n_clusters: 0,
            noise_count: 0,
            labels: ptr::null_mut(),
            probabilities: ptr::null_mut(),
            n_labels: 0,
        };

        let rc = unsafe { insight_hdbscan(data.as_ptr(), 11, 2, 3, 0, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert!(result.n_clusters >= 1);
        assert_eq!(result.n_labels, 11);

        // Verify labels and probs are allocated
        assert!(!result.labels.is_null());
        assert!(!result.probabilities.is_null());

        let probs =
            unsafe { slice::from_raw_parts(result.probabilities, result.n_labels as usize) };
        for &p in probs {
            assert!((0.0..=1.0).contains(&p) || p == 0.0);
        }

        // Clean up
        unsafe {
            insight_free_i32_array(result.labels, result.n_labels);
            insight_free_f64_array(result.probabilities, result.n_labels);
        }
    }

    #[test]
    fn ffi_hdbscan_null_ptr() {
        let mut result = CHdbscanResult {
            n_clusters: 0,
            noise_count: 0,
            labels: ptr::null_mut(),
            probabilities: ptr::null_mut(),
            n_labels: 0,
        };
        let rc = unsafe { insight_hdbscan(ptr::null(), 10, 2, 3, 0, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Correlation FFI tests ─────────────────────────────────────

    #[test]
    fn ffi_correlation_basic() {
        // 3 columns, 5 rows (row-major): c0 and c1 correlated, c2 independent
        let data: Vec<f64> = vec![
            1.0, 2.0, 5.0, 2.0, 4.0, 3.0, 3.0, 6.0, 7.0, 4.0, 8.0, 1.0, 5.0, 10.0, 4.0,
        ];

        let mut result = CCorrelationResult {
            n_vars: 0,
            matrix: ptr::null_mut(),
            n_high_pairs: 0,
        };

        let rc = unsafe {
            insight_correlation(data.as_ptr(), 5, 3, INSIGHT_CORR_PEARSON, &mut result)
        };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_vars, 3);

        let mat = unsafe { slice::from_raw_parts(result.matrix, 9) };
        // Diagonal should be 1.0
        assert!((mat[0] - 1.0).abs() < 1e-10, "r(0,0) = {}", mat[0]);
        assert!((mat[4] - 1.0).abs() < 1e-10, "r(1,1) = {}", mat[4]);
        assert!((mat[8] - 1.0).abs() < 1e-10, "r(2,2) = {}", mat[8]);
        // c0 and c1 are perfectly correlated → r ≈ 1.0
        assert!((mat[1] - 1.0).abs() < 1e-10, "r(0,1) = {}", mat[1]);
        // At least 1 high pair (c0,c1)
        assert!(result.n_high_pairs >= 1);

        unsafe { insight_free_f64_array(result.matrix, 9) };
    }

    #[test]
    fn ffi_correlation_null_ptr() {
        let mut result = CCorrelationResult {
            n_vars: 0,
            matrix: ptr::null_mut(),
            n_high_pairs: 0,
        };
        let rc = unsafe {
            insight_correlation(ptr::null(), 5, 3, INSIGHT_CORR_PEARSON, &mut result)
        };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    #[test]
    fn ffi_correlation_too_few() {
        let data: Vec<f64> = vec![1.0, 2.0]; // 1 row, 2 cols
        let mut result = CCorrelationResult {
            n_vars: 0,
            matrix: ptr::null_mut(),
            n_high_pairs: 0,
        };
        let rc = unsafe {
            insight_correlation(data.as_ptr(), 1, 2, INSIGHT_CORR_PEARSON, &mut result)
        };
        assert_eq!(rc, INSIGHT_ERR_INVALID_INPUT);
    }

    #[test]
    fn ffi_correlation_invalid_method() {
        let data: Vec<f64> = (0..15).map(|i| i as f64).collect();
        let mut result = CCorrelationResult {
            n_vars: 0,
            matrix: ptr::null_mut(),
            n_high_pairs: 0,
        };
        let rc = unsafe {
            insight_correlation(data.as_ptr(), 5, 3, 99, &mut result)
        };
        assert_eq!(rc, INSIGHT_ERR_INVALID_PARAM);
    }

