merge: 修 HUD 停刷与帧率不稳（单槽 latest-wins + GOP 自愈）

- viewer: 无界 mpsc → Mutex 单槽 latest-wins。根因：NVENC 60fps 到达率 > 渲染
  吞吐(~55fps)，无界队列无限积压 → 延迟线性增长 + 渲染 drain 阻塞消息泵 →
  窗口假死、HUD 停刷（软编 20fps 时代掩盖，NVENC 后暴露）
- viewer: 解码帧尺寸与窗口解耦（blit 逐行拷贝，防 NVENC 对齐尺寸不匹配）；
  首帧打印解码尺寸诊断
- agent: NVENC gop_len 0xffffffff→240（60fps 下 4s 周期 IDR 自愈恢复点，
  配合 KeyFrameReq 即时恢复，消除长时间冻结锯齿）
- 回环实测：NVENC 4ms/帧满速、解码 1920x1080 匹配、14s 零错误

B侧设计文档.md

@@ -0,0 +1,117 @@
+# B 侧设计文档（主控端 + 协议 + 网络）
+
+> 负责人：成员 B ｜ 主责 crate：`viewer/`、`common/` ｜ 配合：A（agent/）
+> 按里程碑推进，每个里程碑独立 feature 分支，合入即打 tag。
+
+---
+
+## B1 · 联调基线（1-2 天，和 A 合作完成）
+
+目标：拿到 H.264 版双机实测数据，作为一切优化的对照基线。
+
+1. **HUD 扩展**：新增 丢包率(估算)/抖动(jitter)/帧间隔 p95 三项读数
+   - 抖动 = 相邻帧 E2E 差值的绝对值 EMA；帧间隔 p95 用环形缓冲取 95 分位；
+   - 码流异常计数（解码失败次数）也上 HUD，验证关键帧自愈是否生效。
+2. **测试脚本** `tools/bench.md`：固定三个场景各测 60s —— 静止桌面 / 播放视频 / 拖动窗口，记录 HUD 截图。
+3. **产出**：基线数据表填入本文件末尾。
+
+## B2 · 码率自适应（3-4 天）
+
+目标：弱网自动降码率保流畅，好网缓慢升码率提画质。
+
+### 协议扩展（common/）
+
+```rust
+/// B2 新增控制消息（流 2）
+RateReport {
+    loss_permille: u16,   // 接收端估算丢包率 ‰
+    rtt_ms: u32,          // 最近 RTT
+    e2e_ms: u32,          // 平滑后端到端延迟
+}
+```
+
+viewer 每 2s 汇总一次发送（流 2 复用，无新增握手）。
+
+### 控制算法（先简后精，AIMD 风格）
+
+| 网络状态（连续 3 个周期） | 动作 |
+|---|---|
+| 丢包 > 8% 或 E2E > 400ms | 码率 ×0.6（下限 1Mbps），同时通知编码器降帧率到 20fps |
+| 丢包 3-8% | 码率 ×0.9 |
+| 丢包 < 1% 且 E2E < 150ms 持续 10s | 码率 +500Kbps 步进上探（上限 15Mbps） |
+
+### agent 侧配合点（给 A 的接口需求）
+
+- 需要运行时改码率的入口：优先查 openh264 crate 是否暴露 `SetOption(RC_PARAM)`；
+- 若 crate 不支持 → 方案 B：保留 encoder 实例但按新码率重建（重建成本 ~50ms，降码率不频繁，可接受）；
+- 降级到极限（1Mbps 仍丢包 >15%）→ 主动降分辨率 1080p→720p（A 实现，协议传 `ResChange` 通知，viewer 重配解码器与窗口尺寸）。
+
+## B3 · P2P 打洞 + 信令 + 中继（1.5-2 周，核心里程碑）
+
+目标：跨公网可用。三层结构：
+
+```
+┌─────────┐  ①注册/交换地址(WebSocket)  ┌──────────┐
+│  viewer  │ ───────────────────────────→ │ 信令服务器 │
+└────┬────┘                              └──────────┘
+     │ ②拿到对方候选地址(IP:port×N)            ↑
+     │ ③a 直连(打洞) ────────────────────  agent 注册
+     │ ③b 失败 → 走 TURN 中继
+```
+
+### 信令服务器（B 主建，A 参与评审）
+
+- 最小实现：单二进制 Rust + WebSocket（tokio-tungstenite），无数据库；
+- 消息：`Register{agent_id}` / `Connect{agent_id}` / `Answer{candidates, auth_code_hint}`；
+- agent 开机注册自身 ID（设备码），viewer 输对方 ID 发起；
+- 中立服务器，不碰媒体数据，带宽成本 ≈ 0。
+
+### 候选地址与打洞
+
+1. agent 侧收集候选：本机局域网地址、路由器公网映射（STUN 探测，自建 coturn 的 stun 或公共 stun）；
+2. 双方经信令交换候选后**同时**向对方发探测包（UDP 打洞，锥形 NAT 成功率高；对称 NAT 基本打不通）；
+3. 打通 → 直连（QUIC 直接到该地址）；打不通 → TURN 中继兜底。
+
+### 中继方案（务实路线）
+
+QUIC 本身没有标准 TURN 集成，两条路：
+- **路线一（推荐先做）**：自建中继服务 `relay` —— viewer 把 QUIC 包封装转发给 relay，relay 转给 agent（本质是应用层 TURN，复用信令服务器的部署）；
+- **路线二（二期）**：coturn + 手写 ICE-lite，仅在路线一性能不足时投入。
+
+### 安全（随 B3 一起做，替代现在的跳过校验）
+
+- 预共享验证码升级为 **PAKE（SPAKE2）**：验证码不落服务器、不落网络，双方推导会话密钥；
+- 信令只传地址候选，媒体端到端加密保持 QUIC TLS 不变。
+
+## B4 · 多屏 + 文件传输（各 2-3 天，穿插做）
+
+**多屏**：
+- 协议：`ScreenList{screens: Vec<{id, w, h}>}`（AuthOk 里带）、`SelectScreen{id}`；
+- agent 枚举 DXGI outputs（A 实现），viewer 加屏幕切换下拉；
+- 第二期再做"多屏同显"。
+
+**文件传输**：
+- 首选走已有 QUIC 连接开独立双向流（不引入 aria2 依赖，部署简单）；
+- 协议：`FileOffer{name, size}` → `FileAccept` → 原始分块流（1MB 块 + 校验）；
+- viewer 做拖拽发送 + 保存对话框。
+
+---
+
+## 里程碑与分支规划
+
+| 里程碑 | 分支 | 预估 | 验收标准 |
+|---|---|---|---|
+| B1 联调基线 | feature/hud-baseline | 1-2 天 | 三场景基线数据入库 |
+| B2 码率自适应 | feature/rate-adapt | 3-4 天 | 模拟弱网（clumsy/丢包工具）下自动降码率，画面不断流 |
+| B3 打洞+中继 | feature/p2p-relay | 1.5-2 周 | 两台跨公网机器通过 ID 互联；打洞失败自动走中继 |
+| B4 多屏/文件 | feature/screens-files | 3-5 天 | 多屏切换、拖拽传文件可用 |
+
+顺序建议：B1 → B2 → B3 → B4（B3 是价值最大项，B4 可穿插）。
+
+## 基线数据表（B1 完成后填写）
+
+| 场景 | FPS | RTT | E2E | 码率 | 丢包 | 抖动 |
+|---|---|---|---|---|---|---|
+| 静止桌面 | | | | | | |
+| 视频播放 | | | | | | |
+| 拖动窗口 | | | | | | |

