/**
 * @file   corridor_ekf.c
 * @brief  鲁棒 EKF 走廊相对定位滤波器实现
 *
 *         完整算法流程：
 *
 *         【预测步】Predict(vx, wz, dt)
 *           x_pred = f(x, u)  -- 非线性状态转移
 *           P_pred = F * P * F^T + Q  -- 协方差预测
 *
 *         【更新步】Update(obs)
 *           z     = h(x_pred) -- 观测预测
 *           y     = z_meas - z   -- 新息 (Innovation)
 *           S     = H * P_pred * H^T + R -- 新息协方差
 *           d²    = y^T * S^(-1) * y  -- 马氏距离平方
 *           if d² > χ²_threshold: 拒绝观测 (鲁棒)
 *           K     = P_pred * H^T * S^(-1) -- 卡尔曼增益
 *           x     = x_pred + K * y     -- 状态更新
 *           P     = (I - K * H) * P_pred -- 协方差更新
 */

#include "corridor_ekf.h"
#include <math.h>
#include <string.h>

/* =========================================================
 * 内部静态状态
 * ========================================================= */
static CorridorEKFConfig_t  s_cfg;
static CorridorEKFState_t  s_state;
static bool                s_initialized = false;
static uint32_t            s_last_update_ms = 0U;

/* 协方差上界保护阈值 */
#define P_MAX_DIAG    100.0f

/* =========================================================
 * 内部辅助函数
 * ========================================================= */

/** 限幅 */
static inline float clampf(float val, float lo, float hi)
{
    if (val < lo) return lo;
    if (val > hi) return hi;
    return val;
}

/** 对称矩阵拷贝 + 上三角 symmetrize */
static void symmetrize(float M[3][3])
{
    M[1][0] = M[0][1];
    M[2][0] = M[0][2];
    M[2][1] = M[1][2];
}

/** P 上界保护 */
static void protect_P(float P[3][3])
{
    for (int i = 0; i < 3; i++) {
        if (P[i][i] > P_MAX_DIAG) P[i][i] = P_MAX_DIAG;
        if (P[i][i] < 0.0f)      P[i][i] = 0.0f;
    }
}

/** 计算 2x2 对称矩阵的逆 (原地) */
static bool invert_2x2_sym(float S[2][2])
{
    float det = S[0][0] * S[1][1] - S[0][1] * S[1][0];
    if (fabsf(det) < 1e-8f) {
        return false;  // 奇异矩阵
    }
    float inv_det = 1.0f / det;
    float S00 = S[0][0];
    S[0][0] =  inv_det * S[1][1];
    S[1][1] =  inv_det * S00;
    S[0][1] = -inv_det * S[0][1];
    S[1][0] =  S[0][1];
    return true;
}

/** 2x2 对称矩阵求逆 (原地) */
static bool invert_3x3_cholesky(float S[3][3])
{
    // 使用 Cholesky 分解求逆
    float L[3][3] = {0};
    
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j <= i; j++) {
            float sum = S[i][j];
            for (int k = 0; k < j; k++) {
                sum -= L[i][k] * L[j][k];
            }
            if (i == j) {
                if (sum <= 0.0f) return false;
                L[i][j] = sqrtf(sum);
            } else {
                L[i][j] = sum / L[j][j];
            }
        }
    }

    // 求逆: S_inv = L^(-T) * L^(-1)
    float Linv[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        Linv[i][i] = 1.0f / L[i][i];
        for (int j = i - 1; j >= 0; j--) {
            float sum = 0.0f;
            for (int k = j + 1; k <= i; k++) {
                sum += L[k][j] * Linv[k][i];
            }
            Linv[j][i] = -sum / L[j][j];
        }
    }

    // S_inv = Linv^T * Linv
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            float sum = 0.0f;
            for (int k = 0; k < 3; k++) {
                sum += Linv[k][i] * Linv[k][j];
            }
            S[i][j] = sum;
        }
    }
    return true;
}

