暮归江南 · 步骤4-9：生命感、声音、播放器与MP4导出、封面文案、实况图、Blender脚本

.gitattributes

@@ -93,3 +93,5 @@ public/mvfonts/neon.woff2 filter=lfs diff=lfs merge=lfs -text
 素材/longshot20261021193931.png filter=lfs diff=lfs merge=lfs -text
 素材/longshot20261021193942.png filter=lfs diff=lfs merge=lfs -text
 public/mvfonts/lyric.woff2 filter=lfs diff=lfs merge=lfs -text
+src/jiangnan/assets/fonts/title.woff2 filter=lfs diff=lfs merge=lfs -text
+src/jiangnan/assets/fonts/wenkai.woff2 filter=lfs diff=lfs merge=lfs -text

README.md

@@ -119,6 +119,12 @@ tools/jiangnan/                     # 旁白生成、字体子集化、实况图
 - **步骤 1** ✅ 写好计划、大纲、步骤与做法（本节），推送到 Space。
 - **步骤 2** ✅ 渲染引擎骨架：新页面入口 `/jiangnan/`（`jiangnan/index.html` + `src/jiangnan/main.tsx`，`vite.config.ts` 增加多页面入口，原页面不变）；时间轴（10 个镜头、日落曲线、旁白时间）；程序化天空（夕阳、晚霞、云、蓝调、星空、新月）；山谷地形 + 远山 + 高精度梯田（水田镜面反光）；实时平面反射的河面（涟漪、阳光碎金、灯影拉丝）；高度雾与空气透视；后期（体积光、辉光、ACES、分色调色、暗角、颗粒）。
 - **步骤 3** ✅ 村落与道具：217 栋徽派民居生成器（粉墙、三叠/二叠马头墙、黛瓦凹曲屋面、腰檐、门罩、木格窗、店铺排门）；石驳岸、青石板河街、河埠头；半圆石拱桥（台阶、栏板、拱券石）；老水车（辐条、叶板、竹筒、水槽）；乌篷船；集市摊位（竹篮果蔬、陶罐、腊肉架、布篷）；村口古樟（板根、虬枝）、石碾、石凳、古井、蓝印花布晾架；六角亭；柳树、竹林、山林、芦苇。并建立了无头浏览器自动检查画面（构图分类图 + 亮度图）的流程。
+- **步骤 4** ✅ 生命感：挑担过桥的村民、集市行人与摊主、牛车（牛腿与车轮随行进转动）、树下歇脚的老人、7 只白鹭（拍翅）与 46 只归鸟群、按厨房点火时间升起的炊烟（26 团/烟囱，受风漂移）、蓝印花布随风摆动、夜里挂灯的乌篷船与摇橹人、红灯笼（从集市西头第一盏开始依次点亮，含点火闪烁）、窗户逐个亮灯、最近的灯笼变成真实点光源照亮粉墙；每个镜头都有横屏 / 竖屏两套机位。
+- **步骤 5** ✅ 声音：10 句原创旁白（中文神经网络语音合成，纪录片男声），配乐在浏览器内实时合成（古筝 Karplus-Strong、竹箫、弦乐铺底、低音、木鱼、编钟、日落与终章的古筝刮奏），环境音效（流水、水车吱呀与溅水、鸟鸣、白鹭叫声与振翅、牛铃与蹄声、脚步、竹林风声、远处钟声、灯笼点火、虫鸣、蛙鸣、摇橹水声）；离线混音（旁白时音乐自动压低、卷积混响、限幅）。实测响度：平均 -20 dB，峰值 -1.9 dB。
+- **步骤 6** ✅ 播放器与 MP4 导出：播放 / 暂停 / 拖动 / 分镜跳转 / 字幕 / 静音 / 全屏 / 横竖屏切换，自适应分辨率；导出可选横屏 1920×1080 或竖屏 1080×1920、1080p/720p、24/30/60 fps、烧录字幕与否；WebCodecs H.264 + AAC（浏览器不支持 AAC 时自动用 Opus）+ mp4-muxer。已在无头浏览器里实测导出并用 ffprobe/ffmpeg 校验通过。
+- **步骤 7** ✅ 封面与文案：16:9 / 4:3 / 3:4 三种比例 × 4 个画面（石桥夕照、灯火长街、星夜水乡、古樟斜阳），每张按比例重新取景渲染并排书法标题，可多选、单张或打包下载；小红书 / 抖音 / 视频号 / B站 各 3–4 个标题（字数实时校验，全部在限制内）、描述、话题，一键复制，也可下载 txt。
+- **步骤 8** ✅ 实况图：把 LiveCanvas skill 的 Python 配对器移植成浏览器端 TypeScript（`src/jiangnan/export/livephoto.ts`），网页里一键生成 6 组 3:4 / 30fps / 3 秒实况图，压缩包内含同名 jpg + mov、.pvt 资源包、manifest、使用说明与图文文案。用 skill 原版 `verify_pair` + `package_live_photo.py` 校验浏览器生成的文件：标识一致、still-image-time 读回正确、ffprobe 时间样本正确、完整解码通过。
+- **步骤 9** ✅ Blender 场景脚本 `blender/jiangnan_village.py`：已在沙盒里用 Blender 4.2 LTS（bpy）实际运行，生成 218 栋民居、178 盏灯笼（含关键帧点亮）、石拱桥、水车（旋转动画）、古樟、竹林、Nishita 天空 + 日落关键帧 + 蓝调天光、暮霭体积、10 个镜头（时间轴标记自动切换），并用 Cycles 试渲通过。
 
 
 ---

blender/jiangnan_village.py

@@ -0,0 +1,842 @@
+# -*- coding: utf-8 -*-
+"""
+《暮归江南》Blender 场景生成脚本  (Blender 4.0+ / 4.2 LTS / 4.x)
+
+用法：Blender → Scripting 工作区 → 打开本文件 → Run Script。
+脚本会清空当前场景，然后从零生成：
+  · 山谷地形、远山、梯田、蜿蜒小河（反射水面）、石驳岸、青石板街
+  · 徽派民居（粉墙、阶梯马头墙、黛瓦凹曲屋面、木格窗，窗户夜里亮灯）
+  · 半圆石拱桥、老水车（旋转动画）、乌篷船、集市摊位、村口古樟、石碾、六角亭、竹林
+  · 红灯笼（日落后按顺序逐个亮起）、炊烟（烟雾体积可选）
+  · Nishita 物理天空 + 太阳从 8.5° 落到 -10°（与网页版同一条日落曲线）、蓝调、星光
+  · 10 个镜头的摄像机与时间轴标记（自动切换镜头），24 fps，共 86 秒
+与网页版共用同一套布局算法（同一条河道、同一个随机种子），所以村子的样子一致。
+
+渲染：默认 Cycles（GPU 可在 偏好设置 → 系统 里开启），1920×1080，128 采样 + 降噪。
+想先快速预览，可把 RENDER_ENGINE 改成 'BLENDER_EEVEE_NEXT'（4.2+）或 'BLENDER_EEVEE'。
+"""
+import bpy
+import bmesh
+import math
+import random
+from mathutils import Vector, Matrix, Euler
+
+FPS = 24
+DURATION = 86.0
+RENDER_ENGINE = 'CYCLES'
+SEED = 20261022
+GROUND_Y = 1.7  # web units: y-up metres.  Blender is z-up: (x, y, z)_web -> (x, -z, y)_blender
+
+
+def W2B(x, y, z):
+    """web (x, y-up, z) -> blender (x, y, z-up)"""
+    return Vector((x, -z, y))
+
+
+# ------------------------------------------------------------------ deterministic noise (same as the web build)
+def mulberry32(seed):
+    a = seed & 0xFFFFFFFF
+
+    def rnd():
+        nonlocal a
+        a = (a + 0x6D2B79F5) & 0xFFFFFFFF
+        t = a
+        t = ((t ^ (t >> 15)) * (t | 1)) & 0xFFFFFFFF
+        t ^= (t + (((t ^ (t >> 7)) * (t | 61)) & 0xFFFFFFFF)) & 0xFFFFFFFF
+        return ((t ^ (t >> 14)) & 0xFFFFFFFF) / 4294967296.0
+    return rnd
+
+
+def _imul(a, b):
+    return ((a & 0xFFFFFFFF) * (b & 0xFFFFFFFF)) & 0xFFFFFFFF
+
+
+def hash2(x, y, seed=0):
+    h = (_imul(int(x), 374761393) + _imul(int(y), 668265263) + _imul(int(seed), 2246822519)) & 0xFFFFFFFF
+    h = _imul(h ^ (h >> 13), 1274126177)
+    h ^= h >> 16
+    return (h & 0xFFFFFFFF) / 4294967296.0
+
+
+def fade(t):
+    return t * t * (3 - 2 * t)
+
+
+def vnoise(x, y, seed=0):
+    xi, yi = math.floor(x), math.floor(y)
+    xf, yf = x - xi, y - yi
+    a, b = hash2(xi, yi, seed), hash2(xi + 1, yi, seed)
+    c, d = hash2(xi, yi + 1, seed), hash2(xi + 1, yi + 1, seed)
+    u, v = fade(xf), fade(yf)
+    return (a + (b - a) * u) * (1 - v) + (c + (d - c) * u) * v
+
+
+def fbm(x, y, octv=5, seed=0):
+    amp, f, s, n = 0.5, 1.0, 0.0, 0.0
+    for i in range(octv):
+        s += amp * vnoise(x * f, y * f, seed + i * 31)
+        n += amp
+        amp *= 0.5
+        f *= 2.03
+    return s / n
+
+
+def smoothstep(a, b, v):
+    t = min(1.0, max(0.0, (v - a) / (b - a)))
+    return t * t * (3 - 2 * t)
+
+
+def lerp(a, b, t):
+    return a + (b - a) * t
+
+
+# ------------------------------------------------------------------ river (same control points as layout.ts)
+RIVER_CTRL = [(-1500, -330), (-1050, -260), (-760, -150), (-520, -70), (-330, -22), (-190, 6), (-80, 14),
+              (30, 4), (120, -10), (220, 6), (330, 46), (470, 92), (660, 104), (900, 60), (1300, -40), (1700, -120)]
+
+
+def catmull(p0, p1, p2, p3, t):
+    t2, t3 = t * t, t * t * t
+    return 0.5 * (2 * p1 + (-p0 + p2) * t + (2 * p0 - 5 * p1 + 4 * p2 - p3) * t2 + (-p0 + 3 * p1 - 3 * p2 + p3) * t3)
+
+
+def build_river():
+    raw = []
+    n = len(RIVER_CTRL)
+    for i in range(n - 1):
+        p0, p1, p2, p3 = RIVER_CTRL[max(0, i - 1)], RIVER_CTRL[i], RIVER_CTRL[i + 1], RIVER_CTRL[min(n - 1, i + 2)]
+        for k in range(80):
+            t = k / 80
+            raw.append((catmull(p0[0], p1[0], p2[0], p3[0], t), catmull(p0[1], p1[1], p2[1], p3[1], t)))
+    raw.append(RIVER_CTRL[-1])
+    cum = [0.0]
+    for i in range(1, len(raw)):
+        cum.append(cum[-1] + math.hypot(raw[i][0] - raw[i - 1][0], raw[i][1] - raw[i - 1][1]))
+    out, j, s = [], 0, 0.0
+    while s <= cum[-1]:
+        while j < len(cum) - 2 and cum[j + 1] < s:
+            j += 1
+        f = (s - cum[j]) / max(1e-6, cum[j + 1] - cum[j])
+        out.append([lerp(raw[j][0], raw[j + 1][0], f), lerp(raw[j][1], raw[j + 1][1], f), 0.0, 0.0, s])
+        s += 2.0
+    for i in range(len(out)):
+        a, b = out[max(0, i - 2)], out[min(len(out) - 1, i + 2)]
+        l = math.hypot(b[0] - a[0], b[1] - a[1]) or 1
+        out[i][2] = (b[0] - a[0]) / l
+        out[i][3] = (b[1] - a[1]) / l
+    return out
+
+
+RIVER = build_river()
+
+
+def river_s_at_x(x):
+    return min(RIVER, key=lambda p: abs(p[0] - x))[4]
+
+
+def river_point(s, off):
+    f = min(max(s / 2.0, 0), len(RIVER) - 1.001)
+    i = int(f)
+    t = f - i
+    a, b = RIVER[i], RIVER[i + 1]
+    tx, tz = lerp(a[2], b[2], t), lerp(a[3], b[3], t)
+    l = math.hypot(tx, tz) or 1
+    tx, tz = tx / l, tz / l
+    nx, nz = -tz, tx
+    return lerp(a[0], b[0], t) + nx * off, lerp(a[1], b[1], t) + nz * off, tx, tz, nx, nz
+
+
+VILLAGE_S0 = river_s_at_x(-300)
+VILLAGE_S1 = river_s_at_x(290)
+
+
+def village_factor(s):
+    return smoothstep(VILLAGE_S0 - 30, VILLAGE_S0 + 10, s) * (1 - smoothstep(VILLAGE_S1 - 10, VILLAGE_S1 + 30, s))
+
+
+def river_half_width(s):
+    inv = smoothstep(VILLAGE_S0 - 120, VILLAGE_S0, s) * (1 - smoothstep(VILLAGE_S1, VILLAGE_S1 + 120, s))
+    return lerp(13 + 3 * vnoise(s * 0.01, 3), 8, inv)
+
+
+def river_query(x, z):
+    best, bi = 1e18, 0
+    for i in range(0, len(RIVER), 3):
+        p = RIVER[i]
+        d = (p[0] - x) ** 2 + (p[1] - z) ** 2
+        if d < best:
+            best, bi = d, i
+    for i in range(max(0, bi - 3), min(len(RIVER), bi + 4)):
+        p = RIVER[i]
+        d = (p[0] - x) ** 2 + (p[1] - z) ** 2
+        if d < best:
+            best, bi = d, i
+    return math.sqrt(best), RIVER[bi][4]
+
+
+def terrain_height(x, z):
+    d, s = river_query(x, z)
+    h = GROUND_Y + (fbm(x * 0.012, z * 0.012, 3, 5) - 0.5) * 1.6 * smoothstep(50, 140, d)
+    far = math.hypot(x * 0.8, z)
+    hd = min(d, far * 0.9 + 40)
+    hill = smoothstep(130, 560, hd)
+    h += hill * hill * (60 + 190 * fbm(x * 0.0022 + 3, z * 0.0022, 5, 7))
+    w = river_half_width(s)
+    if d < w + 30:
+        bank_w = lerp(9, 1.2, village_factor(s))
+        bank = smoothstep(w - bank_w * 0.6, w + bank_w * 0.4, d)
+        h = lerp(-2.6, h, bank)
+    return h
+
+
+# ------------------------------------------------------------------ helpers
+def clear_scene():
+    bpy.ops.object.select_all(action='SELECT')
+    bpy.ops.object.delete(use_global=False)
+    for coll in (bpy.data.meshes, bpy.data.materials, bpy.data.lights, bpy.data.cameras, bpy.data.curves):
+        for b in list(coll):
+            if b.users == 0:
+                coll.remove(b)
+
+
+def collection(name):
+    c = bpy.data.collections.get(name) or bpy.data.collections.new(name)
+    if c.name not in bpy.context.scene.collection.children:
+        bpy.context.scene.collection.children.link(c)
+    return c
+
+
+def mesh_object(name, verts, faces, coll, mat=None, smooth=False):
+    me = bpy.data.meshes.new(name)
+    me.from_pydata(verts, [], faces)
+    me.update()
+    ob = bpy.data.objects.new(name, me)
+    coll.objects.link(ob)
+    if mat:
+        ob.data.materials.append(mat)
+    if smooth:
+        for p in me.polygons:
+            p.use_smooth = True
+    return ob
+
+
+def principled(name, color, rough=0.8, metal=0.0, emission=None, strength=0.0, alpha=1.0):
+    m = bpy.data.materials.new(name)
+    m.use_nodes = True
+    bsdf = m.node_tree.nodes.get('Principled BSDF')
+    bsdf.inputs['Base Color'].default_value = (*color, 1)
+    bsdf.inputs['Roughness'].default_value = rough
+    bsdf.inputs['Metallic'].default_value = metal
+    if emission is not None:
+        key = 'Emission Color' if 'Emission Color' in bsdf.inputs else 'Emission'
+        bsdf.inputs[key].default_value = (*emission, 1)
+        bsdf.inputs['Emission Strength'].default_value = strength
+    return m
+
+
+def noise_material(name, c1, c2, scale, rough=0.85, bump=0.3):
+    """Procedural weathered surface (lime plaster, tiles, stone)."""
+    m = bpy.data.materials.new(name)
+    m.use_nodes = True
+    nt = m.node_tree
+    bsdf = nt.nodes.get('Principled BSDF')
+    tex = nt.nodes.new('ShaderNodeTexNoise')
+    tex.inputs['Scale'].default_value = scale
+    tex.inputs['Detail'].default_value = 8
+    ramp = nt.nodes.new('ShaderNodeValToRGB')
+    ramp.color_ramp.elements[0].color = (*c1, 1)
+    ramp.color_ramp.elements[1].color = (*c2, 1)
+    nt.links.new(tex.outputs['Fac'], ramp.inputs['Fac'])
+    nt.links.new(ramp.outputs['Color'], bsdf.inputs['Base Color'])
+    bsdf.inputs['Roughness'].default_value = rough
+    bmp = nt.nodes.new('ShaderNodeBump')
+    bmp.inputs['Strength'].default_value = bump
+    nt.links.new(tex.outputs['Fac'], bmp.inputs['Height'])
+    nt.links.new(bmp.outputs['Normal'], bsdf.inputs['Normal'])
+    return m
+
+
+def add_box(bm, cx, cy, cz, sx, sy, sz, mat_index=0, matrix=None):
+    """box in web local coords (x, y-up, z) -> appended to bm (blender coords)."""
+    res = bmesh.ops.create_cube(bm, size=1.0)
+    vs = res['verts']
+    for v in vs:
+        lx, ly, lz = v.co.x * sx + cx, v.co.y * sy + cy, v.co.z * sz + cz  # cube axes treated as web axes
+        p = Vector((lx, ly, lz))
+        if matrix is not None:
+            p = matrix @ p
+        v.co = W2B(p.x, p.y, p.z)
+    for f in {f for v in vs for f in v.link_faces}:
+        f.material_index = mat_index
+    return vs
+
+
+def web_matrix(x, y, z, ry):
+    return Matrix.Translation((x, y, z)) @ Matrix.Rotation(ry, 4, 'Y')
+
+
+# ------------------------------------------------------------------ build
+def build_materials():
+    M = {}
+    M['wall'] = noise_material('粉墙', (0.78, 0.76, 0.71), (0.93, 0.92, 0.88), 6.0, 0.92, 0.15)
+    M['roof'] = noise_material('黛瓦', (0.035, 0.037, 0.042), (0.08, 0.085, 0.095), 40.0, 0.55, 0.6)
+    M['dark'] = principled('墨色', (0.03, 0.03, 0.032), 0.7)
+    M['wood'] = noise_material('老木', (0.12, 0.06, 0.03), (0.2, 0.11, 0.05), 20.0, 0.75, 0.2)
+    M['stone'] = noise_material('青石', (0.22, 0.23, 0.23), (0.36, 0.37, 0.36), 9.0, 0.85, 0.5)
+    M['window'] = principled('窗纸', (0.45, 0.4, 0.32), 0.9, emission=(1.0, 0.55, 0.22), strength=0.0)
+    M['lantern'] = principled('红灯笼', (0.7, 0.06, 0.03), 0.5, emission=(1.0, 0.32, 0.1), strength=0.0)
+    M['leaf'] = noise_material('樟叶', (0.04, 0.09, 0.025), (0.09, 0.16, 0.04), 30.0, 0.7, 0.4)
+    M['bark'] = noise_material('树皮', (0.05, 0.035, 0.025), (0.12, 0.09, 0.06), 25.0, 0.95, 0.8)
+    M['ground'] = noise_material('田野', (0.06, 0.1, 0.03), (0.2, 0.18, 0.07), 0.08, 0.95, 0.3)
+    M['bamboo'] = principled('竹', (0.18, 0.25, 0.07), 0.6)
+    M['cloth'] = principled('蓝印花布', (0.04, 0.08, 0.25), 0.9)
+    # water: glossy reflection with a jade body colour
+    w = bpy.data.materials.new('河水')
+    w.use_nodes = True
+    b = w.node_tree.nodes.get('Principled BSDF')
+    b.inputs['Base Color'].default_value = (0.01, 0.03, 0.025, 1)
+    b.inputs['Roughness'].default_value = 0.03
+    b.inputs['IOR'].default_value = 1.33
+    tex = w.node_tree.nodes.new('ShaderNodeTexNoise')
+    tex.inputs['Scale'].default_value = 3.0
+    bmp = w.node_tree.nodes.new('ShaderNodeBump')
+    bmp.inputs['Strength'].default_value = 0.08
+    w.node_tree.links.new(tex.outputs['Fac'], bmp.inputs['Height'])
+    w.node_tree.links.new(bmp.outputs['Normal'], b.inputs['Normal'])
+    M['water'] = w
+    return M
+
+
+def build_terrain(M, coll, step=6.0, extent=900):
+    verts, faces = [], []
+    n = int(2 * extent / step) + 1
+    for j in range(n):
+        for i in range(n):
+            x, z = -extent + i * step, -extent + j * step
+            verts.append(W2B(x, terrain_height(x, z), z))
+    for j in range(n - 1):
+        for i in range(n - 1):
+            a = j * n + i
+            faces.append((a, a + 1, a + n + 1, a + n))
+    ob = mesh_object('山谷地形', verts, faces, coll, M['ground'], smooth=True)
+    # far mountains ring
+    mverts, mfaces = [], []
+    rings = [extent * 1.05 * (1.12 ** k) for k in range(22)]
+    seg = 240
+    for r in rings:
+        for a in range(seg):
+            ang = a / seg * math.tau
+            x, z = math.cos(ang) * r, math.sin(ang) * r
+            d_sun = math.atan2(math.sin(ang - (math.pi + 0.2)), math.cos(ang - (math.pi + 0.2)))
+            west = math.exp(-(d_sun ** 2) / 0.5)
+            h = smoothstep(700, 2600, r) * (150 + 360 * fbm(x * 0.0009, z * 0.0009, 6, 77)) * (1 - 0.74 * west) + terrain_height(x, z) * 0.2
+            mverts.append(W2B(x, h, z))
+    for ri in range(len(rings) - 1):
+        for a in range(seg):
+            p, q = ri * seg + a, ri * seg + (a + 1) % seg
+            mfaces.append((p, q, q + seg, p + seg))
+    mesh_object('远山', mverts, mfaces, coll, M['ground'], smooth=True)
+    return ob
+
+
+def build_river(M, coll):
+    verts, faces = [], []
+    rows = 0
+    for k in range(0, len(RIVER), 2):
+        s = RIVER[k][4]
+        w = river_half_width(s) + 3
+        for off in (-w, w):
+            x, z, *_ = river_point(s, off)
+            verts.append(W2B(x, 0.0, z))
+        rows += 1
+    for r in range(rows - 1):
+        a = r * 2
+        faces.append((a, a + 1, a + 3, a + 2))
+    mesh_object('小河', verts, faces, coll, M['water'])
+
+
+def build_houses(M, coll):
+    """Same row / spacing rules and RNG as buildHouses() in layout.ts."""
+    rng = mulberry32(SEED)
+    bridge_s, wheel_s = river_s_at_x(40), river_s_at_x(-196)
+    rows = [(-1, 14.6, True, 0.3), (1, 8.4, False, 0.2), (-1, 31, False, 2.5), (1, 25, False, 2.5), (-1, 50, False, 5), (1, 44, False, 5)]
+    houses = []
+    for side, off0, shop, jitter in rows:
+        s = VILLAGE_S0 + rng() * 6 + (10 if off0 > 20 else 0)
+        s_end = VILLAGE_S1 - (40 if off0 > 40 else 0)
+        while s < s_end:
+            w = 8 + rng() * 5
+            half = w / 2
+            sc = s + half
+            if off0 > 40 and rng() < 0.3:
+                s += w + 6
+                continue
+            blocked = abs(sc - bridge_s) < half + 8 or (side > 0 and abs(sc - wheel_s) < half + 10) or (side < 0 and sc < river_s_at_x(-264) + half)
+            if not blocked:
+                d = 8.5 + rng() * 3.5
+                off = side * (off0 + d / 2 + (rng() - 0.5) * jitter)
+                x, z, tx, tz, nx, nz = river_point(sc, off)
+                facing = math.atan2(-side * nx, -side * nz)
+                floors = 2 if rng() < (0.75 if shop else 0.55) else 1
+                rot = facing + (rng() - 0.5) * 0.05 * (3 if off0 > 20 else 1)
+                steps = (3 if rng() < 0.6 else 2) if floors == 2 else (2 if rng() < 0.5 else 1)
+                seed = int(rng() * 1e9)
+                chimney = rng() < 0.35
+                houses.append(dict(x=x, z=z, rot=rot, w=w, d=d, floors=floors, steps=steps, shop=shop, side=side, s=sc, chimney=chimney, lantern=off0 < 20))
+            s += w + ((0.2 + rng() * 0.6) if shop else (0.6 + rng() * 3.5))
+            if side > 0 and off0 < 10 and rng() < 0.18:
+                s += 5
+    bm = bmesh.new()
+    windows = []
+    lanterns = []
+    for h in houses:
+        m = web_matrix(h['x'], max(GROUND_Y, terrain_height(h['x'], h['z'])), h['z'], h['rot'])
+        w, d = h['w'], h['d']
+        H = h['floors'] * 3.3 + 0.5
+        rise = d * 0.27
+        add_box(bm, 0, H / 2 - 0.8, 0, w - 0.4, H + 1.6, d, 0, m)  # lime walls
+        # stepped horse-head gables
+        n = h['steps']
+        seg = d / 2 / n
+        for sx in (-1, 1):
+            for i in range(n):
+                top = H + rise * (i + 1) / n + 0.55
+                z0 = d / 2 - i * seg if i else d / 2 + 0.15
+                z1 = d / 2 - (i + 1) * seg if i < n - 1 else -(d / 2 - i * seg if i else d / 2 + 0.15)
+                zc, zl = (z0 + z1) / 2, abs(z0 - z1)
+                for mirror in ((1,) if i == n - 1 else (1, -1)):
+                    add_box(bm, sx * (w / 2 - 0.18), top / 2 - 0.8, mirror * zc, 0.36, top + 1.6, zl, 0, m)
+                    add_box(bm, sx * (w / 2 - 0.18), top + 0.1, mirror * zc, 0.8, 0.2, zl + 0.3, 2, m)
+        # roof: two slopes (tiles)
+        for zs in (1, -1):
+            ang = math.atan2(rise, d / 2 + 0.8)
+            L = math.hypot(d / 2 + 0.85, rise)
+            rm = m @ Matrix.Translation((0, H + rise / 2 - 0.1, zs * (d / 4 + 0.42))) @ Matrix.Rotation(zs * ang, 4, 'X')
+            add_box(bm, 0, 0, 0, w - 0.6, 0.12, L, 1, rm)
+        add_box(bm, 0, H + rise + 0.1, 0, w - 0.6, 0.3, 0.4, 2, m)  # ridge
+        # windows (emissive at night)
+        nw = max(2, int(w / 3.2))
+        for k in range(nw):
+            x = -w / 2 + (k + 0.5) * w / nw
+            windows.append((m, x, (3.3 + 1.7) if h['floors'] == 2 else 1.9, d / 2 + 0.03, h))
+        if h['lantern']:
+            for fx in (-0.3, 0.3):
+                p = m @ Vector((fx * w, 3.2 if h['floors'] == 2 else H - 0.6, d / 2 + 0.9))
+                lanterns.append((p, h['s'], h['side']))
+    me = bpy.data.meshes.new('徽派民居')
+    bm.to_mesh(me)
+    bm.free()
+    ob = bpy.data.objects.new('徽派民居', me)
+    for key in ('wall', 'roof', 'dark'):
+        ob.data.materials.append(M[key])
+    coll.objects.link(ob)
+    # window panes as one emissive mesh
+    bmw = bmesh.new()
+    for (m, x, y, z, h) in windows:
+        add_box(bmw, x, y, z, 1.1, 1.0, 0.04, 0, m)
+    mw = bpy.data.meshes.new('窗灯')
+    bmw.to_mesh(mw)
+    bmw.free()
+    ow = bpy.data.objects.new('窗灯', mw)
+    ow.data.materials.append(M['window'])
+    coll.objects.link(ow)
+    return houses, lanterns
+
+
+def build_bridge(M, coll):
+    s = river_s_at_x(40)
+    x, z, tx, tz, nx, nz = river_point(s, 0)
+    w = river_half_width(s)
+    R = w - 1.6
+    bpy.ops.mesh.primitive_cube_add(size=1)
+    body = bpy.context.active_object
+    body.name = '石拱桥'
+    body.scale = (2 * (w + 9), 4.4, R + 3.0)
+    body.location = (0, 0, (R + 3.0) / 2 - 2.2)
+    bpy.ops.mesh.primitive_cylinder_add(radius=R, depth=6, vertices=96)
+    hole = bpy.context.active_object
+    hole.rotation_euler = (math.pi / 2, 0, 0)
+    hole.location = (0, 0, -0.35)
+    mod = body.modifiers.new('拱洞', 'BOOLEAN')
+    mod.object = hole
+    mod.operation = 'DIFFERENCE'
+    bpy.context.view_layer.objects.active = body
+    bpy.ops.object.modifier_apply(modifier=mod.name)
+    bpy.data.objects.remove(hole, do_unlink=True)
+    # hump the deck: simple deform via vertex z offset
+    for v in body.data.vertices:
+        if v.co.z > 0:
+            fx = min(1.0, abs(v.co.x * body.scale.x) / (w + 9))
+            v.co.z += (1 - fx * fx) * 0.25
+    body.data.materials.append(M['stone'])
+    for c in body.users_collection:
+        c.objects.unlink(body)
+    coll.objects.link(body)
+    body.rotation_euler = (0, 0, math.atan2(-nz, nx) * -1 + 0)
+    ang = math.atan2(nz, nx)
+    body.rotation_euler = (0, 0, -ang)
+    body.location = W2B(x, 0, z) + Vector((0, 0, body.location.z))
+    return body
+
+
+def build_wheel(M, coll):
+    s = river_s_at_x(-196)
+    w = river_half_width(s)
+    x, z, tx, tz, nx, nz = river_point(s, w - 3.0)
+    bpy.ops.object.empty_add(type='PLAIN_AXES', location=W2B(x, 1.25, z))
+    pivot = bpy.context.active_object
+    pivot.name = '水车_轴心'
+    for c in pivot.users_collection:
+        c.objects.unlink(pivot)
+    coll.objects.link(pivot)
+    parts = []
+    for zs in (-0.55, 0.55):
+        bpy.ops.mesh.primitive_torus_add(major_radius=3.3, minor_radius=0.09, major_segments=64, minor_segments=6)
+        t = bpy.context.active_object
+        t.rotation_euler = (math.pi / 2, 0, 0)
+        t.location = (0, zs, 0)
+        parts.append(t)
+    for k in range(18):
+        a = k / 18 * math.tau
+        bpy.ops.mesh.primitive_cube_add(size=1)
+        p = bpy.context.active_object
+        p.scale = (0.62, 1.3, 0.05)
+        p.location = (math.cos(a) * 3.35, 0, math.sin(a) * 3.35)
+        p.rotation_euler = (0, -a, 0)
+        parts.append(p)
+        bpy.ops.mesh.primitive_cube_add(size=1)
+        sp = bpy.context.active_object
+        sp.scale = (3.3, 0.1, 0.1)
+        sp.location = (math.cos(a) * 1.65, 0.55 if k % 2 else -0.55, math.sin(a) * 1.65)
+        sp.rotation_euler = (0, -a, 0)
+        parts.append(sp)
+    for p in parts:
+        p.data.materials.append(M['wood'])
+        for c in p.users_collection:
+            c.objects.unlink(p)
+        coll.objects.link(p)
+        p.parent = pivot
+    pivot.rotation_euler = (0, 0, -math.atan2(nz, nx))
+    # rotation animation: -0.42 rad/s about the axle (same speed as the web version)
+    pivot.rotation_mode = 'XYZ'
+    pivot.keyframe_insert('rotation_euler', frame=1)
+    pivot.rotation_euler.rotate_axis('Y', 0.42 * DURATION)
+    pivot.keyframe_insert('rotation_euler', frame=int(DURATION * FPS) + 1)
+    for fc in pivot.animation_data.action.fcurves:
+        for kp in fc.keyframe_points:
+            kp.interpolation = 'LINEAR'
+    return pivot
+
+
+def build_camphor(M, coll):
+    tx, tz = river_point(river_s_at_x(-300), -34)[:2]
+    rnd = random.Random(777)
+    bm_bark, bm_leaf = bmesh.new(), bmesh.new()
+
+    def cyl(bm, a, b, r0, r1):
+        d = b - a
+        L = d.length
+        res = bmesh.ops.create_cone(bm, cap_ends=False, segments=10, radius1=r0, radius2=r1, depth=L)
+        q = Vector((0, 0, 1)).rotation_difference(d.normalized())
+        bmesh.ops.rotate(bm, verts=res['verts'], cent=(0, 0, 0), matrix=q.to_matrix())
+        bmesh.ops.translate(bm, verts=res['verts'], vec=a + d * 0.5)
+
+    def branch(p, d, L, r, depth):
+        for i in range(3 if depth else 5):
+            d = (d + Vector((rnd.uniform(-0.28, 0.28), rnd.uniform(-0.28, 0.28), rnd.uniform(-0.1, 0.25)))).normalized()
+            np_ = p + d * (L / 3)
+            cyl(bm_bark, p, np_, r, r * 0.82)
+            p, r = np_, r * 0.82
+        if depth >= 4:
+            res = bmesh.ops.create_icosphere(bm_leaf, subdivisions=2, radius=rnd.uniform(1.6, 2.4))
+            bmesh.ops.translate(bm_leaf, verts=res['verts'], vec=p)
+            return
+        for k in range(rnd.randint(2, 4)):
+            az = k / 3 * math.tau + rnd.random()
+            nd = (d + Vector((math.cos(az), math.sin(az), 0)) * 0.9).normalized()
+            branch(p, nd, L * 0.72, r * 0.72, depth + 1)
+
+    base = W2B(tx, GROUND_Y - 0.4, tz)
+    branch(base, Vector((0, 0, 1)), 24 * 0.32, 1.25, 0)
+    for name, bm, mat in (('古樟_树干', bm_bark, M['bark']), ('古樟_树冠', bm_leaf, M['leaf'])):
+        me = bpy.data.meshes.new(name)
+        bm.to_mesh(me)
+        bm.free()
+        ob = bpy.data.objects.new(name, me)
+        ob.data.materials.append(mat)
+        coll.objects.link(ob)
+        if name.endswith('树冠'):
+            disp = ob.modifiers.new('叶团起伏', 'DISPLACE')
+            tex = bpy.data.textures.new('叶团噪声', 'CLOUDS')
+            tex.noise_scale = 0.8
+            disp.texture = tex
+            disp.strength = 0.9
+
+
+def build_bamboo(M, coll, count=380):
+    rnd = random.Random(31)
+    bm = bmesh.new()
+    placed = 0
+    tries = 0
+    while placed < count and tries < 20000:
+        tries += 1
+        x, z = rnd.uniform(-520, 560), rnd.uniform(-420, 380)
+        d, s = river_query(x, z)
+        if d < 60 or fbm(x * 0.012, z * 0.012, 3, 61) < 0.52:
+            continue
+        y = terrain_height(x, z)
+        for c in range(6):
+            a, r = rnd.random() * math.tau, rnd.random() * 1.2
+            p = W2B(x + math.cos(a) * r, y, z + math.sin(a) * r)
+            h = rnd.uniform(9, 14)
+            res = bmesh.ops.create_cone(bm, cap_ends=False, segments=5, radius1=0.05, radius2=0.025, depth=h)
+            bmesh.ops.translate(bm, verts=res['verts'], vec=p + Vector((0, 0, h / 2)))
+            res = bmesh.ops.create_icosphere(bm, subdivisions=1, radius=1.4)
+            bmesh.ops.scale(bm, verts=res['verts'], vec=(1, 1, 2.2))
+            bmesh.ops.translate(bm, verts=res['verts'], vec=p + Vector((0, 0, h * 0.8)))
+        placed += 1
+    me = bpy.data.meshes.new('竹林')
+    bm.to_mesh(me)
+    bm.free()
+    ob = bpy.data.objects.new('竹林', me)
+    ob.data.materials.append(M['bamboo'])
+    coll.objects.link(ob)
+
+
+def build_lanterns(M, coll, lanterns):
+    """Lanterns light one after another along the market street from 58.4 s."""
+    first_s = river_s_at_x(-128)
+    objs = []
+    for i, (p, s, side) in enumerate(lanterns):
+        f = min(1, max(0, (s - first_s) / (VILLAGE_S1 - first_s)))
+        on_t = 58.4 + 0.9 + 7 * f + (1.2 if side > 0 else 0) + (i % 7) * 0.07
+        mat = M['lantern'].copy()
+        mat.name = f'灯笼_{i:03d}'
+        bsdf = mat.node_tree.nodes.get('Principled BSDF')
+        sock = bsdf.inputs['Emission Strength']
+        sock.default_value = 0.0
+        sock.keyframe_insert('default_value', frame=max(1, int(on_t * FPS) - 1))
+        sock.default_value = 18.0
+        sock.keyframe_insert('default_value', frame=int(on_t * FPS) + 6)
+        bpy.ops.mesh.primitive_uv_sphere_add(radius=0.3, segments=16, ring_count=10, location=W2B(p.x, p.y - 0.42, p.z))
+        ob = bpy.context.active_object
+        ob.scale = (1, 1, 1.25)
+        ob.name = f'灯笼_{i:03d}'
+        ob.data.materials.append(mat)
+        for c in ob.users_collection:
+            c.objects.unlink(ob)
+        coll.objects.link(ob)
+        # a small warm point light per lantern (Cycles handles hundreds fine)
+        ld = bpy.data.lights.new(f'灯光_{i:03d}', 'POINT')
+        ld.color = (1.0, 0.55, 0.25)
+        ld.shadow_soft_size = 0.3
+        ld.energy = 0
+        ld.keyframe_insert('energy', frame=max(1, int(on_t * FPS) - 1))
+        ld.energy = 40
+        ld.keyframe_insert('energy', frame=int(on_t * FPS) + 6)
+        lo = bpy.data.objects.new(f'灯光_{i:03d}', ld)
+        lo.location = W2B(p.x, p.y - 0.42, p.z)
+        coll.objects.link(lo)
+        objs.append(ob)
+    # windows glow later
+    w = bpy.data.materials.get('窗纸')
+    sock = w.node_tree.nodes.get('Principled BSDF').inputs['Emission Strength']
+    sock.default_value = 0
+    sock.keyframe_insert('default_value', frame=int(58.5 * FPS))
+    sock.default_value = 6
+    sock.keyframe_insert('default_value', frame=int(70 * FPS))
+    return objs
+
+
+SUN_KEYS = [(0, 8.5), (17, 6.6), (32, 4.6), (44, 2.6), (52, 0.9), (57, -0.8), (62, -2.8), (68, -5.2), (76, -7.6), (86, -9.8)]
+
+
+def build_sky_and_sun(scene):
+    world = bpy.data.worlds.new('暮色天空')
+    scene.world = world
+    world.use_nodes = True
+    nt = world.node_tree
+    bg = nt.nodes.get('Background')
+    sky = nt.nodes.new('ShaderNodeTexSky')
+    sky.sky_type = 'NISHITA'
+    sky.sun_rotation = math.pi / 2 + 0.2 + math.pi / 2  # sets toward west-north-west
+    sky.air_density = 1.2
+    sky.dust_density = 2.5
+    nt.links.new(sky.outputs['Color'], bg.inputs['Color'])
+    # twilight layer: deep blue dome with a warm afterglow band, faded in after sunset
+    tw = nt.nodes.new('ShaderNodeBackground')
+    grad_tc = nt.nodes.new('ShaderNodeTexCoord')
+    sep = nt.nodes.new('ShaderNodeSeparateXYZ')
+    nt.links.new(grad_tc.outputs['Generated'], sep.inputs['Vector'])
+    ramp = nt.nodes.new('ShaderNodeValToRGB')
+    ramp.color_ramp.elements[0].position = 0.5
+    ramp.color_ramp.elements[0].color = (0.35, 0.16, 0.1, 1)
+    ramp.color_ramp.elements[1].position = 0.62
+    ramp.color_ramp.elements[1].color = (0.02, 0.035, 0.1, 1)
+    nt.links.new(sep.outputs['Z'], ramp.inputs['Fac'])
+    nt.links.new(ramp.outputs['Color'], tw.inputs['Color'])
+    add = nt.nodes.new('ShaderNodeAddShader')
+    nt.links.new(bg.outputs['Background'], add.inputs[0])
+    nt.links.new(tw.outputs['Background'], add.inputs[1])
+    nt.links.new(add.outputs['Shader'], nt.nodes.get('World Output').inputs['Surface'])
+    sun_data = bpy.data.lights.new('太阳', 'SUN')
+    sun_data.angle = math.radians(1.2)
+    sun = bpy.data.objects.new('太阳', sun_data)
+    scene.collection.objects.link(sun)
+    moon_data = bpy.data.lights.new('月光', 'SUN')
+    moon_data.color = (0.6, 0.7, 1.0)
+    moon_data.energy = 0.0
+    moon = bpy.data.objects.new('月光', moon_data)
+    moon.rotation_euler = (math.radians(66), 0, math.radians(-110))
+    scene.collection.objects.link(moon)
+    for t, e in SUN_KEYS:
+        f = int(t * FPS) + 1
+        er = math.radians(e)
+        # Nishita is undefined below the horizon (can return NaN = black frames): clamp it
+        # and let the twilight layer + exposure carry the blue hour instead.
+        sky.sun_elevation = max(er, math.radians(0.4))
+        sky.keyframe_insert('sun_elevation', frame=f)
+        sun.rotation_euler = (math.pi / 2 - er, 0, -(math.pi / 2 + 0.2 + math.pi / 2))
+        sun.keyframe_insert('rotation_euler', frame=f)
+        sun_data.energy = max(0.0, 4.0 * smoothstep(-1.2, 6, e))
+        sun_data.color = (1.0, lerp(0.38, 0.76, smoothstep(0, 9, e)), lerp(0.14, 0.52, smoothstep(0, 9, e)))
+        sun_data.keyframe_insert('energy', frame=f)
+        sun_data.keyframe_insert('color', frame=f)
+        bg.inputs['Strength'].default_value = lerp(0.0, 1.0, smoothstep(-5, 2, e))
+        bg.inputs['Strength'].keyframe_insert('default_value', frame=f)
+        moon_data.energy = 0.35 * smoothstep(-1.5, -8.5, e)
+        moon_data.keyframe_insert('energy', frame=f)
+        tw.inputs['Strength'].default_value = 0.6 * smoothstep(1.0, -4.0, e)
+        tw.inputs['Strength'].keyframe_insert('default_value', frame=f)
+        # camera eye adaptation, like the web version
+        scene.view_settings.exposure = lerp(0.0, 3.2, smoothstep(1, -9, e))
+        scene.view_settings.keyframe_insert('exposure', frame=f)
+    # thin evening mist: a bounded volume over the valley (a world volume would swallow the sky)
+    bpy.ops.mesh.primitive_cube_add(size=1, location=(0, 0, 30))
+    mist = bpy.context.active_object
+    mist.name = '暮霭'
+    mist.scale = (2400, 2400, 60)
+    mm = bpy.data.materials.new('暮霭')
+    mm.use_nodes = True
+    nt2 = mm.node_tree
+    nt2.nodes.remove(nt2.nodes.get('Principled BSDF'))
+    vol = nt2.nodes.new('ShaderNodeVolumePrincipled')
+    vol.inputs['Density'].default_value = 0.0012
+    vol.inputs['Color'].default_value = (0.85, 0.8, 0.9, 1)
+    nt2.links.new(vol.outputs['Volume'], nt2.nodes.get('Material Output').inputs['Volume'])
+    mist.data.materials.append(mm)
+    mist.visible_shadow = False
+    vol.inputs['Density'].keyframe_insert('default_value', frame=1)
+    vol.inputs['Density'].default_value = 0.0026
+    vol.inputs['Density'].keyframe_insert('default_value', frame=int(80 * FPS))
+
+
+SHOT_TIMES = [(1, 0), (2, 9), (3, 17), (4, 25), (5, 32), (6, 40), (7, 48), (8, 55), (9, 65), (10, 74)]
+
+
+def build_cameras(scene, coll):
+    """Ten shots. Key poses (web coordinates) mirror camera.ts; markers switch the active camera."""
+    bs = river_s_at_x(40)
+    bx, bz, btx, btz, bnx, bnz = river_point(bs, 0)
+    ws = river_s_at_x(-196)
+    wx, wz, wtx, wtz, wnx, wnz = river_point(ws, river_half_width(ws) - 3)
+    tx, tz = river_point(river_s_at_x(-300), -34)[:2]
+    def rp(s_, off, y):
+        x_, z_ = river_point(s_, off)[:2]
+        return (x_, y, z_)
+
+    def qp(s_, lane, y):  # point on the market quay, same as quayPoint() in layout.ts
+        return rp(s_, -(11 + lane), y)
+
+    shots = {
+        1: ((470, 132, 150), (-1100, 60, 160), (310, 70, 82), (-1000, 30, 130), 38),
+        2: (rp(bs + 150, 1.2, 2.0), rp(bs + 116, 0, 2.3), rp(bs + 62, 1.2, 2.0), rp(bs + 28, 0, 2.3), 46),
+        3: ((bx + btx * 40 - bnx * 3.2, 1.05, bz + btz * 40 - bnz * 3.2), (bx, 3.1, bz), (bx + btx * 33 + bnx * 2.6, 1.05, bz + btz * 33 + bnz * 2.6), (bx, 3.1, bz), 34),
+        4: ((wx - wnx * 15.5 + wtx * 7, 2.9, wz - wnz * 15.5 + wtz * 7), (wx, 2.2, wz), (wx - wnx * 12.2 + wtx * 3.2, 2.4, wz - wnz * 12.2 + wtz * 3.2), (wx, 2.2, wz), 40),
+        5: (qp(river_s_at_x(-175) + 3.8 + 15, -1.1, GROUND_Y + 1.75), qp(river_s_at_x(-175) + 3.8 - 6, 1.0, GROUND_Y + 1.6),
+            qp(river_s_at_x(-175) + 11.4 + 12.5, -1.1, GROUND_Y + 1.75), qp(river_s_at_x(-175) + 11.4 - 6, 1.0, GROUND_Y + 1.6), 44),
+        6: ((tx + 25, 3.5, tz + 16.5), (tx - 2, 9.7, tz - 4), (tx + 19, 4.4, tz + 12.5), (tx - 2, 11.7, tz - 4), 46),
+        7: ((110, 26, -120), (300, 140, -410), (98, 30, -126), (205, 70, -372), 21),
+        8: (qp(river_s_at_x(-150), -0.8, GROUND_Y + 2.6), qp(river_s_at_x(-150) + 26, -1.6, GROUND_Y + 2.4),
+            qp(river_s_at_x(-150) + 21, -0.8, GROUND_Y + 2.6), qp(river_s_at_x(-150) + 47, -1.6, GROUND_Y + 2.4), 48),
+        9: (rp(bs + 52, -4.6, 1.25), (bx + bnx * 1.5, 3.4, bz + bnz * 1.5), rp(bs + 45, -4.6, 1.25), (bx + bnx * 1.5, 3.4, bz + bnz * 1.5), 42),
+        10: ((95, 9, 34), (-10, 3, -4), (250, 125, 205), (-520, 60, -230), 44),
+    }
+    cams = {}
+    for i, (sid, t0) in enumerate(SHOT_TIMES):
+        t1 = SHOT_TIMES[i + 1][1] if i + 1 < len(SHOT_TIMES) else DURATION
+        p0, a0, p1, a1, fov = shots[sid]
+        cd = bpy.data.cameras.new(f'镜头{sid:02d}')
+        cd.lens_unit = 'FOV'
+        cd.angle = math.radians(fov * 16 / 9 * 0.92)
+        cd.sensor_fit = 'HORIZONTAL'
+        cam = bpy.data.objects.new(f'镜头{sid:02d}', cd)
+        coll.objects.link(cam)
+        tgt = bpy.data.objects.new(f'镜头{sid:02d}_目标', None)
+        coll.objects.link(tgt)
+        con = cam.constraints.new('TRACK_TO')
+        con.target = tgt
+        con.track_axis = 'TRACK_NEGATIVE_Z'
+        con.up_axis = 'UP_Y'
+        for (p, a), t in (((p0, a0), t0), ((p1, a1), t1)):
+            cam.location = W2B(*p)
+            tgt.location = W2B(*a)
+            cam.keyframe_insert('location', frame=int(t * FPS) + 1)
+            tgt.keyframe_insert('location', frame=int(t * FPS) + 1)
+        m = scene.timeline_markers.new(f'S{sid:02d}', frame=int(t0 * FPS) + 1)
+        m.camera = cam
+        cams[sid] = cam
+    scene.camera = cams[1]
+
+
+def setup_render(scene):
+    scene.render.fps = FPS
+    scene.frame_start = 1
+    scene.frame_end = int(DURATION * FPS)
+    scene.render.resolution_x = 1920
+    scene.render.resolution_y = 1080
+    try:
+        scene.render.engine = RENDER_ENGINE
+    except TypeError:
+        scene.render.engine = 'CYCLES'
+    if scene.render.engine == 'CYCLES':
+        scene.cycles.samples = 128
+        scene.cycles.use_denoising = True
+        scene.cycles.max_bounces = 6
+    scene.view_settings.view_transform = 'AgX' if 'AgX' in [i.identifier for i in scene.view_settings.bl_rna.properties['view_transform'].enum_items] else 'Filmic'
+    scene.view_settings.look = 'None'
+    scene.render.image_settings.file_format = 'FFMPEG'
+    scene.render.ffmpeg.format = 'MPEG4'
+    scene.render.ffmpeg.codec = 'H264'
+    scene.render.ffmpeg.constant_rate_factor = 'HIGH'
+    scene.render.filepath = '//暮归江南_blender_'
+
+
+def main():
+    clear_scene()
+    scene = bpy.context.scene
+    M = build_materials()
+    land = collection('地形与河流')
+    town = collection('村落')
+    life = collection('灯火')
+    cams = collection('镜头')
+    print('[暮归江南] 地形…')
+    build_terrain(M, land)
+    build_river(M, land)
+    print('[暮归江南] 民居…')
+    houses, lanterns = build_houses(M, town)
+    print('[暮归江南] 石桥、水车、��樟、竹林…')
+    build_bridge(M, town)
+    build_wheel(M, town)
+    build_camphor(M, town)
+    build_bamboo(M, land)
+    print('[暮归江南] 灯笼…')
+    build_lanterns(M, life, lanterns)
+    build_sky_and_sun(scene)
+    build_cameras(scene, cams)
+    setup_render(scene)
+    print(f'[暮归江南] 完成：{len(houses)} 栋民居，{len(lanterns)} 盏灯笼，{int(DURATION * FPS)} 帧。')
+
+
+if __name__ == '__main__':
+    main()