Cargo.lock

@@ -9,6 +9,7 @@ dependencies = [
  "anyhow",
  "common",
  "lz4_flex",
+ "nvenc",
  "openh264",
  "quinn",
  "rcgen",
@@ -44,6 +45,27 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "bitfields"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef6e59298da389bc0649c7463856b34c6e17fe542f88939426ede4436c6b1195"
+dependencies = [
+ "bitfields-impl",
+]
+
+[[package]]
+name = "bitfields-impl"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2c044f98f86f15414668d6c8187c7e4fadab1ad2b31680f648703e0fe07c555"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+ "thiserror",
+]
+
 [[package]]
 name = "bitflags"
 version = "1.3.2"
@@ -567,6 +589,17 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
 
+[[package]]
+name = "nvenc"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3227c3b09b1e042dab7d8d82cea77c0404e7d0c2bbb92e7e87cb319749cd20dc"
+dependencies = [
+ "bitfields",
+ "libloading",
+ "windows",
+]
+
 [[package]]
 name = "once_cell"
 version = "1.21.4"
@@ -1512,32 +1545,54 @@ checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
 
 [[package]]
 name = "windows"
-version = "0.58.0"
+version = "0.62.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dd04d41d93c4992d421894c18c8b43496aa748dd4c081bac0dc93eb0489272b6"
+checksum = "527fadee13e0c05939a6a05d5bd6eec6cd2e3dbd648b9f8e447c6518133d8580"
+dependencies = [
+ "windows-collections",
+ "windows-core",
+ "windows-future",
+ "windows-numerics",
+]
+
+[[package]]
+name = "windows-collections"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "23b2d95af1a8a14a3c7367e1ed4fc9c20e0a26e79551b1454d72583c97cc6610"
 dependencies = [
  "windows-core",
- "windows-targets",
 ]
 
 [[package]]
 name = "windows-core"
-version = "0.58.0"
+version = "0.62.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6ba6d44ec8c2591c134257ce647b7ea6b20335bf6379a27dac5f1641fcf59f99"
+checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
 dependencies = [
  "windows-implement",
  "windows-interface",
+ "windows-link",
  "windows-result",
  "windows-strings",
- "windows-targets",
+]
+
+[[package]]
+name = "windows-future"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1d6f90251fe18a279739e78025bd6ddc52a7e22f921070ccdc67dde84c605cb"
+dependencies = [
+ "windows-core",
+ "windows-link",
+ "windows-threading",
 ]
 
 [[package]]
 name = "windows-implement"
-version = "0.58.0"
+version = "0.60.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2bbd5b46c938e506ecbce286b6628a02171d56153ba733b6c741fc627ec9579b"
+checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -1546,9 +1601,9 @@ dependencies = [
 
 [[package]]
 name = "windows-interface"
-version = "0.58.0"
+version = "0.59.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "053c4c462dc91d3b1504c6fe5a726dd15e216ba718e84a0e46a88fbe5ded3515"
+checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -1562,22 +1617,31 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
 
 [[package]]
-name = "windows-result"
+name = "windows-numerics"
-version = "0.2.0"
+version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1d1043d8214f791817bab27572aaa8af63732e11bf84aa21a45a78d6c317ae0e"
+checksum = "6e2e40844ac143cdb44aead537bbf727de9b044e107a0f1220392177d15b0f26"
 dependencies = [
- "windows-targets",
+ "windows-core",
+ "windows-link",
+]
+
+[[package]]
+name = "windows-result"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
+dependencies = [
+ "windows-link",
 ]
 
 [[package]]
 name = "windows-strings"
-version = "0.1.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4cd9b125c486025df0eabcb585e62173c6c9eddcec5d117d3b6e8c30e2ee4d10"
+checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
 dependencies = [
- "windows-result",
+ "windows-link",
- "windows-targets",
 ]
 
 [[package]]
@@ -1614,6 +1678,15 @@ dependencies = [
  "windows_x86_64_msvc",
 ]
 
+[[package]]
+name = "windows-threading"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3949bd5b99cafdf1c7ca86b43ca564028dfe27d66958f2470940f73d86d75b37"
+dependencies = [
+ "windows-link",
+]
+
 [[package]]
 name = "windows_aarch64_gnullvm"
 version = "0.52.6"