/** 计算马氏距离平方 (新息向量 y, 新息协方差 S_inv) */
static float mahalanobis_d2_3dof(const float y[3], const float S_inv[3][3])
{
    // d² = y^T * S_inv * y
    float tmp[3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            tmp[i] += S_inv[i][j] * y[j];
        }
    }
    float d2 = 0.0f;
    for (int i = 0; i < 3; i++) {
        d2 += y[i] * tmp[i];
    }
    return d2;
}

/** 计算马氏距离平方 (1DOF: 只用 e_y) */
static float mahalanobis_d2_1dof(float y_ey, float S_inv_00)
{
    return y_ey * y_ey * S_inv_00;
}

/** 计算马氏距离平方 (2DOF: e_ey + e_th_avg) */
static float mahalanobis_d2_2dof(const float y[2], const float S_inv[2][2])
{
    float tmp[2] = {0};
    for (int i = 0; i < 2; i++) {
        for (int j = 0; j < 2; j++) {
            tmp[i] += S_inv[i][j] * y[j];
        }
    }
    float d2 = 0.0f;
    for (int i = 0; i < 2; i++) {
        d2 += y[i] * tmp[i];
    }
    return d2;
}

/** 零矩阵 */
static void zero_3x3(float M[3][3])
{
    memset(M, 0, sizeof(float) * 9);
}

/** 单位矩阵 */
static void eye_3x3(float M[3][3])
{
    zero_3x3(M);
    M[0][0] = M[1][1] = M[2][2] = 1.0f;
}

/* =========================================================
 * API 实现
 * ========================================================= */

void CorridorEKF_Init(const CorridorEKFConfig_t *config)
{
    s_cfg = *config;
    memset(&s_state, 0, sizeof(s_state));

    /* 初始化状态 */
    s_state.x[0] = 0.0f;  // e_y
    s_state.x[1] = 0.0f;  // e_th
    s_state.x[2] = 0.0f;  // s

    /* 初始化协方差 */
    eye_3x3(s_state.P);
    s_state.P[0][0] = config->P0_diag[0];
    s_state.P[1][1] = config->P0_diag[1];
    s_state.P[2][2] = config->P0_diag[2];

    s_initialized = true;
}

void CorridorEKF_Reset(void)
{
    if (!s_initialized) return;
    
    s_state.x[0] = 0.0f;
    s_state.x[1] = 0.0f;
    s_state.x[2] = 0.0f;
    
    eye_3x3(s_state.P);
    s_state.P[0][0] = s_cfg.P0_diag[0];
    s_state.P[1][1] = s_cfg.P0_diag[1];
    s_state.P[2][2] = s_cfg.P0_diag[2];
    
    s_last_update_ms = 0U;
}

void CorridorEKF_SetProcessNoise(float q_ey, float q_eth, float q_s)
{
    s_cfg.q_ey  = q_ey;
    s_cfg.q_eth = q_eth;
    s_cfg.q_s  = q_s;
}

void CorridorEKF_SetMeasurementNoise(float r_ey, float r_eth)
{
    s_cfg.r_ey  = r_ey;
    s_cfg.r_eth = r_eth;
}

/* =========================================================
 * 预测步 (Predict)
 * ========================================================= */
void CorridorEKF_Predict(float odom_vx, float imu_wz, float dt)
{
    if (!s_initialized || dt <= 0.0f) return;

    float e_y  = s_state.x[0];
    float e_th = s_state.x[1];
    float s    = s_state.x[2];
    float vx   = odom_vx;
    float wz   = imu_wz;

    /* 状态预测: x_pred = f(x, u) */
    float cos_th = cosf(e_th);
    float sin_th = sinf(e_th);

    /* 安全检查: cos_th 不能太小 (防止数值爆炸) */
    if (fabsf(cos_th) < 0.01f) cos_th = (cos_th >= 0.0f) ? 0.01f : -0.01f;

    float e_y_pred  = e_y  + vx * sin_th * dt;
    float e_th_pred = e_th + wz * dt;
    float s_pred    = s    + vx * cos_th * dt;

    s_state.x[0] = e_y_pred;
    s_state.x[1] = e_th_pred;
    s_state.x[2] = s_pred;

