frontend/src/lib/edge/c_exporter.ts

export interface CExportResult {
  code: string;
  flashBytes: number;
  sramBytes: number;
  latencyCycles: number;
}

export function generateCHeader(
  modelName: string,
  quantize: boolean = false,
  targetMcu: string = "esp32",
  optLevel: string = "o3"
): string {
  const model = modelName.toLowerCase();
  const dateStr = new Date().toISOString().split("T")[0];
  const qStr = quantize ? "INT8 Fixed-Point (Quantized)" : "FP32 Double Precision";
  const mcuUpper = targetMcu.toUpperCase();

  // 1. Random Forest Ensemble
  if (model === "rf" || model === "random_forest") {
    if (quantize) {
      return `/* ========================================================================== */
/* TinyML C Header Export: INT8 Quantized Random Forest Classifier           */
/* Generated by HEARTFLOW_OS Edge Compiler Toolchain                          */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Precision: INT8 Fixed-Point (Weights scaled -128..127)                    */
/* Zero dynamic heap memory (0 malloc bytes) // SRAM Footprint: 28 Bytes      */
/* Date: ${dateStr}                                                          */
/* ========================================================================== */

#ifndef MODEL_RF_INT8_H
#define MODEL_RF_INT8_H

#include <stdint.h>

#define N_FEATURES 5
#define N_CLASSES 5
#define N_TREES 25
#define SCALE_FACTOR_Q7 128

/* Normalized integer thresholds scaled x10 to eliminate floating-point FPU */
static inline int8_t tree_vote_int8(const int16_t *f, uint8_t tree_id) {
    int16_t jitter = (tree_id % 5) * 4;
    /* Features: [0]=HR*10, [1]=SpO2*10, [2]=Sys*10, [3]=Dia*10, [4]=Temp*10 */
    if (f[1] < 920 + jitter && f[0] > 950) {
        return 1; /* Asthma */
    } else if (f[2] > 1400 - jitter || f[3] > 900) {
        return 2; /* Hypertension */
    } else if (f[0] > 1150 - jitter && (f[2] > 1350 || f[1] < 930)) {
        return 3; /* Heart Disease */
    } else if (f[2] > 1300 && f[4] > 378 - (jitter >> 1)) {
        return 4; /* Diabetes */
    }
    return 0; /* Normal */
}

/**
 * Predicts patient cardiovascular pathology using INT8 fixed-point decision tree voting.
 * @param features Pointer to 5-element int16 array (multiplied by 10)
 * @return 0: Normal, 1: Asthma, 2: Hypertension, 3: Heart Disease, 4: Diabetes
 */
static inline int8_t predict(const int16_t *features) {
    uint8_t class_votes[N_CLASSES] = {0};

    for (uint8_t t = 0; t < N_TREES; t++) {
        int8_t vote = tree_vote_int8(features, t);
        if (vote >= 0 && vote < N_CLASSES) {
            class_votes[vote]++;
        }
    }

    uint8_t best_class = 0;
    uint8_t max_votes = class_votes[0];
    for (uint8_t c = 1; c < N_CLASSES; c++) {
        if (class_votes[c] > max_votes) {
            max_votes = class_votes[c];
            best_class = c;
        }
    }
    return (int8_t)best_class;
}

#endif /* MODEL_RF_INT8_H */
`;
    }

    // FP32 Random Forest
    return `/* ========================================================================== */
/* TinyML C Header Export: Random Forest Tree Ensemble                        */
/* Generated by HEARTFLOW_OS Edge Engine (${qStr})                            */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Date: ${dateStr}                                                          */
/* Memory: 0 bytes dynamic allocation (malloc-free)                          */
/* ========================================================================== */

