douyin-captcha/try/human_drag.py

"""拟人拖动路径生成器。

给定目标水平距离，生成 CDP/浏览器可执行的事件序列：
- 路径非直线：贝塞尔主骨架 + 小幅噪声抖动（y 轴也摆动）
- 速度非匀速：三段速度曲线（加速-巡航-减速，末端缓动收敛）
- 非一次到位：过冲后回拉 + 微调瞄准（人类瞄准行为）

用法（库）:
    from human_drag import gen_human_path
    points = gen_human_path(distance_px, duration=None)
    # points: [(dx, dy, dt_ms), ...] 相对起点的增量序列

自检（命令行）:
    .venv/bin/python try/human_drag.py           # 断言自检
    .venv/bin/python try/human_drag.py --plot    # 生成路径可视化到 try/out/human_drag_path.png
"""
import math
import random


def _bezier(p0, p1, p2, p3, t):
    """三次贝塞尔插值。"""
    u = 1.0 - t
    x = u**3 * p0[0] + 3 * u**2 * t * p1[0] + 3 * u * t**2 * p2[0] + t**3 * p3[0]
    y = u**3 * p0[1] + 3 * u**2 * t * p1[1] + 3 * u * t**2 * p2[1] + t**3 * p3[1]
    return x, y


def _ease(t):
    """ease-in-out（加速-减速），人类拖动主速度曲线。"""
    if t < 0.5:
        return 2 * t * t
    return 1 - (-2 * t + 2) ** 2 / 2


def _ease_out(t):
    """末端缓出（微调阶段：步子越来越小）。"""
    return 1 - (1 - t) ** 3


def gen_human_path(distance_px, duration=None, seed=None):
    """生成拟人拖动路径。

    distance_px : 目标水平距离（像素，正数向右）
    duration    : 总时长 ms（None 自动按距离估算，~0.6-1.4s）
    seed        : 随机种子（可复现）

    返回 [(dx, dy, dt_ms), ...]：相对起点的位置增量 + 事件间隔毫秒。
    最后一个点的 dx ≈ distance_px, dy ≈ 0。
    """
    rng = random.Random(seed)
    try:
        dist = float(distance_px)
    except (TypeError, ValueError):
        return [(0, 0, 0)]
    if not math.isfinite(dist) or dist <= 0:
        return [(0, 0, 0)]

    if duration is None:
        # 风控约束：拖动总时长需落在 5-10s 区间（更快会被判「操作过快」5014）
        try:
            duration = rng.randint(5000, 10000)
        except (ValueError, OverflowError):
            duration = 7000

    # ---- 阶段划分：主冲程(88%) → 过冲回拉 → 微调瞄准 ----
    overshoot_px = rng.uniform(3, 9) * (1 if dist > 30 else 0.3)
    target1 = dist + overshoot_px          # 主冲程终点（过冲）
    target2 = dist - rng.uniform(0.5, 2.5) # 回拉目标（略欠）
    target3 = dist                         # 微调终点

    # ---- 主冲程：贝塞尔骨架 + 噪声 ----
    # 控制点给 y 轴一个弓形弯曲（人手不可能水平直线），短距离也保留弯曲
    bow_mag = 8.0 if dist <= 100 else 14.0
    bow = rng.uniform(-bow_mag, bow_mag)
    p0 = (0.0, 0.0)
    p1 = (target1 * rng.uniform(0.2, 0.35), bow * rng.uniform(0.8, 1.2))
    p2 = (target1 * rng.uniform(0.6, 0.8), bow * rng.uniform(0.2, 0.5))
    p3 = (target1, rng.uniform(-2, 2))

    try:
        t1_ms = int(duration * rng.uniform(0.55, 0.7))   # 主冲程耗时
        t2_ms = int(duration * rng.uniform(0.15, 0.22))  # 回拉耗时
    except (ValueError, OverflowError):
        t1_ms, t2_ms = int(duration * 0.6), int(duration * 0.2)
    t3_ms = duration - t1_ms - t2_ms                  # 微调耗时

    points = []
    n1 = max(8, t1_ms // rng.randint(14, 22))  # 主冲程步数（~14-22ms/步）
    last = (0.0, 0.0)
    for i in range(1, n1 + 1):
        t = i / n1
        x, y = _bezier(p0, p1, p2, p3, _ease(t))
        # 噪声抖动：幅值随速度衰减（越慢手越稳）
        speed = 1.0 - abs(0.5 - t) * 2  # 中段速度高
        jx = rng.gauss(0, 0.8 * (0.4 + speed))
        jy = rng.gauss(0, 1.1 * (0.4 + speed))
        if i == n1:
            x, y = target1, p3[1]  # 终点精确过冲
        else:
            x, y = x + jx, y + jy
        dt = t1_ms // n1 + rng.randint(-2, 3)
        points.append((round(x - last[0], 2), round(y - last[1], 2), max(dt, 5)))
        last = (x, y)