    /* 雅可比矩阵 F (状态转移的 Jacobian) */
    float F[3][3] = {0};
    F[0][0] = 1.0f;
    F[0][1] = vx * cos_th * dt;  // de_y/de_th
    F[0][2] = 0.0f;

    F[1][0] = 0.0f;
    F[1][1] = 1.0f;
    F[1][2] = 0.0f;

    F[2][0] = 0.0f;
    F[2][1] = -vx * sin_th * dt;  // ds/de_th
    F[2][2] = 1.0f;

    /* 过程噪声协方差 Q (对角) */
    float Q[3][3] = {0};
    Q[0][0] = s_cfg.q_ey  * dt * dt;
    Q[1][1] = s_cfg.q_eth * dt * dt;
    Q[2][2] = s_cfg.q_s   * dt * dt;

    /* 协方差预测: P_pred = F * P * F^T + Q */
    float F_P[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            for (int k = 0; k < 3; k++) {
                F_P[i][j] += F[i][k] * s_state.P[k][j];
            }
        }
    }

    float P_pred[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            for (int k = 0; k < 3; k++) {
                P_pred[i][j] += F_P[i][k] * F[j][k];  // F^T: F[j][k] = F[k][j]
            }
        }
        P_pred[i][i] += Q[i][i];  // 加过程噪声
    }

    symmetrize(P_pred);

    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            s_state.P[i][j] = P_pred[i][j];
        }
    }

    protect_P(s_state.P);
}

/* =========================================================
 * 观测步 (Update) - 鲁棒 EKF
 * ========================================================= */
int CorridorEKF_Update(const CorridorObs_t *obs, CorridorState_t *out_state)
{
    if (!s_initialized) return 0;

    int updated_obs_count = 0;

    /* 清除新息和拒绝掩码 */
    memset(s_state.K, 0, sizeof(s_state.K));
    uint8_t reject_mask = 0U;
    float max_maha_d2 = 0.0f;

    /* ----------------------------------------------------
     * 提取有效观测
     * ---------------------------------------------------- */
    bool left_ok  = ((obs->valid_mask & (1U << 0)) != 0U) &&
                    ((obs->valid_mask & (1U << 1)) != 0U);
    bool right_ok = ((obs->valid_mask & (1U << 2)) != 0U) &&
                    ((obs->valid_mask & (1U << 3)) != 0U);

    /* 左右侧横向平均距离 */
    float d_lf = obs->d_lf, d_lr = obs->d_lr;
    float d_rf = obs->d_rf, d_rr = obs->d_rr;
    float Ls   = s_cfg.sensor_base_length;
    float W    = s_cfg.corridor_width;
    float yoff = s_cfg.y_offset;
    float inset = s_cfg.side_sensor_inset;
    float Rw    = s_cfg.robot_width;

    /* 传感器居中时的理论读数 (考虑车体宽度和传感器内缩)
     *
     * 推导 (以左侧为例)：
     *   沟道宽 W，车体宽 Rw，传感器内缩 inset
     *   居中时：
     *     车体左边缘到左墙 = (W - Rw) / 2
     *     传感器到左墙     = (W - Rw) / 2 + inset  (传感器比边缘更靠内)
     *   
     *   所以 d_center = (W - Rw) / 2 + inset
     *
     * 验证：W=0.40, Rw=0.20, inset=0
     *   d_center = (0.40-0.20)/2 + 0 = 0.10m  ✓ (每边余量10cm)
     *
     * 单侧公式：e_y = d_center - d_left (左侧传感器越近墙，偏差越大)
     * 双侧公式：e_y = [(d_center - d_left) + (d_right - d_center)] / 2
     *               = (d_right - d_left) / 2  (d_center 被消掉)
     *
     * ⚠ 当 inset = 0 且 Rw = 0 时，d_center = W/2，退化回原始行为
     */
    float d_center = (W - Rw) / 2.0f + inset;  /* 传感器居中时的理论读数 */