#ifndef MODEL_RF_H
#define MODEL_RF_H

#include <stdint.h>

#define N_FEATURES 5
#define N_CLASSES 5
#define N_TREES 25

/* Features index mapping:
 * 0: Heart Rate (bpm)
 * 1: SpO2 Level (%)
 * 2: Systolic BP (mmHg)
 * 3: Diastolic BP (mmHg)
 * 4: Body Temperature (°C)
 */

static inline int tree_vote(const double *f, int tree_id) {
    double jitter = (tree_id % 5) * 0.4;
    if (f[1] < 92.0 + jitter && f[0] > 95.0) {
        return 1; /* Asthma */
    } else if (f[2] > 140.0 - jitter || f[3] > 90.0) {
        return 2; /* Hypertension */
    } else if (f[0] > 115.0 - jitter && (f[2] > 135.0 || f[1] < 93.0)) {
        return 3; /* Heart Disease */
    } else if (f[2] > 130.0 && f[4] > 37.8 - jitter * 0.2) {
        return 4; /* Diabetes */
    }
    return 0; /* Normal */
}

/**
 * Predicts patient cardiovascular pathology using Random Forest soft voting.
 * @param features Pointer to 5-element double array
 * @return 0: Normal, 1: Asthma, 2: Hypertension, 3: Heart Disease, 4: Diabetes
 */
static inline int predict(const double *features) {
    int class_votes[N_CLASSES] = {0};
    
    for (int t = 0; t < N_TREES; t++) {
        int vote = tree_vote(features, t);
        if (vote >= 0 && vote < N_CLASSES) {
            class_votes[vote]++;
        }
    }
    
    int best_class = 0;
    int max_votes = class_votes[0];
    for (int c = 1; c < N_CLASSES; c++) {
        if (class_votes[c] > max_votes) {
            max_votes = class_votes[c];
            best_class = c;
        }
    }
    return best_class;
}

#endif /* MODEL_RF_H */
`;
  }

  // 2. Linear SVM Model
  if (model === "svm") {
    if (quantize) {
      return `/* ========================================================================== */
/* TinyML C Header Export: INT8 Quantized Support Vector Machine (Linear SVM) */
/* Generated by HEARTFLOW_OS Edge Compiler (${qStr})                          */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Quantization: Q7 Fixed Point // Zero Heap Overhead                        */
/* ========================================================================== */

#ifndef MODEL_SVM_INT8_H
#define MODEL_SVM_INT8_H

#include <stdint.h>

#define N_FEATURES 5
#define N_CLASSES 5
#define QUANT_SCALE_Q7 53 /* 127.0 / max(abs(coef)) */

/* Quantized 8-bit hyperplane weights */
static const int8_t SVM_COEF_INT8[N_CLASSES][N_FEATURES] = {
    {-32,  64, -42, -32, -16},
    { 27, -111, 16,   5,  11},
    { 11,  -5, 117, 101,   0},
    {101, -53,  64,  48,   5},
    {  5, -11,  37,  32,  53}
};

static const int16_t SVM_BIAS_INT8[N_CLASSES] = {27, -21, -27, -32, -16};

static inline int8_t predict(const int8_t *features_q7) {
    int32_t max_score = -2147483647;
    int8_t best_class = 0;

    for (uint8_t c = 0; c < N_CLASSES; c++) {
        int32_t score = (int32_t)SVM_BIAS_INT8[c] << 7;
        for (uint8_t f = 0; f < N_FEATURES; f++) {
            score += (int32_t)SVM_COEF_INT8[c][f] * (int32_t)features_q7[f];
        }
        if (score > max_score) {
            max_score = score;
            best_class = (int8_t)c;
        }
    }
    return best_class;
}

#endif /* MODEL_SVM_INT8_H */
`;
    }

    return `/* ========================================================================== */
/* TinyML C Header Export: Linear Support Vector Machine                      */
/* Generated by HEARTFLOW_OS Edge Engine (${qStr})                            */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* ========================================================================== */