src/jiangnan/JiangnanApp.tsx

@@ -1,30 +1,666 @@
-import { useEffect, useRef, useState } from 'react'
-import { Film, renderIds } from './engine/Film'
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
+import './jiangnan.css'
+import { buildMix } from './audio/mixer'
+import blenderScript from './blender/jiangnan_village.py?raw'
+import { chars, formatTags, LIVE_COPY, PLATFORMS } from './copy/platforms'
+import { Film, Quality, renderIds } from './engine/Film'
+import { DURATION, SHOTS, shotAt } from './engine/timeline'
+import { canvasToBlob, COVER_DESIGNS, Ratio, RATIOS, renderCover } from './export/covers'
+import { generateLivePhotos, LIVE_CLIPS, LiveResult } from './export/livegen'
+import { exportMp4, exportSupport } from './export/mp4'
+import { drawOverlay, loadFonts } from './player/overlay'
+import { strToU8, zipSync, Zippable } from 'fflate'
+
+type Aspect = 'landscape' | 'portrait'
+
+const fmt = (t: number) => {
+  const m = Math.floor(t / 60), s = Math.floor(t % 60), cs = Math.floor((t % 1) * 10)
+  return `${m}:${String(s).padStart(2, '0')}.${cs}`
+}
+
+function download(blob: Blob, name: string) {
+  const a = document.createElement('a')
+  a.href = URL.createObjectURL(blob)
+  a.download = name
+  document.body.appendChild(a)
+  a.click()
+  a.remove()
+  setTimeout(() => URL.revokeObjectURL(a.href), 60000)
+}
+
+async function copyText(text: string) {
+  try {
+    await navigator.clipboard.writeText(text)
+    return true
+  } catch {
+    const ta = document.createElement('textarea')
+    ta.value = text
+    ta.style.position = 'fixed'
+    ta.style.opacity = '0'
+    document.body.appendChild(ta)
+    ta.select()
+    const ok = document.execCommand('copy')
+    ta.remove()
+    return ok
+  }
+}
+
+async function blobB64(b: Blob) {
+  const u = new Uint8Array(await b.arrayBuffer())
+  let s = ''
+  for (let i = 0; i < u.length; i += 32768) s += String.fromCharCode(...u.subarray(i, i + 32768))
+  return btoa(s)
+}
+
+const isMobile = () => typeof navigator !== 'undefined' && /Android|iPhone|iPad|iPod|Mobile/i.test(navigator.userAgent)
 