    /* 观测值 (测量) */
    float z_ey   = 0.0f;
    float z_eth_L = 0.0f;
    float z_eth_R = 0.0f;

    int valid_sides = 0;

    if (left_ok) {
        z_ey   += d_center - ((d_lf + d_lr) / 2.0f) - yoff;
        z_eth_L = atan2f(d_lr - d_lf, Ls);
        valid_sides++;
    }

    if (right_ok) {
        z_ey   += ((d_rf + d_rr) / 2.0f) - d_center - yoff;
        z_eth_R = atan2f(d_rf - d_rr, Ls);
        valid_sides++;
    }

    if (valid_sides == 0) {
        /* 两边都失效: 协方差持续膨胀，输出预测值 */
        out_state->t_ms          = obs->t_ms;
        out_state->e_y           = s_state.x[0];
        out_state->e_th          = s_state.x[1];
        out_state->s             = s_state.x[2];
        out_state->conf          = clampf(1.0f - (s_state.P[0][0] + s_state.P[1][1]) * 0.1f, 0.0f, 1.0f);
        out_state->obs_reject_mask = 0xFF;
        out_state->mahalanobis_d2 = 0.0f;

        for (int i = 0; i < 3; i++) {
            for (int j = 0; j < 3; j++) {
                out_state->P[i][j] = s_state.P[i][j];
            }
        }

        /* 协方差膨胀 (无观测时的信任衰减) */
        s_state.P[0][0] += s_cfg.q_ey  * 5.0f;
        s_state.P[1][1] += s_cfg.q_eth * 5.0f;
        protect_P(s_state.P);
        return 0;
    }

    /* 横向观测: 两侧平均 */
    if (valid_sides == 2) {
        z_ey /= 2.0f;
    }

    /* ----------------------------------------------------
     * 构建观测向量 z 和预测观测 h(x)
     * ---------------------------------------------------- */
    float e_y   = s_state.x[0];
    float e_th = s_state.x[1];

    /* 预测观测 h(x) */
    float h_ey    = e_y;  // 横向: z_ey ≈ e_y
    float h_eth_L = e_th;  // 左侧航向: z_eth_L ≈ e_th
    float h_eth_R = e_th;  // 右侧航向: z_eth_R ≈ e_th

    /* 新息向量 y = z - h(x) */
    float y[3] = {0};
    int obs_idx = 0;

    /* e_y 观测 */
    y[obs_idx++] = z_ey - h_ey;

    /* e_th_L 观测 */
    if (left_ok)  y[obs_idx++] = z_eth_L - h_eth_L;

    /* e_th_R 观测 */
    if (right_ok) y[obs_idx++] = z_eth_R - h_eth_R;

    /* ----------------------------------------------------
     * 构建 H 矩阵 (Jacobian of h(x))
     * ---------------------------------------------------- */
    /* H 布局:
     *   航向角观测对应列是 1,0,0 (e_y 观测量)
     *   航向角观测对应列是 0,1,0 (e_th 观测量)
     *   s 不被直接观测
     */
    float H[3][3] = {0};
    int H_row = 0;

    /* e_y 行: H = [1, 0, 0] */
    H[H_row][0] = 1.0f; H[H_row][1] = 0.0f; H[H_row][2] = 0.0f;
    H_row++;

    /* e_th_L 行: H = [0, 1, 0] */
    if (left_ok) {
        H[H_row][0] = 0.0f; H[H_row][1] = 1.0f; H[H_row][2] = 0.0f;
        H_row++;
    }

    /* e_th_R 行: H = [0, 1, 0] */
    if (right_ok) {
        H[H_row][0] = 0.0f; H[H_row][1] = 1.0f; H[H_row][2] = 0.0f;
        H_row++;
    }

    int obs_count = H_row;

    /* ----------------------------------------------------
     * 构建观测噪声协方差 R (根据有效侧数量调整)
     * ---------------------------------------------------- */
    float R[3][3] = {0};
    R[0][0] = s_cfg.r_ey;  // e_y 的噪声

    if (left_ok && right_ok) {
        /* 双侧: 航向噪声更小 (两个独立观测平均) */
        R[1][1] = s_cfg.r_eth * 0.5f;  // e_th_L
        R[2][2] = s_cfg.r_eth * 0.5f;  // e_th_R
    } else {
        /* 单侧: 航向噪声较大 */
        if (left_ok) {
            R[1][1] = s_cfg.r_eth;
        }
        if (right_ok) {
            R[1][1] = s_cfg.r_eth;
        }
    }