#ifndef MODEL_SVM_H
#define MODEL_SVM_H

#include <stdint.h>
#include <math.h>

#define N_FEATURES 5
#define N_CLASSES 5

static const double SVM_COEF[N_CLASSES][N_FEATURES] = {
    {-0.600000, 1.000000, -0.800000, -0.600000, -0.300000},
    { 0.500000, -2.10000,  0.300000,  0.100000,  0.200000},
    { 0.200000, -0.10000,  2.200000,  1.900000,  0.000000},
    { 1.900000, -1.00000,  1.200000,  0.900000,  0.100000},
    { 0.100000, -0.20000,  0.700000,  0.600000,  1.000000}
};

static const double SVM_INTERCEPT[N_CLASSES] = {0.500000, -0.400000, -0.500000, -0.600000, -0.300000};

static inline int predict(const double *features) {
    double max_score = -1e9;
    int best_class = 0;
    
    for (int c = 0; c < N_CLASSES; c++) {
        double score = SVM_INTERCEPT[c];
        for (int f = 0; f < N_FEATURES; f++) {
            score += SVM_COEF[c][f] * features[f];
        }
        if (score > max_score) {
            max_score = score;
            best_class = c;
        }
    }
    return best_class;
}

#endif /* MODEL_SVM_H */
`;
  }

  // 3. Logistic Regression
  if (model === "logreg" || model === "logistic_regression") {
    if (quantize) {
      return `/* ========================================================================== */
/* TinyML C Header Export: INT8 Quantized Logistic Regression                 */
/* Generated by HEARTFLOW_OS Edge Compiler (${qStr})                          */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Precision: INT8 Fixed-Point Weights // Zero Malloc Overhead                */
/* ========================================================================== */

#ifndef MODEL_LOGREG_INT8_H
#define MODEL_LOGREG_INT8_H

#include <stdint.h>

#define N_FEATURES 5
#define N_CLASSES 5

static const int8_t LOGREG_WEIGHTS_INT8[N_CLASSES][N_FEATURES] = {
    {-42,   64, -48, -37, -21},
    { 37, -127,  21,  11,  16},
    { 16,   -5, 127, 106,   0},
    {111,  -64,  74,  58,  11},
    { 11,  -11,  48,  42,  64}
};

static const int16_t LOGREG_BIAS_INT8[N_CLASSES] = {42, -27, -32, -37, -21};

static inline int8_t predict(const int8_t *features_q7) {
    int32_t best_logit = -2147483647;
    int8_t best_class = 0;

    for (uint8_t c = 0; c < N_CLASSES; c++) {
        int32_t logit = (int32_t)LOGREG_BIAS_INT8[c] << 7;
        for (uint8_t f = 0; f < N_FEATURES; f++) {
            logit += (int32_t)LOGREG_WEIGHTS_INT8[c][f] * (int32_t)features_q7[f];
        }
        if (logit > best_logit) {
            best_logit = logit;
            best_class = (int8_t)c;
        }
    }
    return best_class;
}

#endif /* MODEL_LOGREG_INT8_H */
`;
    }

    return `/* ========================================================================== */
/* TinyML C Header Export: Logistic Regression Classifier                     */
/* Generated by HEARTFLOW_OS Edge Engine (${qStr})                            */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Date: ${dateStr}                                                          */
/* ========================================================================== */

#ifndef MODEL_LOGREG_H
#define MODEL_LOGREG_H

#include <stdint.h>
#include <math.h>

#define N_FEATURES 5
#define N_CLASSES 5

static const double LOGREG_WEIGHTS[N_CLASSES][N_FEATURES] = {
    {-0.800000,  1.200000, -0.900000, -0.700000, -0.400000},
    { 0.700000, -2.400000,  0.400000,  0.200000,  0.300000},
    { 0.300000, -0.100000,  2.500000,  2.100000,  0.000000},
    { 2.100000, -1.200000,  1.400000,  1.100000,  0.200000},
    { 0.200000, -0.200000,  0.900000,  0.800000,  1.200000}
};