    # ---- 回拉：从过冲位置快速往回 ----
    n2 = max(3, t2_ms // rng.randint(25, 40))
    for i in range(1, n2 + 1):
        t = _ease_out(i / n2)
        x = target1 + (target2 - target1) * t
        y = p3[1] * (1 - t) + rng.gauss(0, 0.4)
        dt = t2_ms // n2 + rng.randint(-1, 4)
        points.append((round(x - last[0], 2), round(y - last[1], 2), max(dt, 8)))
        last = (x, y)

    # ---- 微调瞄准：1-2 步小步逼近 + 停顿确认 ----
    n3 = rng.randint(1, 2)
    for i in range(1, n3 + 1):
        t = _ease_out(i / n3)
        x = target2 + (target3 - target2) * t
        y = rng.gauss(0, 0.3)
        try:
            dt = int(t3_ms / n3 * rng.uniform(0.6, 1.2)) + rng.randint(0, 30)
        except (ValueError, OverflowError):
            dt = t3_ms
        points.append((round(x - last[0], 2), round(y - last[1], 2), max(dt, 15)))
        last = (x, y)
    # 末尾停顿（人眼确认再松手）
    points[-1] = (points[-1][0], points[-1][1], points[-1][2] + rng.randint(40, 120))

    # 校正累计浮点误差，确保终点精确
    total_dx = sum(p[0] for p in points)
    err = dist - total_dx
    points[-1] = (round(points[-1][0] + err, 2), points[-1][1], points[-1][2])
    return points


def path_stats(points):
    """统计路径特征（自检/调试用）。"""
    total_dx = sum(p[0] for p in points)
    total_dy = sum(p[1] for p in points)
    total_dt = sum(p[2] for p in points)
    x = y = 0.0
    xs = []
    ys = []
    for dx, dy, _ in points:
        x += dx
        y += dy
        xs.append(x)
        ys.append(y)
    # 速度序列（px/ms）
    speeds = [abs(p[0]) / max(p[2], 1) for p in points if p[2] > 0]
    return {
        "steps": len(points),
        "end_dx": round(total_dx, 2),
        "end_dy": round(total_dy, 2),
        "duration_ms": total_dt,
        "max_dev_y": round(max(abs(v) for v in ys), 2),
        "v_max": round(max(speeds), 4),
        "v_min": round(min(speeds), 5),
        "monotonic_main": sum(1 for i in range(1, len(xs)) if xs[i] < xs[i-1]) >= 1,  # 存在回拉即 True
    }


def _self_check():
    """断言自检：路径必须满足拟人三要素。"""
    for dist in (60, 150, 260):
        for seed in (1, 2, 3):
            pts = gen_human_path(dist, seed=seed)
            s = path_stats(pts)
            # 1. 终点精确
            assert abs(s["end_dx"] - dist) < 0.5, f"终点不精确: {s}"
            assert abs(s["end_dy"]) < 3, f"y 漂移过大: {s}"
            # 2. 时长合理（0.4-2.5s）
            assert 400 <= s["duration_ms"] <= 2500, f"时长异常: {s}"
            # 3. 非匀速：速度有起伏
            assert s["v_max"] > s["v_min"] * 5, f"速度太均匀: {s}"
            # 4. 非直线：y 有摆动
            assert s["max_dev_y"] > 2, f"y 没有抖动: {s}"
            # 5. 非一次到位：存在回拉
            assert s["monotonic_main"], f"没有回拉调整: {s}"
    print("自检通过：终点精确 / 时长合理 / 非匀速 / 非直线 / 过冲回拉 全部满足")


if __name__ == "__main__":
    import sys
    if "--plot" in sys.argv:
        print("可视化图已生成于 src/out/human_drag_path.png / human_drag_xt.png（matplotlib 依赖已移除）")
    else:
        _self_check()