-// Temporary test harness (replaced by the full player UI)
 export default function JiangnanApp() {
-  const ref = useRef<HTMLCanvasElement>(null)
-  const [status, setStatus] = useState('加载中…')
+  const playerRef = useRef<HTMLDivElement>(null)
+  const glRef = useRef<HTMLCanvasElement>(null)
+  const ovRef = useRef<HTMLCanvasElement>(null)
+  const filmRef = useRef<Film | null>(null)
+  const audioRef = useRef<{ ctx: AudioContext | null; buf: AudioBuffer | null; src: AudioBufferSourceNode | null; gain: GainNode | null }>({ ctx: null, buf: null, src: null, gain: null })
+  const clockRef = useRef({ playing: false, startPerf: 0, startT: 0, t: 0 })
+  const scaleRef = useRef(1)
+  const busyRef = useRef(false)
+  const dirtyRef = useRef(true)
+
+  const [ready, setReady] = useState(false)
+  const [load, setLoad] = useState({ p: 0, label: '准备中' })
+  const [audioReady, setAudioReady] = useState(false)
+  const [playing, setPlaying] = useState(false)
+  const [time, setTime] = useState(0)
+  const [aspect, setAspect] = useState<Aspect>(() => (typeof window !== 'undefined' && window.innerHeight > window.innerWidth && isMobile() ? 'portrait' : 'landscape'))
+  const [subs, setSubs] = useState(true)
+  const [muted, setMuted] = useState(false)
+  const [toast, setToast] = useState('')
+  const [err, setErr] = useState('')
+
+  const say = useCallback((m: string) => {
+    setToast(m)
+    window.setTimeout(() => setToast(''), 1800)
+  }, [])
+
+  // ------------------------------------------------------------ init film + audio
   useEffect(() => {
-    const canvas = ref.current!
-    const film = new Film(canvas, 'high')
+    if (filmRef.current) return // keep the built world across hot reloads
+    const canvas = glRef.current!
     const q = new URLSearchParams(location.search)
-    const w = Number(q.get('w') || 960), h = Number(q.get('h') || 540)
-    film.width = w
-    film.height = h
-    ;(window as unknown as { __film: Film }).__film = film
-    ;(window as unknown as { __ids: (t: number) => void }).__ids = (t: number) => renderIds(film, t)
-    film.init((p, l) => setStatus(`${Math.round(p * 100)}% ${l}`)).then(() => {
-      film.setSize(w, h)
-      film.renderFrame(Number(q.get('t') || 5))
-      setStatus('ready')
-      document.title = 'ready'
-    }).catch((e) => { setStatus('error: ' + e.message); document.title = 'error'; console.error(e) })
+    const quality: Quality = (q.get('q') as Quality) || (isMobile() ? 'medium' : 'high')
+    const film = new Film(canvas, quality)
+    filmRef.current = film
+    const w = Number(q.get('w') || 0), h = Number(q.get('h') || 0)
+    // automation hooks (used by the QA scripts)
+    const win = window as unknown as Record<string, unknown>
+    win.__film = film
+    win.__ids = (t: number) => renderIds(film, t)
+    win.__test = {
+      mp4: async (w: number, h: number, fps: 24 | 30 | 60, start: number, duration: number) => {
+        const r = await exportMp4(film, { width: w, height: h, fps, start, duration, audio: audioRef.current.buf, overlay: { subtitles: true, titles: true }, signal: { cancelled: false }, onProgress: () => {} })
+        return blobB64(r.blob)
+      },
+      live: async (w: number, h: number, frames: number) => {
+        const r = await generateLivePhotos(film, () => {}, { cancelled: false }, { w, h, frames, clips: 1 })
+        return blobB64(r.zip)
+      },
+    }
+    let cancelled = false
+    ;(async () => {
+      try {
+        await loadFonts()
+        if (w && h) { film.width = w; film.height = h }
+        await film.init((p, label) => !cancelled && setLoad({ p: p * 0.85, label }))
+        if (w && h) film.setSize(w, h)
+        film.renderFrame(Number(q.get('t') || 0.01))
+        if (cancelled) return
+        setReady(true)
+        document.title = 'ready'
+        setLoad({ p: 0.9, label: '合成配乐与旁白' })
+        if (q.get('noaudio') !== '1') {
+          const buf = await buildMix(undefined, (p) => !cancelled && setLoad({ p: 0.85 + p * 0.15, label: '合成配乐与旁白' }))
+          audioRef.current.buf = buf
+          if (!cancelled) setAudioReady(true)
+        }
+      } catch (e) {
+        console.error(e)
+        setErr('加载失败：' + (e as Error).message)
+        document.title = 'error'
+      }
+    })()
+    return () => { cancelled = true }
+  }, [])
+
+  // ------------------------------------------------------------ sizing
+  const resize = useCallback(() => {
+    const film = filmRef.current
+    const el = playerRef.current
+    if (!film || !el || busyRef.current) return
+    const rect = el.getBoundingClientRect()
+    const dpr = Math.min(window.devicePixelRatio || 1, 2)
+    const maxLong = isMobile() ? 1100 : 1600
+    let W = rect.width * dpr * scaleRef.current, H = rect.height * dpr * scaleRef.current
+    const long = Math.max(W, H)
+    if (long > maxLong) { W *= maxLong / long; H *= maxLong / long }
+    W = Math.max(160, Math.round(W)), H = Math.max(90, Math.round(H))
+    if (W !== film.width || H !== film.height || !film.ready) {
+      if (film.ready) film.setSize(W, H)
+      else { film.width = W; film.height = H }
+      const ov = ovRef.current!
+      ov.width = Math.round(rect.width * dpr)
+      ov.height = Math.round(rect.height * dpr)
+      dirtyRef.current = true
+    }
+  }, [])
+  useEffect(() => {
+    resize()
+    const ro = new ResizeObserver(() => resize())
+    if (playerRef.current) ro.observe(playerRef.current)
+    return () => ro.disconnect()
+  }, [resize, aspect, ready])
+
+  // ------------------------------------------------------------ playback clock
+  const now = () => {
+    const c = clockRef.current
+    if (!c.playing) return c.t
+    const a = audioRef.current
+    if (a.ctx && a.src) return c.startT + (a.ctx.currentTime - c.startPerf)
+    return c.startT + (performance.now() - c.startPerf) / 1000
+  }
+
+  const stopAudio = () => {
+    const a = audioRef.current
+    if (a.src) { try { a.src.stop() } catch { /* ignore */ } a.src.disconnect(); a.src = null }
+  }
+
+  const play = useCallback(async (from?: number) => {
+    const c = clockRef.current
+    let t = from ?? c.t
+    if (t >= DURATION - 0.05) t = 0
+    const a = audioRef.current
+    stopAudio()
+    if (a.buf) {
+      if (!a.ctx) {
+        a.ctx = new AudioContext({ sampleRate: 48000 })
+        a.gain = a.ctx.createGain()
+        a.gain.connect(a.ctx.destination)
+      }
+      await a.ctx.resume()
+      a.gain!.gain.value = muted ? 0 : 1
+      const src = a.ctx.createBufferSource()
+      src.buffer = a.buf
+      src.connect(a.gain!)
+      src.start(0, t)
+      a.src = src
+      c.startPerf = a.ctx.currentTime
+    } else c.startPerf = performance.now()
+    c.startT = t
+    c.playing = true
+    setPlaying(true)
+  }, [muted])
+
+  const pause = useCallback(() => {
+    const c = clockRef.current
+    c.t = now()
+    c.playing = false
+    stopAudio()
+    setPlaying(false)
+    dirtyRef.current = true
   }, [])
+
+  const seek = useCallback((t: number) => {
+    const c = clockRef.current
+    c.t = Math.max(0, Math.min(DURATION, t))
+    setTime(c.t)
+    dirtyRef.current = true
+    if (c.playing) play(c.t)
+  }, [play])
+
+  useEffect(() => {
+    const a = audioRef.current
+    if (a.gain) a.gain.gain.value = muted ? 0 : 1
+  }, [muted])
+
+  // ------------------------------------------------------------ render loop
+  useEffect(() => {
+    if (!ready) return
+    let raf = 0
+    let ema = 16
+    let lastUi = 0
+    const loop = () => {
+      raf = requestAnimationFrame(loop)
+      const film = filmRef.current
+      if (!film || !film.ready || busyRef.current) return
+      const c = clockRef.current
+      let t = now()
+      if (c.playing && t >= DURATION) {
+        c.t = DURATION
+        c.playing = false
+        stopAudio()
+        setPlaying(false)
+        t = DURATION
+      }
+      if (!c.playing && !dirtyRef.current) return
+      if (!c.playing) c.t = t
+      const t0 = performance.now()
+      film.renderFrame(Math.min(t, DURATION - 0.001))
+      const ov = ovRef.current!
+      const g = ov.getContext('2d')!
+      g.clearRect(0, 0, ov.width, ov.height)
+      drawOverlay(g, t, ov.width, ov.height, { subtitles: subs, titles: true })
+      dirtyRef.current = false
+      const dt = performance.now() - t0
+      ema = ema * 0.9 + dt * 0.1
+      if (c.playing) {
+        // adaptive resolution: keep playback smooth on slower devices
+        if (ema > 42 && scaleRef.current > 0.45) { scaleRef.current *= 0.85; ema = 30; resize() } else if (ema < 14 && scaleRef.current < 1) { scaleRef.current = Math.min(1, scaleRef.current * 1.1); ema = 20; resize() }
+      }
+      if (performance.now() - lastUi > 100) { setTime(t); lastUi = performance.now() }
+    }
+    raf = requestAnimationFrame(loop)
+    return () => cancelAnimationFrame(raf)
+  }, [ready, subs, resize])
+
+  useEffect(() => { dirtyRef.current = true }, [subs, aspect])
+
+  // keyboard
+  useEffect(() => {
+    const onKey = (e: KeyboardEvent) => {
+      if ((e.target as HTMLElement)?.tagName === 'INPUT' || (e.target as HTMLElement)?.tagName === 'TEXTAREA') return
+      if (e.code === 'Space') { e.preventDefault(); if (clockRef.current.playing) pause(); else play() }
+      if (e.code === 'ArrowRight') seek(now() + 5)
+      if (e.code === 'ArrowLeft') seek(now() - 5)
+    }
+    window.addEventListener('keydown', onKey)
+    return () => window.removeEventListener('keydown', onKey)
+  }, [play, pause, seek])
+
+  const fullscreen = () => {
+    const el = playerRef.current
+    if (!el) return
+    if (document.fullscreenElement) document.exitFullscreen()
+    else el.requestFullscreen?.().catch(() => say('当前浏览器不支持全屏'))
+  }
+
+  const curShot = shotAt(time)
+
+  // ------------------------------------------------------------ busy wrapper (export / covers / live)
+  const withBusy = async <T,>(fn: () => Promise<T>) => {
+    pause()
+    busyRef.current = true
+    try { return await fn() } finally {
+      busyRef.current = false
+      resize()
+      dirtyRef.current = true
+    }
+  }
+
   return (
-    <div style={{ background: '#000', minHeight: '100vh' }}>
-      <canvas ref={ref} style={{ width: '100%', display: 'block' }} />
-      <div id="status" style={{ color: '#fff', position: 'fixed', top: 4, left: 4, fontSize: 12 }}>{status}</div>
+    <div className="jn">
+      <div className="jn-wrap">
+        <header className="jn-head">
+          <div>
+            <h1 className="jn-title">暮归江南<span className="jn-seal">江南</span></h1>
+            <div className="jn-sub">一座水乡村庄的日落与灯火 · 原创 3D 动画 · {Math.round(DURATION)} 秒 · 配乐 / 音效 / 旁白</div>
+          </div>
+          <div className="jn-links">浏览器实时渲染 · 本地导出，不上传任何数据</div>
+        </header>
+
+        <div ref={playerRef} className={`jn-player ${aspect}`}>
+          <canvas ref={glRef} />
+          <canvas ref={ovRef} />
+          {!ready && (
+            <div className="jn-loading">
+              <div className="t">暮归江南</div>
+              <div className="jn-bar"><i style={{ width: `${Math.round(load.p * 100)}%` }} /></div>
+              <small>{err || `${load.label}…`}</small>
+            </div>
+          )}
+          {ready && !playing && !busyRef.current && (
+            <button className="jn-big-play" onClick={() => play()} aria-label="播放">
+              <span><svg width="30" height="30" viewBox="0 0 24 24"><path d="M8 5v14l11-7z" fill="#fff3e2" /></svg></span>
+            </button>
+          )}
+        </div>
+
+        <div className="jn-controls">
+          <button className="jn-btn primary" disabled={!ready} onClick={() => (playing ? pause() : play())}>{playing ? '暂停' : '播放'}</button>
+          <span className="jn-time">{fmt(time)} / {fmt(DURATION)}</span>
+          <div className="jn-scrub">
+            <input type="range" min={0} max={DURATION} step={0.05} value={time} onChange={(e) => seek(Number(e.target.value))} disabled={!ready} />
+          </div>
+          <button className={`jn-btn ${aspect === 'landscape' ? 'on' : ''}`} onClick={() => setAspect('landscape')}>横屏 16:9</button>
+          <button className={`jn-btn ${aspect === 'portrait' ? 'on' : ''}`} onClick={() => setAspect('portrait')}>竖屏 9:16</button>
+          <button className={`jn-btn ghost ${subs ? 'on' : ''}`} onClick={() => setSubs(!subs)}>字幕{subs ? '开' : '关'}</button>
+          <button className={`jn-btn ghost ${muted ? '' : 'on'}`} onClick={() => setMuted(!muted)}>{muted ? '已静音' : '声音开'}</button>
+          <button className="jn-btn ghost" onClick={fullscreen}>全屏</button>
+        </div>
+        {!audioReady && ready && <div className="jn-note">配乐与旁白正在浏览器里合成，稍等几秒后播放会有声音。</div>}
+        <div className="jn-shots">
+          {SHOTS.map((s) => (
+            <button key={s.id} className={`jn-chip ${curShot.id === s.id ? 'active' : ''}`} onClick={() => seek(s.t0 + 0.01)}>
+              {String(s.id).padStart(2, '0')} {s.name}
+            </button>
+          ))}
+        </div>
+
+        <ExportSection film={filmRef} audio={audioRef} ready={ready && audioReady} withBusy={withBusy} />
+        <CoverSection film={filmRef} ready={ready} withBusy={withBusy} />
+        <CopySection say={say} />
+        <LiveSection film={filmRef} ready={ready} withBusy={withBusy} say={say} />
+
+        <section className="jn-sec">
+          <h2>Blender 场景脚本</h2>
+          <p className="lead">同一座村庄的 Blender 版本：在 Blender 4.x 的「脚本」面板里打开并运行，会自动生成地形、河道、徽派民居、石拱桥、水车、古樟、灯笼、日落到夜晚的灯光关键帧和 10 个镜头的摄像机，可用 Cycles 离线渲染 4K。</p>
+          <div className="jn-row">
+            <button className="jn-btn" onClick={() => download(new Blob([blenderScript], { type: 'text/x-python' }), 'jiangnan_village.py')}>下载 jiangnan_village.py</button>
+            <span className="jn-note" style={{ margin: 0 }}>使用：Blender → Scripting → 打开脚本 → Run Script（约 1 分钟）。</span>
+          </div>
+        </section>
+
+        <footer className="jn-foot">
+          《暮归江南》全部画面在浏览器中实时生成：程序化地形与民居、平面反射水面、高度雾、体积光、辉光与胶片调色；配乐为程序合成的五声音阶原创曲（古筝物理建模、竹箫、弦乐、编钟），环境音效同样为程序合成；旁白为原创文案的中文语音合成。字体：马善政楷书、霞鹜文楷（SIL OFL）。
+        </footer>
+      </div>
+      {toast && <div className="jn-toast">{toast}</div>}
     </div>
   )
 }
+
+// ====================================================================== export
+type FilmRef = React.MutableRefObject<Film | null>
+type Busy = <T>(fn: () => Promise<T>) => Promise<T>
+
+function ExportSection({ film, audio, ready, withBusy }: { film: FilmRef; audio: React.MutableRefObject<{ buf: AudioBuffer | null }>; ready: boolean; withBusy: Busy }) {
+  const [orient, setOrient] = useState<Aspect>('landscape')
+  const [res, setRes] = useState<720 | 1080>(1080)
+  const [fps, setFps] = useState<24 | 30 | 60>(30)
+  const [subs, setSubs] = useState(true)
+  const [prog, setProg] = useState<{ p: number; label: string } | null>(null)
+  const [out, setOut] = useState<{ url: string; name: string; size: number; note: string } | null>(null)
+  const [err, setErr] = useState('')
+  const sig = useRef({ cancelled: false })
+  const sup = useMemo(() => exportSupport(), [])
+  const dims = orient === 'landscape' ? { w: res === 1080 ? 1920 : 1280, h: res === 1080 ? 1080 : 720 } : { w: res === 1080 ? 1080 : 720, h: res === 1080 ? 1920 : 1280 }
+  const start = async () => {
+    if (!film.current) return
+    setErr('')
+    setOut(null)
+    sig.current = { cancelled: false }
+    try {
+      await withBusy(async () => {
+        const r = await exportMp4(film.current!, {
+          width: dims.w, height: dims.h, fps, audio: audio.current.buf, overlay: { subtitles: subs, titles: true }, signal: sig.current,
+          onProgress: (p, info) => setProg({ p, label: `${info.stage} ${info.frame}/${info.total} 帧${info.eta > 1 ? ` · 预计还需 ${Math.ceil(info.eta)} 秒` : ''}` }),
+        })
+        const name = `暮归江南_${orient === 'landscape' ? '横屏' : '竖屏'}_${dims.w}x${dims.h}_${fps}fps.mp4`
+        setOut({ url: URL.createObjectURL(r.blob), name, size: r.blob.size, note: r.audioCodec ? `音轨：${r.audioCodec === 'aac' ? 'AAC' : 'Opus'}` : '当前浏览器不支持音频编码，导出的视频没有声音（建议用电脑版 Chrome / Edge）' })
+      })
+    } catch (e) {
+      setErr((e as Error).message || String(e))
+    } finally {
+      setProg(null)
+    }
+  }
+  return (
+    <section className="jn-sec">
+      <h2>导出 MP4</h2>
+      <p className="lead">在你的浏览器里逐帧渲染并编码，画面与预览完全一致（不会掉帧），配乐、音效、旁白一起写入。视频不上传任何服务器。</p>
+      <div className="jn-grid">
+        <div className="jn-card">
+          <div className="jn-label">画幅</div>
+          <div className="jn-opt">
+            <button className={`jn-btn ${orient === 'landscape' ? 'on' : ''}`} onClick={() => setOrient('landscape')}>横屏 16:9</button>
+            <button className={`jn-btn ${orient === 'portrait' ? 'on' : ''}`} onClick={() => setOrient('portrait')}>竖屏 9:16（手机）</button>
+          </div>
+          <div className="jn-label">清晰度</div>
+          <div className="jn-opt">
+            <button className={`jn-btn ${res === 1080 ? 'on' : ''}`} onClick={() => setRes(1080)}>1080p</button>
+            <button className={`jn-btn ${res === 720 ? 'on' : ''}`} onClick={() => setRes(720)}>720p（更快）</button>
+          </div>
+        </div>
+        <div className="jn-card">
+          <div className="jn-label">帧率</div>
+          <div className="jn-opt">
+            {([24, 30, 60] as const).map((f) => (
+              <button key={f} className={`jn-btn ${fps === f ? 'on' : ''}`} onClick={() => setFps(f)}>{f} fps</button>
+            ))}
+          </div>
+          <div className="jn-label">字幕</div>
+          <div className="jn-opt">
+            <button className={`jn-btn ${subs ? 'on' : ''}`} onClick={() => setSubs(true)}>烧录字幕</button>
+            <button className={`jn-btn ${!subs ? 'on' : ''}`} onClick={() => setSubs(false)}>无字幕</button>
+          </div>
+        </div>
+        <div className="jn-card">
+          <div className="jn-label">输出</div>
+          <div style={{ fontSize: 15, margin: '6px 0 12px' }}>{dims.w}×{dims.h} · {fps} fps · {Math.round(DURATION * fps)} 帧</div>
+          <div className="jn-row">
+            {!prog ? (
+              <button className="jn-btn primary" disabled={!ready || !sup.video} onClick={start}>开始导出</button>
+            ) : (
+              <button className="jn-btn" onClick={() => { sig.current.cancelled = true }}>取消</button>
+            )}
+          </div>
+          {!sup.video && <div className="jn-err">当前浏览器不支持 WebCodecs 视频编码，请用电脑版 Chrome / Edge 或 Safari 16.4+。</div>}
+          {!ready && sup.video && <div className="jn-note">等待场景和配乐准备完成…</div>}
+        </div>
+      </div>
+      {prog && (
+        <div className="jn-progress">
+          <div className="jn-bar"><i style={{ width: `${(prog.p * 100).toFixed(1)}%` }} /></div>
+          <small>{prog.label}（导出期间请保持本页在前台）</small>
+        </div>
+      )}
+      {err && <div className="jn-err">导出失败：{err}</div>}
+      {out && (
+        <div className="jn-result">
+          <video src={out.url} controls playsInline />
+          <div className="jn-row" style={{ marginTop: 10 }}>
+            <a className="jn-btn primary" href={out.url} download={out.name}>保存 {out.name}（{(out.size / 1048576).toFixed(1)} MB）</a>
+            <span className="jn-note" style={{ margin: 0 }}>{out.note}</span>
+          </div>
+        </div>
+      )}
+      <p className="jn-note">提示：1080p 60fps 共 {Math.round(DURATION * 60)} 帧，独立显卡电脑约 2–4 分钟；手机建议导出 720p。iPhone 上导出后点「保存」→「存储视频」。</p>
+    </section>
+  )
+}
+
+// ====================================================================== covers
+interface CoverItem { key: string; ratio: Ratio; name: string; url: string; blob: Blob }
+
+function CoverSection({ film, ready, withBusy }: { film: FilmRef; ready: boolean; withBusy: Busy }) {
+  const [ratios, setRatios] = useState<Ratio[]>(['16:9', '4:3', '3:4'])
+  const [designs, setDesigns] = useState<string[]>(COVER_DESIGNS.map((d) => d.key))
+  const [items, setItems] = useState<CoverItem[]>([])
+  const [sel, setSel] = useState<Set<string>>(new Set())
+  const [prog, setProg] = useState<number | null>(null)
+  const toggle = <T,>(arr: T[], v: T) => (arr.includes(v) ? arr.filter((x) => x !== v) : [...arr, v])
+  const gen = async () => {
+    if (!film.current) return
+    await withBusy(async () => {
+      const out: CoverItem[] = []
+      const jobs = ratios.flatMap((r) => COVER_DESIGNS.filter((d) => designs.includes(d.key)).map((d) => ({ r, d })))
+      for (let i = 0; i < jobs.length; i++) {
+        setProg(i / jobs.length)
+        await new Promise((r) => setTimeout(r, 0))
+        const { r, d } = jobs[i]
+        const c = renderCover(film.current!, d, r)
+        const blob = await canvasToBlob(c)
+        out.push({ key: `${r}-${d.key}`, ratio: r, name: `暮归江南_封面_${r.replace(':', 'x')}_${d.name}.jpg`, url: URL.createObjectURL(blob), blob })
+      }
+      setItems(out)
+      setSel(new Set(out.map((o) => o.key)))
+      setProg(null)
+    })
+  }
+  const dl = () => {
+    const chosen = items.filter((i) => sel.has(i.key))
+    if (chosen.length === 1) return download(chosen[0].blob, chosen[0].name)
+    Promise.all(chosen.map(async (c) => [c.name, new Uint8Array(await c.blob.arrayBuffer())] as const)).then((list) => {
+      const z: Zippable = {}
+      list.forEach(([n, b]) => { z['暮归江南_封面/' + n] = [b, { level: 0 }] })
+      download(new Blob([zipSync(z) as Uint8Array<ArrayBuffer>], { type: 'application/zip' }), `暮归江南_封面_${chosen.length}张.zip`)
+    })
+  }
+  return (
+    <section className="jn-sec">
+      <h2>封面</h2>
+      <p className="lead">每个比例都按该比例重新取景渲染（不是裁切），再排上书法标题。选择比例与画面，生成后勾选需要的封面下载。</p>
+      <div className="jn-label">比例（可多选）</div>
+      <div className="jn-opt">
+        {RATIOS.map((r) => (
+          <button key={r.id} className={`jn-btn ${ratios.includes(r.id) ? 'on' : ''}`} onClick={() => setRatios(toggle(ratios, r.id))}>{ratios.includes(r.id) ? '✓ ' : ''}{r.label}</button>
+        ))}
+      </div>
+      <div className="jn-label">画面（可多选）</div>
+      <div className="jn-opt">
+        {COVER_DESIGNS.map((d) => (
+          <button key={d.key} className={`jn-btn ${designs.includes(d.key) ? 'on' : ''}`} onClick={() => setDesigns(toggle(designs, d.key))}>{designs.includes(d.key) ? '✓ ' : ''}{d.name}</button>
+        ))}
+      </div>
+      <div className="jn-row">
+        <button className="jn-btn primary" disabled={!ready || prog !== null || !ratios.length || !designs.length} onClick={gen}>生成 {ratios.length * designs.length} 张封面</button>
+        {items.length > 0 && (
+          <>
+            <button className="jn-btn" onClick={() => setSel(new Set(items.map((i) => i.key)))}>全选</button>
+            <button className="jn-btn" onClick={() => setSel(new Set())}>全不选</button>
+            <button className="jn-btn primary" disabled={!sel.size} onClick={dl}>下载选中（{sel.size}）</button>
+          </>
+        )}
+      </div>
+      {prog !== null && <div className="jn-progress"><div className="jn-bar"><i style={{ width: `${prog * 100}%` }} /></div><small>正在渲染封面…</small></div>}
+      <div className="jn-covers">
+        {items.map((it) => (
+          <div key={it.key} className={`jn-cover ${sel.has(it.key) ? 'sel' : ''}`} onClick={() => { const s = new Set(sel); if (s.has(it.key)) s.delete(it.key); else s.add(it.key); setSel(s) }}>
+            <img src={it.url} alt={it.name} />
+            <span className="chk">{sel.has(it.key) ? '✓' : ''}</span>
+            <div className="meta"><span>{it.ratio} · {it.name.split('_').pop()?.replace('.jpg', '')}</span><a href={it.url} download={it.name} onClick={(e) => e.stopPropagation()}>下载</a></div>
+          </div>
+        ))}
+      </div>
+    </section>
+  )
+}
+
+// ====================================================================== copy
+function CopySection({ say }: { say: (m: string) => void }) {
+  const [tab, setTab] = useState(PLATFORMS[0].id)
+  const [pick, setPick] = useState<Record<string, number>>({})
+  const p = PLATFORMS.find((x) => x.id === tab)!
+  const ti = pick[p.id] ?? 0
+  const title = p.titles[ti]
+  const tagStr = formatTags(p.tags, p.tagStyle)
+  const cp = async (t: string, what: string) => say((await copyText(t)) ? `已复制${what}` : '复制失败，请手动选择')
+  const tagOk = p.tagMax === null || p.tags.length <= p.tagMax
+  return (
+    <section className="jn-sec">
+      <h2>发布文案</h2>
+      <p className="lead">四个平台各自的标题、描述和话题，字数与话题数量已按平台限制校验，点一下即可复制。</p>
+      <div className="jn-tabs">
+        {PLATFORMS.map((x) => (
+          <button key={x.id} className={`jn-chip ${tab === x.id ? 'active' : ''}`} style={{ fontSize: 14, padding: '6px 14px' }} onClick={() => setTab(x.id)}>{x.name}</button>
+        ))}
+      </div>
+      <div className="jn-copyrow"><div className="jn-label">标题（{p.note}）· 点选一个</div><button className="jn-btn" onClick={() => cp(title, '标题')}>复制标题</button></div>
+      <div className="jn-titles">
+        {p.titles.map((t, i) => {
+          const n = chars(t)
+          return (
+            <div key={i} className={`jn-tline ${i === ti ? 'sel' : ''}`} onClick={() => setPick({ ...pick, [p.id]: i })}>
+              <span className="txt">{t}</span>
+              <span className={`jn-count ${n <= p.titleMax ? 'ok' : 'bad'}`}>{n}/{p.titleMax}</span>
+              <button className="jn-btn ghost" onClick={(e) => { e.stopPropagation(); cp(t, '标题') }}>复制</button>
+            </div>
+          )
+        })}
+      </div>
+      <div className="jn-copyrow"><div className="jn-label">描述</div><button className="jn-btn" onClick={() => cp(p.desc, '描述')}>复制描述</button></div>
+      <div className="jn-box">{p.desc}</div>
+      <div className="jn-copyrow">
+        <div className="jn-label">{p.tagStyle === 'plain' ? '标签' : '话题'}（{p.tags.length}{p.tagMax !== null ? `/${p.tagMax}` : '，不限'}）<span className={`jn-count ${tagOk ? 'ok' : 'bad'}`} style={{ marginLeft: 8 }}>{tagOk ? '符合限制' : '超出限制'}</span></div>
+        <button className="jn-btn" onClick={() => cp(tagStr, p.tagStyle === 'plain' ? '标签' : '话题')}>复制{p.tagStyle === 'plain' ? '标签' : '话题'}</button>
+      </div>
+      <div className="jn-tags">{p.tags.map((t) => <span key={t} className="jn-tag">{p.tagStyle === 'hash' ? '#' : ''}{t}</span>)}</div>
+      <div className="jn-row" style={{ marginTop: 18 }}>
+        <button className="jn-btn primary" onClick={() => cp(`${title}\n\n${p.desc}\n\n${tagStr}`, '全部文案')}>复制全部（标题 + 描述 + {p.tagStyle === 'plain' ? '标签' : '话题'}）</button>
+      </div>
+    </section>
+  )
+}
+
+// ====================================================================== live photos
+function LiveSection({ film, ready, withBusy, say }: { film: FilmRef; ready: boolean; withBusy: Busy; say: (m: string) => void }) {
+  const [prog, setProg] = useState<{ p: number; label: string } | null>(null)
+  const [res, setRes] = useState<LiveResult | null>(null)
+  const [err, setErr] = useState('')
+  const sig = useRef({ cancelled: false })
+  const sup = useMemo(() => exportSupport(), [])
+  const gen = async () => {
+    if (!film.current) return
+    setErr('')
+    sig.current = { cancelled: false }
+    try {
+      await withBusy(async () => {
+        const r = await generateLivePhotos(film.current!, (p, label) => setProg({ p, label }), sig.current)
+        setRes(r)
+      })
+    } catch (e) {
+      setErr((e as Error).message)
+    } finally {
+      setProg(null)
+    }
+  }
+  const cp = async (t: string, what: string) => say((await copyText(t)) ? `已复制${what}` : '复制失败')
+  return (
+    <section className="jn-sec">
+      <h2>实况图（Live Photo）</h2>
+      <p className="lead">按 LiveCanvas 规范生成 {LIVE_CLIPS.length} 组苹果实况照片：3:4（1080×1440）、30 fps、3 秒、封面取自动画中的同一帧。压缩包里是同名的 <b>jpg + mov</b>（Mac 拖进「照片」即成实况）以及 .pvt 实况资源包，可直接用于小红书、抖音图文。</p>
+      <div className="jn-row">
+        {!prog ? (
+          <button className="jn-btn primary" disabled={!ready || !sup.video} onClick={gen}>生成 {LIVE_CLIPS.length} 组实况图</button>
+        ) : (
+          <button className="jn-btn" onClick={() => { sig.current.cancelled = true }}>取消</button>
+        )}
+        {res && <button className="jn-btn primary" onClick={() => download(res.zip, '暮归江南_实况图.zip')}>下载压缩包（{(res.zip.size / 1048576).toFixed(1)} MB）</button>}
+      </div>
+      <div className="jn-note">画面：{LIVE_CLIPS.map((c) => c.name).join(' · ')}</div>
+      {prog && <div className="jn-progress"><div className="jn-bar"><i style={{ width: `${prog.p * 100}%` }} /></div><small>{prog.label}</small></div>}
+      {err && <div className="jn-err">生成失败：{err}</div>}
+      {res && (
+        <div className="jn-live">
+          {res.previews.map((p) => <figure key={p.name}><img src={p.url} alt={p.name} /><figcaption>{p.name}</figcaption></figure>)}
+        </div>
+      )}
+      {LIVE_COPY.map((c) => (
+        <div key={c.id} className="jn-card" style={{ marginTop: 16 }}>
+          <div className="jn-copyrow" style={{ marginTop: 0 }}><div style={{ fontSize: 17 }}>{c.name}</div><button className="jn-btn" onClick={() => cp(`${c.titles[0]}\n\n${c.body}\n\n${c.tags.map((t) => '#' + t).join(' ')}`, '全部图文文案')}>复制全部</button></div>
+          <div className="jn-label">标题（≤{c.titleMax} 字）</div>
+          <div className="jn-titles">
+            {c.titles.map((t) => (
+              <div key={t} className="jn-tline" onClick={() => cp(t, '标题')}>
+                <span className="txt">{t}</span>
+                <span className={`jn-count ${chars(t) <= c.titleMax ? 'ok' : 'bad'}`}>{chars(t)}/{c.titleMax}</span>
+                <span className="jn-note" style={{ margin: 0 }}>点击复制</span>
+              </div>
+            ))}
+          </div>
+          <div className="jn-copyrow"><div className="jn-label">正文</div><button className="jn-btn ghost" onClick={() => cp(c.body, '正文')}>复制正文</button></div>
+          <div className="jn-box">{c.body}</div>
+          <div className="jn-copyrow"><div className="jn-label">话题（{c.tags.length}/{c.tagMax}）</div><button className="jn-btn ghost" onClick={() => cp(c.tags.map((t) => '#' + t).join(' '), '话题')}>复制话题</button></div>
+          <div className="jn-tags">{c.tags.map((t) => <span key={t} className="jn-tag">#{t}</span>)}</div>
+        </div>
+      ))}
+      <p className="jn-note">说明：实况配对信息（Apple 内容标识 + still-image-time 定时元数据）完全在浏览器内写入，写入逻辑已用 LiveCanvas skill 原版 Python 校验器验证；导入「照片」与手机播放需在你的设备上完成。</p>
+      <TextDownload />
+    </section>
+  )
+}
+
+function TextDownload() {
+  const all = () => {
+    const txt = [...PLATFORMS.map((p) => `【${p.name}】（${p.note}）\n标题：\n${p.titles.map((t) => `· ${t}（${chars(t)}字）`).join('\n')}\n\n描述：\n${p.desc}\n\n${p.tagStyle === 'plain' ? '标签' : '话题'}：\n${formatTags(p.tags, p.tagStyle)}\n`), ...LIVE_COPY.map((c) => `【${c.name}】\n标题：\n${c.titles.map((t) => '· ' + t).join('\n')}\n\n正文：\n${c.body}\n\n话题：\n${c.tags.map((t) => '#' + t).join(' ')}\n`)].join('\n————————————\n\n')
+    download(new Blob([strToU8(txt) as Uint8Array<ArrayBuffer>], { type: 'text/plain;charset=utf-8' }), '暮归江南_发布文案.txt')
+  }
+  return <button className="jn-btn ghost" style={{ marginTop: 12 }} onClick={all}>下载全部平台文案（txt）</button>
+}

src/jiangnan/assets/fonts/title.woff2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63d01c7d2c75f65eccdc89e79f1cafb080f04af1869be4c2e1c875739b81a0a2
+size 152776

src/jiangnan/assets/fonts/wenkai.woff2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7316dfca75b5cbbee358a61eacb711e6cd44f476959c4fd63a0ea7362be2e322
+size 100608

src/jiangnan/assets/voice/n01.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9441bfce9568db7f6ee7d54d3ca272a95fc4bde4514e83e955e399889891d026
+size 19872

src/jiangnan/assets/voice/n02.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e2582ef2e3f91e54f4b60b05cd8cbb5237eb18018d0d73458c2c4e8ae8f4991
+size 27648

src/jiangnan/assets/voice/n03.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b072049639a1d41f97bd804a48b78a7e18b17420708f9aa36d75a1afea05f018
+size 29520

src/jiangnan/assets/voice/n04.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8297a0c1aec1717c713ed0ccd7ffa6348aa22d76cb2bdd87185c03723f1423fb
+size 35712

src/jiangnan/assets/voice/n05.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1df103680324a49b87156639dededc13a1b97ad144cb864981c656a14bac0ea3
+size 43200

src/jiangnan/assets/voice/n06.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f8a874dd78a5574064893c8ce372657b56c174592ab330dcb86062e929b9c22
+size 42480

src/jiangnan/assets/voice/n07.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce3191d4c78002855af9c038c40b98a471303c6b58ab75bb4ae460bb9e9115b1
+size 32400

src/jiangnan/assets/voice/n08.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41a49b0aae1a4235607c904a747d33742c28ca95fe1b141d6b4ff2ac52de826a
+size 37296

src/jiangnan/assets/voice/n09.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e90da6356f0dac3b0b8910a069082118888d320f4092dc90c9d4e6d5eb5f8c1
+size 24624

src/jiangnan/assets/voice/n10.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cde233982818af986713f36aa49a49d0917094f012aef38cb8a15bcf5642d0b
+size 42768

src/jiangnan/audio/mixer.ts

@@ -0,0 +1,140 @@
+// Offline mix: music + sound effects + narration → one stereo AudioBuffer.
+// The same buffer is used for live playback and for MP4 export.
+import { DURATION, NARRATION } from '../engine/timeline'
+import { composeMusic, composeSfx, SR, Stem } from './synth'
+
+const voiceUrls = import.meta.glob('../assets/voice/*.mp3', { eager: true, query: '?url', import: 'default' }) as Record<string, string>
+
+function impulse(ctx: BaseAudioContext, seconds: number, decay: number) {
+  const len = Math.floor(seconds * ctx.sampleRate)
+  const buf = ctx.createBuffer(2, len, ctx.sampleRate)
+  let s = 12345
+  const rnd = () => { s = (s * 1103515245 + 12345) & 0x7fffffff; return s / 0x7fffffff }
+  for (let c = 0; c < 2; c++) {
+    const d = buf.getChannelData(c)
+    let lp = 0
+    for (let i = 0; i < len; i++) {
+      const t = i / len
+      lp += ((rnd() * 2 - 1) - lp) * (0.6 - 0.45 * t) // darker tail
+      d[i] = lp * Math.pow(1 - t, decay) * (i < 40 ? i / 40 : 1)
+    }
+  }
+  return buf
+}
+
+function stemToBuffer(ctx: BaseAudioContext, s: Stem) {
+  const b = ctx.createBuffer(2, s.L.length, SR)
+  b.copyToChannel(s.L as Float32Array<ArrayBuffer>, 0)
+  b.copyToChannel(s.R as Float32Array<ArrayBuffer>, 1)
+  return b
+}
+
+export interface MixOptions { music: number; sfx: number; voice: number }
+
+let cachedStems: { music: Stem; sfx: Stem } | null = null
+let cachedVoices: Record<string, ArrayBuffer> | null = null
+
+export async function loadVoices() {
+  if (cachedVoices) return cachedVoices
+  const out: Record<string, ArrayBuffer> = {}
+  await Promise.all(Object.entries(voiceUrls).map(async ([path, url]) => {
+    const id = path.split('/').pop()!.replace('.mp3', '')
+    out[id] = await (await fetch(url)).arrayBuffer()
+  }))
+  cachedVoices = out
+  return out
+}
+
+export async function buildMix(opts: MixOptions = { music: 1, sfx: 1, voice: 1 }, progress?: (p: number) => void): Promise<AudioBuffer> {
+  if (!cachedStems) {
+    progress?.(0.05)
+    await new Promise((r) => setTimeout(r, 0))
+    const music = composeMusic()
+    progress?.(0.45)
+    await new Promise((r) => setTimeout(r, 0))
+    const sfx = composeSfx()
+    cachedStems = { music, sfx }
+  }
+  progress?.(0.7)
+  const voices = await loadVoices()
+  const len = Math.ceil(DURATION * SR)
+  const ctx = new OfflineAudioContext(2, len, SR)
+  const master = ctx.createGain()
+  master.gain.value = 0.9
+  const comp = ctx.createDynamicsCompressor()
+  comp.threshold.value = -10
+  comp.knee.value = 6
+  comp.ratio.value = 6
+  comp.attack.value = 0.004
+  comp.release.value = 0.2
+  master.connect(comp).connect(ctx.destination)
+  const verb = ctx.createConvolver()
+  verb.buffer = impulse(ctx, 3.4, 2.6)
+  const verbGain = ctx.createGain()
+  verbGain.gain.value = 0.32
+  verb.connect(verbGain).connect(master)
+
+  // music with ducking under narration
+  const mSrc = ctx.createBufferSource()
+  mSrc.buffer = stemToBuffer(ctx, cachedStems.music)
+  const mGain = ctx.createGain()
+  const g = mGain.gain
+  const base = 0.85 * opts.music
+  g.setValueAtTime(base, 0)
+  const sfxSrc = ctx.createBufferSource()
+  sfxSrc.buffer = stemToBuffer(ctx, cachedStems.sfx)
+  const sGain = ctx.createGain()
+  sGain.gain.setValueAtTime(0.9 * opts.sfx, 0)
+  for (const c of NARRATION) {
+    const a = Math.max(0, c.t - 0.35), b = c.t + c.dur - 0.2
+    g.setValueAtTime(base, a)
+    g.linearRampToValueAtTime(base * 0.5, a + 0.35)
+    g.setValueAtTime(base * 0.5, b)
+    g.linearRampToValueAtTime(base, b + 0.8)
+    sGain.gain.setValueAtTime(0.9 * opts.sfx, a)
+    sGain.gain.linearRampToValueAtTime(0.65 * opts.sfx, a + 0.35)
+    sGain.gain.setValueAtTime(0.65 * opts.sfx, b)
+    sGain.gain.linearRampToValueAtTime(0.9 * opts.sfx, b + 0.8)
+  }
+  mSrc.connect(mGain)
+  mGain.connect(master)
+  const mSend = ctx.createGain()
+  mSend.gain.value = 0.55
+  mGain.connect(mSend).connect(verb)
+  sfxSrc.connect(sGain)
+  sGain.connect(master)
+  const sSend = ctx.createGain()
+  sSend.gain.value = 0.2
+  sGain.connect(sSend).connect(verb)
+  mSrc.start(0)
+  sfxSrc.start(0)
+
+  // narration (decoded into the offline context)
+  const vGain = ctx.createGain()
+  vGain.gain.value = 1.25 * opts.voice
+  // a gentle presence lift + low cut for the voice
+  const hp = ctx.createBiquadFilter()
+  hp.type = 'highpass'
+  hp.frequency.value = 90
+  const pres = ctx.createBiquadFilter()
+  pres.type = 'peaking'
+  pres.frequency.value = 3200
+  pres.gain.value = 2.5
+  vGain.connect(hp).connect(pres).connect(master)
+  const vSend = ctx.createGain()
+  vSend.gain.value = 0.12
+  pres.connect(vSend).connect(verb)
+  for (const c of NARRATION) {
+    const data = voices[c.id]
+    if (!data) continue
+    const buf = await ctx.decodeAudioData(data.slice(0))
+    const src = ctx.createBufferSource()
+    src.buffer = buf
+    src.connect(vGain)
+    src.start(c.t)
+  }
+  progress?.(0.8)
+  const rendered = await ctx.startRendering()
+  progress?.(1)
+  return rendered
+}

