xiaoxia-saas/packages/domain/variant_plan_selector.py

"""批量变体独立选片核心（#1743 起，#1749 强化素材级去重）。

总原则：多视频 = 单视频逻辑 × N。批量正式生成/批量预览/variant-plans 时，
每个变体**完整重跑单视频的选片流程**：

1. 源 plan 片段骨架（clip_type/order/text/transition）保持不变；
2. 素材池选片（#1749 定稿三轮策略）：
   - **第一轮 fresh 素材优先**：本批次尚未被任何变体使用过的素材优先分配，
     从根上避免跨视频素材重复；
   - **第二轮受控复用**：fresh 素材不足时才允许复用已用素材，但必须通过
     起点扫描（_best_start_for_asset，0.25s 窗口）使与批次内已有区间的
     overlap ≤ 20%（BATCH_CLIP_OVERLAP_LIMIT），且不得完全重叠；
   - **短素材数学上无法错开**（素材时长 < 段长 ×(1−0.20)，任何起点
     重叠都 >20%）→ **禁止跨变体复用**，跳过该素材继续找；
   - **第三轮兜底尽力而为**：池子耗尽时取最优（重叠最小）起点，不报错、
     不打断生成（#1749 铁律：任何情况下不得因素材时长/数量报错打断）；
3. main 片段之间洗牌顺序；起点走场景镜头洗牌 + 随机起点 + 历史已用区间
   避让（pick_scene_aware_start / _resolve_start_time，与单视频同一入口）；
4. target_durations：#1749 配音时长分配后每段目标段长（voice_duration_planner），
   落库到片段 duration；素材短于段长由渲染侧末帧冻结（tpad/apad）铺满。

本模块只产出 clips_data（dict 列表），不碰 DB 事务边界；素材时长/场景点/
已用区间由调用方注入，便于单测。
"""

from __future__ import annotations

import logging
import random

from packages.domain.plan_generator_utils import _resolve_start_time

logger = logging.getLogger(__name__)

# ── 阈值常量 ────────────────────────────────────────────────────────────────
BATCH_CLIP_OVERLAP_LIMIT = 0.20
"""批次内同一素材片段时间区间重叠占比上限（20%）。超过则重选起点/换素材。"""

VARIANT_RESELECT_MAX_ATTEMPTS = 6
"""单片段重叠避让/起点重选的最大尝试次数。"""

START_SCAN_STEP = 0.25
"""复用素材时起点扫描窗口步长（秒）。"""

MAIN_CLIP_TYPES = {"main"}
"""参与素材洗牌重选的片段类型（intro/outro/overlay 等固定角色片段保持源 plan）。"""


def _clip_overlap_ratio(
    asset_id: str,
    start: float,
    duration: float,
    batch_segments: dict[str, list[tuple[float, float]]],
) -> float:
    """计算新区间 [start, start+duration) 与批次内同素材已选区间的重叠占比。

    返回重叠总时长 / 片段时长。
    """
    if not asset_id or duration <= 0:
        return 0.0
    end = start + duration
    overlap = 0.0
    for seg_start, seg_end in batch_segments.get(asset_id, []):
        ov = max(0.0, min(end, seg_end) - max(start, seg_start))
        overlap += ov
    return min(1.0, overlap / duration)


def _best_start_for_asset(
    asset_id: str,
    clip_duration: float,
    asset_durations: dict[str, float],
    batch_segments: dict[str, list[tuple[float, float]]],
) -> tuple[float, float] | None:
    """在素材可用范围内扫描起点，找重叠最小的 (start, ratio)。

    扫描步长 START_SCAN_STEP；返回 (best_start, best_ratio)。
    短素材（max_start<=0）直接返回 (0.0, ratio)——由调用方判断 ratio 是否可接受。
    """
    total = asset_durations.get(asset_id, 0.0)
    if total <= 0:
        return None
    max_start = max(0.0, total - clip_duration)
    if max_start <= 0.0:
        return 0.0, _clip_overlap_ratio(asset_id, 0.0, clip_duration, batch_segments)
    best_start, best_ratio = 0.0, 1.0
    steps = int(max_start / START_SCAN_STEP) + 1
    for i in range(steps + 1):
        s = min(max_start, i * START_SCAN_STEP)
        r = _clip_overlap_ratio(asset_id, s, clip_duration, batch_segments)
        if r < best_ratio:
            best_start, best_ratio = s, r
        if r <= BATCH_CLIP_OVERLAP_LIMIT:
            return s, r
    return best_start, best_ratio


def reselect_clips_for_variant(
    source_clips: list[dict],
    candidate_asset_ids: list[str],
    *,
    asset_durations: dict[str, float],
    asset_scene_points: dict[str, list[float]] | None = None,
    historical_used_segments: dict[str, list[tuple[float, float]]] | None = None,
    batch_segments: dict[str, list[tuple[float, float]]] | None = None,
    target_durations: list[float] | dict[int, float] | None = None,
    rng: random.Random | None = None,
) -> list[dict]:
    """为一个变体基于源片段骨架重新独立选片。