Cargo.toml

@@ -11,5 +11,5 @@ quinn = "0.11"
 rustls = { version = "0.23", default-features = false, features = ["ring", "std", "logging"] }
 rcgen = { version = "0.13", default-features = false, features = ["ring", "pem"] }
 lz4_flex = "0.11"
-windows = "0.58"
+windows = "0.62"
 minifb = "0.27"

agent/Cargo.toml

@@ -12,6 +12,7 @@ rustls.workspace = true
 rcgen.workspace = true
 lz4_flex.workspace = true
 openh264 = { version = "0.9.9", features = ["source"] }
+nvenc = "0.1.0"
 windows = { workspace = true, features = [
     "Win32_Foundation",
     "Win32_Graphics_Direct3D",

agent/src/bin/encbench.rs

@@ -0,0 +1,59 @@
+//! 编码基准：1080p 噪声帧（最恶劣压缩工况）对比单线程与 4 线程 openh264。
+//! 运行：cargo run --release -p agent --bin encbench
+
+use openh264::encoder::{
+    BitRate, Encoder, EncoderConfig, FrameRate, IntraFramePeriod, RateControlMode, UsageType,
+};
+use openh264::formats::{BgraSliceU8, YUVBuffer};
+use openh264::OpenH264API;
+use std::time::Instant;
+
+fn bench(threads: u16, frames: usize) -> f64 {
+    let mut enc = Encoder::with_api_config(
+        OpenH264API::from_source(),
+        EncoderConfig::new()
+            .bitrate(BitRate::from_bps(5_000_000))
+            .max_frame_rate(FrameRate::from_hz(30.0))
+            .rate_control_mode(RateControlMode::Bitrate)
+            .usage_type(UsageType::ScreenContentRealTime)
+            .intra_frame_period(IntraFramePeriod::from_num_frames(300))
+            .num_threads(threads),
+    )
+    .expect("encoder");
+
+    // 预生成 10 张不同噪声帧轮换（保证每帧 P 帧都真实编码，避免运动估计空转）
+    let mut frames_in = Vec::with_capacity(10);
+    for f in 0..10u32 {
+        let mut frame = vec![0u8; 1920 * 1080 * 4];
+        let mut seed = 12345u32.wrapping_add(f * 7919);
+        for px in frame.chunks_exact_mut(4) {
+            seed = seed.wrapping_mul(1103515245).wrapping_add(12345);
+            px.copy_from_slice(&seed.to_le_bytes());
+        }
+        frames_in.push(YUVBuffer::from_bgra8_source(BgraSliceU8::new(&frame, (1920, 1080))));
+    }
+
+    // 预热 5 帧（首帧 IDR、线程池初始化）
+    for (i, y) in frames_in.iter().enumerate().take(5) {
+        let _ = enc.encode(y).unwrap_or_else(|_| panic!("frame {i}"));
+    }
+    let t = Instant::now();
+    for (i, y) in frames_in.iter().cycle().take(frames).enumerate() {
+        let _ = enc.encode(y).unwrap_or_else(|_| panic!("frame {i}"));
+    }
+    t.elapsed().as_millis() as f64 / frames as f64
+}
+
+fn main() {
+    println!("1080p 噪声帧编码基准（含 BGRA→YUV 转换）");
+    let t = Instant::now();
+    let _ = YUVBuffer::from_bgra8_source(BgraSliceU8::new(
+        &vec![7u8; 1920 * 1080 * 4],
+        (1920, 1080),
+    ));
+    println!("BGRA→YUV 转换: {}ms/帧", t.elapsed().as_millis());
+    let s = bench(1, 20);
+    println!("单线程: {s:.1}ms/帧 → {}fps", (1000.0 / s) as u32);
+    let m = bench(4, 20);
+    println!("4 线程: {m:.1}ms/帧 → {}fps", (1000.0 / m) as u32);
+}

agent/src/capture.rs

@@ -1,9 +1,17 @@
 //! DXGI Desktop Duplication 采集（Windows 8+）。
-//! 帧处理：BGRA → openh264 H.264 编码（屏幕内容实时档，5Mbps）→ 线格式。
+//! 编码器自动选择：NVENC 硬编（N 卡：GPU 内 BGRA→H.264 零转换零拷贝直通）
+//! → openh264 软编兜底（无 N 卡/驱动异常时自动启用）。
 //! 支持主控端 KeyFrameReq 强制关键帧（解码异常恢复）。
 
 use anyhow::{anyhow, Result};
 use common::{encode_video_frame, VideoFrameHeader};
+use nvenc::bitstream::BitStream;
+use nvenc::encoder::RegisteredResource;
+use nvenc::session::{InitParams, Session};
+use nvenc::sys::enums::{
+    NVencBufferFormat, NVencPicStruct, NVencPicType, NVencParamsRcMode, NVencTuningInfo,
+};
+use nvenc::sys::guids::{NV_ENC_CODEC_H264_GUID, NV_ENC_PRESET_P3_GUID};
 use openh264::encoder::{
     BitRate, Encoder, EncoderConfig, FrameRate, IntraFramePeriod, RateControlMode, UsageType,
 };
@@ -19,8 +27,9 @@ use windows::Win32::Graphics::Direct3D::{
 };
 use windows::Win32::Graphics::Direct3D11::{
     D3D11CreateDevice, ID3D11Device, ID3D11DeviceContext, ID3D11Resource, ID3D11Texture2D,
-    D3D11_CREATE_DEVICE_FLAG, D3D11_CPU_ACCESS_READ, D3D11_MAP, D3D11_MAP_READ,
+    D3D11_BIND_RENDER_TARGET, D3D11_BIND_SHADER_RESOURCE, D3D11_CREATE_DEVICE_FLAG,
-    D3D11_MAPPED_SUBRESOURCE, D3D11_SDK_VERSION, D3D11_TEXTURE2D_DESC, D3D11_USAGE_STAGING,
+    D3D11_CPU_ACCESS_READ, D3D11_MAP_READ, D3D11_MAPPED_SUBRESOURCE,
+    D3D11_SDK_VERSION, D3D11_TEXTURE2D_DESC, D3D11_USAGE_DEFAULT, D3D11_USAGE_STAGING,
 };
 use windows::Win32::Graphics::Dxgi::Common::{DXGI_FORMAT_B8G8R8A8_UNORM, DXGI_SAMPLE_DESC};
 use windows::Win32::Graphics::Dxgi::{
@@ -31,9 +40,12 @@ use windows::Win32::Graphics::Dxgi::{
 use std::time::{Duration, Instant};
 