    /* ----------------------------------------------------
     * 计算新息协方差 S = H * P * H^T + R
     * ---------------------------------------------------- */
    float HP[3][3] = {0};
    for (int i = 0; i < obs_count; i++) {
        for (int j = 0; j < 3; j++) {
            for (int k = 0; k < 3; k++) {
                HP[i][j] += H[i][k] * s_state.P[k][j];
            }
        }
    }

    float S[3][3] = {0};
    for (int i = 0; i < obs_count; i++) {
        for (int j = 0; j < 3; j++) {
            for (int k = 0; k < 3; k++) {
                S[i][j] += HP[i][k] * H[j][k];  // H^T: H[j][k] = H[k][j]
            }
        }
        S[i][i] += R[i][i];  // 加观测噪声
    }

    /* ----------------------------------------------------
     * 计算 S 的逆
     * ---------------------------------------------------- */
    float S_inv[3][3] = {0};
    if (obs_count == 1) {
        /* 1 观测: 标量 */
        if (fabsf(S[0][0]) < 1e-8f) {
            goto output_result;
        }
        S_inv[0][0] = 1.0f / S[0][0];
    } else if (obs_count == 2) {
        /* 2 观测：2x2 - 拷贝到局部矩阵 */
        float S_2x2[2][2];
        S_2x2[0][0] = S[0][0]; S_2x2[0][1] = S[0][1];
        S_2x2[1][0] = S[1][0]; S_2x2[1][1] = S[1][1];
        
        if (!invert_2x2_sym(S_2x2)) {
            goto output_result;
        }
        
        /* 拷贝回 S_inv */
        S_inv[0][0] = S_2x2[0][0]; S_inv[0][1] = S_2x2[0][1];
        S_inv[1][0] = S_2x2[1][0]; S_inv[1][1] = S_2x2[1][1];
    } else {
        /* 3 观测: 3x3 */
        memcpy(S_inv, S, sizeof(float) * 9);
        if (!invert_3x3_cholesky(S_inv)) {
            goto output_result;
        }
    }

    /* ----------------------------------------------------
     * χ² 马氏距离检验 (鲁棒拒绝)
     * ---------------------------------------------------- */
    float d2_total = 0.0f;
    if (obs_count == 1) {
        d2_total = mahalanobis_d2_1dof(y[0], S_inv[0][0]);
    } else if (obs_count == 2) {
        d2_total = mahalanobis_d2_2dof(y, (const float (*)[2])S_inv);
    } else {
        d2_total = mahalanobis_d2_3dof(y, S_inv);
    }
    max_maha_d2 = d2_total;

    /* 单自由度检验: e_y 单独检验 */
    float d2_ey = mahalanobis_d2_1dof(y[0], S_inv[0][0]);
    if (d2_ey > s_cfg.chi2_1dof) {
        reject_mask |= (1U << 0);  // 拒绝 e_y
    }

    /* 1 DOF 门限约 3.84 (95%), 约 6.63 (99%) */
    /* 如果整体 d² 过大，拒绝最可疑的观测 */
    if (obs_count >= 2) {
        /* 检验 e_th_L */
        if (left_ok && !(reject_mask & (1U << 0))) {
            /* 需要重新计算不含 e_y 的马氏距离 */
            /* 简化: 用 y[1]^2 / S[1][1] 作为 1DOF 近似 */
            if (fabsf(S[1][1]) > 1e-8f) {
                float d2_eth_L = y[1] * y[1] / S[1][1];
                if (d2_eth_L > s_cfg.chi2_1dof) {
                    reject_mask |= (1U << 1);  // 拒绝 e_th_L
                }
            }
        }

        /* 检验 e_th_R */
        if (right_ok && !(reject_mask & (1U << 0)) && obs_count >= 3) {
            if (fabsf(S[2][2]) > 1e-8f) {
                float d2_eth_R = y[2] * y[2] / S[2][2];
                if (d2_eth_R > s_cfg.chi2_1dof) {
                    reject_mask |= (1U << 2);  // 拒绝 e_th_R
                }
            }
        }
    }