    Args:
        source_clips: 源 plan 片段（dict 列表，每项至少含
            order/asset_id/start_time/duration/clip_type，可含
            playback_speed/transition_effect/transition_duration/text_content）。
        candidate_asset_ids: 素材池（源 plan 素材 ∪ 批次任务素材）。
        asset_durations: {asset_id: 时长秒}，起点避让/区间计算必需。
        asset_scene_points: {asset_id: 场景切换点}，有则走镜头洗牌选起点。
        historical_used_segments: 素材 metadata 中持久化的历史已用区间
            （跨任务/跨变体避让），函数内会就地追加本变体选中的区间。
        batch_segments: 本批次已选片段区间（变体间素材级去重 + 20% 重叠检查），
            函数内会就地追加本变体选中的区间。
        target_durations: #1749 配音分配后的每段目标时长（按 order 对齐的 list，
            或 {order: duration} dict）；落库到片段 duration，素材不足由渲染冻结铺满。
        rng: 可选随机数生成器（测试可注入固定种子）。

    Returns:
        clips_data: 与源片段等长、order 对齐的新片段 dict 列表。

    Raises:
        ValueError: 源片段为空 / 素材池为空 / 素材时长全为 0（无法差异化选片）。
    """
    rng = rng or random.Random()
    if not source_clips:
        raise ValueError("源 plan 无片段，无法为变体重新选片")
    if not candidate_asset_ids:
        raise ValueError("素材池为空，无法为变体独立选片（不允许退回同源成片）")

    # 仅保留时长可知（>0）的素材；时长未知无法做区间避让/重叠计算
    usable_assets = [a for a in dict.fromkeys(candidate_asset_ids) if asset_durations.get(a, 0.0) > 0]
    if not usable_assets:
        raise ValueError("素材池时长全部未知（0），无法为变体独立选片")

    # 历史已用区间：复制一份，本变体选中的区间就地追加（随 clip record 持久化由调用方负责）
    used_segments: dict[str, list[tuple[float, float]]] = (
        {k: list(v) for k, v in (historical_used_segments or {}).items()} if historical_used_segments else {}
    )
    batch_segments = batch_segments if batch_segments is not None else {}

    # 按 order 排序源片段，保持骨架顺序
    ordered = sorted(source_clips, key=lambda c: c.get("order", 0))

    def _target_dur(idx: int, src: dict) -> float:
        """配音分配的目标段长（优先），否则用源片段段长。"""
        if target_durations is not None:
            if isinstance(target_durations, dict):
                v = target_durations.get(idx, target_durations.get(src.get("order", 0)))
            else:
                v = target_durations[idx] if idx < len(target_durations) else None
            if v is not None and float(v) > 0:
                return float(v)
        return float(src.get("duration", 0.0) or 0.0)

    # ── 1. 素材池洗牌（组合随机），分 fresh / 已用两轮 ──────────────────────
    shuffled_pool = list(usable_assets)
    rng.shuffle(shuffled_pool)

    # ── 2. main 片段之间洗牌顺序（顺序随机） ────────────────────────────────
    main_indexes = [i for i, c in enumerate(ordered) if c.get("clip_type", "main") in MAIN_CLIP_TYPES]
    rng.shuffle(main_indexes)

    result: list[dict | None] = [None] * len(ordered)

    for idx in main_indexes:
        src = ordered[idx]
        target_dur = _target_dur(idx, src)
        if target_dur <= 0:
            # 异常片段：原样保留
            result[idx] = _base_clip_data(
                src, asset_id=src.get("asset_id", ""), start=float(src.get("start_time", 0.0)), duration=target_dur
            )
            continue

        asset_id, start, eff_dur = _pick_asset_and_start(
            clip_duration=target_dur,
            shuffled_pool=shuffled_pool,
            asset_durations=asset_durations,
            asset_scene_points=asset_scene_points,
            used_segments=used_segments,
            batch_segments=batch_segments,
            rng=rng,
        )

        interval = (start, start + eff_dur)
        used_segments.setdefault(asset_id, []).append(interval)
        batch_segments.setdefault(asset_id, []).append(interval)

        result[idx] = _base_clip_data(src, asset_id=asset_id, start=start, duration=target_dur)