 pub struct Duplication {
-    _device: ID3D11Device,
+    pub device: ID3D11Device,
     context: ID3D11DeviceContext,
     staging_res: ID3D11Resource,
+    /// GPU 内中间纹理（NVENC 注册目标；DEFAULT usage，BGRA 与采集一致）
+    gpu_res: ID3D11Resource,
+    gpu_tex: ID3D11Texture2D,
     dup: IDXGIOutputDuplication,
     pub width: u32,
     pub height: u32,
@@ -47,13 +59,13 @@ impl Duplication {
             D3D11CreateDevice(
                 None,
                 D3D_DRIVER_TYPE_HARDWARE,
-                HMODULE::default(),
+                HMODULE(std::ptr::null_mut()),
                 D3D11_CREATE_DEVICE_FLAG(0),
                 Some(&[D3D_FEATURE_LEVEL_11_0]),
                 D3D11_SDK_VERSION,
-                Some(&mut device),
+                Some(&raw mut device),
                 None,
-                Some(&mut context),
+                Some(&raw mut context),
             )
             .map_err(|e| anyhow!("D3D11CreateDevice failed: {e}"))?;
             let device = device.ok_or_else(|| anyhow!("no d3d device"))?;
@@ -73,6 +85,7 @@ impl Duplication {
             let (width, height) = (desc.ModeDesc.Width, desc.ModeDesc.Height);
             anyhow::ensure!(width > 0 && height > 0, "bad desktop size");
 
+            // staging：软编路径 Map 读内存
             let sdesc = D3D11_TEXTURE2D_DESC {
                 Width: width,
                 Height: height,
@@ -95,10 +108,26 @@ impl Duplication {
             let staging = staging_opt.ok_or_else(|| anyhow!("no staging texture"))?;
             let staging_res: ID3D11Resource = staging.cast()?;
 
+            // GPU 中间纹理：NVENC 路径的注册目标（DEFAULT usage，GPU 可读）
+            let gdesc = D3D11_TEXTURE2D_DESC {
+                Usage: D3D11_USAGE_DEFAULT,
+                BindFlags: (D3D11_BIND_RENDER_TARGET.0 | D3D11_BIND_SHADER_RESOURCE.0) as u32,
+                CPUAccessFlags: 0,
+                ..sdesc
+            };
+            let mut gpu_opt: Option<ID3D11Texture2D> = None;
+            device
+                .CreateTexture2D(&gdesc, None, Some(&mut gpu_opt))
+                .map_err(|e| anyhow!("CreateTexture2D(gpu) failed: {e}"))?;
+            let gpu_tex = gpu_opt.ok_or_else(|| anyhow!("no gpu texture"))?;
+            let gpu_res: ID3D11Resource = gpu_tex.cast()?;
+
             Ok(Self {
-                _device: device,
+                device,
                 context,
                 staging_res,
+                gpu_res,
+                gpu_tex,
                 dup,
                 width,
                 height,
@@ -106,7 +135,8 @@ impl Duplication {
         }
     }
 
-    /// true=拿到新帧并已复制到 staging；false=超时无更新
+    /// true=拿到新帧并已复制；false=超时无更新。
+    /// 同时复制到 staging（软编读内存）与 GPU 中间纹理（NVENC 注册输入）。
     pub fn acquire(&mut self, timeout_ms: u32) -> Result<bool> {
         unsafe {
             let mut info = DXGI_OUTDUPL_FRAME_INFO::default();
@@ -116,6 +146,7 @@ impl Duplication {
                     let res = resource.ok_or_else(|| anyhow!("no desktop resource"))?;
                     let src: ID3D11Resource = res.cast()?;
                     self.context.CopyResource(&self.staging_res, &src);
+                    self.context.CopyResource(&self.gpu_res, &src);
                     let _ = self.dup.ReleaseFrame();
                     Ok(true)
                 }
@@ -125,8 +156,8 @@ impl Duplication {
         }
     }
 
-    /// 把 staging 纹理映射到内存并去掉行间距
+    /// 在 staging 映射窗口内执行回调（软编路径）。回调参数 = (像素数据, 行距)。
-    pub fn copy_frame(&mut self) -> Result<Vec<u8>> {
+    pub fn with_mapped<R>(&mut self, f: impl FnOnce(&[u8], usize) -> R) -> Result<R> {
         unsafe {
             let mut mapped = D3D11_MAPPED_SUBRESOURCE::default();
             self.context
@@ -135,16 +166,107 @@ impl Duplication {
             let (w, h) = (self.width as usize, self.height as usize);
             let pitch = mapped.RowPitch as usize;
             let src = std::slice::from_raw_parts(mapped.pData as *const u8, pitch * h);
-            let mut out = Vec::with_capacity(w * h * 4);
+            let r = f(src, pitch);
-            for row in 0..h {
-                out.extend_from_slice(&src[row * pitch..row * pitch + w * 4]);
-            }
             self.context.Unmap(&self.staging_res, 0);
-            Ok(out)
+            Ok(r)
         }
     }
 }
 