    #[test]
    fn ffi_correlation_kendall() {
        // 3 columns × 5 rows: c0 perfectly monotonic with c1, anti-monotonic with c2.
        let data: Vec<f64> = vec![
            1.0, 2.0, 5.0, // row 0
            2.0, 4.0, 4.0, // row 1
            3.0, 6.0, 3.0, // row 2
            4.0, 8.0, 2.0, // row 3
            5.0, 10.0, 1.0, // row 4
        ];
        let mut result = CCorrelationResult {
            n_vars: 0,
            matrix: ptr::null_mut(),
            n_high_pairs: 0,
        };
        let rc = unsafe {
            insight_correlation(data.as_ptr(), 5, 3, INSIGHT_CORR_KENDALL, &mut result)
        };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_vars, 3);

        let mat = unsafe { slice::from_raw_parts(result.matrix, 9) };
        // c0 and c1 perfectly concordant → tau ≈ 1.0
        assert!((mat[1] - 1.0).abs() < 1e-10, "r(0,1) = {}", mat[1]);
        // c0 and c2 perfectly discordant → tau ≈ -1.0
        assert!((mat[2] + 1.0).abs() < 1e-10, "r(0,2) = {}", mat[2]);

        unsafe { insight_free_f64_array(result.matrix, 9) };
    }

    // ── Regression FFI tests ──────────────────────────────────────

    #[test]
    fn ffi_regression_basic() {
        // y = 2x + 1, perfect linear
        let x: Vec<f64> = vec![1.0, 2.0, 3.0, 4.0, 5.0];
        let y: Vec<f64> = vec![3.0, 5.0, 7.0, 9.0, 11.0];

        let mut result = CRegressionResult {
            intercept: 0.0,
            slope: 0.0,
            r_squared: 0.0,
            adj_r_squared: 0.0,
            f_p_value: 0.0,
        };

        let rc = unsafe { insight_regression(x.as_ptr(), y.as_ptr(), 5, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert!(
            (result.intercept - 1.0).abs() < 1e-6,
            "intercept = {}",
            result.intercept
        );
        assert!(
            (result.slope - 2.0).abs() < 1e-6,
            "slope = {}",
            result.slope
        );
        assert!(
            (result.r_squared - 1.0).abs() < 1e-6,
            "R² = {}",
            result.r_squared
        );
    }

    #[test]
    fn ffi_regression_null_ptr() {
        let mut result = CRegressionResult {
            intercept: 0.0,
            slope: 0.0,
            r_squared: 0.0,
            adj_r_squared: 0.0,
            f_p_value: 0.0,
        };
        let rc = unsafe { insight_regression(ptr::null(), ptr::null(), 5, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    #[test]
    fn ffi_regression_too_few() {
        let x: Vec<f64> = vec![1.0, 2.0];
        let y: Vec<f64> = vec![3.0, 5.0];
        let mut result = CRegressionResult {
            intercept: 0.0,
            slope: 0.0,
            r_squared: 0.0,
            adj_r_squared: 0.0,
            f_p_value: 0.0,
        };
        let rc = unsafe { insight_regression(x.as_ptr(), y.as_ptr(), 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_INVALID_INPUT);
    }

    // ── Mahalanobis FFI tests ─────────────────────────────────────

    #[test]
    fn ffi_mahalanobis_basic() {
        // 8 inliers + 1 outlier, 2D
        let mut data: Vec<f64> = Vec::new();
        for i in 0..8 {
            let x = (i as f64) * 0.5 + (i as f64 * 1.3).sin() * 0.3;
            let y = (i as f64) * 0.4 + (i as f64 * 0.7).cos() * 0.2;
            data.push(x);
            data.push(y);
        }
        data.push(100.0);
        data.push(100.0);

        let mut result = CMahalanobisResult {
            distances: ptr::null_mut(),
            anomalies: ptr::null_mut(),
            n: 0,
            threshold: 0.0,
            outlier_count: 0,
        };

        let rc = unsafe { insight_mahalanobis(data.as_ptr(), 9, 2, 0.975, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n, 9);
        assert!(result.threshold > 0.0);