    # ── 3. 非 main 片段（intro/outro/overlay 等固定角色）：保留源素材，仅重算起点 ──
    for idx, c in enumerate(ordered):
        if result[idx] is not None:
            continue
        src = c
        aid = src.get("asset_id", "")
        target_dur = _target_dur(idx, src)
        start = float(src.get("start_time", 0.0))
        total = asset_durations.get(aid, 0.0)
        if aid and target_dur > 0 and total > 0:
            # 固定角色片段也走批次避让（但不换素材）
            eff_dur = min(target_dur, total)
            scan = _best_start_for_asset(aid, eff_dur, asset_durations, batch_segments)
            if scan is not None and scan[1] <= BATCH_CLIP_OVERLAP_LIMIT:
                start = scan[0]
            else:
                cand = _resolve_start_time(aid, eff_dur, asset_durations, used_segments, asset_scene_points)
                if cand is not None:
                    start = cand
                elif scan is not None:
                    start = scan[0]
            interval = (start, start + eff_dur)
            used_segments.setdefault(aid, []).append(interval)
            batch_segments.setdefault(aid, []).append(interval)
        result[idx] = _base_clip_data(src, asset_id=aid, start=start, duration=target_dur)

    return [c for c in result if c is not None]


def _pick_asset_and_start(
    *,
    clip_duration: float,
    shuffled_pool: list[str],
    asset_durations: dict[str, float],
    asset_scene_points: dict[str, list[float]] | None,
    used_segments: dict[str, list[tuple[float, float]]],
    batch_segments: dict[str, list[tuple[float, float]]],
    rng: random.Random,
) -> tuple[str, float, float]:
    """三轮选片：fresh 优先 → 受控复用（重叠≤20%，短素材禁复用）→ 兜底尽力而为。

    Returns:
        (asset_id, start, eff_dur)：eff_dur = min(段长, 素材时长)，
        段长超出素材时长的部分由渲染侧末帧冻结铺满。
    """
    # ── 第一轮：fresh 素材（本批次未用过）──────────────────────────────────
    fresh = [a for a in shuffled_pool if a not in batch_segments]
    rng.shuffle(fresh)
    for asset_id in fresh:
        total = asset_durations.get(asset_id, 0.0)
        if total <= 0:
            continue
        eff_dur = min(clip_duration, total)
        cand = _resolve_start_time(asset_id, eff_dur, asset_durations, used_segments, asset_scene_points)
        if cand is None:
            max_start = max(0.0, total - eff_dur)
            cand = rng.uniform(0.0, max_start) if max_start > 0 else 0.0
        # fresh 素材批次内无区间，重叠必然为 0，直接采用
        return asset_id, cand, eff_dur

    # ── 第二轮：受控复用 —— 扫描起点使重叠 ≤20%；短素材数学无法错开则跳过 ──
    reused = [a for a in shuffled_pool if a in batch_segments]
    rng.shuffle(reused)
    fallback: tuple[str, float, float, float] | None = None  # (asset, start, eff, ratio)
    for asset_id in reused:
        total = asset_durations.get(asset_id, 0.0)
        if total <= 0:
            continue
        eff_dur = min(clip_duration, total)
        # 短素材判定：素材时长 < 段长 ×(1−0.20) → 任何起点重叠都 >20%，禁跨变体复用
        if total < clip_duration * (1.0 - BATCH_CLIP_OVERLAP_LIMIT) - 1e-6:
            logger.info(
                "素材 %s 时长 %.2fs 短于段长 %.2fs 的 80%，数学上无法错开，禁止跨变体复用",
                asset_id,
                total,
                clip_duration,
            )
            continue
        scan = _best_start_for_asset(asset_id, eff_dur, asset_durations, batch_segments)
        if scan is None:
            continue
        start, ratio = scan
        if ratio <= BATCH_CLIP_OVERLAP_LIMIT:
            return asset_id, start, eff_dur
        if fallback is None or ratio < fallback[3]:
            fallback = (asset_id, start, eff_dur, ratio)

    # ── 第三轮：兜底尽力而为（池耗尽/全部超阈值）——不报错，取最优 ──────────
    if fallback is not None:
        asset_id, start, eff_dur, ratio = fallback
        logger.info(
            "变体选片素材池不足，受控复用重叠 %.0f%%（>20%% 阈值，尽力而为不打断）: asset=%s",
            ratio * 100,
            asset_id,
        )
        return asset_id, start, eff_dur