+/// NVENC 硬编封装：H.264 High、CBR 码率控制、UltraLowLatency 调优、无 B 帧。
+/// 输入 = 注册的 GPU 纹理（BGRA 与 DXGI 采集字节序一致，零转换）。
+struct NvencEncoder {
+    encoder: nvenc::encoder::Encoder,
+    registered: RegisteredResource,
+    bitstream: BitStream,
+    frame_count: usize,
+}
+
+impl NvencEncoder {
+    fn new(dup: &mut Duplication, bitrate_bps: u32) -> Result<Self> {
+        // 打开会话（无 N 卡 / 驱动不含编码库时在此失败）
+        let session: Session<_> = Session::open_dx(&dup.device)
+            .map_err(|e| anyhow!("open_dx: {e:?}"))?;
+        let codecs = session
+            .get_encode_codecs()
+            .map_err(|e| anyhow!("codecs: {e:?}"))?;
+        anyhow::ensure!(codecs.contains(&NV_ENC_CODEC_H264_GUID), "NVENC no H264 codec");
+        let (session, mut config) = session
+            .get_encode_preset_config_ex(
+                NV_ENC_CODEC_H264_GUID,
+                NV_ENC_PRESET_P3_GUID,
+                NVencTuningInfo::UltraLowLatency,
+            )
+            .map_err(|e| anyhow!("preset_config: {e:?}"))?;
+        // CBR 桌面实时档：码率恒定，延迟可预测（对标 UU 3.2Mbps@60fps）
+        config.preset_cfg.rc_params.rate_control_mode = NVencParamsRcMode::CBR;
+        config.preset_cfg.rc_params.average_bit_rate = bitrate_bps;
+        config.preset_cfg.gop_len = 240; // 周期 IDR（60fps 下 4s）：网络异常后的自愈恢复点，
+                                         // 配合 KeyFrameReq 即时恢复，避免长时间冻结
+        config.preset_cfg.frame_interval_p = 1; // 无 B 帧
+
+        let init = InitParams {
+            encode_guid: NV_ENC_CODEC_H264_GUID,
+            preset_guid: NV_ENC_PRESET_P3_GUID,
+            resolution: [dup.width, dup.height],
+            aspect_ratio: [dup.width, dup.height],
+            frame_rate: [60, 1],
+            tuning_info: NVencTuningInfo::UltraLowLatency,
+            buffer_format: NVencBufferFormat::ARGB,
+            encode_config: &mut config.preset_cfg,
+            enable_ptd: true,
+            max_encoder_resolution: [dup.width, dup.height],
+        };
+        let encoder = session
+            .init_encoder(init)
+            .map_err(|e| anyhow!("init_encoder: {e:?}"))?;
+        let registered = encoder
+            .register_resource_dx11(&dup.gpu_tex, NVencBufferFormat::ARGB, 0)
+            .map_err(|e| anyhow!("register_resource: {e:?}"))?;
+        let bitstream = encoder
+            .create_bitstream_buffer()
+            .map_err(|e| anyhow!("bitstream: {e:?}"))?;
+        Ok(Self {
+            encoder,
+            registered,
+            bitstream,
+            frame_count: 0,
+        })
+    }
+
+    /// 编码当前 GPU 纹理内容，返回 Annex B H.264 码流。
+    fn encode(&mut self, force_idr: bool, ts_ms: u64) -> Result<Vec<u8>> {
+        let ty = if force_idr {
+            NVencPicType::IDR
+        } else {
+            NVencPicType::P
+        };
+        self.encoder.encode_picture(
+            &self.registered,
+            &self.bitstream,
+            self.frame_count,
+            ts_ms,
+            NVencBufferFormat::ARGB,
+            NVencPicStruct::Frame,
+            ty,
+            None,
+        )
+        .map_err(|e| anyhow!("encode_picture: {e:?}"))?;
+        self.frame_count += 1;
+        // 同步读出码流（NVENC 单帧编码 1-3ms，阻塞等待可接受）
+        let lock = self
+            .bitstream
+            .try_lock(true)
+            .map_err(|e| anyhow!("bitstream lock: {e:?}"))?;
+        let out = lock.as_slice().to_vec();
+        drop(lock);
+        if out.is_empty() {
+            anyhow::bail!("nvenc empty bitstream");
+        }
+        Ok(out)
+    }
+}
+
 #[allow(dead_code)]
 pub fn desktop_size() -> Result<(u32, u32)> {
     let d = Duplication::new()?;
@@ -169,78 +291,163 @@ pub fn run(tx: Sender<Option<Vec<u8>>>, keyreq: Arc<AtomicBool>) {
             dup.width, dup.height
         );
 
-        // H.264 编码器：屏幕内容实时档 + 码率控制（对标 UU：3.2Mbps@60fps）
+        // ---- 编码器自动选择：NVENC 硬编优先，失败落回 openh264 软编 ----
-        let mut encoder = match Encoder::with_api_config(
+        enum EncoderKind {
-            OpenH264API::from_source(),
+            Nvenc(Box<NvencEncoder>),
-            EncoderConfig::new()
+            Soft(Box<openh264::encoder::Encoder>),
-                .bitrate(BitRate::from_bps(5_000_000))
+        }
-                .max_frame_rate(FrameRate::from_hz(30.0))
+        let mut encoder: EncoderKind = match NvencEncoder::new(&mut dup, 5_000_000) {
-                .rate_control_mode(RateControlMode::Bitrate)
+            Ok(nv) => {
-                .usage_type(UsageType::ScreenContentRealTime)
+                eprintln!("[capture] encoder = NVENC hardware (CBR 5Mbps@60fps, ARGB direct)");
-                .intra_frame_period(IntraFramePeriod::from_num_frames(300)),
+                EncoderKind::Nvenc(Box::new(nv))
-        ) {
+            }
-            Ok(e) => e,
             Err(e) => {
-                eprintln!("[capture] encoder init failed: {e}");
+                eprintln!("[capture] NVENC unavailable ({e}); fallback to openh264 software");
-                return;
+                match Encoder::with_api_config(
+                    OpenH264API::from_source(),
+                    EncoderConfig::new()
+                        .bitrate(BitRate::from_bps(5_000_000))
+                        .max_frame_rate(FrameRate::from_hz(60.0))
+                        .rate_control_mode(RateControlMode::Bitrate)
+                        .usage_type(UsageType::ScreenContentRealTime)
+                        .intra_frame_period(IntraFramePeriod::from_num_frames(300))
+                        .num_threads(4),
+                ) {
+                    Ok(e) => {
+                        eprintln!("[capture] encoder = openh264 software (5Mbps@60fps, 4T)");
+                        EncoderKind::Soft(Box::new(e))
+                    }
+                    Err(e) => {
+                        eprintln!("[capture] encoder init failed: {e}");
+                        return;
+                    }
+                }
             }
         };
 