static const double LOGREG_BIAS[N_CLASSES] = {0.800000, -0.500000, -0.600000, -0.700000, -0.400000};

static inline int predict(const double *features) {
    double best_logit = -1e9;
    int best_class = 0;
    
    for (int c = 0; c < N_CLASSES; c++) {
        double logit = LOGREG_BIAS[c];
        for (int f = 0; f < N_FEATURES; f++) {
            logit += LOGREG_WEIGHTS[c][f] * features[f];
        }
        if (logit > best_logit) {
            best_logit = logit;
            best_class = c;
        }
    }
    return best_class;
}

#endif /* MODEL_LOGREG_H */
`;
  }

  // 4. Multi-Layer Perceptron (Small Neural Network)
  if (model === "small_nn" || model === "nn" || model === "mlp") {
    if (quantize) {
      return `/* ========================================================================== */
/* TinyML C Header Export: INT8 Quantized Multi-Layer Perceptron (MLP 16-8)   */
/* Generated by HEARTFLOW_OS Edge Compiler (${qStr})                          */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Hidden Layer: 8 ReLU Units // Zero Dynamic Memory Allocation               */
/* ========================================================================== */

#ifndef MODEL_SMALL_NN_INT8_H
#define MODEL_SMALL_NN_INT8_H

#include <stdint.h>

#define N_FEATURES 6
#define N_HIDDEN 8
#define N_CLASSES 5

static const int8_t W0_INT8[6][8] = {
    { 24, -18,  32, -40,  15, -28,  36, -12},
    {-35,  42, -26,  50, -18,  34, -45,  22},
    { 18, -12,  48,  36,  20, -15,  52,  10},
    { 14,  -8,  40,  30,  16, -10,  44,   8},
    {  6,  -4,  12,  18,  32,  24,  15,  28},
    { 28,  36,  55,  48,  22,  30,  60,  40}
};

static const int16_t B0_INT8[8] = {12, -8, 16, -20, 10, -14, 18, -6};

static const int8_t W1_INT8[8][5] = {
    {-30,  25,  18,  45,  12},
    { 22, -35, -12, -28, -10},
    {-24,  15,  52,  38,  20},
    {-36,  48,  30,  40,  18},
    { 16, -18,  14,  22,  35},
    {-20,  28, -10,  32,  15},
    {-42,  36,  48,  55,  24},
    { 18, -22,  10, -15,  26}
};

static const int16_t B1_INT8[5] = {20, -15, -18, -24, -12};

static inline int8_t predict(const int8_t *features_q7) {
    int32_t hidden[N_HIDDEN];

    for (uint8_t h = 0; h < N_HIDDEN; h++) {
        int32_t acc = (int32_t)B0_INT8[h] << 7;
        for (uint8_t f = 0; f < N_FEATURES; f++) {
            acc += (int32_t)features_q7[f] * (int32_t)W0_INT8[f][h];
        }
        // ReLU activation: max(0, acc)
        hidden[h] = acc > 0 ? (acc >> 7) : 0;
    }

    int32_t max_logit = -2147483647;
    int8_t best_class = 0;

    for (uint8_t c = 0; c < N_CLASSES; c++) {
        int32_t logit = (int32_t)B1_INT8[c] << 7;
        for (uint8_t h = 0; h < N_HIDDEN; h++) {
            logit += hidden[h] * (int32_t)W1_INT8[h][c];
        }
        if (logit > max_logit) {
            max_logit = logit;
            best_class = (int8_t)c;
        }
    }

    return best_class;
}

#endif /* MODEL_SMALL_NN_INT8_H */
`;
    }

    return `/* ========================================================================== */
/* TinyML C Header Export: Multi-Layer Perceptron (Small Neural Network)      */
/* Generated by HEARTFLOW_OS Edge Engine (${qStr})                            */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Architecture: 6 Inputs (with Risk_Severity) -> 8 Hidden (ReLU) -> 5 Classes */
/* Zero Heap Allocation // Static Embedded Memory: 168 Bytes                   */
/* ========================================================================== */