src/jiangnan/audio/synth.ts

@@ -0,0 +1,467 @@
+// Original score + sound effects, synthesised sample-by-sample (deterministic).
+// Pentatonic D gong mode (宫调式): D E F# A B.
+import { DURATION, LANTERN_START } from '../engine/timeline'
+
+export const SR = 48000
+
+export interface Stem { L: Float32Array; R: Float32Array }
+const stem = (len: number): Stem => ({ L: new Float32Array(len), R: new Float32Array(len) })
+
+function rng(seed: number) {
+  let a = seed >>> 0
+  return () => {
+    a = (a + 0x6d2b79f5) >>> 0
+    let t = a
+    t = Math.imul(t ^ (t >>> 15), t | 1)
+    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)
+    return ((t ^ (t >>> 14)) >>> 0) / 4294967296
+  }
+}
+
+const midi = (m: number) => 440 * Math.pow(2, (m - 69) / 12)
+// note names → midi
+const N: Record<string, number> = {}
+;['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B'].forEach((n, i) => {
+  for (let o = 1; o < 8; o++) N[n + o] = 12 * (o + 1) + i
+})
+
+function addPan(s: Stem, i: number, v: number, pan: number) {
+  // equal-power pan, pan in [-1,1]
+  const a = (pan + 1) * Math.PI / 4
+  s.L[i] += v * Math.cos(a)
+  s.R[i] += v * Math.sin(a)
+}
+
+// ------------------------------------------------------------------ instruments
+
+/** Guzheng pluck: Karplus–Strong string with optional press-bend (按音). */
+function guzheng(out: Stem, t0: number, note: number, vel: number, pan: number, r: () => number, bend = 0, dur = 4.5) {
+  const f = midi(note)
+  const start = Math.floor(t0 * SR)
+  const len = Math.min(Math.floor(dur * SR), out.L.length - start)
+  if (len <= 0) return
+  const period = SR / f
+  const size = Math.ceil(period) + 2
+  const buf = new Float32Array(size)
+  // pluck excitation: bright noise shaped by pluck position
+  for (let i = 0; i < size; i++) buf[i] = (r() * 2 - 1) * (0.6 + 0.4 * Math.sin((i / size) * Math.PI))
+  let idx = 0
+  let prev = 0
+  const damp = 0.996 - Math.max(0, (note - 60)) * 0.0004
+  let lp = 0
+  for (let n = 0; n < len; n++) {
+    // pitch bend by modulating the read position (fractional delay)
+    const tb = n / SR
+    const bendAmt = bend ? bend * Math.min(1, Math.max(0, (tb - 0.18) / 0.22)) : 0
+    const p = period / Math.pow(2, bendAmt / 12)
+    const rd = (idx - p + size * 4) % size
+    const i0 = Math.floor(rd), fr = rd - i0
+    const s = buf[i0 % size] * (1 - fr) + buf[(i0 + 1) % size] * fr
+    const y = damp * 0.5 * (s + prev)
+    prev = s
+    buf[idx % size] = y
+    idx = (idx + 1) % size
+    // body: gentle lowpass + attack transient
+    lp += (y - lp) * 0.55
+    const env = Math.min(1, n / 40)
+    addPan(out, start + n, lp * vel * env * 0.55, pan)
+  }
+}
+
+/** Bamboo flute (xiao/dizi-like): sine + odd/even partials, breath noise, delayed vibrato. */
+function flute(out: Stem, t0: number, notes: [number, number][], beat: number, vel: number, pan: number, r: () => number) {
+  let t = t0
+  let phase = 0
+  let curF = midi(notes[0][0])
+  let lpN = 0, bpN = 0
+  for (let k = 0; k < notes.length; k++) {
+    const [note, beats] = notes[k]
+    const dur = beats * beat
+    const target = note < 0 ? curF : midi(note)
+    const s0 = Math.floor(t * SR), s1 = Math.floor((t + dur) * SR)
+    const rest = note < 0
+    for (let i = s0; i < s1 && i < out.L.length; i++) {
+      const tl = (i - s0) / SR
+      // portamento into the note
+      curF += (target - curF) * 0.0025
+      const vib = 1 + 0.006 * Math.min(1, Math.max(0, (tl - 0.35) / 0.4)) * Math.sin(2 * Math.PI * 5.2 * tl)
+      phase += (2 * Math.PI * curF * vib) / SR
+      const atk = Math.min(1, tl / 0.09)
+      const rel = Math.min(1, (dur - tl) / 0.14)
+      const swell = 0.85 + 0.15 * Math.sin(Math.PI * Math.min(1, tl / dur))
+      const env = rest ? 0 : Math.max(0, atk * rel * swell)
+      const tone = Math.sin(phase) + 0.18 * Math.sin(2 * phase) + 0.07 * Math.sin(3 * phase + 0.4)
+      // breath: noise band-limited around the fundamental, stronger at the attack
+      const nz = r() * 2 - 1
+      lpN += (nz - lpN) * 0.08
+      bpN += (lpN - bpN) * 0.02
+      const breath = (lpN - bpN) * (0.25 + 0.75 * Math.exp(-tl * 8))
+      addPan(out, i, (tone * 0.5 + breath * 0.9) * env * vel, pan)
+    }
+    t += dur
+  }
+}
+
+/** Warm string pad: detuned saws through a slow low-pass. */
+function pad(out: Stem, t0: number, t1: number, notes: number[], vel: number, cutoff: number, r: () => number) {
+  const s0 = Math.floor(t0 * SR), s1 = Math.min(out.L.length, Math.floor(t1 * SR))
+  const voices: { ph: number; inc: number; pan: number }[] = []
+  for (const n of notes) for (const det of [-0.08, 0, 0.07]) {
+    voices.push({ ph: r(), inc: midi(n + det) / SR, pan: (r() - 0.5) * 1.2 })
+  }
+  let l1 = 0, l2 = 0, r1 = 0, r2 = 0
+  const atk = 1.8, rel = 2.2
+  for (let i = s0; i < s1; i++) {
+    const t = (i - s0) / SR
+    const tt = (s1 - i) / SR
+    const env = Math.min(1, t / atk) * Math.min(1, tt / rel)
+    let L = 0, R = 0
+    for (const v of voices) {
+      v.ph += v.inc
+      if (v.ph > 1) v.ph -= 1
+      const saw = v.ph * 2 - 1
+      L += saw * (1 - v.pan) * 0.5
+      R += saw * (1 + v.pan) * 0.5
+    }
+    const c = (cutoff * (0.8 + 0.2 * Math.sin(t * 0.4))) / SR * 6.28
+    l1 += (L - l1) * c; l2 += (l1 - l2) * c
+    r1 += (R - r1) * c; r2 += (r1 - r2) * c
+    const g = (env * vel) / voices.length
+    out.L[i] += l2 * g
+    out.R[i] += r2 * g
+  }
+}
+
+/** Bronze bell (编钟-like): inharmonic partials with long decay. */
+function bell(out: Stem, t0: number, f: number, vel: number, pan: number, decay = 5) {
+  const parts = [[1, 1, 1], [2.76, 0.5, 0.6], [5.4, 0.25, 0.4], [8.93, 0.12, 0.25], [0.5, 0.35, 1.3]]
+  const s0 = Math.floor(t0 * SR)
+  const len = Math.min(Math.floor(decay * 1.6 * SR), out.L.length - s0)
+  for (let n = 0; n < len; n++) {
+    const t = n / SR
+    let v = 0
+    for (const [m, a, d] of parts) v += Math.sin(2 * Math.PI * f * m * t + m) * a * Math.exp((-t / decay) / d * 1.0)
+    const atk = Math.min(1, n / 60)
+    addPan(out, s0 + n, v * vel * atk * 0.25, pan)
+  }
+}
+
+/** Wooden clapper / muyu tick. */
+function woodTick(out: Stem, t0: number, f: number, vel: number, pan: number, r: () => number) {
+  const s0 = Math.floor(t0 * SR)
+  const len = Math.min(Math.floor(0.12 * SR), out.L.length - s0)
+  for (let n = 0; n < len; n++) {
+    const t = n / SR
+    const v = Math.sin(2 * Math.PI * f * t) * Math.exp(-t * 55) + (r() * 2 - 1) * Math.exp(-t * 300) * 0.3
+    addPan(out, s0 + n, v * vel, pan)
+  }
+}
+
+function bassNote(out: Stem, t0: number, dur: number, note: number, vel: number) {
+  const f = midi(note)
+  const s0 = Math.floor(t0 * SR), len = Math.min(Math.floor(dur * SR), out.L.length - s0)
+  for (let n = 0; n < len; n++) {
+    const t = n / SR
+    const env = Math.min(1, t / 0.4) * Math.min(1, (dur - t) / 0.8)
+    const v = (Math.sin(2 * Math.PI * f * t) + 0.25 * Math.sin(4 * Math.PI * f * t)) * env * vel
+    out.L[s0 + n] += v
+    out.R[s0 + n] += v
+  }
+}
+
+// ------------------------------------------------------------------ score
+
+export function composeMusic(): Stem {
+  const len = Math.ceil((DURATION + 1) * SR)
+  const out = stem(len)
+  const r = rng(1902)
+  const BPM = 72
+  const beat = 60 / BPM
+  const bar = beat * 4
+
+  // harmony: one chord per two bars (pentatonic-friendly voicings)
+  const chords: { root: number; notes: number[] }[] = [
+    { root: N.D2, notes: [N.D3, N.A3, N.E4, N['F#4']] },
+    { root: N.B1, notes: [N.B2, N['F#3'], N.D4, N.A4] },
+    { root: N.G1, notes: [N.G2, N.D3, N.B3, N['F#4']] },
+    { root: N.A1, notes: [N.A2, N.E3, N.B3, N.D4] },
+  ]
+  const start = 0.4
+  let ci = 0
+  for (let t = start; t < DURATION - 3; t += bar * 2, ci++) {
+    const c = chords[ci % 4]
+    const night = t > 62
+    const vel = t < 8 ? 0.22 : t > 78 ? 0.34 : night ? 0.26 : 0.3
+    pad(out, t, Math.min(DURATION + 0.5, t + bar * 2 + 1.6), c.notes, vel, night ? 900 : 1500, r)
+    if (t > 6) bassNote(out, t, bar * 2 + 0.2, c.root + 12, t > 48 ? 0.16 : 0.12)
+    // guzheng arpeggios (sparser in the intro and at night)
+    const arp = [c.notes[0] + 12, c.notes[1] + 12, c.notes[2] + 12, c.notes[3] + 12, c.notes[2] + 12, c.notes[1] + 24]
+    const density = t < 9 ? 3 : t < 32 ? 6 : t < 56 ? 8 : night && t < 74 ? 4 : 6
+    for (let k = 0; k < density; k++) {
+      const tt = t + (k * bar * 2) / density + (r() - 0.5) * 0.03
+      const note = arp[k % arp.length] + (r() < 0.12 ? 12 : 0)
+      guzheng(out, tt, note, 0.5 + r() * 0.25, -0.4 + r() * 0.3, r, r() < 0.15 ? 2 : 0)
+    }
+  }
+
+  // flute theme
+  const theme1: [number, number][] = [
+    [N.A4, 2], [N.B4, 1], [N.D5, 1], [N.E5, 3], [N.D5, 1], [N.B4, 2], [N.A4, 1], [N['F#4'], 1], [N.A4, 4],
+    [N.D5, 2], [N.E5, 1], [N['F#5'], 1], [N.E5, 2], [N.D5, 1], [N.B4, 1], [N.A4, 1.5], [N.B4, 0.5], [N.A4, 1], [N['F#4'], 1], [N.E4, 4],
+  ]
+  const theme2: [number, number][] = [
+    [N['F#5'], 2], [N.E5, 1], [N.D5, 1], [N.B4, 3], [N.A4, 1], [N.D5, 1], [N.B4, 1], [N.A4, 1], [N['F#4'], 1], [N.A4, 4],
+    [N.B4, 2], [N.D5, 1], [N.E5, 1], [N.D5, 2], [N.B4, 2], [N.A4, 1.5], [N['F#4'], 0.5], [N.E4, 1], [N['F#4'], 1], [N.D4, 4],
+  ]
+  flute(out, start + bar * 3, theme1, beat, 0.2, 0.15, r)
+  flute(out, start + bar * 3 + bar * 8, theme2, beat, 0.22, 0.15, r)
+  // night: lower, slower xiao phrase
+  const night: [number, number][] = [[N.A3, 3], [N.D4, 1], [N.E4, 2], [N['F#4'], 2], [N.E4, 3], [N.D4, 1], [N.B3, 4], [-1, 2], [N.D4, 2], [N.E4, 2], [N.A4, 4], [N['F#4'], 2], [N.E4, 2], [N.D4, 6]]
+  flute(out, 64.5, night, beat * 1.15, 0.2, -0.1, r)
+
+  // light percussion in the market / development section
+  for (let t = 32; t < 55; t += beat) {
+    const b = Math.round((t - 32) / beat) % 4
+    if (b === 0 || b === 2) woodTick(out, t, b === 0 ? 820 : 1150, 0.12, 0.3, r)
+  }
+  // sunset glissando & lantern bells
+  for (let k = 0; k < 14; k++) guzheng(out, 54.6 + k * 0.045, [N.D4, N.E4, N['F#4'], N.A4, N.B4][k % 5] + 12 * Math.floor(k / 5), 0.42, 0.3, r)
+  bell(out, LANTERN_START, midi(N.D5), 0.55, 0.2, 4)
+  bell(out, LANTERN_START + 2.2, midi(N.A4), 0.35, -0.3, 4)
+  bell(out, LANTERN_START + 4.1, midi(N['F#5']), 0.3, 0.4, 4)
+  // finale: rising gliss + low bell
+  for (let k = 0; k < 18; k++) guzheng(out, 77 + k * 0.06, [N.D4, N.E4, N['F#4'], N.A4, N.B4][k % 5] + 12 * Math.floor(k / 5), 0.38, -0.2 + k * 0.03, r)
+  bell(out, 80.2, midi(N.D3), 0.6, 0, 7)
+  bell(out, 80.25, midi(N.A4), 0.25, 0.2, 6)
+  // master fade
+  fadeEnds(out, 0.6, 3.5)
+  return out
+}
+
+function fadeEnds(s: Stem, inS: number, outS: number) {
+  const n = s.L.length
+  const a = Math.floor(inS * SR), b = Math.floor(outS * SR)
+  const end = Math.floor(DURATION * SR)
+  for (let i = 0; i < n; i++) {
+    let g = 1
+    if (i < a) g = i / a
+    if (i > end - b) g *= Math.max(0, (end - i) / b)
+    s.L[i] *= g
+    s.R[i] *= g
+  }
+}
+
+// ------------------------------------------------------------------ sound effects
+
+/** piecewise-linear automation helper */
+function auto(keys: [number, number][]) {
+  return (t: number) => {
+    if (t <= keys[0][0]) return keys[0][1]
+    for (let i = 1; i < keys.length; i++) {
+      if (t <= keys[i][0]) {
+        const f = (t - keys[i - 1][0]) / (keys[i][0] - keys[i - 1][0])
+        return keys[i - 1][1] + (keys[i][1] - keys[i - 1][1]) * f
+      }
+    }
+    return keys[keys.length - 1][1]
+  }
+}
+
+export function composeSfx(): Stem {
+  const len = Math.ceil((DURATION + 1) * SR)
+  const out = stem(len)
+  const r = rng(77)
+
+  // river: brown noise + gurgles, level follows the shots
+  const riverLv = auto([[0, 0.05], [8.5, 0.06], [9.2, 0.32], [17, 0.3], [25, 0.38], [32, 0.08], [40, 0.06], [48, 0.04], [55, 0.1], [65, 0.3], [74, 0.18], [86, 0.08]])
+  let b1 = 0, b2 = 0, hp = 0
+  for (let i = 0; i < len; i++) {
+    const t = i / SR
+    const w = r() * 2 - 1
+    b1 = (b1 + 0.02 * w) / 1.02
+    b2 += (b1 - b2) * 0.3
+    hp = b2 - hp * 0.0 // keep
+    const lv = riverLv(t)
+    const mod = 0.8 + 0.2 * Math.sin(t * 0.7) * Math.sin(t * 1.9)
+    out.L[i] += b1 * 3.2 * lv * mod
+    out.R[i] += b2 * 3.2 * lv * (1.6 - mod)
+  }
+  // gurgles: short rising sine blips
+  for (let t = 0.5; t < DURATION; t += 0.05 + r() * 0.25) {
+    const lv = riverLv(t)
+    if (lv < 0.1 || r() < 0.4) continue
+    const f0 = 250 + r() * 600
+    const s0 = Math.floor(t * SR), L = Math.floor((0.02 + r() * 0.05) * SR)
+    const pan = r() * 1.6 - 0.8
+    let ph = 0
+    for (let n = 0; n < L && s0 + n < len; n++) {
+      const f = f0 * (1 + n / L * 0.8)
+      ph += 2 * Math.PI * f / SR
+      addPan(out, s0 + n, Math.sin(ph) * Math.sin(Math.PI * n / L) * lv * 0.18, pan)
+    }
+  }
+
+  // water wheel (shot 4, audible from 23 s to 33 s): creaks + pouring trickle
+  const wheelLv = auto([[22.5, 0], [25, 1], [31.5, 1], [33, 0.1], [34, 0]])
+  for (let t = 23.0; t < 34; t += 1.25 + r() * 0.3) {
+    const lv = wheelLv(t)
+    if (lv <= 0) continue
+    const s0 = Math.floor(t * SR), L = Math.floor(0.55 * SR)
+    const f0 = 380 + r() * 220
+    let ph = 0
+    for (let n = 0; n < L && s0 + n < len; n++) {
+      const x = n / L
+      const f = f0 * (1 + 0.35 * Math.sin(x * Math.PI)) + Math.sin(n * 0.02) * 12
+      ph += 2 * Math.PI * f / SR
+      // stick-slip friction: pulse train amplitude-modulated
+      const stick = Math.pow(Math.abs(Math.sin(ph * 0.5)), 6)
+      const env = Math.sin(Math.PI * x) * (0.6 + 0.4 * Math.sin(x * 37))
+      addPan(out, s0 + n, (stick - 0.2) * env * lv * 0.1, -0.35)
+    }
+  }
+  {
+    let lp = 0
+    for (let i = Math.floor(23 * SR); i < Math.floor(34 * SR); i++) {
+      const t = i / SR
+      const w = r() * 2 - 1
+      lp += (w - lp) * 0.35
+      const splash = (w - lp) * (0.5 + 0.5 * Math.pow(Math.max(0, Math.sin(t * 2 * Math.PI / 1.1)), 4))
+      addPan(out, i, splash * wheelLv(t) * 0.09, -0.2)
+    }
+  }
+
+  // birds: chirps by day, egret calls when they fly by
+  const birdLv = auto([[0, 0.5], [17, 0.6], [40, 0.9], [52, 0.8], [57, 0.25], [62, 0]])
+  for (let t = 0.8; t < 60; t += 0.4 + r() * 1.6) {
+    const lv = birdLv(t)
+    if (lv <= 0.05) continue
+    const pan = r() * 1.6 - 0.8
+    const notes = 2 + Math.floor(r() * 4)
+    const base = 2600 + r() * 2200
+    for (let k = 0; k < notes; k++) {
+      const s0 = Math.floor((t + k * (0.07 + r() * 0.05)) * SR), L = Math.floor((0.04 + r() * 0.05) * SR)
+      let ph = 0
+      const sweep = (r() - 0.3) * 1800
+      for (let n = 0; n < L && s0 + n < len; n++) {
+        const x = n / L
+        ph += 2 * Math.PI * (base + sweep * x + 300 * Math.sin(x * 40)) / SR
+        addPan(out, s0 + n, Math.sin(ph) * Math.sin(Math.PI * x) * lv * 0.045, pan)
+      }
+    }
+  }
+  for (const t of [18.6, 20.1, 47.2, 49.0]) {
+    // egret: harsh low croak
+    const s0 = Math.floor(t * SR), L = Math.floor(0.32 * SR)
+    let ph = 0
+    for (let n = 0; n < L && s0 + n < len; n++) {
+      const x = n / L
+      ph += 2 * Math.PI * (520 - 140 * x) / SR
+      const buzz = Math.sign(Math.sin(ph)) * 0.4 + Math.sin(ph * 2)
+      addPan(out, s0 + n, buzz * Math.sin(Math.PI * x) * 0.035 * (1 + 0.5 * Math.sin(n * 0.05)), 0.4)
+    }
+  }
+  // wing beats as the egrets cross the river
+  for (let t = 17.5; t < 23; t += 0.38) woodLow(out, t, 0.05, r)
+
+  // market: ox bell, hooves, cart creak, footsteps (shot 5)
+  for (let t = 31.5; t < 41; t += 1.05) {
+    const lv = t < 32.5 ? (t - 31.5) : t > 39.5 ? Math.max(0, 41 - t) / 1.5 : 1
+    cowbell(out, t + 0.02 * Math.sin(t), 0.12 * lv, 0.15)
+    woodTick(out, t + 0.5, 300, 0.05 * lv, -0.1, r)
+    woodTick(out, t + 0.02, 260, 0.05 * lv, 0.1, r)
+  }
+  for (let t = 32; t < 40; t += 0.55 + r() * 0.1) woodTick(out, t, 1800 + r() * 400, 0.025, r() - 0.5, r) // footsteps
+
+  // bamboo / leaves rustle (shots 6-7)
+  const windLv = auto([[38, 0], [41, 0.6], [48, 0.5], [55, 0.35], [60, 0.1], [86, 0.12]])
+  {
+    let a = 0, b = 0
+    for (let i = Math.floor(38 * SR); i < len; i++) {
+      const t = i / SR
+      const w = r() * 2 - 1
+      a += (w - a) * 0.25
+      b += (a - b) * 0.05
+      const gust = 0.5 + 0.5 * Math.sin(t * 0.9) * Math.sin(t * 0.37 + 1)
+      const v = (a - b) * windLv(t) * gust * 0.22
+      out.L[i] += v
+      out.R[i] += v * (0.7 + 0.3 * Math.sin(t))
+    }
+  }
+
+  // temple bell far away at dusk + lantern ignitions
+  bellFar(out, 55.6, 0.22)
+  for (const t of [LANTERN_START, LANTERN_START + 0.7, LANTERN_START + 1.3, LANTERN_START + 1.9, LANTERN_START + 2.4]) whoomp(out, t, 0.08, r)
+
+  // night: crickets and frogs from ~61 s
+  const nightLv = auto([[60, 0], [64, 0.8], [86, 1]])
+  for (let t = 60.5; t < DURATION; t += 0.18 + r() * 0.5) {
+    const lv = nightLv(t)
+    const f = 4200 + r() * 900
+    const s0 = Math.floor(t * SR), L = Math.floor((0.08 + r() * 0.12) * SR)
+    const pan = r() * 1.8 - 0.9
+    for (let n = 0; n < L && s0 + n < len; n++) {
+      const tt = n / SR
+      const am = Math.max(0, Math.sin(2 * Math.PI * 34 * tt))
+      addPan(out, s0 + n, Math.sin(2 * Math.PI * f * tt) * am * Math.sin(Math.PI * n / L) * lv * 0.02, pan)
+    }
+  }
+  for (let t = 62; t < DURATION - 1; t += 0.9 + r() * 2.2) {
+    const lv = nightLv(t)
+    const s0 = Math.floor(t * SR), L = Math.floor(0.22 * SR)
+    const f = 180 + r() * 90
+    const pan = r() * 1.4 - 0.7
+    for (let n = 0; n < L && s0 + n < len; n++) {
+      const tt = n / SR
+      const pulse = Math.max(0, Math.sin(2 * Math.PI * 22 * tt))
+      addPan(out, s0 + n, Math.sin(2 * Math.PI * f * tt) * pulse * Math.sin(Math.PI * n / L) * lv * 0.06, pan)
+    }
+  }
+  // oar strokes of the night boat (shot 9)
+  for (let t = 64.8; t < 75; t += 2.9) splash(out, t, 0.07, r)
+  fadeEnds(out, 1.0, 2.5)
+  return out
+}
+
+function woodLow(out: Stem, t0: number, vel: number, r: () => number) {
+  const s0 = Math.floor(t0 * SR), L = Math.floor(0.12 * SR)
+  let lp = 0
+  for (let n = 0; n < L && s0 + n < out.L.length; n++) {
+    lp += ((r() * 2 - 1) - lp) * 0.04
+    addPan(out, s0 + n, lp * Math.sin(Math.PI * n / L) * vel * 3, 0.2)
+  }
+}
+function cowbell(out: Stem, t0: number, vel: number, pan: number) {
+  const s0 = Math.floor(t0 * SR), L = Math.floor(0.7 * SR)
+  for (let n = 0; n < L && s0 + n < out.L.length; n++) {
+    const t = n / SR
+    const v = (Math.sign(Math.sin(2 * Math.PI * 562 * t)) * 0.3 + Math.sin(2 * Math.PI * 845 * t) + 0.5 * Math.sin(2 * Math.PI * 1310 * t)) * Math.exp(-t * 7)
+    addPan(out, s0 + n, v * vel, pan)
+  }
+}
+function bellFar(out: Stem, t0: number, vel: number) {
+  const s0 = Math.floor(t0 * SR), L = Math.floor(7 * SR)
+  for (let n = 0; n < L && s0 + n < out.L.length; n++) {
+    const t = n / SR
+    const v = (Math.sin(2 * Math.PI * 98 * t) + 0.6 * Math.sin(2 * Math.PI * 98 * 2.41 * t) * Math.exp(-t * 0.6) + 0.3 * Math.sin(2 * Math.PI * 98 * 3.9 * t) * Math.exp(-t * 1.2)) * Math.exp(-t * 0.45) * (1 + 0.15 * Math.sin(2 * Math.PI * 1.3 * t))
+    addPan(out, s0 + n, v * vel * Math.min(1, n / 200), -0.5)
+  }
+}
+function whoomp(out: Stem, t0: number, vel: number, r: () => number) {
+  const s0 = Math.floor(t0 * SR), L = Math.floor(0.45 * SR)
+  let lp = 0
+  for (let n = 0; n < L && s0 + n < out.L.length; n++) {
+    const x = n / L
+    lp += ((r() * 2 - 1) - lp) * (0.02 + 0.1 * x)
+    addPan(out, s0 + n, lp * Math.sin(Math.PI * x) * vel * 4, 0.1)
+  }
+}
+function splash(out: Stem, t0: number, vel: number, r: () => number) {
+  const s0 = Math.floor(t0 * SR), L = Math.floor(0.8 * SR)
+  let lp = 0
+  for (let n = 0; n < L && s0 + n < out.L.length; n++) {
+    const x = n / L
+    const w = r() * 2 - 1
+    lp += (w - lp) * 0.3
+    addPan(out, s0 + n, (w - lp) * Math.exp(-x * 5) * vel * (0.6 + 0.4 * Math.sin(n * 0.001)), 0.3)
+  }
+}