        // Outlier should have largest distance
        let dists = unsafe { slice::from_raw_parts(result.distances, 9) };
        let max_idx = dists
            .iter()
            .enumerate()
            .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
            .unwrap()
            .0;
        assert_eq!(max_idx, 8);

        unsafe {
            insight_free_f64_array(result.distances, result.n);
            insight_free_i32_array(result.anomalies, result.n);
        }
    }

    #[test]
    fn ffi_mahalanobis_null() {
        let mut result = CMahalanobisResult {
            distances: ptr::null_mut(),
            anomalies: ptr::null_mut(),
            n: 0,
            threshold: 0.0,
            outlier_count: 0,
        };
        let rc = unsafe { insight_mahalanobis(ptr::null(), 5, 2, 0.975, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Cramér's V FFI tests ──────────────────────────────────────

    #[test]
    fn ffi_cramers_v_basic() {
        let table: Vec<f64> = vec![50.0, 0.0, 0.0, 50.0]; // 2x2 perfect
        let mut result = CCramersVResult {
            v: 0.0,
            chi_squared: 0.0,
            p_value: 0.0,
        };

        let rc = unsafe { insight_cramers_v(table.as_ptr(), 2, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert!(result.v > 0.9, "V = {}", result.v);
    }

    #[test]
    fn ffi_cramers_v_null() {
        let mut result = CCramersVResult {
            v: 0.0,
            chi_squared: 0.0,
            p_value: 0.0,
        };
        let rc = unsafe { insight_cramers_v(ptr::null(), 2, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── ANOVA Selection FFI tests ─────────────────────────────────

    #[test]
    fn ffi_anova_select_basic() {
        // 6 points, 2 features (row-major), 2 classes
        let data: Vec<f64> = vec![1.0, 3.0, 1.1, 3.1, 1.2, 2.9, 5.0, 3.0, 5.1, 3.1, 5.2, 2.9];
        let target: Vec<u32> = vec![0, 0, 0, 1, 1, 1];

        let mut result = CAnovaSelectionResult {
            features: ptr::null_mut(),
            n_features: 0,
            n_selected: 0,
        };

        let rc = unsafe {
            insight_anova_select(data.as_ptr(), 6, 2, target.as_ptr(), 0.05, &mut result)
        };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_features, 2);
        assert!(result.n_selected >= 1); // at least f0 should be significant

        unsafe { insight_free_anova_features(result.features, result.n_features) };
    }

    #[test]
    fn ffi_anova_select_null() {
        let mut result = CAnovaSelectionResult {
            features: ptr::null_mut(),
            n_features: 0,
            n_selected: 0,
        };
        let rc = unsafe { insight_anova_select(ptr::null(), 6, 2, ptr::null(), 0.05, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Mutual Information FFI tests ────────────────────────────────

    #[test]
    fn ffi_mutual_info_basic() {
        // 6 points, 2 features, 2 classes; feature 0 separates classes, feature 1 does not
        let data: Vec<f64> = vec![1.0, 5.0, 1.1, 5.1, 1.2, 4.9, 5.0, 5.0, 5.1, 5.1, 5.2, 4.9];
        let target: Vec<u32> = vec![0, 0, 0, 1, 1, 1];

        let mut result = CMutualInfoResult {
            features: ptr::null_mut(),
            n_features: 0,
        };

        let rc =
            unsafe { insight_mutual_info(data.as_ptr(), 6, 2, target.as_ptr(), 0, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_features, 2);

        unsafe { insight_free_mi_features(result.features, result.n_features) };
    }

    #[test]
    fn ffi_mutual_info_null() {
        let mut result = CMutualInfoResult {
            features: ptr::null_mut(),
            n_features: 0,
        };
        let rc = unsafe { insight_mutual_info(ptr::null(), 6, 2, ptr::null(), 0, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Mini-Batch K-Means FFI tests ────────────────────────────────