    /* ----------------------------------------------------
     * 计算卡尔曼增益 K = P * H^T * S^(-1)
     * P1 修复: 必须做完整矩阵乘法 (P*H^T) * S_inv，
     * 而不能只乘 S_inv 的对角项 S_inv[j][j]。
     * 后者会忽略 S 的非对角元素（观测间相关性），
     * 导致卡尔曼增益不正确，影响滤波收敛性。
     * ---------------------------------------------------- */
    float HT[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < obs_count; j++) {
            HT[i][j] = H[j][i];  // H^T
        }
    }

    /* Step 1: PH^T = P * H^T, 结果为 3×obs_count */
    float PHT[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < obs_count; j++) {
            for (int k = 0; k < 3; k++) {
                PHT[i][j] += s_state.P[i][k] * HT[k][j];
            }
        }
    }

    /* Step 2: K = PHT * S_inv, 结果为 3×obs_count */
    float K[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < obs_count; j++) {
            for (int k = 0; k < obs_count; k++) {
                K[i][j] += PHT[i][k] * S_inv[k][j];
            }
        }
    }

    /* ----------------------------------------------------
     * 跳过被拒绝的观测，更新剩余观测
     * ---------------------------------------------------- */
    int used_obs = 0;
    for (int i = 0; i < obs_count; i++) {
        uint8_t bit = (i == 0) ? (1U << 0) : ((i == 1) ? (1U << 1) : (1U << 2));
        if (reject_mask & bit) {
            /*
             * 关键：被拒绝的观测不仅不能更新状态，
             * 也不能参与后续 KH/P 更新，否则会错误收缩协方差。
             */
            K[0][i] = 0.0f;
            K[1][i] = 0.0f;
            K[2][i] = 0.0f;
            continue;
        }

        /* 状态更新: x += K[:, i] * y[i] */
        s_state.x[0] += K[0][i] * y[i];
        s_state.x[1] += K[1][i] * y[i];
        s_state.x[2] += K[2][i] * y[i];
        used_obs++;
    }

    /* ----------------------------------------------------
     * 协方差更新: P = (I - K * H) * P_pred
     * 简化 Joseph 形式: P = (I - K*H) * P * (I - K*H)^T + K * R * K^T
     * 这里使用简化形式: P = (I - K*H) * P
     * ---------------------------------------------------- */
    float KH[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            for (int k = 0; k < obs_count; k++) {
                KH[i][j] += K[i][k] * H[k][j];
            }
        }
    }

    float I_KH[3][3];
    I_KH[0][0] = 1.0f - KH[0][0]; I_KH[0][1] = -KH[0][1]; I_KH[0][2] = -KH[0][2];
    I_KH[1][0] = -KH[1][0]; I_KH[1][1] = 1.0f - KH[1][1]; I_KH[1][2] = -KH[1][2];
    I_KH[2][0] = -KH[2][0]; I_KH[2][1] = -KH[2][1]; I_KH[2][2] = 1.0f - KH[2][2];

    float P_tmp[3][3] = {0};
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            for (int k = 0; k < 3; k++) {
                P_tmp[i][j] += I_KH[i][k] * s_state.P[k][j];
            }
        }
    }

    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            s_state.P[i][j] = P_tmp[i][j];
        }
    }

    symmetrize(s_state.P);
    protect_P(s_state.P);

    updated_obs_count = used_obs;

output_result:
    /* ----------------------------------------------------
     * 填充输出
     * ---------------------------------------------------- */
    out_state->t_ms = obs->t_ms;
    out_state->e_y  = s_state.x[0];
    out_state->e_th = s_state.x[1];
    out_state->s    = s_state.x[2];

    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            out_state->P[i][j] = s_state.P[i][j];
        }
    }

    out_state->mahalanobis_d2 = max_maha_d2;
    out_state->obs_reject_mask = reject_mask;

    /* 置信度: 基于协方差迹和拒绝比例 */
    float P_trace = s_state.P[0][0] + s_state.P[1][1] + s_state.P[2][2];
    float conf_from_P = clampf(1.0f - P_trace * 0.5f, 0.0f, 1.0f);
    