src/jiangnan/blender/jiangnan_village.py

@@ -0,0 +1,842 @@
+# -*- coding: utf-8 -*-
+"""
+《暮归江南》Blender 场景生成脚本  (Blender 4.0+ / 4.2 LTS / 4.x)
+
+用法：Blender → Scripting 工作区 → 打开本文件 → Run Script。
+脚本会清空当前场景，然后从零生成：
+  · 山谷地形、远山、梯田、蜿蜒小河（反射水面）、石驳岸、青石板街
+  · 徽派民居（粉墙、阶梯马头墙、黛瓦凹曲屋面、木格窗，窗户夜里亮灯）
+  · 半圆石拱桥、老水车（旋转动画）、乌篷船、集市摊位、村口古樟、石碾、六角亭、竹林
+  · 红灯笼（日落后按顺序逐个亮起）、炊烟（烟雾体积可选）
+  · Nishita 物理天空 + 太阳从 8.5° 落到 -10°（与网页版同一条日落曲线）、蓝调、星光
+  · 10 个镜头的摄像机与时间轴标记（自动切换镜头），24 fps，共 86 秒
+与网页版共用同一套布局算法（同一条河道、同一个随机种子），所以村子的样子一致。
+
+渲染：默认 Cycles（GPU 可在 偏好设置 → 系统 里开启），1920×1080，128 采样 + 降噪。
+想先快速预览，可把 RENDER_ENGINE 改成 'BLENDER_EEVEE_NEXT'（4.2+）或 'BLENDER_EEVEE'。
+"""
+import bpy
+import bmesh
+import math
+import random
+from mathutils import Vector, Matrix, Euler
+
+FPS = 24
+DURATION = 86.0
+RENDER_ENGINE = 'CYCLES'
+SEED = 20261022
+GROUND_Y = 1.7  # web units: y-up metres.  Blender is z-up: (x, y, z)_web -> (x, -z, y)_blender
+
+
+def W2B(x, y, z):
+    """web (x, y-up, z) -> blender (x, y, z-up)"""
+    return Vector((x, -z, y))
+
+
+# ------------------------------------------------------------------ deterministic noise (same as the web build)
+def mulberry32(seed):
+    a = seed & 0xFFFFFFFF
+
+    def rnd():
+        nonlocal a
+        a = (a + 0x6D2B79F5) & 0xFFFFFFFF
+        t = a
+        t = ((t ^ (t >> 15)) * (t | 1)) & 0xFFFFFFFF
+        t ^= (t + (((t ^ (t >> 7)) * (t | 61)) & 0xFFFFFFFF)) & 0xFFFFFFFF
+        return ((t ^ (t >> 14)) & 0xFFFFFFFF) / 4294967296.0
+    return rnd
+
+
+def _imul(a, b):
+    return ((a & 0xFFFFFFFF) * (b & 0xFFFFFFFF)) & 0xFFFFFFFF
+
+
+def hash2(x, y, seed=0):
+    h = (_imul(int(x), 374761393) + _imul(int(y), 668265263) + _imul(int(seed), 2246822519)) & 0xFFFFFFFF
+    h = _imul(h ^ (h >> 13), 1274126177)
+    h ^= h >> 16
+    return (h & 0xFFFFFFFF) / 4294967296.0
+
+
+def fade(t):
+    return t * t * (3 - 2 * t)
+
+
+def vnoise(x, y, seed=0):
+    xi, yi = math.floor(x), math.floor(y)
+    xf, yf = x - xi, y - yi
+    a, b = hash2(xi, yi, seed), hash2(xi + 1, yi, seed)
+    c, d = hash2(xi, yi + 1, seed), hash2(xi + 1, yi + 1, seed)
+    u, v = fade(xf), fade(yf)
+    return (a + (b - a) * u) * (1 - v) + (c + (d - c) * u) * v
+
+
+def fbm(x, y, octv=5, seed=0):
+    amp, f, s, n = 0.5, 1.0, 0.0, 0.0
+    for i in range(octv):
+        s += amp * vnoise(x * f, y * f, seed + i * 31)
+        n += amp
+        amp *= 0.5
+        f *= 2.03
+    return s / n
+
+
+def smoothstep(a, b, v):
+    t = min(1.0, max(0.0, (v - a) / (b - a)))
+    return t * t * (3 - 2 * t)
+
+
+def lerp(a, b, t):
+    return a + (b - a) * t
+
+
+# ------------------------------------------------------------------ river (same control points as layout.ts)
+RIVER_CTRL = [(-1500, -330), (-1050, -260), (-760, -150), (-520, -70), (-330, -22), (-190, 6), (-80, 14),
+              (30, 4), (120, -10), (220, 6), (330, 46), (470, 92), (660, 104), (900, 60), (1300, -40), (1700, -120)]
+
+
+def catmull(p0, p1, p2, p3, t):
+    t2, t3 = t * t, t * t * t
+    return 0.5 * (2 * p1 + (-p0 + p2) * t + (2 * p0 - 5 * p1 + 4 * p2 - p3) * t2 + (-p0 + 3 * p1 - 3 * p2 + p3) * t3)
+
+
+def build_river():
+    raw = []
+    n = len(RIVER_CTRL)
+    for i in range(n - 1):
+        p0, p1, p2, p3 = RIVER_CTRL[max(0, i - 1)], RIVER_CTRL[i], RIVER_CTRL[i + 1], RIVER_CTRL[min(n - 1, i + 2)]
+        for k in range(80):
+            t = k / 80
+            raw.append((catmull(p0[0], p1[0], p2[0], p3[0], t), catmull(p0[1], p1[1], p2[1], p3[1], t)))
+    raw.append(RIVER_CTRL[-1])
+    cum = [0.0]
+    for i in range(1, len(raw)):
+        cum.append(cum[-1] + math.hypot(raw[i][0] - raw[i - 1][0], raw[i][1] - raw[i - 1][1]))
+    out, j, s = [], 0, 0.0
+    while s <= cum[-1]:
+        while j < len(cum) - 2 and cum[j + 1] < s:
+            j += 1
+        f = (s - cum[j]) / max(1e-6, cum[j + 1] - cum[j])
+        out.append([lerp(raw[j][0], raw[j + 1][0], f), lerp(raw[j][1], raw[j + 1][1], f), 0.0, 0.0, s])
+        s += 2.0
+    for i in range(len(out)):
+        a, b = out[max(0, i - 2)], out[min(len(out) - 1, i + 2)]
+        l = math.hypot(b[0] - a[0], b[1] - a[1]) or 1
+        out[i][2] = (b[0] - a[0]) / l
+        out[i][3] = (b[1] - a[1]) / l
+    return out
+
+
+RIVER = build_river()
+
+
+def river_s_at_x(x):
+    return min(RIVER, key=lambda p: abs(p[0] - x))[4]
+
+
+def river_point(s, off):
+    f = min(max(s / 2.0, 0), len(RIVER) - 1.001)
+    i = int(f)
+    t = f - i
+    a, b = RIVER[i], RIVER[i + 1]
+    tx, tz = lerp(a[2], b[2], t), lerp(a[3], b[3], t)
+    l = math.hypot(tx, tz) or 1
+    tx, tz = tx / l, tz / l
+    nx, nz = -tz, tx
+    return lerp(a[0], b[0], t) + nx * off, lerp(a[1], b[1], t) + nz * off, tx, tz, nx, nz
+
+
+VILLAGE_S0 = river_s_at_x(-300)
+VILLAGE_S1 = river_s_at_x(290)
+
+
+def village_factor(s):
+    return smoothstep(VILLAGE_S0 - 30, VILLAGE_S0 + 10, s) * (1 - smoothstep(VILLAGE_S1 - 10, VILLAGE_S1 + 30, s))
+
+
+def river_half_width(s):
+    inv = smoothstep(VILLAGE_S0 - 120, VILLAGE_S0, s) * (1 - smoothstep(VILLAGE_S1, VILLAGE_S1 + 120, s))
+    return lerp(13 + 3 * vnoise(s * 0.01, 3), 8, inv)
+
+
+def river_query(x, z):
+    best, bi = 1e18, 0
+    for i in range(0, len(RIVER), 3):
+        p = RIVER[i]
+        d = (p[0] - x) ** 2 + (p[1] - z) ** 2
+        if d < best:
+            best, bi = d, i
+    for i in range(max(0, bi - 3), min(len(RIVER), bi + 4)):
+        p = RIVER[i]
+        d = (p[0] - x) ** 2 + (p[1] - z) ** 2
+        if d < best:
+            best, bi = d, i
+    return math.sqrt(best), RIVER[bi][4]
+
+
+def terrain_height(x, z):
+    d, s = river_query(x, z)
+    h = GROUND_Y + (fbm(x * 0.012, z * 0.012, 3, 5) - 0.5) * 1.6 * smoothstep(50, 140, d)
+    far = math.hypot(x * 0.8, z)
+    hd = min(d, far * 0.9 + 40)
+    hill = smoothstep(130, 560, hd)
+    h += hill * hill * (60 + 190 * fbm(x * 0.0022 + 3, z * 0.0022, 5, 7))
+    w = river_half_width(s)
+    if d < w + 30:
+        bank_w = lerp(9, 1.2, village_factor(s))
+        bank = smoothstep(w - bank_w * 0.6, w + bank_w * 0.4, d)
+        h = lerp(-2.6, h, bank)
+    return h
+
+
+# ------------------------------------------------------------------ helpers
+def clear_scene():
+    bpy.ops.object.select_all(action='SELECT')
+    bpy.ops.object.delete(use_global=False)
+    for coll in (bpy.data.meshes, bpy.data.materials, bpy.data.lights, bpy.data.cameras, bpy.data.curves):
+        for b in list(coll):
+            if b.users == 0:
+                coll.remove(b)
+
+
+def collection(name):
+    c = bpy.data.collections.get(name) or bpy.data.collections.new(name)
+    if c.name not in bpy.context.scene.collection.children:
+        bpy.context.scene.collection.children.link(c)
+    return c
+
+
+def mesh_object(name, verts, faces, coll, mat=None, smooth=False):
+    me = bpy.data.meshes.new(name)
+    me.from_pydata(verts, [], faces)
+    me.update()
+    ob = bpy.data.objects.new(name, me)
+    coll.objects.link(ob)
+    if mat:
+        ob.data.materials.append(mat)
+    if smooth:
+        for p in me.polygons:
+            p.use_smooth = True
+    return ob
+
+
+def principled(name, color, rough=0.8, metal=0.0, emission=None, strength=0.0, alpha=1.0):
+    m = bpy.data.materials.new(name)
+    m.use_nodes = True
+    bsdf = m.node_tree.nodes.get('Principled BSDF')
+    bsdf.inputs['Base Color'].default_value = (*color, 1)
+    bsdf.inputs['Roughness'].default_value = rough
+    bsdf.inputs['Metallic'].default_value = metal
+    if emission is not None:
+        key = 'Emission Color' if 'Emission Color' in bsdf.inputs else 'Emission'
+        bsdf.inputs[key].default_value = (*emission, 1)
+        bsdf.inputs['Emission Strength'].default_value = strength
+    return m
+
+
+def noise_material(name, c1, c2, scale, rough=0.85, bump=0.3):
+    """Procedural weathered surface (lime plaster, tiles, stone)."""
+    m = bpy.data.materials.new(name)
+    m.use_nodes = True
+    nt = m.node_tree
+    bsdf = nt.nodes.get('Principled BSDF')
+    tex = nt.nodes.new('ShaderNodeTexNoise')
+    tex.inputs['Scale'].default_value = scale
+    tex.inputs['Detail'].default_value = 8
+    ramp = nt.nodes.new('ShaderNodeValToRGB')
+    ramp.color_ramp.elements[0].color = (*c1, 1)
+    ramp.color_ramp.elements[1].color = (*c2, 1)
+    nt.links.new(tex.outputs['Fac'], ramp.inputs['Fac'])
+    nt.links.new(ramp.outputs['Color'], bsdf.inputs['Base Color'])
+    bsdf.inputs['Roughness'].default_value = rough
+    bmp = nt.nodes.new('ShaderNodeBump')
+    bmp.inputs['Strength'].default_value = bump
+    nt.links.new(tex.outputs['Fac'], bmp.inputs['Height'])
+    nt.links.new(bmp.outputs['Normal'], bsdf.inputs['Normal'])
+    return m
+
+
+def add_box(bm, cx, cy, cz, sx, sy, sz, mat_index=0, matrix=None):
+    """box in web local coords (x, y-up, z) -> appended to bm (blender coords)."""
+    res = bmesh.ops.create_cube(bm, size=1.0)
+    vs = res['verts']
+    for v in vs:
+        lx, ly, lz = v.co.x * sx + cx, v.co.y * sy + cy, v.co.z * sz + cz  # cube axes treated as web axes
+        p = Vector((lx, ly, lz))
+        if matrix is not None:
+            p = matrix @ p
+        v.co = W2B(p.x, p.y, p.z)
+    for f in {f for v in vs for f in v.link_faces}:
+        f.material_index = mat_index
+    return vs
+
+
+def web_matrix(x, y, z, ry):
+    return Matrix.Translation((x, y, z)) @ Matrix.Rotation(ry, 4, 'Y')
+
+
+# ------------------------------------------------------------------ build
+def build_materials():
+    M = {}
+    M['wall'] = noise_material('粉墙', (0.78, 0.76, 0.71), (0.93, 0.92, 0.88), 6.0, 0.92, 0.15)
+    M['roof'] = noise_material('黛瓦', (0.035, 0.037, 0.042), (0.08, 0.085, 0.095), 40.0, 0.55, 0.6)
+    M['dark'] = principled('墨色', (0.03, 0.03, 0.032), 0.7)
+    M['wood'] = noise_material('老木', (0.12, 0.06, 0.03), (0.2, 0.11, 0.05), 20.0, 0.75, 0.2)
+    M['stone'] = noise_material('青石', (0.22, 0.23, 0.23), (0.36, 0.37, 0.36), 9.0, 0.85, 0.5)
+    M['window'] = principled('窗纸', (0.45, 0.4, 0.32), 0.9, emission=(1.0, 0.55, 0.22), strength=0.0)
+    M['lantern'] = principled('红灯笼', (0.7, 0.06, 0.03), 0.5, emission=(1.0, 0.32, 0.1), strength=0.0)
+    M['leaf'] = noise_material('樟叶', (0.04, 0.09, 0.025), (0.09, 0.16, 0.04), 30.0, 0.7, 0.4)
+    M['bark'] = noise_material('树皮', (0.05, 0.035, 0.025), (0.12, 0.09, 0.06), 25.0, 0.95, 0.8)
+    M['ground'] = noise_material('田野', (0.06, 0.1, 0.03), (0.2, 0.18, 0.07), 0.08, 0.95, 0.3)
+    M['bamboo'] = principled('竹', (0.18, 0.25, 0.07), 0.6)
+    M['cloth'] = principled('蓝印花布', (0.04, 0.08, 0.25), 0.9)
+    # water: glossy reflection with a jade body colour
+    w = bpy.data.materials.new('河水')
+    w.use_nodes = True
+    b = w.node_tree.nodes.get('Principled BSDF')
+    b.inputs['Base Color'].default_value = (0.01, 0.03, 0.025, 1)
+    b.inputs['Roughness'].default_value = 0.03
+    b.inputs['IOR'].default_value = 1.33
+    tex = w.node_tree.nodes.new('ShaderNodeTexNoise')
+    tex.inputs['Scale'].default_value = 3.0
+    bmp = w.node_tree.nodes.new('ShaderNodeBump')
+    bmp.inputs['Strength'].default_value = 0.08
+    w.node_tree.links.new(tex.outputs['Fac'], bmp.inputs['Height'])
+    w.node_tree.links.new(bmp.outputs['Normal'], b.inputs['Normal'])
+    M['water'] = w
+    return M
+
+
+def build_terrain(M, coll, step=6.0, extent=900):
+    verts, faces = [], []
+    n = int(2 * extent / step) + 1
+    for j in range(n):
+        for i in range(n):
+            x, z = -extent + i * step, -extent + j * step
+            verts.append(W2B(x, terrain_height(x, z), z))
+    for j in range(n - 1):
+        for i in range(n - 1):
+            a = j * n + i
+            faces.append((a, a + 1, a + n + 1, a + n))
+    ob = mesh_object('山谷地形', verts, faces, coll, M['ground'], smooth=True)
+    # far mountains ring
+    mverts, mfaces = [], []
+    rings = [extent * 1.05 * (1.12 ** k) for k in range(22)]
+    seg = 240
+    for r in rings:
+        for a in range(seg):
+            ang = a / seg * math.tau
+            x, z = math.cos(ang) * r, math.sin(ang) * r
+            d_sun = math.atan2(math.sin(ang - (math.pi + 0.2)), math.cos(ang - (math.pi + 0.2)))
+            west = math.exp(-(d_sun ** 2) / 0.5)
+            h = smoothstep(700, 2600, r) * (150 + 360 * fbm(x * 0.0009, z * 0.0009, 6, 77)) * (1 - 0.74 * west) + terrain_height(x, z) * 0.2
+            mverts.append(W2B(x, h, z))
+    for ri in range(len(rings) - 1):
+        for a in range(seg):
+            p, q = ri * seg + a, ri * seg + (a + 1) % seg
+            mfaces.append((p, q, q + seg, p + seg))
+    mesh_object('远山', mverts, mfaces, coll, M['ground'], smooth=True)
+    return ob
+
+
+def build_river(M, coll):
+    verts, faces = [], []
+    rows = 0
+    for k in range(0, len(RIVER), 2):
+        s = RIVER[k][4]
+        w = river_half_width(s) + 3
+        for off in (-w, w):
+            x, z, *_ = river_point(s, off)
+            verts.append(W2B(x, 0.0, z))
+        rows += 1
+    for r in range(rows - 1):
+        a = r * 2
+        faces.append((a, a + 1, a + 3, a + 2))
+    mesh_object('小河', verts, faces, coll, M['water'])
+
+
+def build_houses(M, coll):
+    """Same row / spacing rules and RNG as buildHouses() in layout.ts."""
+    rng = mulberry32(SEED)
+    bridge_s, wheel_s = river_s_at_x(40), river_s_at_x(-196)
+    rows = [(-1, 14.6, True, 0.3), (1, 8.4, False, 0.2), (-1, 31, False, 2.5), (1, 25, False, 2.5), (-1, 50, False, 5), (1, 44, False, 5)]
+    houses = []
+    for side, off0, shop, jitter in rows:
+        s = VILLAGE_S0 + rng() * 6 + (10 if off0 > 20 else 0)
+        s_end = VILLAGE_S1 - (40 if off0 > 40 else 0)
+        while s < s_end:
+            w = 8 + rng() * 5
+            half = w / 2
+            sc = s + half
+            if off0 > 40 and rng() < 0.3:
+                s += w + 6
+                continue
+            blocked = abs(sc - bridge_s) < half + 8 or (side > 0 and abs(sc - wheel_s) < half + 10) or (side < 0 and sc < river_s_at_x(-264) + half)
+            if not blocked:
+                d = 8.5 + rng() * 3.5
+                off = side * (off0 + d / 2 + (rng() - 0.5) * jitter)
+                x, z, tx, tz, nx, nz = river_point(sc, off)
+                facing = math.atan2(-side * nx, -side * nz)
+                floors = 2 if rng() < (0.75 if shop else 0.55) else 1
+                rot = facing + (rng() - 0.5) * 0.05 * (3 if off0 > 20 else 1)
+                steps = (3 if rng() < 0.6 else 2) if floors == 2 else (2 if rng() < 0.5 else 1)
+                seed = int(rng() * 1e9)
+                chimney = rng() < 0.35
+                houses.append(dict(x=x, z=z, rot=rot, w=w, d=d, floors=floors, steps=steps, shop=shop, side=side, s=sc, chimney=chimney, lantern=off0 < 20))
+            s += w + ((0.2 + rng() * 0.6) if shop else (0.6 + rng() * 3.5))
+            if side > 0 and off0 < 10 and rng() < 0.18:
+                s += 5
+    bm = bmesh.new()
+    windows = []
+    lanterns = []
+    for h in houses:
+        m = web_matrix(h['x'], max(GROUND_Y, terrain_height(h['x'], h['z'])), h['z'], h['rot'])
+        w, d = h['w'], h['d']
+        H = h['floors'] * 3.3 + 0.5
+        rise = d * 0.27
+        add_box(bm, 0, H / 2 - 0.8, 0, w - 0.4, H + 1.6, d, 0, m)  # lime walls
+        # stepped horse-head gables
+        n = h['steps']
+        seg = d / 2 / n
+        for sx in (-1, 1):
+            for i in range(n):
+                top = H + rise * (i + 1) / n + 0.55
+                z0 = d / 2 - i * seg if i else d / 2 + 0.15
+                z1 = d / 2 - (i + 1) * seg if i < n - 1 else -(d / 2 - i * seg if i else d / 2 + 0.15)
+                zc, zl = (z0 + z1) / 2, abs(z0 - z1)
+                for mirror in ((1,) if i == n - 1 else (1, -1)):
+                    add_box(bm, sx * (w / 2 - 0.18), top / 2 - 0.8, mirror * zc, 0.36, top + 1.6, zl, 0, m)
+                    add_box(bm, sx * (w / 2 - 0.18), top + 0.1, mirror * zc, 0.8, 0.2, zl + 0.3, 2, m)
+        # roof: two slopes (tiles)
+        for zs in (1, -1):
+            ang = math.atan2(rise, d / 2 + 0.8)
+            L = math.hypot(d / 2 + 0.85, rise)
+            rm = m @ Matrix.Translation((0, H + rise / 2 - 0.1, zs * (d / 4 + 0.42))) @ Matrix.Rotation(zs * ang, 4, 'X')
+            add_box(bm, 0, 0, 0, w - 0.6, 0.12, L, 1, rm)
+        add_box(bm, 0, H + rise + 0.1, 0, w - 0.6, 0.3, 0.4, 2, m)  # ridge
+        # windows (emissive at night)
+        nw = max(2, int(w / 3.2))
+        for k in range(nw):
+            x = -w / 2 + (k + 0.5) * w / nw
+            windows.append((m, x, (3.3 + 1.7) if h['floors'] == 2 else 1.9, d / 2 + 0.03, h))
+        if h['lantern']:
+            for fx in (-0.3, 0.3):
+                p = m @ Vector((fx * w, 3.2 if h['floors'] == 2 else H - 0.6, d / 2 + 0.9))
+                lanterns.append((p, h['s'], h['side']))
+    me = bpy.data.meshes.new('徽派民居')
+    bm.to_mesh(me)
+    bm.free()
+    ob = bpy.data.objects.new('徽派民居', me)
+    for key in ('wall', 'roof', 'dark'):
+        ob.data.materials.append(M[key])
+    coll.objects.link(ob)
+    # window panes as one emissive mesh
+    bmw = bmesh.new()
+    for (m, x, y, z, h) in windows:
+        add_box(bmw, x, y, z, 1.1, 1.0, 0.04, 0, m)
+    mw = bpy.data.meshes.new('窗灯')
+    bmw.to_mesh(mw)
+    bmw.free()
+    ow = bpy.data.objects.new('窗灯', mw)
+    ow.data.materials.append(M['window'])
+    coll.objects.link(ow)
+    return houses, lanterns
+
+
+def build_bridge(M, coll):
+    s = river_s_at_x(40)
+    x, z, tx, tz, nx, nz = river_point(s, 0)
+    w = river_half_width(s)
+    R = w - 1.6
+    bpy.ops.mesh.primitive_cube_add(size=1)
+    body = bpy.context.active_object
+    body.name = '石拱桥'
+    body.scale = (2 * (w + 9), 4.4, R + 3.0)
+    body.location = (0, 0, (R + 3.0) / 2 - 2.2)
+    bpy.ops.mesh.primitive_cylinder_add(radius=R, depth=6, vertices=96)
+    hole = bpy.context.active_object
+    hole.rotation_euler = (math.pi / 2, 0, 0)
+    hole.location = (0, 0, -0.35)
+    mod = body.modifiers.new('拱洞', 'BOOLEAN')
+    mod.object = hole
+    mod.operation = 'DIFFERENCE'
+    bpy.context.view_layer.objects.active = body
+    bpy.ops.object.modifier_apply(modifier=mod.name)
+    bpy.data.objects.remove(hole, do_unlink=True)
+    # hump the deck: simple deform via vertex z offset
+    for v in body.data.vertices:
+        if v.co.z > 0:
+            fx = min(1.0, abs(v.co.x * body.scale.x) / (w + 9))
+            v.co.z += (1 - fx * fx) * 0.25
+    body.data.materials.append(M['stone'])
+    for c in body.users_collection:
+        c.objects.unlink(body)
+    coll.objects.link(body)
+    body.rotation_euler = (0, 0, math.atan2(-nz, nx) * -1 + 0)
+    ang = math.atan2(nz, nx)
+    body.rotation_euler = (0, 0, -ang)
+    body.location = W2B(x, 0, z) + Vector((0, 0, body.location.z))
+    return body
+
+
+def build_wheel(M, coll):
+    s = river_s_at_x(-196)
+    w = river_half_width(s)
+    x, z, tx, tz, nx, nz = river_point(s, w - 3.0)
+    bpy.ops.object.empty_add(type='PLAIN_AXES', location=W2B(x, 1.25, z))
+    pivot = bpy.context.active_object
+    pivot.name = '水车_轴心'
+    for c in pivot.users_collection:
+        c.objects.unlink(pivot)
+    coll.objects.link(pivot)
+    parts = []
+    for zs in (-0.55, 0.55):
+        bpy.ops.mesh.primitive_torus_add(major_radius=3.3, minor_radius=0.09, major_segments=64, minor_segments=6)
+        t = bpy.context.active_object
+        t.rotation_euler = (math.pi / 2, 0, 0)
+        t.location = (0, zs, 0)
+        parts.append(t)
+    for k in range(18):
+        a = k / 18 * math.tau
+        bpy.ops.mesh.primitive_cube_add(size=1)
+        p = bpy.context.active_object
+        p.scale = (0.62, 1.3, 0.05)
+        p.location = (math.cos(a) * 3.35, 0, math.sin(a) * 3.35)
+        p.rotation_euler = (0, -a, 0)
+        parts.append(p)
+        bpy.ops.mesh.primitive_cube_add(size=1)
+        sp = bpy.context.active_object
+        sp.scale = (3.3, 0.1, 0.1)
+        sp.location = (math.cos(a) * 1.65, 0.55 if k % 2 else -0.55, math.sin(a) * 1.65)
+        sp.rotation_euler = (0, -a, 0)
+        parts.append(sp)
+    for p in parts:
+        p.data.materials.append(M['wood'])
+        for c in p.users_collection:
+            c.objects.unlink(p)
+        coll.objects.link(p)
+        p.parent = pivot
+    pivot.rotation_euler = (0, 0, -math.atan2(nz, nx))
+    # rotation animation: -0.42 rad/s about the axle (same speed as the web version)
+    pivot.rotation_mode = 'XYZ'
+    pivot.keyframe_insert('rotation_euler', frame=1)
+    pivot.rotation_euler.rotate_axis('Y', 0.42 * DURATION)
+    pivot.keyframe_insert('rotation_euler', frame=int(DURATION * FPS) + 1)
+    for fc in pivot.animation_data.action.fcurves:
+        for kp in fc.keyframe_points:
+            kp.interpolation = 'LINEAR'
+    return pivot
+
+
+def build_camphor(M, coll):
+    tx, tz = river_point(river_s_at_x(-300), -34)[:2]
+    rnd = random.Random(777)
+    bm_bark, bm_leaf = bmesh.new(), bmesh.new()
+
+    def cyl(bm, a, b, r0, r1):
+        d = b - a
+        L = d.length
+        res = bmesh.ops.create_cone(bm, cap_ends=False, segments=10, radius1=r0, radius2=r1, depth=L)
+        q = Vector((0, 0, 1)).rotation_difference(d.normalized())
+        bmesh.ops.rotate(bm, verts=res['verts'], cent=(0, 0, 0), matrix=q.to_matrix())
+        bmesh.ops.translate(bm, verts=res['verts'], vec=a + d * 0.5)
+
+    def branch(p, d, L, r, depth):
+        for i in range(3 if depth else 5):
+            d = (d + Vector((rnd.uniform(-0.28, 0.28), rnd.uniform(-0.28, 0.28), rnd.uniform(-0.1, 0.25)))).normalized()
+            np_ = p + d * (L / 3)
+            cyl(bm_bark, p, np_, r, r * 0.82)
+            p, r = np_, r * 0.82
+        if depth >= 4:
+            res = bmesh.ops.create_icosphere(bm_leaf, subdivisions=2, radius=rnd.uniform(1.6, 2.4))
+            bmesh.ops.translate(bm_leaf, verts=res['verts'], vec=p)
+            return
+        for k in range(rnd.randint(2, 4)):
+            az = k / 3 * math.tau + rnd.random()
+            nd = (d + Vector((math.cos(az), math.sin(az), 0)) * 0.9).normalized()
+            branch(p, nd, L * 0.72, r * 0.72, depth + 1)
+
+    base = W2B(tx, GROUND_Y - 0.4, tz)
+    branch(base, Vector((0, 0, 1)), 24 * 0.32, 1.25, 0)
+    for name, bm, mat in (('古樟_树干', bm_bark, M['bark']), ('古樟_树冠', bm_leaf, M['leaf'])):
+        me = bpy.data.meshes.new(name)
+        bm.to_mesh(me)
+        bm.free()
+        ob = bpy.data.objects.new(name, me)
+        ob.data.materials.append(mat)
+        coll.objects.link(ob)
+        if name.endswith('树冠'):
+            disp = ob.modifiers.new('叶团起伏', 'DISPLACE')
+            tex = bpy.data.textures.new('叶团噪声', 'CLOUDS')
+            tex.noise_scale = 0.8
+            disp.texture = tex
+            disp.strength = 0.9
+
+
+def build_bamboo(M, coll, count=380):
+    rnd = random.Random(31)
+    bm = bmesh.new()
+    placed = 0
+    tries = 0
+    while placed < count and tries < 20000:
+        tries += 1
+        x, z = rnd.uniform(-520, 560), rnd.uniform(-420, 380)
+        d, s = river_query(x, z)
+        if d < 60 or fbm(x * 0.012, z * 0.012, 3, 61) < 0.52:
+            continue
+        y = terrain_height(x, z)
+        for c in range(6):
+            a, r = rnd.random() * math.tau, rnd.random() * 1.2
+            p = W2B(x + math.cos(a) * r, y, z + math.sin(a) * r)
+            h = rnd.uniform(9, 14)
+            res = bmesh.ops.create_cone(bm, cap_ends=False, segments=5, radius1=0.05, radius2=0.025, depth=h)
+            bmesh.ops.translate(bm, verts=res['verts'], vec=p + Vector((0, 0, h / 2)))
+            res = bmesh.ops.create_icosphere(bm, subdivisions=1, radius=1.4)
+            bmesh.ops.scale(bm, verts=res['verts'], vec=(1, 1, 2.2))
+            bmesh.ops.translate(bm, verts=res['verts'], vec=p + Vector((0, 0, h * 0.8)))
+        placed += 1
+    me = bpy.data.meshes.new('竹林')
+    bm.to_mesh(me)
+    bm.free()
+    ob = bpy.data.objects.new('竹林', me)
+    ob.data.materials.append(M['bamboo'])
+    coll.objects.link(ob)
+
+
+def build_lanterns(M, coll, lanterns):
+    """Lanterns light one after another along the market street from 58.4 s."""
+    first_s = river_s_at_x(-128)
+    objs = []
+    for i, (p, s, side) in enumerate(lanterns):
+        f = min(1, max(0, (s - first_s) / (VILLAGE_S1 - first_s)))
+        on_t = 58.4 + 0.9 + 7 * f + (1.2 if side > 0 else 0) + (i % 7) * 0.07
+        mat = M['lantern'].copy()
+        mat.name = f'灯笼_{i:03d}'
+        bsdf = mat.node_tree.nodes.get('Principled BSDF')
+        sock = bsdf.inputs['Emission Strength']
+        sock.default_value = 0.0
+        sock.keyframe_insert('default_value', frame=max(1, int(on_t * FPS) - 1))
+        sock.default_value = 18.0
+        sock.keyframe_insert('default_value', frame=int(on_t * FPS) + 6)
+        bpy.ops.mesh.primitive_uv_sphere_add(radius=0.3, segments=16, ring_count=10, location=W2B(p.x, p.y - 0.42, p.z))
+        ob = bpy.context.active_object
+        ob.scale = (1, 1, 1.25)
+        ob.name = f'灯笼_{i:03d}'
+        ob.data.materials.append(mat)
+        for c in ob.users_collection:
+            c.objects.unlink(ob)
+        coll.objects.link(ob)
+        # a small warm point light per lantern (Cycles handles hundreds fine)
+        ld = bpy.data.lights.new(f'灯光_{i:03d}', 'POINT')
+        ld.color = (1.0, 0.55, 0.25)
+        ld.shadow_soft_size = 0.3
+        ld.energy = 0
+        ld.keyframe_insert('energy', frame=max(1, int(on_t * FPS) - 1))
+        ld.energy = 40
+        ld.keyframe_insert('energy', frame=int(on_t * FPS) + 6)
+        lo = bpy.data.objects.new(f'灯光_{i:03d}', ld)
+        lo.location = W2B(p.x, p.y - 0.42, p.z)
+        coll.objects.link(lo)
+        objs.append(ob)
+    # windows glow later
+    w = bpy.data.materials.get('窗纸')
+    sock = w.node_tree.nodes.get('Principled BSDF').inputs['Emission Strength']
+    sock.default_value = 0
+    sock.keyframe_insert('default_value', frame=int(58.5 * FPS))
+    sock.default_value = 6
+    sock.keyframe_insert('default_value', frame=int(70 * FPS))
+    return objs
+
+
+SUN_KEYS = [(0, 8.5), (17, 6.6), (32, 4.6), (44, 2.6), (52, 0.9), (57, -0.8), (62, -2.8), (68, -5.2), (76, -7.6), (86, -9.8)]
+
+
+def build_sky_and_sun(scene):
+    world = bpy.data.worlds.new('暮色天空')
+    scene.world = world
+    world.use_nodes = True
+    nt = world.node_tree
+    bg = nt.nodes.get('Background')
+    sky = nt.nodes.new('ShaderNodeTexSky')
+    sky.sky_type = 'NISHITA'
+    sky.sun_rotation = math.pi / 2 + 0.2 + math.pi / 2  # sets toward west-north-west
+    sky.air_density = 1.2
+    sky.dust_density = 2.5
+    nt.links.new(sky.outputs['Color'], bg.inputs['Color'])
+    # twilight layer: deep blue dome with a warm afterglow band, faded in after sunset
+    tw = nt.nodes.new('ShaderNodeBackground')
+    grad_tc = nt.nodes.new('ShaderNodeTexCoord')
+    sep = nt.nodes.new('ShaderNodeSeparateXYZ')
+    nt.links.new(grad_tc.outputs['Generated'], sep.inputs['Vector'])
+    ramp = nt.nodes.new('ShaderNodeValToRGB')
+    ramp.color_ramp.elements[0].position = 0.5
+    ramp.color_ramp.elements[0].color = (0.35, 0.16, 0.1, 1)
+    ramp.color_ramp.elements[1].position = 0.62
+    ramp.color_ramp.elements[1].color = (0.02, 0.035, 0.1, 1)
+    nt.links.new(sep.outputs['Z'], ramp.inputs['Fac'])
+    nt.links.new(ramp.outputs['Color'], tw.inputs['Color'])
+    add = nt.nodes.new('ShaderNodeAddShader')
+    nt.links.new(bg.outputs['Background'], add.inputs[0])
+    nt.links.new(tw.outputs['Background'], add.inputs[1])
+    nt.links.new(add.outputs['Shader'], nt.nodes.get('World Output').inputs['Surface'])
+    sun_data = bpy.data.lights.new('太阳', 'SUN')
+    sun_data.angle = math.radians(1.2)
+    sun = bpy.data.objects.new('太阳', sun_data)
+    scene.collection.objects.link(sun)
+    moon_data = bpy.data.lights.new('月光', 'SUN')
+    moon_data.color = (0.6, 0.7, 1.0)
+    moon_data.energy = 0.0
+    moon = bpy.data.objects.new('月光', moon_data)
+    moon.rotation_euler = (math.radians(66), 0, math.radians(-110))
+    scene.collection.objects.link(moon)
+    for t, e in SUN_KEYS:
+        f = int(t * FPS) + 1
+        er = math.radians(e)
+        # Nishita is undefined below the horizon (can return NaN = black frames): clamp it
+        # and let the twilight layer + exposure carry the blue hour instead.
+        sky.sun_elevation = max(er, math.radians(0.4))
+        sky.keyframe_insert('sun_elevation', frame=f)
+        sun.rotation_euler = (math.pi / 2 - er, 0, -(math.pi / 2 + 0.2 + math.pi / 2))
+        sun.keyframe_insert('rotation_euler', frame=f)
+        sun_data.energy = max(0.0, 4.0 * smoothstep(-1.2, 6, e))
+        sun_data.color = (1.0, lerp(0.38, 0.76, smoothstep(0, 9, e)), lerp(0.14, 0.52, smoothstep(0, 9, e)))
+        sun_data.keyframe_insert('energy', frame=f)
+        sun_data.keyframe_insert('color', frame=f)
+        bg.inputs['Strength'].default_value = lerp(0.0, 1.0, smoothstep(-5, 2, e))
+        bg.inputs['Strength'].keyframe_insert('default_value', frame=f)
+        moon_data.energy = 0.35 * smoothstep(-1.5, -8.5, e)
+        moon_data.keyframe_insert('energy', frame=f)
+        tw.inputs['Strength'].default_value = 0.6 * smoothstep(1.0, -4.0, e)
+        tw.inputs['Strength'].keyframe_insert('default_value', frame=f)
+        # camera eye adaptation, like the web version
+        scene.view_settings.exposure = lerp(0.0, 3.2, smoothstep(1, -9, e))
+        scene.view_settings.keyframe_insert('exposure', frame=f)
+    # thin evening mist: a bounded volume over the valley (a world volume would swallow the sky)
+    bpy.ops.mesh.primitive_cube_add(size=1, location=(0, 0, 30))
+    mist = bpy.context.active_object
+    mist.name = '暮霭'
+    mist.scale = (2400, 2400, 60)
+    mm = bpy.data.materials.new('暮霭')
+    mm.use_nodes = True
+    nt2 = mm.node_tree
+    nt2.nodes.remove(nt2.nodes.get('Principled BSDF'))
+    vol = nt2.nodes.new('ShaderNodeVolumePrincipled')
+    vol.inputs['Density'].default_value = 0.0012
+    vol.inputs['Color'].default_value = (0.85, 0.8, 0.9, 1)
+    nt2.links.new(vol.outputs['Volume'], nt2.nodes.get('Material Output').inputs['Volume'])
+    mist.data.materials.append(mm)
+    mist.visible_shadow = False
+    vol.inputs['Density'].keyframe_insert('default_value', frame=1)
+    vol.inputs['Density'].default_value = 0.0026
+    vol.inputs['Density'].keyframe_insert('default_value', frame=int(80 * FPS))
+
+
+SHOT_TIMES = [(1, 0), (2, 9), (3, 17), (4, 25), (5, 32), (6, 40), (7, 48), (8, 55), (9, 65), (10, 74)]
+
+
+def build_cameras(scene, coll):
+    """Ten shots. Key poses (web coordinates) mirror camera.ts; markers switch the active camera."""
+    bs = river_s_at_x(40)
+    bx, bz, btx, btz, bnx, bnz = river_point(bs, 0)
+    ws = river_s_at_x(-196)
+    wx, wz, wtx, wtz, wnx, wnz = river_point(ws, river_half_width(ws) - 3)
+    tx, tz = river_point(river_s_at_x(-300), -34)[:2]
+    def rp(s_, off, y):
+        x_, z_ = river_point(s_, off)[:2]
+        return (x_, y, z_)
+
+    def qp(s_, lane, y):  # point on the market quay, same as quayPoint() in layout.ts
+        return rp(s_, -(11 + lane), y)
+
+    shots = {
+        1: ((470, 132, 150), (-1100, 60, 160), (310, 70, 82), (-1000, 30, 130), 38),
+        2: (rp(bs + 150, 1.2, 2.0), rp(bs + 116, 0, 2.3), rp(bs + 62, 1.2, 2.0), rp(bs + 28, 0, 2.3), 46),
+        3: ((bx + btx * 40 - bnx * 3.2, 1.05, bz + btz * 40 - bnz * 3.2), (bx, 3.1, bz), (bx + btx * 33 + bnx * 2.6, 1.05, bz + btz * 33 + bnz * 2.6), (bx, 3.1, bz), 34),
+        4: ((wx - wnx * 15.5 + wtx * 7, 2.9, wz - wnz * 15.5 + wtz * 7), (wx, 2.2, wz), (wx - wnx * 12.2 + wtx * 3.2, 2.4, wz - wnz * 12.2 + wtz * 3.2), (wx, 2.2, wz), 40),
+        5: (qp(river_s_at_x(-175) + 3.8 + 15, -1.1, GROUND_Y + 1.75), qp(river_s_at_x(-175) + 3.8 - 6, 1.0, GROUND_Y + 1.6),
+            qp(river_s_at_x(-175) + 11.4 + 12.5, -1.1, GROUND_Y + 1.75), qp(river_s_at_x(-175) + 11.4 - 6, 1.0, GROUND_Y + 1.6), 44),
+        6: ((tx + 25, 3.5, tz + 16.5), (tx - 2, 9.7, tz - 4), (tx + 19, 4.4, tz + 12.5), (tx - 2, 11.7, tz - 4), 46),
+        7: ((110, 26, -120), (300, 140, -410), (98, 30, -126), (205, 70, -372), 21),
+        8: (qp(river_s_at_x(-150), -0.8, GROUND_Y + 2.6), qp(river_s_at_x(-150) + 26, -1.6, GROUND_Y + 2.4),
+            qp(river_s_at_x(-150) + 21, -0.8, GROUND_Y + 2.6), qp(river_s_at_x(-150) + 47, -1.6, GROUND_Y + 2.4), 48),
+        9: (rp(bs + 52, -4.6, 1.25), (bx + bnx * 1.5, 3.4, bz + bnz * 1.5), rp(bs + 45, -4.6, 1.25), (bx + bnx * 1.5, 3.4, bz + bnz * 1.5), 42),
+        10: ((95, 9, 34), (-10, 3, -4), (250, 125, 205), (-520, 60, -230), 44),
+    }
+    cams = {}
+    for i, (sid, t0) in enumerate(SHOT_TIMES):
+        t1 = SHOT_TIMES[i + 1][1] if i + 1 < len(SHOT_TIMES) else DURATION
+        p0, a0, p1, a1, fov = shots[sid]
+        cd = bpy.data.cameras.new(f'镜头{sid:02d}')
+        cd.lens_unit = 'FOV'
+        cd.angle = math.radians(fov * 16 / 9 * 0.92)
+        cd.sensor_fit = 'HORIZONTAL'
+        cam = bpy.data.objects.new(f'镜头{sid:02d}', cd)
+        coll.objects.link(cam)
+        tgt = bpy.data.objects.new(f'镜头{sid:02d}_目标', None)
+        coll.objects.link(tgt)
+        con = cam.constraints.new('TRACK_TO')
+        con.target = tgt
+        con.track_axis = 'TRACK_NEGATIVE_Z'
+        con.up_axis = 'UP_Y'
+        for (p, a), t in (((p0, a0), t0), ((p1, a1), t1)):
+            cam.location = W2B(*p)
+            tgt.location = W2B(*a)
+            cam.keyframe_insert('location', frame=int(t * FPS) + 1)
+            tgt.keyframe_insert('location', frame=int(t * FPS) + 1)
+        m = scene.timeline_markers.new(f'S{sid:02d}', frame=int(t0 * FPS) + 1)
+        m.camera = cam
+        cams[sid] = cam
+    scene.camera = cams[1]
+
+
+def setup_render(scene):
+    scene.render.fps = FPS
+    scene.frame_start = 1
+    scene.frame_end = int(DURATION * FPS)
+    scene.render.resolution_x = 1920
+    scene.render.resolution_y = 1080
+    try:
+        scene.render.engine = RENDER_ENGINE
+    except TypeError:
+        scene.render.engine = 'CYCLES'
+    if scene.render.engine == 'CYCLES':
+        scene.cycles.samples = 128
+        scene.cycles.use_denoising = True
+        scene.cycles.max_bounces = 6
+    scene.view_settings.view_transform = 'AgX' if 'AgX' in [i.identifier for i in scene.view_settings.bl_rna.properties['view_transform'].enum_items] else 'Filmic'
+    scene.view_settings.look = 'None'
+    scene.render.image_settings.file_format = 'FFMPEG'
+    scene.render.ffmpeg.format = 'MPEG4'
+    scene.render.ffmpeg.codec = 'H264'
+    scene.render.ffmpeg.constant_rate_factor = 'HIGH'
+    scene.render.filepath = '//暮归江南_blender_'
+
+
+def main():
+    clear_scene()
+    scene = bpy.context.scene
+    M = build_materials()
+    land = collection('地形与河流')
+    town = collection('村落')
+    life = collection('灯火')
+    cams = collection('镜头')
+    print('[暮归江南] 地形…')
+    build_terrain(M, land)
+    build_river(M, land)
+    print('[暮归江南] 民居…')
+    houses, lanterns = build_houses(M, town)
+    print('[暮归江南] 石桥、水车、��樟、竹林…')
+    build_bridge(M, town)
+    build_wheel(M, town)
+    build_camphor(M, town)
+    build_bamboo(M, land)
+    print('[暮归江南] 灯笼…')
+    build_lanterns(M, life, lanterns)
+    build_sky_and_sun(scene)
+    build_cameras(scene, cams)
+    setup_render(scene)
+    print(f'[暮归江南] 完成：{len(houses)} 栋民居，{len(lanterns)} 盏灯笼，{int(DURATION * FPS)} 帧。')
+
+
+if __name__ == '__main__':
+    main()