-        let mut last_frame: Option<Vec<u8>> = None;
+        let mut last_frame: Option<Vec<u8>> = None; // 软编静止重发
         let mut last_sent = Instant::now() - Duration::from_secs(10);
-        let min_frame = Duration::from_millis(33); // 30fps 上限
+        let min_frame = Duration::from_millis(16); // 60fps 上限
+        let mut scratch: Vec<u8> = Vec::new(); // 软编：行距不连续时的复用缓冲
+        // 编码耗时统计（每 120 帧输出一次均值，ms）
+        let mut stat_frames = 0u64;
+        let mut stat_convert_ms = 0u64;
+        let mut stat_encode_ms = 0u64;
+
+        macro_rules! emit {
+            ($payload:expr) => {{
+                let header = VideoFrameHeader {
+                    width: dup.width,
+                    height: dup.height,
+                    ts_ms: start.elapsed().as_millis() as u64,
+                };
+                let frame = encode_video_frame(&header, &$payload);
+                if tx.send(Some(frame.clone())).is_err() {
+                    return; // 会话结束
+                }
+                last_frame = Some(frame);
+                last_sent = Instant::now();
+                stat_frames += 1;
+            }};
+        }
+
         loop {
             let iter_start = Instant::now();
             match dup.acquire(80) {
-                Ok(true) => match dup.copy_frame() {
+                Ok(true) => {
-                    Ok(raw) => {
+                    // 主控端请求关键帧（解码异常恢复）
-                        // 主控端请求关键帧（解码异常恢复）
+                    let force_key = keyreq.swap(false, Ordering::Relaxed);
-                        if keyreq.swap(false, Ordering::Relaxed) {
+                    match &mut encoder {
-                            encoder.force_intra_frame();
+                        EncoderKind::Nvenc(enc) => {
+                            let enc_start = Instant::now();
+                            let r = enc.encode(force_key, start.elapsed().as_millis() as u64);
+                            stat_encode_ms += enc_start.elapsed().as_millis() as u64;
+                            match r {
+                                Ok(payload) => emit!(payload),
+                                Err(e) => eprintln!("[capture] nvenc encode failed: {e}"),
+                            }
                         }
-                        let header = VideoFrameHeader {
+                        EncoderKind::Soft(enc) => {
-                            width: dup.width,
+                            if force_key {
-                            height: dup.height,
+                                enc.force_intra_frame();
-                            ts_ms: start.elapsed().as_millis() as u64,
+                            }
-                        };
+                            // 映射窗口内直接做 BGRA→YUV（行距连续时零拷贝，否则复用 scratch）
-                        let yuv = YUVBuffer::from_bgra8_source(BgraSliceU8::new(
+                            let conv_start = Instant::now();
-                            &raw,
+                            let dims = (dup.width as usize, dup.height as usize);
-                            (dup.width as usize, dup.height as usize),
+                            let conv = dup.with_mapped(|src, pitch| {
-                        ));
+                                if pitch == dims.0 * 4 {
-                        match encoder.encode(&yuv) {
+                                    YUVBuffer::from_bgra8_source(BgraSliceU8::new(src, dims))
-                            Ok(ebs) => {
+                                } else {
-                                let payload = ebs.to_vec();
+                                    scratch.resize(dims.0 * dims.1 * 4, 0);
-                                let frame = encode_video_frame(&header, &payload);
+                                    for (row, dst) in
-                                if tx.send(Some(frame.clone())).is_err() {
+                                        scratch.chunks_exact_mut(dims.0 * 4).enumerate()
-                                    return; // 会话结束
+                                    {
+                                        dst.copy_from_slice(
+                                            &src[row * pitch..row * pitch + dims.0 * 4],
+                                        );
+                                    }
+                                    YUVBuffer::from_bgra8_source(BgraSliceU8::new(&scratch, dims))
+                                }
+                            });
+                            match conv {
+                                Ok(yuv) => {
+                                    stat_convert_ms +=
+                                        conv_start.elapsed().as_millis() as u64;
+                                    let enc_start = Instant::now();
+                                    let r = enc.encode(&yuv);
+                                    stat_encode_ms += enc_start.elapsed().as_millis() as u64;
+                                    match r {
+                                        Ok(ebs) => emit!(ebs.to_vec()),
+                                        Err(e) => {
+                                            eprintln!("[capture] encode failed: {e}");
+                                        }
+                                    }
+                                }
+                                Err(e) => {
+                                    eprintln!("[capture] frame convert failed: {e}");
+                                    break;
                                 }
-                                last_frame = Some(frame);
-                                last_sent = Instant::now();
-                            }
-                            Err(e) => {
-                                eprintln!("[capture] encode failed: {e}");
                             }
                         }
-                        // 帧间隔节流到 ~30fps
-                        let el = iter_start.elapsed();
-                        if el < min_frame {
-                            std::thread::sleep(min_frame - el);
-                        }
                     }
-                    Err(e) => {
+                    if stat_frames > 0 && stat_frames % 120 == 0 {
-                        eprintln!("[capture] frame copy failed: {e}");
+                        eprintln!(
-                        break; // 重建 duplication
+                            "[capture] avg convert {}ms, encode {}ms/frame (last 120)",
+                            stat_convert_ms / 120,
+                            stat_encode_ms / 120
+                        );
+                        stat_convert_ms = 0;
+                        stat_encode_ms = 0;
                     }
-                },
+                    // 帧间隔节流到 ~60fps
+                    let el = iter_start.elapsed();
+                    if el < min_frame {
+                        std::thread::sleep(min_frame - el);
+                    }
+                }
                 Ok(false) => {
-                    // 桌面静止：低频重发最后一帧（重复 P 帧，解码端画面不变）
+                    // 桌面静止：低频重发（丢包后恢复 + 会话保活）。
+                    // NVENC 重编当前 GPU 帧（输出极小 P 帧）；软编重发最后一帧副本。
                     if last_sent.elapsed() > Duration::from_millis(500) {
-                        if let Some(f) = last_frame.as_ref() {
+                        match &mut encoder {
-                            if tx.send(Some(f.clone())).is_err() {
+                            EncoderKind::Nvenc(enc) => {
-                                return;
+                                // 重编当前 GPU 帧（内容未变，输出极小 P 帧）
+                                let r = enc.encode(false, start.elapsed().as_millis() as u64);
+                                match r {
+                                    Ok(payload) => emit!(payload),
+                                    Err(e) => {
+                                        eprintln!("[capture] nvenc idle encode failed: {e}")
+                                    }
+                                }
+                            }
+                            EncoderKind::Soft(_) => {
+                                // 软编：直接重发最后一帧码流副本
+                                if let Some(f) = last_frame.as_ref() {
+                                    let _ = tx.send(Some(f.clone()));
+                                }
                             }
-                            last_sent = Instant::now();
                         }
+                        last_sent = Instant::now();
                     } else {
                         std::thread::sleep(Duration::from_millis(8));
                     }