    /* 根据有效侧数加权 */
    float side_factor = (valid_sides == 2) ? 1.0f : 0.7f;
    
    /* 根据拒绝比例降低置信度 */
    float reject_ratio = 0.0f;
    if (obs_count > 0) {
        int rejected = 0;
        if (reject_mask & (1U << 0)) rejected++;
        if (reject_mask & (1U << 1)) rejected++;
        if (reject_mask & (1U << 2)) rejected++;
        reject_ratio = (float)rejected / (float)obs_count;
    }
    
    out_state->conf = clampf(conf_from_P * side_factor * (1.0f - reject_ratio * 0.5f), 0.0f, 1.0f);

    return updated_obs_count;
}

/* =========================================================
 * 辅助 API
 * ========================================================= */

/* =========================================================
 * IMU 航向观测更新 (独立 1DOF 标量 EKF 更新)
 *
 * 观测方程: z_eth_imu = imu_yaw_rad - imu_yaw_ref_rad
 * 对应状态: e_th (x[1])
 * H = [0, 1, 0]  (只观测 e_th)
 *
 * 设计意图：
 *   - IMU 内部维护的 yaw 经过模块校准，比外部积分 wz 更稳定
 *   - 在侧墙观测丢失 (转弯/单侧退化) 时提供航向约束
 *   - 使用较大 R 值，让侧墙观测在有效时主导
 * ========================================================= */
void CorridorEKF_UpdateIMUYaw(float imu_yaw_rad, float imu_yaw_ref_rad, bool valid)
{
    if (!s_initialized || !valid) return;

    /* 观测值: IMU 相对于走廊参考方向的航向偏差 */
    float z_eth_imu = imu_yaw_rad - imu_yaw_ref_rad;

    /* 新息: y = z - h(x), h(x) = e_th = x[1] */
    float y_imu = z_eth_imu - s_state.x[1];

    /* H = [0, 1, 0] → S = P[1][1] + R_imu */
    float R_imu = s_cfg.r_eth_imu;
    float S_imu = s_state.P[1][1] + R_imu;

    if (fabsf(S_imu) < 1e-8f) return;

    /* χ² 1DOF 检验: d² = y² / S */
    float d2_imu = y_imu * y_imu / S_imu;
    if (d2_imu > s_cfg.chi2_1dof) return;  /* 拒绝异常观测 */

    /* 卡尔曼增益: K = P * H^T / S = P[:][1] / S */
    float K[3];
    float S_inv = 1.0f / S_imu;
    K[0] = s_state.P[0][1] * S_inv;
    K[1] = s_state.P[1][1] * S_inv;
    K[2] = s_state.P[2][1] * S_inv;

    /* 状态更新: x += K * y */
    s_state.x[0] += K[0] * y_imu;
    s_state.x[1] += K[1] * y_imu;
    s_state.x[2] += K[2] * y_imu;

    /* 协方差更新: P = (I - K*H) * P
     * H = [0, 1, 0], 所以 KH[i][j] = K[i] * H[j] = K[i] * δ(j,1) */
    float P_new[3][3];
    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            P_new[i][j] = s_state.P[i][j] - K[i] * s_state.P[1][j];
        }
    }

    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            s_state.P[i][j] = P_new[i][j];
        }
    }

    symmetrize(s_state.P);
    protect_P(s_state.P);
}

void CorridorEKF_GetState(CorridorState_t *out)
{
    if (!s_initialized || out == NULL) return;

    out->t_ms = s_last_update_ms;
    out->e_y  = s_state.x[0];
    out->e_th = s_state.x[1];
    out->s    = s_state.x[2];

    for (int i = 0; i < 3; i++) {
        for (int j = 0; j < 3; j++) {
            out->P[i][j] = s_state.P[i][j];
        }
    }

    out->mahalanobis_d2 = 0.0f;
    out->obs_reject_mask = 0U;

    float P_trace = s_state.P[0][0] + s_state.P[1][1] + s_state.P[2][2];
    out->conf = clampf(1.0f - P_trace * 0.5f, 0.0f, 1.0f);
}