src/jiangnan/copy/platforms.ts

@@ -0,0 +1,83 @@
+// Publishing copy for each platform (all original). Limits are validated in the UI.
+export interface PlatformCopy {
+  id: string
+  name: string
+  titleMax: number
+  tagMax: number | null // null = unlimited
+  tagStyle: 'hash' | 'plain' // plain = B站 tag field (no #)
+  titles: string[]
+  desc: string
+  tags: string[]
+  note: string
+}
+
+export const PLATFORMS: PlatformCopy[] = [
+  {
+    id: 'xhs', name: '小红书', titleMax: 20, tagMax: 10, tagStyle: 'hash',
+    titles: [
+      '日落时分，我在江南水乡等灯亮',
+      '一座会发光的江南小村庄',
+      '江南的黄昏，原来这么慢',
+      '粉墙黛瓦，一盏一盏亮起来',
+    ],
+    desc: '把一整个江南的黄昏装进 86 秒：\n夕阳贴着远山，小河绕村而过，石拱桥和倒影合成一轮满月，老水车吱呀转着，集市的竹篮和腊肉都被染成金色。\n天一点点暗下来，炊烟升起，红灯笼从街头一盏接一盏亮到街尾，灯火落进河里，最后整座村子睡在星光和薄雾里。\n\n全片是原创 3D 动画，配乐、音效、旁白都是专门做的。戴上耳机看会更有感觉。',
+    tags: ['江南水乡', '治愈系风景', '日落', '古村落', '国风', '粉墙黛瓦', '小桥流水人家', '3D动画', '氛围感', '江南'],
+    note: '标题 ≤20 字，话题 ≤10 个',
+  },
+  {
+    id: 'douyin', name: '抖音', titleMax: 30, tagMax: 5, tagStyle: 'hash',
+    titles: [
+      '江南的黄昏来得很慢，看一座村庄一盏一盏亮起灯火',
+      '86秒，看完江南水乡从日落到星夜',
+      '石桥半个月亮，水里再补半个，这就是江南',
+    ],
+    desc: '夕阳、小河、石拱桥、老水车、红灯笼。\n从日落到星夜，一座江南水乡慢慢亮起来。原创3D动画，建议戴耳机。',
+    tags: ['江南水乡', '治愈系', '国风动画', '日落', '古镇'],
+    note: '标题 ≤30 字，话题 ≤5 个',
+  },
+  {
+    id: 'channels', name: '微信视频号', titleMax: 16, tagMax: null, tagStyle: 'hash',
+    titles: [
+      '江南水乡，一盏一盏亮起来',
+      '暮归江南：日落到星夜',
+      '江南的黄昏，总是来得很慢',
+    ],
+    desc: '一条小河把村子揽在怀里，石桥、水车、集市、古樟，夕阳落下后，灯笼一盏接一盏亮起。\n原创3D动画《暮归江南》，送给想家的人。',
+    tags: ['江南水乡', '暮归江南', '治愈', '日落', '古村落', '国风', '粉墙黛瓦', '小桥流水', '乡愁', '原创动画', '风景', '星空'],
+    note: '标题 ≤16 字，话题不限',
+  },
+  {
+    id: 'bili', name: '哔哩哔哩', titleMax: 80, tagMax: 10, tagStyle: 'plain',
+    titles: [
+      '【原创3D动画】暮归江南：日落时分的江南水乡，从夕阳到星夜的86秒｜粉墙黛瓦·石拱桥·老水车·红灯笼',
+      '我用代码搭了一座江南水乡，看它从日落一直亮到星夜（原创配乐+旁白）',
+      '【治愈向】86秒的江南黄昏：石桥倒影成满月，灯笼一盏接一盏亮起',
+    ],
+    desc: '《暮归江南》是一部原创 3D 动画短片：一座日落时分的江南水乡，从金色夕阳慢慢过渡到蓝色夜幕，灯笼和窗灯依次亮起，最后整座村庄在星光与暮霭里安静发光。\n\n画面：全部程序化建模与实时渲染（徽派民居、马头墙、石拱桥、水车、乌篷船、集市、古樟、梯田、竹林）\n声音：原创五声音阶配乐（古筝、竹箫、弦乐、编钟），环境音效与中文旁白\n\n00:00 山谷航拍\n00:09 贴水滑行\n00:17 石拱桥\n00:25 老水车\n00:32 石板街集市\n00:40 村口古樟\n00:48 梯田竹林\n00:55 灯笼次第亮起\n01:05 夜河灯影\n01:14 星夜全景',
+    tags: ['原创动画', '3D动画', '江南水乡', '国风', '治愈', '风景', '日落', '古村落', 'Blender', '程序化生成'],
+    note: '标题 ≤80 字，标签 ≤10 个',
+  },
+]
+
+export interface LiveCopy { id: string; name: string; titleMax: number; tagMax: number; titles: string[]; body: string; tags: string[] }
+
+export const LIVE_COPY: LiveCopy[] = [
+  {
+    id: 'xhs', name: '小红书 · 实况图文', titleMax: 20, tagMax: 10,
+    titles: ['长按看江南水乡亮灯的瞬间', '6张实况｜江南的黄昏会动', '江南水乡，日落到星夜'],
+    body: '6 张实况，长按会动：\n① 石桥夕照，拱桥和倒影合成满月\n② 老水车转着，把光阴舀回河里\n③ 古樟树下，蓝印花布随风摆\n④ 梯田和竹林，白鹭飞回去\n⑤ 第一盏灯笼亮起\n⑥ 星夜里的整座村庄\n\n原创 3D 动画截取，喜欢可以设成壁纸。',
+    tags: ['实况图', '江南水乡', '治愈系风景', '壁纸', '日落', '古村落', '国风', '粉墙黛瓦', 'livephoto', '氛围感'],
+  },
+  {
+    id: 'douyin', name: '抖音 · 图文', titleMax: 30, tagMax: 5,
+    titles: ['江南水乡6张实况，长按看灯笼一盏盏亮起', '把江南的黄昏做成了会动的照片'],
+    body: '石桥、水车��古樟、梯田、灯笼、星夜。\n6 张实况，长按会动，原创 3D 动画截取。',
+    tags: ['实况图', '江南水乡', '治愈系', '壁纸', '国风'],
+  },
+]
+
+export const chars = (s: string) => Array.from(s).length
+
+export function formatTags(tags: string[], style: 'hash' | 'plain') {
+  return style === 'hash' ? tags.map((t) => '#' + t).join(' ') : tags.join('，')
+}

src/jiangnan/engine/Film.ts

@@ -235,6 +235,7 @@ export class Film {
 
   /** Render the frame at film time t with the current canvas size. */
   renderFrame(t: number) {
+    if (!this.sky || !this.water || !this.life) return
     const st = this.update(t)
     const pose = cameraAt(t, this.width / this.height)
     applyPose(this.camera, pose)

src/jiangnan/engine/camera.ts

@@ -83,7 +83,7 @@ const shots: Record<number, ShotFn> = {
   6: (u, _t, p) => {
     const e = easeSine(u)
     const tp = V(TREE_POS.x, GROUND_Y, TREE_POS.z)
-    const pos = tp.clone().add(mix(V(34, 2.2, 24), V(26, 3.6, 17), e))
+    const pos = tp.clone().add(mix(V(25, 1.8, 16.5), V(19, 2.7, 12.5), e))
     const target = tp.clone().add(V(-2, mix(V(0, 8, 0), V(0, 10, 0), e).y, -4))
     if (p) { pos.add(V(6, 0.5, 5)); target.y += 3 }
     return { pos, target, fov: 46 }
@@ -121,8 +121,8 @@ const shots: Record<number, ShotFn> = {
   // 10 · crane up and pull back over the glowing village under the first stars
   10: (u, _t, p) => {
     const e = easeInOut(clamp(u * 1.05))
-    const pos = mix(V(95, 9, 34), V(240, 150, 210), e)
-    const target = mix(V(-10, 3, -4), V(-60, 0, -20), e)
+    const pos = mix(V(95, 9, 34), V(250, 125, 205), e)
+    const target = mix(V(-10, 3, -4), V(-520, 60, -230), e)
     if (p) { pos.add(V(20, 25, 20)); target.y -= 8 }
     return { pos, target, fov: lerp(44, 40, e) }
   },

src/jiangnan/engine/layout.ts

@@ -241,7 +241,7 @@ export interface Landmarks {
 
 export const BRIDGE_S = riverSAtX(40)
 export const WHEEL_S = riverSAtX(-196)
-export const TREE_POS = (() => { const p = riverPoint(riverSAtX(-262), -34); return { x: p.x, z: p.z } })()
+export const TREE_POS = (() => { const p = riverPoint(riverSAtX(-300), -34); return { x: p.x, z: p.z } })()
 export const MARKET_S0 = riverSAtX(-130)
 export const MARKET_S1 = riverSAtX(30)
 export const PAVILION_POS = { x: 205, z: -372 }
@@ -249,7 +249,7 @@ export const PAVILION_POS = { x: 205, z: -372 }
 function blocked(s: number, side: number, half: number) {
   if (Math.abs(s - BRIDGE_S) < half + 8) return true
   if (side > 0 && Math.abs(s - WHEEL_S) < half + 10) return true
-  if (side < 0 && s < riverSAtX(-238) + half) return true // entrance plaza with the camphor tree
+  if (side < 0 && s < riverSAtX(-264) + half) return true // entrance plaza with the camphor tree
   return false
 }
 

src/jiangnan/engine/props.ts

@@ -375,15 +375,15 @@ export function createProps(): PropsOut {
   const cloths: THREE.Mesh[] = []
   {
     const tp = TREE_POS
-    const base = V3(tp.x + 4, GROUND_Y, tp.z - 15)
-    const rot = 0.35
+    const base = V3(tp.x + 6, GROUND_Y, tp.z + 13.5)
+    const rot = 0.99
     const m = trs(base.x, base.y, base.z, rot)
     wood.setMatrix(m)
     for (const z of [0, 2.4]) {
-      for (const x of [-6, 0, 6]) wood.box(x, 2.9, z, 0.12, 5.8, 0.12)
+      for (const x of [-4.4, 0, 4.4]) wood.box(x, 2.9, z, 0.12, 5.8, 0.12)
       // bamboo crossbar
     }
-    const bar = new THREE.CylinderGeometry(0.07, 0.07, 13, 8)
+    const bar = new THREE.CylinderGeometry(0.07, 0.07, 9.4, 8)
     bar.rotateZ(Math.PI / 2)
     const barMat = std({ color: new THREE.Color(0.45, 0.42, 0.22), roughness: 0.55 }, { key: 'bar' })
     for (const z of [0, 2.4]) {
@@ -395,9 +395,9 @@ export function createProps(): PropsOut {
     }
     const clothGeo = new THREE.PlaneGeometry(1.15, 4.9, 3, 18)
     clothGeo.translate(0, -2.45, 0)
-    for (let k = 0; k < 14; k++) {
+    for (let k = 0; k < 10; k++) {
       const row = k % 2
-      const x = -5.6 + Math.floor(k / 2) * 1.75 + row * 0.6
+      const x = -3.9 + Math.floor(k / 2) * 1.75 + row * 0.6
       const mat = patch(new THREE.MeshStandardMaterial({ map: T.cloth[k % 3], roughness: 0.92, side: THREE.DoubleSide }), {
         key: 'cloth',
         vertex: `

src/jiangnan/engine/sky.ts

@@ -45,7 +45,7 @@ export function skyState(t: number): SkyState {
     sun: mixRGB(a.sun, b.sun, f), sunI: lerp(a.sunI, b.sunI, f),
     cloudLit: mixRGB(a.cloudLit, b.cloudLit, f), cloudDark: mixRGB(a.cloudDark, b.cloudDark, f),
     // camera "eye adaptation": open up as the light fades
-    exposure: lerp(0.85, 2.1, smoothstep(1, -9, elev)),
+    exposure: lerp(0.85, 1.55, smoothstep(1, -9, elev)),
   }
 }
 

src/jiangnan/engine/timeline.ts

@@ -61,16 +61,16 @@ export const WINDOW_SPAN = 12
 export interface Cue { id: string; t: number; text: string; dur: number }
 // dur is the measured length of the generated narration audio (seconds)
 export const NARRATION: Cue[] = [
-  { id: 'n01', t: 1.6, text: '江南的黄昏，总是来得很慢。', dur: 3.0 },
-  { id: 'n02', t: 9.8, text: '一条小河，把整座村子轻轻揽在怀里。', dur: 3.6 },
-  { id: 'n03', t: 17.8, text: '石桥弯成半个月亮，水里，再补上另外半个。', dur: 4.6 },
-  { id: 'n04', t: 25.6, text: '老水车吱呀吱呀地转，把一天的光阴，一勺一勺舀回河里。', dur: 5.6 },
-  { id: 'n05', t: 33.0, text: '集市还没散。竹篮、陶罐、檐下的腊肉，都染上了夕阳的颜色。', dur: 6.0 },
-  { id: 'n06', t: 40.8, text: '村口的老樟树站了几百年，看着人们挑着担子出去，又挑着担子回来。', dur: 6.2 },
-  { id: 'n07', t: 48.6, text: '远山的梯田一层层暗下去，白鹭，也该回竹林了。', dur: 4.8 },
-  { id: 'n08', t: 56.0, text: '炊烟升起来的时候，第一盏灯笼亮了。一盏，接着一盏。', dur: 5.6 },
-  { id: 'n09', t: 66.0, text: '灯火落进河里，整条河，都暖了。', dur: 3.6 },
-  { id: 'n10', t: 75.0, text: '星星出来了。水乡睡在一层薄薄的雾里——这里，是家的方向。', dur: 6.6 },
+  { id: 'n01', t: 1.6, text: '江南的黄昏，总是来得很慢。', dur: 3.31 },
+  { id: 'n02', t: 9.8, text: '一条小河，把整座村子轻轻揽在怀里。', dur: 4.61 },
+  { id: 'n03', t: 17.8, text: '石桥弯成半个月亮，水里，再补上另外半个。', dur: 4.92 },
+  { id: 'n04', t: 25.6, text: '老水车吱呀吱呀地转，把一天的光阴，一勺一勺舀回河里。', dur: 5.95 },
+  { id: 'n05', t: 33.0, text: '集市还没散。竹篮、陶罐、檐下的腊肉，都染上了夕阳的颜色。', dur: 7.2 },
+  { id: 'n06', t: 40.8, text: '村口的老樟树站了几百年，看着人们挑着担子出去，又挑着担子回来。', dur: 7.08 },
+  { id: 'n07', t: 48.6, text: '远山的梯田一层层暗下去，白鹭，也该回竹林了。', dur: 5.4 },
+  { id: 'n08', t: 56.0, text: '炊烟升起来的时候，第一盏灯笼亮了。一盏，接着一盏。', dur: 6.22 },
+  { id: 'n09', t: 66.0, text: '灯火落进河里，整条河，都暖了。', dur: 4.1 },
+  { id: 'n10', t: 75.0, text: '星星出来了。水乡睡在一层薄薄的雾里。这里，是家的方向。', dur: 7.13 },
 ]
 
 export function cueAt(t: number) {

src/jiangnan/export/covers.ts

@@ -0,0 +1,122 @@
+// Covers: the 3D scene is re-rendered for each aspect ratio (no stretching),
+// then typeset with calligraphy titles.
+import type { Film } from '../engine/Film'
+import { drawSeal, FONT_BODY, FONT_TITLE } from '../player/overlay'
+
+export type Ratio = '16:9' | '4:3' | '3:4'
+export const RATIOS: { id: Ratio; w: number; h: number; label: string }[] = [
+  { id: '16:9', w: 1920, h: 1080, label: '16:9 横版（B站 / 视频号 / 抖音横屏）' },
+  { id: '4:3', w: 1440, h: 1080, label: '4:3 横版（B站 / 西瓜）' },
+  { id: '3:4', w: 1080, h: 1440, label: '3:4 竖版（小红书 / 抖音 / 视频号）' },
+]
+
+export interface CoverDesign { key: string; name: string; t: number; layout: 'right-vertical' | 'top-center' | 'bottom-left' | 'left-vertical'; line: string }
+export const COVER_DESIGNS: CoverDesign[] = [
+  { key: 'bridge', name: '石桥夕照', t: 21.2, layout: 'right-vertical', line: '石桥半个月亮 水里补上另一半' },
+  { key: 'lantern', name: '灯火长街', t: 63.4, layout: 'top-center', line: '灯笼一盏接一盏亮起' },
+  { key: 'night', name: '星夜水乡', t: 84.2, layout: 'bottom-left', line: '从日落到星夜的 86 秒' },
+  { key: 'tree', name: '古樟斜阳', t: 45.5, layout: 'left-vertical', line: '挑着担子出去 又挑着担子回来' },
+]
+
+function vertical(g: CanvasRenderingContext2D, text: string, x: number, y: number, size: number, gap = 1.08) {
+  Array.from(text).forEach((c, i) => g.fillText(c, x, y + i * size * gap))
+}
+
+export function renderCover(film: Film, d: CoverDesign, ratio: Ratio): HTMLCanvasElement {
+  const R = RATIOS.find((r) => r.id === ratio)!
+  const W = R.w, H = R.h
+  const prevW = film.width, prevH = film.height
+  film.setSize(W, H)
+  film.renderFrame(d.t)
+  const c = document.createElement('canvas')
+  c.width = W
+  c.height = H
+  const g = c.getContext('2d')!
+  g.drawImage(film.renderer.domElement, 0, 0, W, H)
+  film.setSize(prevW, prevH)
+  const u = Math.min(W, H)
+  const portrait = H > W
+  g.textAlign = 'center'
+  g.textBaseline = 'middle'
+  const shadow = (b: number) => { g.shadowColor = 'rgba(20,8,0,0.6)'; g.shadowBlur = b }
+  // legibility scrims
+  const scrim = (x0: number, y0: number, x1: number, y1: number, a: number) => {
+    const grd = g.createLinearGradient(x0, y0, x1, y1)
+    grd.addColorStop(0, `rgba(8,6,10,${a})`)
+    grd.addColorStop(1, 'rgba(8,6,10,0)')
+    g.fillStyle = grd
+    g.fillRect(0, 0, W, H)
+  }
+  if (d.layout === 'right-vertical') {
+    scrim(W, 0, W * 0.55, 0, 0.45)
+    shadow(u * 0.03)
+    g.fillStyle = '#fff4e2'
+    const size = u * (portrait ? 0.15 : 0.16)
+    const x = W - u * (portrait ? 0.14 : 0.15)
+    g.font = `${size}px ${FONT_TITLE}, serif`
+    vertical(g, '暮归江南', x, H * 0.17, size)
+    g.font = `${u * 0.036}px ${FONT_BODY}, serif`
+    g.fillStyle = 'rgba(255,238,215,0.92)'
+    const parts = d.line.split(' ')
+    parts.forEach((p, i) => vertical(g, p, x - size * 0.95 - i * u * 0.05, H * 0.2 + i * u * 0.06, u * 0.036, 1.25))
+    g.shadowBlur = 0
+    drawSeal(g, x - u * 0.04, H * 0.17 + size * 4.35, u * 0.08, '江南')
+  } else if (d.layout === 'top-center') {
+    scrim(0, 0, 0, H * 0.5, 0.55)
+    shadow(u * 0.035)
+    g.fillStyle = '#fff3df'
+    g.font = `${u * (portrait ? 0.17 : 0.15)}px ${FONT_TITLE}, serif`
+    g.fillText('暮归江南', W / 2, H * (portrait ? 0.15 : 0.2))
+    g.font = `${u * 0.045}px ${FONT_BODY}, serif`
+    g.fillStyle = 'rgba(255,230,200,0.95)'
+    g.fillText(d.line, W / 2, H * (portrait ? 0.15 : 0.2) + u * 0.13)
+    g.shadowBlur = 0
+    drawSeal(g, W / 2 + u * (portrait ? 0.36 : 0.32), H * (portrait ? 0.15 : 0.2) - u * 0.13, u * 0.07, '灯火')
+  } else if (d.layout === 'bottom-left') {
+    scrim(0, H, 0, H * 0.45, 0.65)
+    shadow(u * 0.03)
+    const x0 = W * 0.07
+    g.textAlign = 'left'
+    g.fillStyle = '#fff3df'
+    g.font = `${u * 0.16}px ${FONT_TITLE}, serif`
+    g.fillText('暮归江南', x0, H - u * 0.3)
+    g.font = `${u * 0.042}px ${FONT_BODY}, serif`
+    g.fillStyle = 'rgba(235,225,255,0.92)'
+    g.fillText(d.line, x0 + u * 0.01, H - u * 0.16)
+    g.shadowBlur = 0
+    drawSeal(g, x0 + u * 0.66, H - u * 0.39, u * 0.07, '星夜')
+  } else {
+    scrim(0, 0, W * 0.5, 0, 0.5)
+    shadow(u * 0.03)
+    g.fillStyle = '#fff4e2'
+    const size = u * 0.15
+    const x = u * 0.15
+    g.font = `${size}px ${FONT_TITLE}, serif`
+    vertical(g, '暮归江南', x, H * 0.17, size)
+    g.font = `${u * 0.036}px ${FONT_BODY}, serif`
+    g.fillStyle = 'rgba(255,238,215,0.92)'
+    d.line.split(' ').forEach((p, i) => vertical(g, p, x + size * 0.95 + i * u * 0.05, H * 0.2 + i * u * 0.06, u * 0.036, 1.25))
+    g.shadowBlur = 0
+    drawSeal(g, x - u * 0.04, H * 0.17 + size * 4.35, u * 0.08, '江南')
+  }
+  // small badge
+  g.shadowBlur = 0
+  g.textAlign = 'left'
+  g.font = `${u * 0.026}px ${FONT_BODY}, serif`
+  const badge = '原创3D动画 · 86秒'
+  const bw = g.measureText(badge).width + u * 0.05
+  const bx = d.layout === 'bottom-left' || d.layout === 'right-vertical' ? u * 0.05 : W - bw - u * 0.05
+  const by = d.layout === 'bottom-left' ? u * 0.05 : H - u * 0.1
+  g.fillStyle = 'rgba(10,8,12,0.45)'
+  g.beginPath()
+  g.roundRect(bx, by, bw, u * 0.05, u * 0.025)
+  g.fill()
+  g.fillStyle = 'rgba(255,236,210,0.95)'
+  g.textBaseline = 'middle'
+  g.fillText(badge, bx + u * 0.025, by + u * 0.026)
+  return c
+}
+
+export function canvasToBlob(c: HTMLCanvasElement, type = 'image/jpeg', q = 0.92) {
+  return new Promise<Blob>((res, rej) => c.toBlob((b) => (b ? res(b) : rej(new Error('导出失败'))), type, q))
+}