agent/src/main.rs

@@ -43,12 +43,12 @@ async fn main() -> Result<()> {
 
     // 传输参数调优：
     // - 收发窗口放大，避免高 RTT 下吞吐被 窗口/RTT 死锁（默认 512KB 流窗口）
-    // - 发送缓冲上限 2MB：QUIC 内部排队有界（约 1 帧多），配合上游 latest-wins 丢帧
+    // - 发送缓冲上限 512KB：QUIC 内部排队 ≈ 3 帧 @5Mbps，E2E 延迟天花板 ~150ms；
-    //   把端到端排队延迟压在 ~0.4s 内
+    //   配合上游 latest-wins 丢帧，超量部分在进 QUIC 前就被丢弃
     let mut transport = quinn::TransportConfig::default();
     transport.receive_window(quinn::VarInt::from_u32(16 * 1024 * 1024));
     transport.stream_receive_window(quinn::VarInt::from_u32(8 * 1024 * 1024));
-    transport.send_window(2 * 1024 * 1024);
+    transport.send_window(512 * 1024);
     let mut server_config = quinn::ServerConfig::with_single_cert(vec![cert], key.into())?;
     server_config.transport_config(Arc::new(transport));
 
@@ -93,8 +93,10 @@ async fn main() -> Result<()> {
     let (ctl_tx, mut ctl_rx) = mpsc::unbounded_channel::<ControlMsg>();
     let t0 = std::time::Instant::now();
 
-    // 读循环（流 2）：输入事件 + Ping
+    // 读循环（流 2）：输入事件 + Ping；Pong 附带视频下行方向的 QUIC 丢包统计
+    // （conn.stats().path 统计的是本端发送方向，agent 的发送方向即视频下行）
     let keyreq_reader = keyreq.clone();
+    let conn_stats = conn.clone();
     let reader = tokio::spawn(async move {
         loop {
             match read_control(&mut irecv).await {
@@ -104,7 +106,13 @@ async fn main() -> Result<()> {
                 }
                 Ok(ControlMsg::Ping(t)) => {
                     let agent_ts = t0.elapsed().as_millis() as u64;
-                    let _ = ctl_tx.send(ControlMsg::Pong { t, agent_ts });
+                    let p = conn_stats.stats().path;
+                    let _ = ctl_tx.send(ControlMsg::Pong {
+                        t,
+                        agent_ts,
+                        lost: p.lost_packets,
+                        sent: p.sent_packets,
+                    });
                 }
                 Ok(_) => {}
                 Err(_) => break,

common/src/lib.rs

@@ -15,7 +15,8 @@ pub enum ControlMsg {
     Ping(u64),
     /// t=主控端发送时刻（主控时钟），agent_ts=被控端收到时刻（被控时钟）
     /// 两者配合可估计时钟偏移，进而用帧头 ts_ms 计算端到端延迟
-    Pong { t: u64, agent_ts: u64 },
+    /// lost/sent = agent 发送方向（视频下行）的 QUIC 丢包/发包累计计数（conn.stats().path）
+    Pong { t: u64, agent_ts: u64, lost: u64, sent: u64 },
     Input(InputEvent),
 }
 

tools/bench.md

@@ -0,0 +1,37 @@
+# 双机测试基线脚本（B1）
+
+每个场景测试 60 秒，记录 HUD 读数填入《B侧设计文档.md》末尾的基线数据表。
+
+## 准备
+
+- 两端都用最新 release：`agent.exe` + `viewer.exe`（同目录带 3 个 mingw DLL）
+- 被控机：`agent.exe --port 47800 --code 123456`
+- 主控机：`viewer.exe --addr <被控机IP> --port 47800 --code 123456`
+- 测试前确认：被控机亮屏有桌面；同一网络（记下是有线还是 WiFi 频段）
+
+## 场景
+
+| # | 场景 | 被控机操作 | 观察点 |
+|---|---|---|---|
+| 1 | 静止桌面 | 打开一个空白文档，不动 60s | FPS 应低（静止检测），E2E/JIT 应绿 |
+| 2 | 视频播放 | 全屏播放任意 1080p 视频 | 码率峰值、FPS、丢包（LOSS 行） |
+| 3 | 拖动窗口 | 持续快速拖动一个大窗口 60s | 最恶劣工况：E2E/P95/JIT 是否变红、DECFAIL 是否增长 |
+
+## 记录方法
+
+每个场景进行中截一张 HUD 图（Win+Shift+S 框选右上角），或肉眼逐行抄录：
+
+```
+FPS=  RTT=  E2E=  JIT=  P95=  LOSS=  DECFAIL=  RATE=
+```
+
+## 附加记录
+
+- 网络环境：有线 / WiFi 5GHz / WiFi 2.4GHz，信号强度
+- 被控机 CPU 占用（任务管理器，agent.exe 一行）
+- 编码器类型：当前 openh264 软编（A 侧换 NVENC 后需重测一轮对比）
+
+## 注意
+
+- LOSS 显示 `--` 表示第一个 Pong 还没回来或 Δsent=0（静止场景发送少，属正常）
+- DECFAIL 在连接建立初期可能有 1-2 次（等第一个关键帧），后续不应增长；持续增长 = 关键帧自愈有问题，报给 A

viewer/src/main.rs

@@ -81,8 +81,15 @@ fn now_ms() -> u64 {
     T0.get().unwrap().elapsed().as_millis() as u64
 }
 