    #[test]
    fn ffi_mini_batch_kmeans_basic() {
        // Two obvious clusters
        let data: Vec<f64> = vec![0.0, 0.0, 0.1, 0.1, 0.2, 0.0, 5.0, 5.0, 5.1, 5.1, 5.2, 5.0];
        let mut result = CKMeansResult {
            k: 0,
            wcss: 0.0,
            iterations: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
        };

        let rc =
            unsafe { insight_mini_batch_kmeans(data.as_ptr(), 6, 2, 2, 3, 50, 42, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.k, 2);
        assert_eq!(result.n_labels, 6);

        unsafe { insight_free_labels(result.labels, result.n_labels) };
    }

    #[test]
    fn ffi_mini_batch_kmeans_null() {
        let mut result = CKMeansResult {
            k: 0,
            wcss: 0.0,
            iterations: 0,
            labels: ptr::null_mut(),
            n_labels: 0,
        };
        let rc = unsafe { insight_mini_batch_kmeans(ptr::null(), 6, 2, 2, 3, 50, 42, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Gap Statistic FFI tests ─────────────────────────────────────

    #[test]
    fn ffi_gap_statistic_basic() {
        // Two clusters
        let data: Vec<f64> = vec![
            0.0, 0.0, 0.1, 0.1, 0.2, 0.0, 0.0, 0.2, 5.0, 5.0, 5.1, 5.1, 5.2, 5.0, 5.0, 5.2,
        ];
        let mut result = CGapStatResult {
            best_k: 0,
            n_values: 0,
            gap_values: ptr::null_mut(),
            std_errors: ptr::null_mut(),
        };

        let rc = unsafe { insight_gap_statistic(data.as_ptr(), 8, 2, 1, 4, 3, 42, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert!(result.best_k >= 1 && result.best_k <= 4);
        assert!(result.n_values > 0);

        unsafe {
            insight_free_f64_array(result.gap_values, result.n_values * 2);
            insight_free_f64_array(result.std_errors, result.n_values * 2);
        }
    }

    #[test]
    fn ffi_gap_statistic_null() {
        let mut result = CGapStatResult {
            best_k: 0,
            n_values: 0,
            gap_values: ptr::null_mut(),
            std_errors: ptr::null_mut(),
        };
        let rc = unsafe { insight_gap_statistic(ptr::null(), 8, 2, 1, 4, 3, 42, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── Permutation Importance FFI tests ────────────────────────────

    #[test]
    fn ffi_permutation_importance_basic() {
        // y = 2*x0 + noise; x1 = noise
        let data: Vec<f64> = vec![1.0, 0.5, 2.0, 0.3, 3.0, 0.8, 4.0, 0.1, 5.0, 0.9];
        let target: Vec<f64> = vec![2.1, 4.0, 6.2, 7.9, 10.1];

        let mut result = CPermImportanceResult {
            baseline_score: 0.0,
            features: ptr::null_mut(),
            n_features: 0,
        };

        let rc = unsafe {
            insight_permutation_importance(data.as_ptr(), 5, 2, target.as_ptr(), 5, 42, &mut result)
        };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_features, 2);
        assert!(result.baseline_score > 0.5); // should have decent R²

        unsafe { insight_free_perm_features(result.features, result.n_features) };
    }

    #[test]
    fn ffi_permutation_importance_null() {
        let mut result = CPermImportanceResult {
            baseline_score: 0.0,
            features: ptr::null_mut(),
            n_features: 0,
        };
        let rc = unsafe {
            insight_permutation_importance(ptr::null(), 5, 2, ptr::null(), 5, 42, &mut result)
        };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    // ── PELT tests ───────────────────────────────────────────────────

    #[test]
    fn ffi_pelt_single_changepoint() {
        let mut data = vec![0.0_f64; 50];
        data.extend(vec![5.0; 50]);