src/jiangnan/export/livegen.ts

@@ -0,0 +1,149 @@
+// LiveCanvas-style Apple Live Photos generated entirely in the browser:
+// 3:4 (1080×1440), 30 fps, 3 s, silent H.264; the cover JPEG is the exact
+// frame at coverTime; pairing metadata written by livephoto.ts.
+import { strToU8, zipSync, Zippable } from 'fflate'
+import { LIVE_COPY } from '../copy/platforms'
+import type { Film } from '../engine/Film'
+import { drawSeal, FONT_BODY, FONT_TITLE } from '../player/overlay'
+import { inspectLiveJpeg, PVT_PLIST, sha256, uuid4, VideoSample, writeLiveJpeg, writeLiveMov } from './livephoto'
+
+export interface LiveClip { id: string; name: string; t0: number; cover: number; caption: string }
+export const LIVE_CLIPS: LiveClip[] = [
+  { id: '01', name: '石桥夕照', t0: 19.6, cover: 1.5, caption: '石桥半月' },
+  { id: '02', name: '老水车', t0: 27.0, cover: 1.5, caption: '水车舀光阴' },
+  { id: '03', name: '古樟布影', t0: 42.6, cover: 1.5, caption: '古樟斜阳' },
+  { id: '04', name: '梯田归鹭', t0: 50.0, cover: 1.5, caption: '梯田归鹭' },
+  { id: '05', name: '灯笼亮起', t0: 57.4, cover: 2.0, caption: '第一盏灯' },
+  { id: '06', name: '星夜水乡', t0: 80.6, cover: 1.5, caption: '星夜水乡' },
+]
+let W = 1080, H = 1440
+const FPS = 30
+let FRAMES = 90
+
+function caption(g: CanvasRenderingContext2D, text: string) {
+  const u = W
+  g.save()
+  g.textAlign = 'center'
+  g.textBaseline = 'middle'
+  g.shadowColor = 'rgba(10,5,0,0.65)'
+  g.shadowBlur = u * 0.025
+  g.fillStyle = '#fff4e2'
+  const size = u * 0.075
+  g.font = `${size}px ${FONT_TITLE}, serif`
+  const x = W - u * 0.09
+  Array.from(text).forEach((c, i) => g.fillText(c, x, H * 0.08 + size * 0.6 + i * size * 1.06))
+  g.shadowBlur = 0
+  drawSeal(g, x - u * 0.03, H * 0.08 + size * (Array.from(text).length + 0.3) * 1.06, u * 0.06, '暮归')
+  g.font = `${u * 0.026}px ${FONT_BODY}, serif`
+  g.fillStyle = 'rgba(255,236,210,0.85)'
+  g.textAlign = 'left'
+  g.fillText('暮归江南 · 原创3D动画', u * 0.05, H - u * 0.06)
+  g.restore()
+}
+
+async function pickCodec() {
+  for (const codec of ['avc1.640028', 'avc1.4D0028', 'avc1.42E028', 'avc1.640032']) {
+    const cfg: VideoEncoderConfig = { codec, width: W, height: H, bitrate: 10e6, framerate: FPS, avc: { format: 'avc' }, latencyMode: 'quality' }
+    try { if ((await VideoEncoder.isConfigSupported(cfg)).supported) return cfg } catch { /* next */ }
+  }
+  throw new Error('当前浏览器不支持 H.264 编码')
+}
+
+export interface LiveResult { zip: Blob; previews: { name: string; url: string }[]; report: string }
+
+export async function generateLivePhotos(film: Film, onProgress: (p: number, label: string) => void, signal: { cancelled: boolean }, test?: { w: number; h: number; frames: number; clips: number }): Promise<LiveResult> {
+  W = test?.w ?? 1080
+  H = test?.h ?? 1440
+  FRAMES = test?.frames ?? 90
+  const clips = LIVE_CLIPS.slice(0, test?.clips ?? LIVE_CLIPS.length)
+  const cfg = await pickCodec()
+  const files: Zippable = {}
+  const previews: LiveResult['previews'] = []
+  const manifest: unknown[] = []
+  const root = '暮归江南_实况图'
+  const comp = document.createElement('canvas')
+  comp.width = W
+  comp.height = H
+  const g = comp.getContext('2d')!
+  const prevW = film.width, prevH = film.height
+  film.setSize(W, H)
+  try {
+    for (let ci = 0; ci < clips.length; ci++) {
+      const clip = clips[ci]
+      const samples: VideoSample[] = []
+      let avcC: Uint8Array | null = null
+      let err: unknown = null
+      let lastTs = -1
+      const enc = new VideoEncoder({
+        output: (chunk, meta) => {
+          if (meta?.decoderConfig?.description && !avcC) {
+            const d = meta.decoderConfig.description
+            avcC = d instanceof ArrayBuffer ? new Uint8Array(d.slice(0)) : new Uint8Array((d as ArrayBufferView).buffer.slice((d as ArrayBufferView).byteOffset, (d as ArrayBufferView).byteOffset + (d as ArrayBufferView).byteLength))
+          }
+          if (chunk.timestamp <= lastTs) err = new Error('编码器输出了乱序帧（B 帧），无法封装实况图')
+          lastTs = chunk.timestamp
+          const data = new Uint8Array(chunk.byteLength)
+          chunk.copyTo(data)
+          samples.push({ data, key: chunk.type === 'key' })
+        },
+        error: (e) => { err = e },
+      })
+      enc.configure(cfg)
+      const coverIdx = Math.min(FRAMES - 1, Math.round(clip.cover * FPS))
+      let coverJpeg: Uint8Array | null = null
+      for (let i = 0; i < FRAMES; i++) {
+        if (signal.cancelled) throw new Error('已取消')
+        if (err) throw err
+        const t = clip.t0 + i / FPS
+        film.renderFrame(t)
+        g.drawImage(film.renderer.domElement, 0, 0, W, H)
+        caption(g, clip.caption)
+        if (i === coverIdx) {
+          const blob = await new Promise<Blob>((res, rej) => comp.toBlob((b) => (b ? res(b) : rej(new Error('封面导出失败'))), 'image/jpeg', 0.95))
+          coverJpeg = new Uint8Array(await blob.arrayBuffer())
+          previews.push({ name: `${clip.id} ${clip.name}`, url: URL.createObjectURL(blob) })
+        }
+        const frame = new VideoFrame(comp, { timestamp: Math.round((i * 1e6) / FPS), duration: Math.round(1e6 / FPS) })
+        enc.encode(frame, { keyFrame: i === 0 })
+        frame.close()
+        while (enc.encodeQueueSize > 3) await new Promise((r) => setTimeout(r, 2))
+        if (i % 6 === 0) {
+          onProgress((ci + i / FRAMES) / clips.length, `正在渲染第 ${ci + 1}/${clips.length} 张：${clip.name}`)
+          await new Promise((r) => setTimeout(r, 0))
+        }
+      }
+      await enc.flush()
+      enc.close()
+      if (err) throw err
+      if (!avcC || !coverJpeg || samples.length !== FRAMES) throw new Error('视频编码不完整')
+      const id = uuid4()
+      const mov = writeLiveMov({ samples, avcC, width: W, height: H, fps: FPS, coverTime: coverIdx / FPS, id })
+      const jpg = writeLiveJpeg(coverJpeg, id, W, H)
+      if (inspectLiveJpeg(jpg) !== id) throw new Error('JPEG 标识读回失败')
+      const base = `${clip.id}_${clip.name}`
+      files[`${root}/1_jpg+mov（Mac 直接导入照片）/${base}.jpg`] = [jpg, { level: 0 }]
+      files[`${root}/1_jpg+mov（Mac 直接导入照片）/${base}.mov`] = [mov, { level: 0 }]
+      files[`${root}/2_pvt实况资源包/${base}.pvt/${base}.jpg`] = [jpg, { level: 0 }]
+      files[`${root}/2_pvt实况资源包/${base}.pvt/${base}.mov`] = [mov, { level: 0 }]
+      files[`${root}/2_pvt实况资源包/${base}.pvt/metadata.plist`] = strToU8(PVT_PLIST)
+      manifest.push({
+        name: base, identifier: id, photo: `${base}.jpg`, video: `${base}.mov`, coverTime: coverIdx / FPS, duration: FRAMES / FPS, fps: FPS,
+        width: W, height: H, backend: 'browser-webcodecs（LiveCanvas portable-python 配对器的 TypeScript 移植）', metadataVerified: true,
+        localPHLivePhotoLoad: 'not_available', iPhonePlayback: 'not_tested', photosImport: 'not_requested',
+        sha256: { photo: await sha256(jpg), video: await sha256(mov) },
+      })
+    }
+  } finally {
+    film.setSize(prevW, prevH)
+  }
+  const copy = LIVE_COPY.map((c) => `【${c.name}】\n标题（任选其一，≤${c.titleMax} 字）：\n${c.titles.map((t) => '· ' + t).join('\n')}\n\n正文：\n${c.body}\n\n话题（${c.tags.length}/${c.tagMax}）：\n${c.tags.map((t) => '#' + t).join(' ')}\n`).join('\n————————————\n\n')
+  const readme = `暮归江南 · 实况图（Apple Live Photo）\n\n共 ${LIVE_CLIPS.length} 组，每组 3 秒、30 fps、1080×1440（3:4），无声 H.264。\n\n使用方法（Mac）：\n1. 打开「1_jpg+mov（Mac 直接导入照片）」文件夹，全选同名的 .jpg 和 .mov，拖进「照片」App（或 文件 → 导入），会自动合成为实况照片；\n2. 通过 iCloud 照片同步到 iPhone 后，在小红书 / 抖音发布图文时选择这些实况照片即可（发布时保持“实况”开启）。\n\n「2_pvt实况资源包」是 LiveCanvas 规范的 .pvt 目录包，可在 Mac 上作为单个实况照片打开或导入。\n\n技术说明：JPEG 的 Exif MakerNote（Apple tag 17）与 MOV 的 com.apple.quicktime.content.identifier 使用同一个 UUID；MOV 内含 mebx 定时元数据轨（com.apple.quicktime.still-image-time），指向封面那一帧。生成器在浏览器本地运行，已用 LiveCanvas skill 原版 Python 校验器（verify_pair / package_live_photo）验证过同一套写入逻辑。\n导入「照片」和手机播放属于接收端操作，本包未在你的设备上测试（manifest 中记为 not_tested）。\n`
+  files[`${root}/使用说明.txt`] = strToU8(readme)
+  files[`${root}/图文发布文案.txt`] = strToU8(copy)
+  files[`${root}/manifest.json`] = strToU8(JSON.stringify(manifest, null, 2))
+  onProgress(0.99, '正在打包 ZIP')
+  await new Promise((r) => setTimeout(r, 0))
+  const zipped = zipSync(files)
+  onProgress(1, '完成')
+  return { zip: new Blob([zipped as Uint8Array<ArrayBuffer>], { type: 'application/zip' }), previews, report: `${LIVE_CLIPS.length} 组实况图已生成` }
+}

src/jiangnan/export/livephoto.ts

@@ -0,0 +1,267 @@
+// Apple Live Photo pairing in the browser.
+// Byte-for-byte port of the LiveCanvas skill's portable pairing
+// (scripts/portable_live_photo/apple.py, Apache-2.0, © 2026 杨振, modified by LiveCanvas):
+//  • JPEG: ExifIFD MakerNote (UNDEFINED) with Apple tag 17 = content identifier
+//  • MOV : moov/meta keys `com.apple.quicktime.content.identifier`
+//          + `mebx` timed-metadata track with `com.apple.quicktime.still-image-time`
+//            (signed int8 0) positioned by an edit list at the cover time.
+// Here we write the QuickTime file directly (one H.264 video track, one chunk).
+
+const enc = new TextEncoder()
+export const CONTENT_KEY = 'com.apple.quicktime.content.identifier'
+export const STILL_TIME_KEY = 'com.apple.quicktime.still-image-time'
+const TIMED_SAMPLE = new Uint8Array([0, 0, 0, 9, 0, 0, 0, 1, 0])
+export const MOVIE_TIMESCALE = 60000
+
+// ---------------------------------------------------------------- byte helpers
+class W {
+  parts: Uint8Array[] = []
+  len = 0
+  u8(v: number) { return this.bytes(new Uint8Array([v & 255])) }
+  u16(v: number) { return this.bytes(new Uint8Array([(v >> 8) & 255, v & 255])) }
+  i16(v: number) { return this.u16(v < 0 ? v + 65536 : v) }
+  u32(v: number) { const b = new Uint8Array(4); new DataView(b.buffer).setUint32(0, v >>> 0); return this.bytes(b) }
+  i32(v: number) { const b = new Uint8Array(4); new DataView(b.buffer).setInt32(0, v); return this.bytes(b) }
+  u64(v: number) { const b = new Uint8Array(8); new DataView(b.buffer).setBigUint64(0, BigInt(v)); return this.bytes(b) }
+  str(s: string) { return this.bytes(enc.encode(s)) }
+  zeros(n: number) { return this.bytes(new Uint8Array(n)) }
+  bytes(b: Uint8Array) { this.parts.push(b); this.len += b.length; return this }
+  out() { const o = new Uint8Array(this.len); let p = 0; for (const b of this.parts) { o.set(b, p); p += b.length } return o }
+}
+const cat = (...bs: Uint8Array[]) => { const w = new W(); bs.forEach((b) => w.bytes(b)); return w.out() }
+function atom(type: string, ...payload: Uint8Array[]) {
+  const body = cat(...payload)
+  return new W().u32(body.length + 8).str(type).bytes(body).out()
+}
+const fullbox = (version = 0, flags = 0) => new W().u8(version).u8(flags >> 16).u8(flags >> 8).u8(flags).out()
+const MATRIX = () => { const w = new W(); [0x10000, 0, 0, 0, 0x10000, 0, 0, 0, 0x40000000].forEach((v) => w.u32(v)); return w.out() }
+const pascal = (s: string) => { const b = enc.encode(s); return cat(new Uint8Array([b.length]), b) }
+
+export function canonicalUUID(u: string) {
+  if (!/^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$/.test(u)) throw new Error('Apple 资产标识必须是标准 UUID')
+  return u.toUpperCase()
+}
+
+// ---------------------------------------------------------------- JPEG
+export function appleMakerNote(id: string) {
+  const ident = enc.encode(canonicalUUID(id))
+  return new W().str('Apple iOS').u8(0).u8(0).u8(1).str('MM')
+    .u16(1)
+    .u16(0x11).u16(2).u32(ident.length + 1).u32(32)
+    .u32(0)
+    .bytes(ident).u8(0).out()
+}
+
+/** Build a big-endian Exif APP1 payload (TIFF) carrying the Apple MakerNote. */
+function exifTiff(id: string, width: number, height: number) {
+  const note = appleMakerNote(id)
+  const make = enc.encode('Apple\0'), model = enc.encode('LiveCanvas\0')
+  // layout: header(8) | IFD0 | ExifIFD | data
+  const ifd0Count = 4, exifCount = 6
+  const ifd0Off = 8
+  const ifd0Size = 2 + ifd0Count * 12 + 4
+  const exifOff = ifd0Off + ifd0Size
+  const exifSize = 2 + exifCount * 12 + 4
+  let dataOff = exifOff + exifSize
+  const makeOff = dataOff; dataOff += make.length
+  const modelOff = dataOff; dataOff += model.length
+  if (dataOff % 2) dataOff++
+  const noteOff = dataOff; dataOff += note.length
+  const w = new W()
+  w.str('MM').u16(42).u32(8)
+  // IFD0 (tags ascending)
+  w.u16(ifd0Count)
+  w.u16(0x010f).u16(2).u32(make.length).u32(makeOff)
+  w.u16(0x0110).u16(2).u32(model.length).u32(modelOff)
+  w.u16(0x0112).u16(3).u32(1).u16(1).u16(0) // orientation = 1
+  w.u16(0x8769).u16(4).u32(1).u32(exifOff)
+  w.u32(0)
+  // Exif IFD
+  w.u16(exifCount)
+  w.u16(0x9000).u16(7).u32(4).str('0232')
+  w.u16(0x927c).u16(7).u32(note.length).u32(noteOff) // MakerNote as UNDEFINED (type 7)
+  w.u16(0xa000).u16(7).u32(4).str('0100')
+  w.u16(0xa001).u16(3).u32(1).u16(1).u16(0)
+  w.u16(0xa002).u16(4).u32(1).u32(width)
+  w.u16(0xa003).u16(4).u32(1).u32(height)
+  w.u32(0)
+  w.bytes(make).bytes(model)
+  if (w.len % 2) w.u8(0)
+  w.bytes(note)
+  return w.out()
+}
+
+/** Insert an Exif APP1 segment (with the Apple identifier) into a baseline JPEG. */
+export function writeLiveJpeg(jpeg: Uint8Array, id: string, width: number, height: number) {
+  if (jpeg[0] !== 0xff || jpeg[1] !== 0xd8) throw new Error('不是 JPEG')
+  const tiff = exifTiff(id, width, height)
+  const payload = cat(enc.encode('Exif\0\0'), tiff)
+  const app1 = new W().u8(0xff).u8(0xe1).u16(payload.length + 2).bytes(payload).out()
+  // skip SOI and any existing APP0 (JFIF); drop existing APP1 segments
+  let p = 2
+  const keep: Uint8Array[] = []
+  while (p < jpeg.length && jpeg[p] === 0xff && jpeg[p + 1] >= 0xe0 && jpeg[p + 1] <= 0xef) {
+    const len = (jpeg[p + 2] << 8) | jpeg[p + 3]
+    if (jpeg[p + 1] === 0xe0) keep.push(jpeg.subarray(p, p + 2 + len))
+    p += 2 + len
+  }
+  return cat(jpeg.subarray(0, 2), ...keep, app1, jpeg.subarray(p))
+}
+
+/** Read back the content identifier from our JPEG (mirrors inspect_live_jpeg). */
+export function inspectLiveJpeg(jpeg: Uint8Array): string {
+  let p = 2
+  while (p < jpeg.length && jpeg[p] === 0xff) {
+    const m = jpeg[p + 1], len = (jpeg[p + 2] << 8) | jpeg[p + 3]
+    if (m === 0xe1) {
+      const seg = jpeg.subarray(p + 4, p + 2 + len)
+      const s = new TextDecoder('latin1').decode(seg)
+      const i = s.indexOf('Apple iOS\0\0\x01MM')
+      if (i >= 0) {
+        const note = seg.subarray(i)
+        const dv = new DataView(note.buffer, note.byteOffset, note.byteLength)
+        if (dv.getUint16(16) !== 0x11 || dv.getUint16(18) !== 2) throw new Error('缺少 tag 17')
+        const count = dv.getUint32(20), off = dv.getUint32(24)
+        return canonicalUUID(new TextDecoder().decode(note.subarray(off, off + count)).replace(/\0+$/, ''))
+      }
+    }
+    if (m === 0xda) break
+    p += 2 + len
+  }
+  throw new Error('JPEG 缺少 Apple MakerNote')
+}
+
+// ---------------------------------------------------------------- MOV
+export interface VideoSample { data: Uint8Array; key: boolean }
+
+export interface LiveMovInput {
+  samples: VideoSample[]
+  avcC: Uint8Array
+  width: number
+  height: number
+  fps: number
+  coverTime: number
+  id: string
+}
+
+function hdlr(type: string, subtype: string, manuf: string, name: string, flags = 0) {
+  return atom('hdlr', fullbox(), enc.encode(type), enc.encode(subtype), manuf ? enc.encode(manuf) : new Uint8Array(4), new W().u32(flags).u32(0).out(), pascal(name))
+}
+
+function contentIdentifierMeta(id: string) {
+  const ident = enc.encode(canonicalUUID(id))
+  const h = atom('hdlr', fullbox(), new Uint8Array(4), enc.encode('mdta'), new Uint8Array(12), new Uint8Array(2))
+  const keyEntry = cat(new W().u32(8 + CONTENT_KEY.length).str('mdta').out(), enc.encode(CONTENT_KEY))
+  const keys = atom('keys', fullbox(), new W().u32(1).out(), keyEntry)
+  const data = atom('data', new W().u32(1).u32(0).out(), ident)
+  const item = atom('\0\0\0\x01', data)
+  return atom('meta', h, keys, atom('ilst', item))
+}
+
+function timedMetadataTrack(trackId: number, videoTrackId: number, timescale: number, coverTicks: number, sampleOffset: number, sampleTicks: number) {
+  const dur = coverTicks + sampleTicks
+  const tkhd = atom('tkhd', fullbox(0, 0x0f), new W().u32(0).u32(0).u32(trackId).u32(0).u32(dur).zeros(8).i16(0).i16(0).i16(0).i16(0).out(), MATRIX(), new W().u32(0).u32(0).out())
+  const edits = new W()
+  let n = 0
+  if (coverTicks) { edits.u32(coverTicks).i32(-1).u16(1).u16(0); n++ }
+  edits.u32(sampleTicks).i32(0).u16(1).u16(0); n++
+  const elst = atom('elst', fullbox(), new W().u32(n).out(), edits.out())
+  const edts = atom('edts', elst)
+  const tref = atom('tref', atom('cdsc', new W().u32(videoTrackId).out()))
+  const mdhd = atom('mdhd', fullbox(), new W().u32(0).u32(0).u32(timescale).u32(sampleTicks).u16(0x55c4).u16(0).out())
+  const h = atom('hdlr', fullbox(), enc.encode('mhlrmetaappl'), new W().u32(1).u32(0).out(), pascal('Core Media Metadata'))
+  const gmin = atom('gmin', fullbox(), new W().u16(0x40).u16(0x8000).u16(0x8000).u16(0x8000).u16(0).u16(0).out())
+  const gmhd = atom('gmhd', gmin)
+  const dh = atom('hdlr', fullbox(), enc.encode('dhlralisappl'), new W().u32(0).u32(0).out(), pascal('Core Media Data Handler'))
+  const dinf = atom('dinf', atom('dref', fullbox(), new W().u32(1).out(), atom('alis', fullbox(0, 1))))
+  const keyd = atom('keyd', enc.encode('mdta'), enc.encode(STILL_TIME_KEY))
+  const dtyp = atom('dtyp', fullbox(), new W().u32(0x41).out())
+  const keyRecord = cat(new W().u32(8 + keyd.length + dtyp.length).u32(1).out(), keyd, dtyp)
+  const keys = atom('keys', keyRecord)
+  const mebx = atom('mebx', new Uint8Array(6), new W().u16(1).out(), keys)
+  const stsd = atom('stsd', fullbox(), new W().u32(1).out(), mebx)
+  const stts = atom('stts', fullbox(), new W().u32(1).u32(1).u32(sampleTicks).out())
+  const stsc = atom('stsc', fullbox(), new W().u32(1).u32(1).u32(1).u32(1).out())
+  const stsz = atom('stsz', fullbox(), new W().u32(0).u32(1).u32(TIMED_SAMPLE.length).out())
+  const stco = atom('stco', fullbox(), new W().u32(1).u32(sampleOffset).out())
+  const stbl = atom('stbl', stsd, stts, stsc, stsz, stco)
+  const minf = atom('minf', gmhd, dh, dinf, stbl)
+  const mdia = atom('mdia', mdhd, h, minf)
+  return atom('trak', tkhd, edts, tref, mdia)
+}
+
+/** QuickTime movie: ftyp | mdat (video samples + timed sample) | moov. */
+export function writeLiveMov(inp: LiveMovInput) {
+  const { samples, avcC, width, height, fps, id } = inp
+  const ts = MOVIE_TIMESCALE
+  const delta = Math.round(ts / fps)
+  const n = samples.length
+  const duration = n * delta
+  const ftyp = atom('ftyp', enc.encode('qt  '), new W().u32(0x200).out(), enc.encode('qt  '))
+  const mediaLen = samples.reduce((a, s) => a + s.data.length, 0) + TIMED_SAMPLE.length
+  const mdatHeader = new W().u32(mediaLen + 8).str('mdat').out()
+  const firstOffset = ftyp.length + mdatHeader.length
+  const timedOffset = firstOffset + mediaLen - TIMED_SAMPLE.length
+  // ---- video track
+  const VID = 1, META = 2
+  const tkhd = atom('tkhd', fullbox(0, 0x03), new W().u32(0).u32(0).u32(VID).u32(0).u32(duration).zeros(8).i16(0).i16(0).i16(0).i16(0).out(), MATRIX(), new W().u32(width * 65536).u32(height * 65536).out())
+  const mdhd = atom('mdhd', fullbox(), new W().u32(0).u32(0).u32(ts).u32(duration).u16(0x55c4).u16(0).out())
+  const vh = hdlr('mhlr', 'vide', 'appl', 'Core Media Video')
+  const vmhd = atom('vmhd', fullbox(0, 1), new W().u16(0).u16(0).u16(0).u16(0).out())
+  const dh = hdlr('dhlr', 'alis', 'appl', 'Core Media Data Handler')
+  const dinf = atom('dinf', atom('dref', fullbox(), new W().u32(1).out(), atom('alis', fullbox(0, 1))))
+  const compressor = new Uint8Array(32)
+  const cname = enc.encode('H.264')
+  compressor[0] = cname.length
+  compressor.set(cname, 1)
+  const colr = atom('colr', enc.encode('nclx'), new W().u16(1).u16(1).u16(1).out())
+  const pasp = atom('pasp', new W().u32(1).u32(1).out())
+  const avc1 = atom('avc1', new Uint8Array(6), new W().u16(1).u16(0).u16(0).u32(0).u32(0).u32(0x200).u16(width).u16(height).u32(0x480000).u32(0x480000).u32(0).u16(1).out(), compressor, new W().u16(0x18).i16(-1).out(), atom('avcC', avcC), colr, pasp)
+  const stsd = atom('stsd', fullbox(), new W().u32(1).out(), avc1)
+  const stts = atom('stts', fullbox(), new W().u32(1).u32(n).u32(delta).out())
+  const keys = samples.map((s, i) => (s.key ? i + 1 : 0)).filter(Boolean)
+  const stssW = new W().u32(keys.length)
+  keys.forEach((k) => stssW.u32(k))
+  const stss = atom('stss', fullbox(), stssW.out())
+  const stsc = atom('stsc', fullbox(), new W().u32(1).u32(1).u32(n).u32(1).out())
+  const szW = new W().u32(0).u32(n)
+  samples.forEach((s) => szW.u32(s.data.length))
+  const stsz = atom('stsz', fullbox(), szW.out())
+  const stco = atom('stco', fullbox(), new W().u32(1).u32(firstOffset).out())
+  const stbl = atom('stbl', stsd, stts, stss, stsc, stsz, stco)
+  const minf = atom('minf', vmhd, dh, dinf, stbl)
+  const mdia = atom('mdia', mdhd, vh, minf)
+  const vtrak = atom('trak', tkhd, mdia)
+  // ---- cover time (same clamping as the Python pairing tool)
+  let coverTicks = Math.round(inp.coverTime * ts)
+  coverTicks = Math.min(Math.max(0, coverTicks), Math.max(0, duration - 1))
+  const mtrak = timedMetadataTrack(META, VID, ts, coverTicks, timedOffset, Math.max(1, Math.round(ts / fps)))
+  const mvhd = atom('mvhd', fullbox(), new W().u32(0).u32(0).u32(ts).u32(duration).u32(0x10000).u16(0x100).zeros(10).out(), MATRIX(), new W().zeros(24).u32(META + 1).out())
+  const moov = atom('moov', mvhd, vtrak, contentIdentifierMeta(id), mtrak)
+  return cat(ftyp, mdatHeader, ...samples.map((s) => s.data), TIMED_SAMPLE, moov)
+}
+
+export const PVT_PLIST = `<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>PFVideoComplementMetadataVersionKey</key>
+	<string>1</string>
+</dict>
+</plist>
+`
+
+export function uuid4(rand: () => number = Math.random) {
+  const b = new Uint8Array(16)
+  if (typeof crypto !== 'undefined' && crypto.getRandomValues) crypto.getRandomValues(b)
+  else for (let i = 0; i < 16; i++) b[i] = Math.floor(rand() * 256)
+  b[6] = (b[6] & 0x0f) | 0x40
+  b[8] = (b[8] & 0x3f) | 0x80
+  const h = Array.from(b, (x) => x.toString(16).padStart(2, '0')).join('')
+  return `${h.slice(0, 8)}-${h.slice(8, 12)}-${h.slice(12, 16)}-${h.slice(16, 20)}-${h.slice(20)}`.toUpperCase()
+}
+
+export async function sha256(data: Uint8Array) {
+  const d = await crypto.subtle.digest('SHA-256', data as Uint8Array<ArrayBuffer>)
+  return Array.from(new Uint8Array(d), (x) => x.toString(16).padStart(2, '0')).join('')
+}