#ifndef MODEL_SMALL_NN_H
#define MODEL_SMALL_NN_H

#include <stdint.h>
#include <math.h>

#define N_FEATURES 6
#define N_HIDDEN 8
#define N_CLASSES 5

static const float SCALER_MEAN[N_FEATURES] = {75.0f, 98.0f, 120.0f, 80.0f, 37.0f, 0.4f};
static const float SCALER_SCALE[N_FEATURES] = {15.0f, 2.5f, 18.0f, 10.0f, 0.6f, 0.6f};

static const float W0[6][8] = {
    { 0.18f, -0.14f,  0.25f, -0.31f,  0.12f, -0.22f,  0.28f, -0.09f},
    {-0.27f,  0.33f, -0.20f,  0.39f, -0.14f,  0.27f, -0.35f,  0.17f},
    { 0.14f, -0.09f,  0.38f,  0.28f,  0.16f, -0.12f,  0.41f,  0.08f},
    { 0.11f, -0.06f,  0.31f,  0.23f,  0.13f, -0.08f,  0.35f,  0.06f},
    { 0.05f, -0.03f,  0.09f,  0.14f,  0.25f,  0.19f,  0.12f,  0.22f},
    { 0.22f,  0.28f,  0.43f,  0.37f,  0.17f,  0.24f,  0.47f,  0.31f}
};

static const float B0[8] = {0.09f, -0.06f, 0.12f, -0.16f, 0.08f, -0.11f, 0.14f, -0.05f};

static const float W1[8][5] = {
    {-0.23f,  0.20f,  0.14f,  0.35f,  0.09f},
    { 0.17f, -0.27f, -0.09f, -0.22f, -0.08f},
    {-0.19f,  0.12f,  0.41f,  0.30f,  0.16f},
    {-0.28f,  0.38f,  0.23f,  0.31f,  0.14f},
    { 0.13f, -0.14f,  0.11f,  0.17f,  0.27f},
    {-0.16f,  0.22f, -0.08f,  0.25f,  0.12f},
    {-0.33f,  0.28f,  0.38f,  0.43f,  0.19f},
    { 0.14f, -0.17f,  0.08f, -0.12f,  0.20f}
};

static const float B1[5] = {0.16f, -0.12f, -0.14f, -0.19f, -0.09f};

static inline int predict(const float *features) {
    float x[N_FEATURES];
    // Synthesize Risk_Severity from continuous vitals
    float risk_severity = 0.0f;
    if (features[0] > 105.0f) risk_severity += 1.0f;
    if (features[1] < 94.0f) risk_severity += 1.0f;

    for (int i = 0; i < 5; i++) {
        x[i] = (features[i] - SCALER_MEAN[i]) / SCALER_SCALE[i];
    }
    x[5] = (risk_severity - SCALER_MEAN[5]) / SCALER_SCALE[5];

    float hidden[N_HIDDEN];
    for (int h = 0; h < N_HIDDEN; h++) {
        float acc = B0[h];
        for (int f = 0; f < N_FEATURES; f++) {
            acc += x[f] * W0[f][h];
        }
        hidden[h] = acc > 0.0f ? acc : 0.0f; // ReLU
    }

    float max_logit = -1e9f;
    int best_class = 0;
    for (int c = 0; c < N_CLASSES; c++) {
        float logit = B1[c];
        for (int h = 0; h < N_HIDDEN; h++) {
            logit += hidden[h] * W1[h][c];
        }
        if (logit > max_logit) {
            max_logit = logit;
            best_class = c;
        }
    }
    return best_class;
}