        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };

        let rc = unsafe { insight_pelt(data.as_ptr(), 100, 0, 0.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_changepoints, 1);
        assert!(!result.changepoints.is_null());

        let cp = unsafe { *result.changepoints };
        assert!(
            (cp as i64 - 50).unsigned_abs() <= 2,
            "changepoint near 50, got {}",
            cp
        );

        unsafe { insight_free_pelt_result(&mut result) };
    }

    #[test]
    fn ffi_pelt_no_changepoint() {
        let data = vec![5.0_f64; 100];

        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };

        let rc = unsafe { insight_pelt(data.as_ptr(), 100, 0, 0.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_changepoints, 0);
        assert!(result.changepoints.is_null());
    }

    #[test]
    fn ffi_pelt_null_pointer() {
        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };
        let rc = unsafe { insight_pelt(ptr::null(), 10, 0, 0.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }

    #[test]
    fn ffi_pelt_invalid_cost() {
        let data = [1.0_f64; 10];
        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };
        let rc = unsafe { insight_pelt(data.as_ptr(), 10, 99, 0.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_INVALID_PARAM);
    }

    #[test]
    fn ffi_pelt_multi_two_channels() {
        // Row-major [n_samples=100, n_channels=2]:
        // sample s, channel c → data[s * 2 + c]
        let mut data = Vec::with_capacity(200);
        for s in 0..100 {
            // Channel 0: [0..50]=0, [50..100]=5
            data.push(if s < 50 { 0.0_f64 } else { 5.0 });
            // Channel 1: [0..50]=0, [50..100]=3
            data.push(if s < 50 { 0.0_f64 } else { 3.0 });
        }

        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };

        // n_samples=100, n_channels=2
        let rc = unsafe { insight_pelt_multi(data.as_ptr(), 100, 2, 0, 0.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_changepoints, 1);

        let cp = unsafe { *result.changepoints };
        assert!(
            (cp as i64 - 50).unsigned_abs() <= 2,
            "changepoint near 50, got {}",
            cp
        );

        unsafe { insight_free_pelt_result(&mut result) };
    }

    #[test]
    fn ffi_pelt_multi_asymmetric_layout() {
        // Asymmetric shape (n_samples=80 ≠ n_channels=3) so any swapped
        // dimensions or off-by-one transpose surfaces immediately.
        // Channel 0: step at sample 40, level jump 0 → 4
        // Channel 1: step at sample 40, level jump 0 → 6
        // Channel 2: step at sample 40, level jump 0 → 2
        let n_samples = 80usize;
        let n_channels = 3usize;
        let mut data = Vec::with_capacity(n_samples * n_channels);
        for s in 0..n_samples {
            let pre = s < 40;
            data.push(if pre { 0.0_f64 } else { 4.0 });
            data.push(if pre { 0.0_f64 } else { 6.0 });
            data.push(if pre { 0.0_f64 } else { 2.0 });
        }

        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };

        let rc = unsafe {
            insight_pelt_multi(
                data.as_ptr(),
                n_samples as u32,
                n_channels as u32,
                0,
                0.0,
                2,
                &mut result,
            )
        };
        assert_eq!(rc, INSIGHT_OK);
        assert_eq!(result.n_changepoints, 1);

        let cp = unsafe { *result.changepoints };
        assert!(
            (cp as i64 - 40).unsigned_abs() <= 2,
            "changepoint near 40, got {}",
            cp
        );

        unsafe { insight_free_pelt_result(&mut result) };
    }

    #[test]
    fn ffi_pelt_multi_null() {
        let mut result = CPeltResult {
            changepoints: ptr::null_mut(),
            n_changepoints: 0,
        };
        // n_samples=50, n_channels=2
        let rc = unsafe { insight_pelt_multi(ptr::null(), 50, 2, 0, 0.0, 2, &mut result) };
        assert_eq!(rc, INSIGHT_ERR_NULL_PTR);
    }
}