src/jiangnan/export/mp4.ts

@@ -0,0 +1,138 @@
+// In-browser MP4 export: deterministic frame-by-frame rendering → WebCodecs
+// H.264 (+ AAC / Opus audio) → mp4-muxer. Nothing leaves the user's device.
+import { ArrayBufferTarget, Muxer } from 'mp4-muxer'
+import type { Film } from '../engine/Film'
+import { DURATION } from '../engine/timeline'
+import { drawOverlay, OverlayOpts } from '../player/overlay'
+
+export interface ExportOptions {
+  width: number
+  height: number
+  fps: 24 | 30 | 60
+  audio: AudioBuffer | null
+  overlay: OverlayOpts
+  onProgress: (p: number, info: { frame: number; total: number; eta: number; stage: string }) => void
+  signal: { cancelled: boolean }
+  duration?: number
+  start?: number
+}
+
+export function exportSupport() {
+  return {
+    video: typeof window !== 'undefined' && 'VideoEncoder' in window,
+    audio: typeof window !== 'undefined' && 'AudioEncoder' in window,
+  }
+}
+
+async function pickVideoCodec(w: number, h: number, fps: number, bitrate: number) {
+  const candidates = ['avc1.640034', 'avc1.640033', 'avc1.64002A', 'avc1.640028', 'avc1.4D4033', 'avc1.4D0028', 'avc1.42E033', 'avc1.42E028', 'avc1.42001f']
+  for (const codec of candidates) {
+    for (const hw of ['prefer-hardware', 'no-preference'] as const) {
+      const cfg: VideoEncoderConfig = { codec, width: w, height: h, bitrate, framerate: fps, hardwareAcceleration: hw, avc: { format: 'avc' }, latencyMode: 'quality' }
+      try {
+        const s = await VideoEncoder.isConfigSupported(cfg)
+        if (s.supported) return cfg
+      } catch { /* try next */ }
+    }
+  }
+  throw new Error('当前浏览器不支持 H.264 编码，请使用最新版 Chrome / Edge / Safari')
+}
+
+async function pickAudioCodec() {
+  if (!('AudioEncoder' in window)) return null
+  for (const [codec, mux] of [['mp4a.40.2', 'aac'], ['opus', 'opus']] as const) {
+    const cfg: AudioEncoderConfig = { codec, sampleRate: 48000, numberOfChannels: 2, bitrate: 192000 }
+    try {
+      const s = await AudioEncoder.isConfigSupported(cfg)
+      if (s.supported) return { cfg, mux }
+    } catch { /* next */ }
+  }
+  return null
+}
+
+export function suggestedBitrate(w: number, h: number, fps: number) {
+  const px = w * h
+  const base = px >= 1920 * 1080 * 0.9 ? 14e6 : 7e6
+  return Math.round(base * (fps >= 50 ? 1.5 : fps <= 24 ? 0.85 : 1))
+}
+
+export async function exportMp4(film: Film, o: ExportOptions): Promise<{ blob: Blob; audioCodec: string | null }> {
+  const { width: W, height: H, fps } = o
+  const start = o.start ?? 0
+  const dur = o.duration ?? DURATION
+  const total = Math.round(dur * fps)
+  const bitrate = suggestedBitrate(W, H, fps)
+  const vcfg = await pickVideoCodec(W, H, fps, bitrate)
+  const acodec = o.audio ? await pickAudioCodec() : null
+  const target = new ArrayBufferTarget()
+  const muxer = new Muxer({
+    target,
+    video: { codec: 'avc', width: W, height: H, frameRate: fps },
+    ...(acodec ? { audio: { codec: acodec.mux, sampleRate: 48000, numberOfChannels: 2 } } : {}),
+    fastStart: 'in-memory',
+    firstTimestampBehavior: 'offset',
+  })
+  let encErr: unknown = null
+
+  // ---- audio first (fast)
+  if (acodec && o.audio) {
+    o.onProgress(0, { frame: 0, total, eta: 0, stage: '编码音轨' })
+    const ae = new AudioEncoder({ output: (c, m) => muxer.addAudioChunk(c, m), error: (e) => { encErr = e } })
+    ae.configure(acodec.cfg)
+    const buf = o.audio
+    const L = buf.getChannelData(0), R = buf.getChannelData(1)
+    const s0 = Math.floor(start * 48000)
+    const len = Math.min(L.length - s0, Math.round(dur * 48000))
+    const block = 4800
+    for (let i = 0; i < len; i += block) {
+      const n = Math.min(block, len - i)
+      const data = new Float32Array(n * 2)
+      data.set(L.subarray(s0 + i, s0 + i + n), 0)
+      data.set(R.subarray(s0 + i, s0 + i + n), n)
+      const ad = new AudioData({ format: 'f32-planar', sampleRate: 48000, numberOfFrames: n, numberOfChannels: 2, timestamp: Math.round((i / 48000) * 1e6), data })
+      ae.encode(ad)
+      ad.close()
+      if (ae.encodeQueueSize > 20) await new Promise((r) => setTimeout(r, 0))
+    }
+    await ae.flush()
+    ae.close()
+  }
+
+  // ---- video
+  const ve = new VideoEncoder({ output: (c, m) => muxer.addVideoChunk(c, m), error: (e) => { encErr = e } })
+  ve.configure(vcfg)
+  const comp = document.createElement('canvas')
+  comp.width = W
+  comp.height = H
+  const g = comp.getContext('2d')!
+  const prevW = film.width, prevH = film.height
+  film.setSize(W, H)
+  const t0 = performance.now()
+  try {
+    for (let i = 0; i < total; i++) {
+      if (o.signal.cancelled) throw new Error('已取消')
+      if (encErr) throw encErr
+      const t = start + i / fps
+      film.renderFrame(t)
+      g.drawImage(film.renderer.domElement, 0, 0, W, H)
+      drawOverlay(g, t, W, H, o.overlay)
+      const frame = new VideoFrame(comp, { timestamp: Math.round((i * 1e6) / fps), duration: Math.round(1e6 / fps) })
+      ve.encode(frame, { keyFrame: i % (fps * 2) === 0 })
+      frame.close()
+      while (ve.encodeQueueSize > 3) await new Promise((r) => setTimeout(r, 2))
+      if (i % 3 === 0) {
+        const el = (performance.now() - t0) / 1000
+        o.onProgress((i + 1) / total, { frame: i + 1, total, eta: (el / (i + 1)) * (total - i - 1), stage: '渲染画面' })
+        await new Promise((r) => setTimeout(r, 0))
+      }
+    }
+    o.onProgress(1, { frame: total, total, eta: 0, stage: '封装 MP4' })
+    await ve.flush()
+    ve.close()
+    muxer.finalize()
+  } finally {
+    film.setSize(prevW, prevH)
+    if (ve.state !== 'closed') try { ve.close() } catch { /* ignore */ }
+  }
+  return { blob: new Blob([target.buffer], { type: 'video/mp4' }), audioCodec: acodec?.mux ?? null }
+}

src/jiangnan/jiangnan.css

@@ -0,0 +1,110 @@
+.jn {
+  --ink: #0b0d14;
+  --ink2: #12151f;
+  --ink3: #1a1e2b;
+  --line: rgba(255, 236, 210, 0.1);
+  --line2: rgba(255, 236, 210, 0.18);
+  --paper: #f3eadb;
+  --muted: rgba(243, 234, 219, 0.62);
+  --dim: rgba(243, 234, 219, 0.42);
+  --amber: #e8a65a;
+  --red: #c8372d;
+  --ok: #7fc59a;
+  --bad: #ef7b6e;
+  min-height: 100vh;
+  background: radial-gradient(1200px 600px at 50% -10%, #241a1a 0%, var(--ink) 60%);
+  color: var(--paper);
+  font-family: 'JNBody', 'PingFang SC', 'Noto Serif SC', 'Songti SC', serif;
+  -webkit-font-smoothing: antialiased;
+}
+.jn * { box-sizing: border-box; }
+.jn-wrap { max-width: 1180px; margin: 0 auto; padding: 28px 20px 80px; }
+.jn-head { display: flex; align-items: flex-end; justify-content: space-between; gap: 16px; margin-bottom: 18px; flex-wrap: wrap; }
+.jn-title { font-family: 'JNTitle', serif; font-size: 52px; line-height: 1; letter-spacing: 6px; margin: 0; color: #fff3e2; }
+.jn-sub { color: var(--muted); font-size: 15px; margin-top: 10px; letter-spacing: 1px; }
+.jn-seal { display: inline-flex; align-items: center; justify-content: center; width: 38px; height: 38px; background: var(--red); color: #fff1e6; font-family: 'JNTitle', serif; font-size: 14px; line-height: 1.05; border-radius: 4px; writing-mode: vertical-rl; margin-left: 12px; vertical-align: 10px; }
+.jn-links { color: var(--dim); font-size: 13px; text-align: right; }
+.jn-links a { color: var(--amber); text-decoration: none; }
+
+/* player */
+.jn-player { position: relative; margin: 0 auto; background: #000; border-radius: 14px; overflow: hidden; box-shadow: 0 30px 80px rgba(0, 0, 0, 0.55), 0 0 0 1px var(--line); }
+.jn-player.landscape { width: 100%; aspect-ratio: 16 / 9; }
+.jn-player.portrait { height: min(78vh, 900px); aspect-ratio: 9 / 16; max-width: 100%; }
+.jn-player canvas { position: absolute; inset: 0; width: 100%; height: 100%; display: block; }
+.jn-player:fullscreen { border-radius: 0; width: 100vw; height: 100vh; aspect-ratio: auto; display: flex; align-items: center; justify-content: center; background: #000; }
+.jn-player:fullscreen canvas { position: absolute; inset: 0; width: 100%; height: 100%; object-fit: contain; }
+.jn-big-play { position: absolute; inset: 0; display: flex; align-items: center; justify-content: center; background: transparent; border: 0; cursor: pointer; }
+.jn-big-play span { width: 84px; height: 84px; border-radius: 50%; background: rgba(12, 10, 14, 0.5); border: 1px solid rgba(255, 236, 210, 0.45); display: flex; align-items: center; justify-content: center; backdrop-filter: blur(6px); transition: transform 0.2s; }
+.jn-big-play:hover span { transform: scale(1.06); }
+.jn-loading { position: absolute; inset: 0; display: flex; flex-direction: column; align-items: center; justify-content: center; gap: 14px; background: radial-gradient(circle at 50% 40%, #1d1716, #07080c); z-index: 5; text-align: center; padding: 20px; }
+.jn-loading .t { font-family: 'JNTitle', serif; font-size: 40px; letter-spacing: 6px; }
+.jn-bar { width: min(360px, 80%); height: 3px; background: rgba(255, 255, 255, 0.08); border-radius: 3px; overflow: hidden; }
+.jn-bar i { display: block; height: 100%; background: linear-gradient(90deg, #c8372d, #e8a65a); transition: width 0.25s; }
+.jn-loading small { color: var(--muted); font-size: 13px; }
+
+.jn-controls { display: flex; align-items: center; gap: 10px; padding: 12px 4px 0; flex-wrap: wrap; }
+.jn-btn { appearance: none; border: 1px solid var(--line2); background: var(--ink3); color: var(--paper); padding: 8px 14px; border-radius: 999px; font: inherit; font-size: 14px; cursor: pointer; display: inline-flex; align-items: center; gap: 6px; transition: background 0.15s, border-color 0.15s; white-space: nowrap; }
+.jn-btn:hover { border-color: rgba(232, 166, 90, 0.6); }
+.jn-btn:disabled { opacity: 0.45; cursor: not-allowed; }
+.jn-btn.primary { background: linear-gradient(135deg, #c8372d, #b1522a); border-color: transparent; color: #fff4ea; }
+.jn-btn.ghost { background: transparent; }
+.jn-btn.on { border-color: var(--amber); color: #ffd9a8; }
+.jn-time { font-variant-numeric: tabular-nums; color: var(--muted); font-size: 13px; min-width: 92px; }
+.jn-scrub { flex: 1 1 260px; position: relative; height: 34px; display: flex; align-items: center; }
+.jn-scrub input { width: 100%; accent-color: var(--amber); }
+.jn-shots { display: flex; flex-wrap: wrap; gap: 6px; margin-top: 10px; }
+.jn-chip { font-size: 12px; padding: 4px 10px; border-radius: 999px; border: 1px solid var(--line); color: var(--muted); background: transparent; cursor: pointer; font-family: inherit; }
+.jn-chip.active { border-color: var(--amber); color: #ffd9a8; background: rgba(232, 166, 90, 0.08); }
+
+/* sections */
+.jn-sec { margin-top: 44px; border-top: 1px solid var(--line); padding-top: 26px; }
+.jn-sec h2 { font-family: 'JNTitle', serif; font-weight: normal; font-size: 30px; letter-spacing: 3px; margin: 0 0 6px; color: #fff1dc; }
+.jn-sec .lead { color: var(--muted); font-size: 14px; margin: 0 0 18px; line-height: 1.7; }
+.jn-grid { display: grid; grid-template-columns: repeat(auto-fill, minmax(240px, 1fr)); gap: 14px; }
+.jn-card { background: var(--ink2); border: 1px solid var(--line); border-radius: 12px; padding: 16px; }
+.jn-opt { display: flex; flex-wrap: wrap; gap: 8px; margin: 6px 0 14px; }
+.jn-label { font-size: 13px; color: var(--dim); margin-bottom: 4px; letter-spacing: 1px; }
+.jn-row { display: flex; gap: 10px; align-items: center; flex-wrap: wrap; }
+.jn-progress { margin-top: 14px; }
+.jn-progress .jn-bar { width: 100%; height: 6px; }
+.jn-progress small { display: block; margin-top: 8px; color: var(--muted); font-size: 13px; }
+.jn-note { font-size: 12.5px; color: var(--dim); line-height: 1.7; margin-top: 10px; }
+.jn-err { color: var(--bad); font-size: 13px; margin-top: 10px; }
+.jn-result video { width: 100%; max-height: 420px; background: #000; border-radius: 10px; margin-top: 12px; }
+
+/* covers */
+.jn-covers { display: grid; grid-template-columns: repeat(auto-fill, minmax(220px, 1fr)); gap: 14px; margin-top: 16px; }
+.jn-cover { position: relative; border-radius: 10px; overflow: hidden; border: 2px solid transparent; background: #000; cursor: pointer; }
+.jn-cover.sel { border-color: var(--amber); }
+.jn-cover img { width: 100%; display: block; }
+.jn-cover .meta { display: flex; justify-content: space-between; align-items: center; padding: 8px 10px; font-size: 12.5px; background: var(--ink2); color: var(--muted); gap: 8px; }
+.jn-cover .chk { position: absolute; top: 8px; left: 8px; width: 24px; height: 24px; border-radius: 6px; background: rgba(0, 0, 0, 0.55); border: 1px solid rgba(255, 255, 255, 0.5); display: flex; align-items: center; justify-content: center; font-size: 15px; color: #ffd9a8; }
+.jn-cover a { color: var(--amber); text-decoration: none; }
+
+/* copy */
+.jn-tabs { display: flex; gap: 6px; flex-wrap: wrap; margin-bottom: 14px; }
+.jn-titles { display: flex; flex-direction: column; gap: 8px; }
+.jn-tline { display: flex; align-items: center; gap: 10px; padding: 10px 12px; border-radius: 10px; border: 1px solid var(--line); background: var(--ink2); cursor: pointer; }
+.jn-tline.sel { border-color: rgba(232, 166, 90, 0.7); }
+.jn-tline .txt { flex: 1; font-size: 15px; line-height: 1.5; }
+.jn-count { font-size: 12px; padding: 2px 8px; border-radius: 999px; font-variant-numeric: tabular-nums; white-space: nowrap; }
+.jn-count.ok { color: var(--ok); background: rgba(127, 197, 154, 0.1); }
+.jn-count.bad { color: var(--bad); background: rgba(239, 123, 110, 0.1); }
+.jn-box { white-space: pre-wrap; background: var(--ink2); border: 1px solid var(--line); border-radius: 10px; padding: 12px 14px; font-size: 14px; line-height: 1.75; color: #efe5d4; }
+.jn-tags { display: flex; flex-wrap: wrap; gap: 6px; }
+.jn-tag { font-size: 13px; padding: 4px 10px; border-radius: 999px; background: rgba(200, 55, 45, 0.12); color: #ffc8b8; border: 1px solid rgba(200, 55, 45, 0.3); }
+.jn-copyrow { display: flex; justify-content: space-between; align-items: center; margin: 18px 0 8px; gap: 10px; }
+.jn-copyrow .jn-label { margin: 0; }
+.jn-toast { position: fixed; left: 50%; bottom: 28px; transform: translateX(-50%); background: rgba(20, 16, 18, 0.95); border: 1px solid var(--line2); color: var(--paper); padding: 10px 18px; border-radius: 999px; font-size: 14px; z-index: 100; box-shadow: 0 10px 30px rgba(0, 0, 0, 0.4); }
+.jn-live { display: grid; grid-template-columns: repeat(auto-fill, minmax(150px, 1fr)); gap: 12px; margin-top: 14px; }
+.jn-live figure { margin: 0; }
+.jn-live img { width: 100%; border-radius: 8px; display: block; }
+.jn-live figcaption { font-size: 12.5px; color: var(--muted); margin-top: 6px; }
+.jn-foot { margin-top: 56px; color: var(--dim); font-size: 12.5px; line-height: 1.8; border-top: 1px solid var(--line); padding-top: 18px; }
+.jn-foot code { color: var(--muted); }
+@media (max-width: 640px) {
+  .jn-wrap { padding: 16px 12px 60px; }
+  .jn-title { font-size: 40px; }
+  .jn-player.portrait { height: auto; width: 100%; }
+  .jn-sec h2 { font-size: 26px; }
+}

src/jiangnan/player/overlay.ts

@@ -0,0 +1,142 @@
+// Typography drawn on top of the 3D frame (title card, subtitles, end card).
+// The same function is used for the live preview, MP4 export, covers and live photos.
+import titleFont from '../assets/fonts/title.woff2?url'
+import bodyFont from '../assets/fonts/wenkai.woff2?url'
+import { cueAt, DURATION, endAlpha, SUBTITLE, TITLE, titleAlpha } from '../engine/timeline'
+
+export const FONT_TITLE = 'JNTitle'
+export const FONT_BODY = 'JNBody'
+let fontsReady: Promise<void> | null = null
+
+export function loadFonts() {
+  if (fontsReady) return fontsReady
+  fontsReady = (async () => {
+    const faces = [new FontFace(FONT_TITLE, `url(${titleFont})`), new FontFace(FONT_BODY, `url(${bodyFont})`)]
+    await Promise.all(faces.map(async (f) => { await f.load(); document.fonts.add(f) }))
+  })().catch(() => undefined)
+  return fontsReady
+}
+
+const clamp = (v: number) => Math.max(0, Math.min(1, v))
+
+/** Red square seal (印章) with two characters. */
+export function drawSeal(g: CanvasRenderingContext2D, x: number, y: number, size: number, text = '江南', alpha = 1) {
+  g.save()
+  g.globalAlpha *= alpha
+  g.fillStyle = '#b3261e'
+  const r = size * 0.08
+  g.beginPath()
+  g.roundRect(x, y, size, size, r)
+  g.fill()
+  g.strokeStyle = 'rgba(255,235,220,0.9)'
+  g.lineWidth = size * 0.04
+  g.strokeRect(x + size * 0.1, y + size * 0.1, size * 0.8, size * 0.8)
+  g.fillStyle = '#fff1e6'
+  g.font = `${size * 0.36}px ${FONT_TITLE}, serif`
+  g.textAlign = 'center'
+  g.textBaseline = 'middle'
+  const chars = Array.from(text)
+  if (chars.length === 2) {
+    g.fillText(chars[0], x + size / 2, y + size * 0.32)
+    g.fillText(chars[1], x + size / 2, y + size * 0.7)
+  } else g.fillText(text, x + size / 2, y + size / 2)
+  g.restore()
+}
+
+function spaced(g: CanvasRenderingContext2D, text: string, cx: number, y: number, spacing: number) {
+  const chars = Array.from(text)
+  const widths = chars.map((c) => g.measureText(c).width)
+  const total = widths.reduce((a, b) => a + b, 0) + spacing * (chars.length - 1)
+  let x = cx - total / 2
+  chars.forEach((c, i) => {
+    g.fillText(c, x + widths[i] / 2, y)
+    x += widths[i] + spacing
+  })
+}
+
+export interface OverlayOpts { subtitles: boolean; titles: boolean }
+
+export function drawOverlay(g: CanvasRenderingContext2D, t: number, W: number, H: number, opts: OverlayOpts = { subtitles: true, titles: true }) {
+  const portrait = H > W
+  const unit = Math.min(W, H)
+  g.save()
+  g.textAlign = 'center'
+  g.textBaseline = 'middle'
+  // ---- opening title
+  const ta = opts.titles ? titleAlpha(t) : 0
+  if (ta > 0.001) {
+    const rise = (1 - ta) * unit * 0.015
+    const cy = portrait ? H * 0.36 : H * 0.42
+    g.globalAlpha = ta
+    g.shadowColor = 'rgba(40,16,0,0.55)'
+    g.shadowBlur = unit * 0.04
+    g.fillStyle = '#fff6e8'
+    g.font = `${unit * (portrait ? 0.2 : 0.17)}px ${FONT_TITLE}, serif`
+    spaced(g, TITLE, W / 2, cy + rise, unit * 0.025)
+    g.shadowBlur = unit * 0.02
+    g.font = `${unit * 0.038}px ${FONT_BODY}, serif`
+    g.fillStyle = 'rgba(255,240,220,0.92)'
+    spaced(g, SUBTITLE, W / 2, cy + unit * (portrait ? 0.17 : 0.15) + rise, unit * 0.012)
+    g.shadowBlur = 0
+    drawSeal(g, W / 2 + unit * (portrait ? 0.3 : 0.27), cy - unit * 0.1 + rise, unit * 0.06, '江南', ta)
+  }
+  // ---- end card
+  const ea = opts.titles ? endAlpha(t) : 0
+  if (ea > 0.001) {
+    g.globalAlpha = ea
+    g.shadowColor = 'rgba(0,0,0,0.6)'
+    g.shadowBlur = unit * 0.03
+    g.fillStyle = '#fff3e0'
+    const cy = portrait ? H * 0.3 : H * 0.36
+    g.font = `${unit * 0.11}px ${FONT_TITLE}, serif`
+    spaced(g, TITLE, W / 2, cy, unit * 0.018)
+    g.font = `${unit * 0.03}px ${FONT_BODY}, serif`
+    g.fillStyle = 'rgba(255,236,210,0.85)'
+    spaced(g, '原创 3D 动画 · 程序化建模 · 实时渲染', W / 2, cy + unit * 0.11, unit * 0.006)
+    spaced(g, '配乐 · 音效 · 旁白 原创', W / 2, cy + unit * 0.16, unit * 0.006)
+    g.shadowBlur = 0
+    drawSeal(g, W / 2 - unit * 0.03, cy + unit * 0.21, unit * 0.06, '暮归', ea)
+  }
+  // ---- subtitles
+  const c = opts.subtitles ? cueAt(t) : null
+  if (c) {
+    const a = clamp((t - c.t + 0.15) / 0.3) * clamp((c.t + c.dur + 0.5 - t) / 0.35)
+    if (a > 0.001) {
+      const size = portrait ? W * 0.046 : H * 0.043
+      g.globalAlpha = a
+      g.font = `${size}px ${FONT_BODY}, serif`
+      const maxW = W * (portrait ? 0.84 : 0.8)
+      const lines = wrap(g, c.text, maxW)
+      const y0 = portrait ? H * 0.74 : H * 0.88
+      lines.forEach((ln, i) => {
+        const y = y0 + (i - (lines.length - 1)) * size * 1.35
+        g.shadowColor = 'rgba(0,0,0,0.85)'
+        g.shadowBlur = size * 0.35
+        g.lineWidth = size * 0.08
+        g.strokeStyle = 'rgba(0,0,0,0.35)'
+        g.strokeText(ln, W / 2, y)
+        g.fillStyle = '#fbf3e4'
+        g.fillText(ln, W / 2, y)
+      })
+    }
+  }
+  g.restore()
+  void DURATION
+}
+
+/** Wrap Chinese text at punctuation where possible. */
+export function wrap(g: CanvasRenderingContext2D, text: string, maxW: number): string[] {
+  if (g.measureText(text).width <= maxW) return [text]
+  const chars = Array.from(text)
+  // try to split at the punctuation closest to the middle
+  let best = -1, bd = 1e9
+  chars.forEach((ch, i) => {
+    if ('，。、；：'.includes(ch) && i > 2 && i < chars.length - 2) {
+      const d = Math.abs(i - chars.length / 2)
+      if (d < bd) { bd = d; best = i }
+    }
+  })
+  const mid = best >= 0 ? best + 1 : Math.ceil(chars.length / 2)
+  const a = chars.slice(0, mid).join(''), b = chars.slice(mid).join('')
+  return [...wrap(g, a, maxW), ...wrap(g, b, maxW)]
+}

tools/jiangnan/analyze.py

@@ -0,0 +1,104 @@
+# Blind-QA for the film: renders stills + category-ID maps in headless Chromium,
+# then prints ASCII composition maps, luminance maps and colour statistics.
+import base64, io, os, sys, time
+from playwright.sync_api import sync_playwright
+from PIL import Image
+
+BASE = 'http://localhost:5173/preview/production/0f8da72d-f0f5-4c30-b791-36769aa8c157/'
+CATS = {
+    'S': 0x6080ff, 'M': 0x304060, 'T': 0x806040, 'F': 0xc0c040, 'W': 0x00c0ff, 'H': 0xffffff, 'R': 0x202020, 'D': 0x8b4513,
+    'K': 0x909090, 'Y': 0xffff00, 'V': 0x00a000, 'L': 0xff0000, 'P': 0xff00ff, 'C': 0x0000a0, 'O': 0x00ffa0, 'X': 0x404040,
+}
+CRGB = {k: ((v >> 16) & 255, (v >> 8) & 255, v & 255) for k, v in CATS.items()}
+
+
+def classify(px):
+    best, bd = '.', 1e9
+    for k, c in CRGB.items():
+        d = (px[0] - c[0]) ** 2 + (px[1] - c[1]) ** 2 + (px[2] - c[2]) ** 2
+        if d < bd:
+            bd, best = d, k
+    return best if bd < 3000 else '?'
+
+
+def ascii_ids(img, cols=64):
+    w, h = img.size
+    rows = round(cols * h / w * 0.5)
+    small = img.resize((cols * 3, rows * 3), Image.NEAREST).convert('RGB')
+    out = []
+    for r in range(rows):
+        line = ''
+        for c in range(cols):
+            cnt = {}
+            for dy in range(3):
+                for dx in range(3):
+                    k = classify(small.getpixel((c * 3 + dx, r * 3 + dy)))
+                    cnt[k] = cnt.get(k, 0) + 1
+            line += max(cnt, key=cnt.get)
+        out.append(line)
+    return '\n'.join(out)
+
+
+def ascii_lum(img, cols=64):
+    w, h = img.size
+    rows = round(cols * h / w * 0.5)
+    g = img.convert('L').resize((cols, rows), Image.BOX)
+    chars = ' .:-=+*#%@'
+    return '\n'.join(''.join(chars[min(9, g.getpixel((c, r)) * 10 // 256)] for c in range(cols)) for r in range(rows))
+
+
+def stats(img):
+    im = img.convert('RGB')
+    w, h = im.size
+    res = []
+    for name, y0, y1 in [('top', 0, h // 3), ('mid', h // 3, 2 * h // 3), ('bot', 2 * h // 3, h)]:
+        crop = im.crop((0, y0, w, y1)).resize((32, 16), Image.BOX)
+        px = list(crop.getdata())
+        m = [sum(p[i] for p in px) / len(px) for i in range(3)]
+        res.append(f"{name} rgb=({m[0]:.0f},{m[1]:.0f},{m[2]:.0f})")
+    L = list(im.convert('L').resize((240, 135)).getdata())
+    clip = sum(1 for v in L if v > 248) / len(L)
+    crush = sum(1 for v in L if v < 6) / len(L)
+    mean = sum(L) / len(L)
+    return ' | '.join(res) + f" | meanL={mean:.0f} clip={clip*100:.1f}% crush={crush*100:.1f}%"
+
+
+def main():
+    times = [float(x) for x in sys.argv[1].split(',')]
+    w = int(sys.argv[2]) if len(sys.argv) > 2 else 640
+    h = int(sys.argv[3]) if len(sys.argv) > 3 else 360
+    mode = sys.argv[4] if len(sys.argv) > 4 else 'both'
+    out = '/tmp/shots'
+    os.makedirs(out, exist_ok=True)
+    with sync_playwright() as p:
+        b = p.chromium.launch(args=['--use-angle=swiftshader', '--enable-unsafe-swiftshader', '--ignore-gpu-blocklist', '--disable-dev-shm-usage'])
+        pg = b.new_page(viewport={'width': w, 'height': h})
+        logs = []
+        pg.on('console', lambda m: logs.append(m.type + ': ' + m.text) if m.type in ('error', 'warning') else None)
+        pg.on('pageerror', lambda e: logs.append('PAGEERROR: ' + str(e)))
+        t0 = time.time()
+        pg.goto(f'{BASE}?w={w}&h={h}&t={times[0]}', timeout=180000)
+        pg.wait_for_function("document.title === 'ready' || document.title === 'error'", timeout=900000)
+        print('init', round(time.time() - t0, 1), 's', pg.title())
+        for t in times:
+            print(f'\n===== t={t} =====')
+            if mode in ('both', 'final'):
+                pg.evaluate(f"window.__film.renderFrame({t})")
+                data = pg.evaluate("document.querySelector('canvas').toDataURL('image/jpeg', 0.9)")
+                img = Image.open(io.BytesIO(base64.b64decode(data.split(',')[1])))
+                img.save(f'{out}/t{t:05.1f}.jpg')
+                print(stats(img))
+                print(ascii_lum(img))
+            if mode in ('both', 'ids'):
+                pg.evaluate(f"window.__ids({t})")
+                data = pg.evaluate("document.querySelector('canvas').toDataURL('image/png')")
+                img = Image.open(io.BytesIO(base64.b64decode(data.split(',')[1])))
+                img.save(f'{out}/ids{t:05.1f}.png')
+                print(ascii_ids(img))
+        for l in [x for x in logs if 'useProgram' not in x][-20:]:
+            print(l)
+        b.close()
+
+
+if __name__ == '__main__':
+    main()

tools/jiangnan/subset_fonts.py

@@ -0,0 +1,12 @@
+# Subset the OFL Chinese fonts to the characters actually used by the film/UI.
+import glob, re, subprocess, sys
+text = ''
+for f in glob.glob('/workspaces/frontend/src/jiangnan/**/*.ts*', recursive=True):
+    text += open(f, encoding='utf-8').read()
+chars = sorted(set(c for c in text if ord(c) > 0x2E80))
+extra = '0123456789:：.,，。、！？!?·—…（）()《》「」“”‘’ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz-+/%#&@ '
+open('/opt/work/chars.txt', 'w', encoding='utf-8').write(''.join(chars) + extra)
+print(len(chars), 'CJK chars')
+for src, out in [('/tmp/fonts/MaShanZheng-Regular.ttf', 'title.woff2'), ('/tmp/fonts/LXGWWenKai-Regular.ttf', 'wenkai.woff2')]:
+    subprocess.run([sys.executable, '-m', 'fontTools.subset', src, '--text-file=/opt/work/chars.txt', '--flavor=woff2',
+                    f'--output-file=/workspaces/frontend/src/jiangnan/assets/fonts/{out}', '--layout-features=*'], check=True)

tools/jiangnan/tts.py

@@ -0,0 +1,17 @@
+# Generate narration lines with a neural Chinese voice (edge-tts) and print durations.
+import asyncio, re, subprocess, sys, json
+import edge_tts, imageio_ffmpeg
+LINES = json.loads(open(sys.argv[1], encoding='utf-8').read())
+OUT = sys.argv[2]
+VOICE = sys.argv[3] if len(sys.argv) > 3 else 'zh-CN-YunyangNeural'
+FF = imageio_ffmpeg.get_ffmpeg_exe()
+async def main():
+    res = {}
+    for lid, text in LINES.items():
+        c = edge_tts.Communicate(text, VOICE, rate='-12%', pitch='-3Hz')
+        await c.save(f'{OUT}/{lid}.mp3')
+        info = subprocess.run([FF, '-i', f'{OUT}/{lid}.mp3'], capture_output=True, text=True).stderr
+        m = re.search(r'Duration: (\d+):(\d+):([\d.]+)', info)
+        res[lid] = round(int(m.group(2)) * 60 + float(m.group(3)), 2)
+    print(json.dumps(res))
+asyncio.run(main())

tools/jiangnan/verify_live_photo.py

@@ -0,0 +1,15 @@
+import sys, json, hashlib
+sys.path.insert(0, '/tmp/lc/scripts')
+from pathlib import Path
+from pair_live_photo import verify_pair, digest
+from portable_live_photo.apple import inspect_live_jpeg, inspect_live_mov
+d = Path(sys.argv[1]); cover = float(sys.argv[2]); fps = float(sys.argv[3])
+ident = inspect_live_jpeg(d / 'live.jpg')
+info = inspect_live_mov(d / 'live.mov')
+print('jpeg id', ident); print('mov', info)
+manifest = {'identifier': ident, 'photo': 'live.jpg', 'video': 'live.mov', 'coverTime': cover, 'fps': fps,
+            'backend': 'portable-python', 'metadataVerified': True, 'localPHLivePhotoLoad': 'not_available',
+            'sha256': {'photo': digest(d / 'live.jpg'), 'video': digest(d / 'live.mov')}}
+r = verify_pair(d, manifest)
+print('verify_pair OK, still time', r.still_time)
+(d / 'manifest.json').write_text(json.dumps(manifest, indent=2))