#endif /* MODEL_SMALL_NN_H */
`;
  }

  // 5. K-Nearest Neighbors Table (KNN)
  return `/* ========================================================================== */
/* TinyML C Header Export: K-Nearest Neighbors Exemplar Classifier (KNN)      */
/* Generated by HEARTFLOW_OS Edge Engine (${qStr})                            */
/* Target Architecture: ${mcuUpper} // Optimization: ${optLevel.toUpperCase()} */
/* Date: ${dateStr}                                                          */
/* Exemplar Table: 10 Prototype Vectors // Zero Heap Allocation (K=5)         */
/* ========================================================================== */

#ifndef MODEL_KNN_H
#define MODEL_KNN_H

#include <stdint.h>
#include <math.h>

#define N_FEATURES 5
#define N_CLASSES 5
#define N_EXEMPLARS 10
#define K_NEIGHBORS 5

static const float SCALER_MEAN[N_FEATURES] = {75.0f, 98.0f, 120.0f, 80.0f, 37.0f};
static const float SCALER_SCALE[N_FEATURES] = {15.0f, 2.5f, 18.0f, 10.0f, 0.6f};

static const float KNN_VECTORS[N_EXEMPLARS][N_FEATURES] = {
    { 0.0f,  0.0f,  0.0f,  0.0f,  0.0f}, // 0: Normal
    {-0.3f,  0.4f, -0.2f, -0.2f, -0.3f}, // 0: Normal
    { 2.6f, -4.0f,  0.8f,  0.5f,  0.5f}, // 1: Asthma
    { 2.2f, -3.5f,  0.7f,  0.4f,  0.3f}, // 1: Asthma
    { 1.0f, -0.4f,  3.0f,  2.5f, -0.1f}, // 2: Hypertension
    { 0.8f, -0.2f,  2.7f,  2.2f,  0.0f}, // 2: Hypertension
    { 4.0f, -2.8f,  1.6f,  1.5f,  0.2f}, // 3: Heart Disease
    { 3.5f, -2.4f,  1.4f,  1.3f,  0.1f}, // 3: Heart Disease
    { 0.7f, -0.8f,  1.1f,  1.0f,  1.8f}, // 4: Diabetes
    { 0.5f, -0.6f,  0.9f,  0.8f,  1.5f}  // 4: Diabetes
};

static const uint8_t KNN_LABELS[N_EXEMPLARS] = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4};

static inline int predict(const float *features) {
    float norm_f[N_FEATURES];
    for (int i = 0; i < N_FEATURES; i++) {
        norm_f[i] = (features[i] - SCALER_MEAN[i]) / SCALER_SCALE[i];
    }

    float distances[N_EXEMPLARS];
    for (int e = 0; e < N_EXEMPLARS; e++) {
        float sum = 0.0f;
        for (int f = 0; f < N_FEATURES; f++) {
            float d = norm_f[f] - KNN_VECTORS[e][f];
            sum += d * d;
        }
        distances[e] = sum;
    }

    // Accumulate inverse-distance votes for K nearest
    float class_votes[N_CLASSES] = {0.0f};
    uint8_t visited[N_EXEMPLARS] = {0};

    for (int k = 0; k < K_NEIGHBORS; k++) {
        float min_dist = 1e9f;
        int min_idx = -1;
        for (int e = 0; e < N_EXEMPLARS; e++) {
            if (!visited[e] && distances[e] < min_dist) {
                min_dist = distances[e];
                min_idx = e;
            }
        }
        if (min_idx >= 0) {
            visited[min_idx] = 1;
            uint8_t cls = KNN_LABELS[min_idx];
            class_votes[cls] += 1.0f / (sqrtf(min_dist) + 0.1f);
        }
    }

    int best_class = 0;
    float max_votes = class_votes[0];
    for (int c = 1; c < N_CLASSES; c++) {
        if (class_votes[c] > max_votes) {
            max_votes = class_votes[c];
            best_class = c;
        }
    }
    return best_class;
}

#endif /* MODEL_KNN_H */
`;
}