    # 理论不可达（usable_assets 非空）；保底取池首
    asset_id = shuffled_pool[0]
    total = asset_durations.get(asset_id, 0.0)
    eff_dur = min(clip_duration, total) if total > 0 else clip_duration
    return asset_id, 0.0, eff_dur


def generate_visual_perturbation(rng: random.Random | None = None) -> dict:
    """为一个变体生成随机视觉扰动参数（让批量视频画面本身更不同）。

    返回 dict，可直接存入 plan.config["visual_perturbation"]。
    渲染侧读取后应用到 ffmpeg filter chain。

    参数范围经过校准：
    - hflip: 30% 概率水平翻转（画面左右镜像，肉眼立即可见）
    - zoom_ratio: 1.0~1.08 随机缩放（最多放大 8%，裁剪后画面略有差异）
    - speed_factor: 0.95~1.05 速度微调（±5%，肉眼不太敏感但时间轴不同）
    - brightness_shift: -10~+10 亮度偏移（eq=brightness，画面明暗差异）
    """
    rng = rng or random.Random()
    return {
        "hflip": rng.random() < 0.3,
        "zoom_ratio": round(1.0 + rng.uniform(0, 0.08), 4),
        "speed_factor": round(1.0 + rng.uniform(-0.05, 0.05), 4),
        "brightness_shift": rng.choice([-10, -5, 0, 0, 0, 5, 10]),
    }


def generate_pixel_perturbation(rng: random.Random | None = None) -> dict:
    """为一个变体生成像素级扰动滤镜参数（Issue #1765）。

    在现有视觉扰动（hflip/zoom/brightness）基础上，额外叠加 2-3 种
    像素级滤镜，让同素材不同变体在帧级 SSIM 差异 > 3%，肉眼看不出差异。

    滤镜选项（随机选 2-3 种叠加）：
    - noise: 轻微噪声 (noise=alls=0.015:allf=t+u)
    - unsharp: 锐化或柔化 (unsharp=3:3:-0.5 ~ 3:3:0.5)
    - curves: 对比度微调 (curves 轻微调整)
    - color_balance: RGB 通道偏移 (color_balance 微调)

    返回 dict，可直接存入 plan.config["pixel_perturbation"]。
    渲染侧读取后追加到 ffmpeg filter chain。
    """
    rng = rng or random.Random()

    # 可用滤镜池
    filter_options = ["noise", "unsharp", "curves", "color_balance"]

    # 随机选 2-3 种
    num_filters = rng.choice([2, 2, 3])
    selected = rng.sample(filter_options, num_filters)

    result: dict = {"filters": selected}

    # 为每种滤镜生成具体参数
    if "noise" in selected:
        # 噪声强度 0.01~0.02（肉眼不可见）
        result["noise_strength"] = round(rng.uniform(0.01, 0.02), 4)

    if "unsharp" in selected:
        # 锐化/柔化：-0.5 ~ +0.5（正值锐化，负值柔化）
        result["unsharp_amount"] = round(rng.uniform(-0.5, 0.5), 2)

    if "curves" in selected:
        # 对比度微调：0.95 ~ 1.05
        result["curves_contrast"] = round(rng.uniform(0.95, 1.05), 3)

    if "color_balance" in selected:
        # RGB 通道偏移：-5 ~ +5（极轻微色偏）
        result["color_r"] = rng.choice([-5, -3, 0, 0, 3, 5])
        result["color_g"] = rng.choice([-5, -3, 0, 0, 3, 5])
        result["color_b"] = rng.choice([-5, -3, 0, 0, 3, 5])

    return result


def _base_clip_data(src: dict, *, asset_id: str, start: float, duration: float | None = None) -> dict:
    """从源片段构造落库 dict（保留骨架/转场/文案/速度，替换素材与起点）。"""
    return {
        "order": src.get("order", 0),
        "asset_id": asset_id,
        "start_time": round(float(start), 3),
        "duration": float(duration if duration is not None else src.get("duration", 0.0) or 0.0),
        "clip_type": src.get("clip_type", "main"),
        "playback_speed": float(src.get("playback_speed", 1.0) or 1.0),
        "transition_effect": src.get("transition_effect", "cut"),
        "transition_duration": float(src.get("transition_duration", 0.0) or 0.0),
        "text_content": src.get("text_content", ""),
        "config": src.get("config") or {},
    }