+/// 帧到达间隔环形缓冲（取 p95 用）
+struct GapRing {
+    last: Option<Instant>,
+    ring: [u32; 64],
+    idx: usize,
+    filled: usize,
+}
+
 /// 会话统计（HUD 数据源，全部原子量，跨任务共享）
-#[derive(Default)]
 struct Stats {
     /// 最近 RTT ms（0=未知）
     rtt: AtomicU64,
@@ -95,6 +102,80 @@ struct Stats {
     /// 累计收帧数 / 字节数
     frames: AtomicU64,
     bytes: AtomicU64,
+    /// 抖动：相邻帧 |ΔE2E| 的 EMA（ms×16 定点存储避免浮点原子）
+    jitter_x16: AtomicU64,
+    prev_e2e: AtomicI64,
+    /// agent 发送侧（视频下行）QUIC 丢包/发包累计（最近一次 Pong）
+    agent_lost: AtomicU64,
+    agent_sent: AtomicU64,
+    /// H.264 解码失败计数（关键帧自愈验证指标）
+    dec_fail: AtomicU64,
+    /// 帧到达间隔环形缓冲
+    gaps: std::sync::Mutex<GapRing>,
+}
+
+impl Default for Stats {
+    fn default() -> Self {
+        Self {
+            rtt: AtomicU64::new(0),
+            offset: AtomicI64::new(0),
+            offset_valid: AtomicBool::new(false),
+            e2e: AtomicU64::new(0),
+            e2e_valid: AtomicBool::new(false),
+            frames: AtomicU64::new(0),
+            bytes: AtomicU64::new(0),
+            jitter_x16: AtomicU64::new(0),
+            prev_e2e: AtomicI64::new(-1),
+            agent_lost: AtomicU64::new(0),
+            agent_sent: AtomicU64::new(0),
+            dec_fail: AtomicU64::new(0),
+            gaps: std::sync::Mutex::new(GapRing {
+                last: None,
+                ring: [0; 64],
+                idx: 0,
+                filled: 0,
+            }),
+        }
+    }
+}
+
+impl Stats {
+    /// 记录一次帧到达：更新间隔环形缓冲与抖动 EMA
+    fn on_frame_arrival(&self, est: Option<i64>) {
+        let now = Instant::now();
+        if let Ok(mut g) = self.gaps.lock() {
+            if let Some(t) = g.last {
+                let gap = (now - t).as_millis() as u32;
+                let i = g.idx;
+                g.ring[i] = gap;
+                g.idx = (i + 1) % g.ring.len();
+                g.filled = (g.filled + 1).min(g.ring.len());
+            }
+            g.last = Some(now);
+        }
+        // 抖动 = |本帧E2E - 上帧E2E| 的 EMA（RFC3550 风格，÷16 衰减）
+        if let Some(e) = est {
+            let prev = self.prev_e2e.load(Ordering::Relaxed);
+            self.prev_e2e.store(e, Ordering::Relaxed);
+            if prev >= 0 {
+                let d = (e - prev).unsigned_abs() as u64;
+                let old = self.jitter_x16.load(Ordering::Relaxed);
+                let sm = if old == 0 { d * 16 } else { old + d * 16 / 16 - old / 16 };
+                self.jitter_x16.store(sm, Ordering::Relaxed);
+            }
+        }
+    }
+
+    /// 帧间隔 p95（ms），样本不足返回 None
+    fn gap_p95(&self) -> Option<u32> {
+        let g = self.gaps.lock().ok()?;
+        if g.filled < 8 {
+            return None;
+        }
+        let mut v: Vec<u32> = g.ring[..g.filled].to_vec();
+        v.sort_unstable();
+        v.get(v.len() * 95 / 100).copied()
+    }
 }
 
 fn main() -> Result<()> {
@@ -160,48 +241,70 @@ fn session(
     let (w, h) = (width as usize, height as usize);
     let stats = Arc::new(Stats::default());
 
-    // 读任务（流 1）：H.264 解码 → RGB → mpsc；同时统计端到端延迟与流量
+    // 读任务（流 1）：H.264 解码 → RGB → 单槽 latest-wins；同时统计端到端延迟与流量
-    let (frame_tx, frame_rx) = std::sync::mpsc::channel::<Vec<u8>>();
+    // 帧格式 (src_w, src_h, rgb)；(0,0,_) = 解码失败占位，触发 KeyFrameReq
+    // 单槽而非无界队列：渲染跟不上 60fps 到达率时旧帧直接被覆盖，绝不积压
+    //（无界 mpsc 在 NVENC 60fps 下会无限积压 → 延迟线性增长 + 渲染 drain 时
+    //  长时间阻塞窗口消息泵 → 窗口假死、HUD 停刷新）
+    let latest: Arc<std::sync::Mutex<Option<(u32, u32, Vec<u8>)>>> =
+        Arc::new(std::sync::Mutex::new(None));
     {
         let stats = stats.clone();
+        let latest = latest.clone();
         let mut decoder = openh264::decoder::Decoder::new()
             .map_err(|e| anyhow::anyhow!("decoder init failed: {e}"))?;
         tokio::spawn(async move {
+            let mut size_logged = false;
             loop {
                 match common::read_video_frame(&mut recv).await {
                     Ok((hdr, payload)) => {
                         stats.bytes.fetch_add(payload.len() as u64, Ordering::Relaxed);
                         stats.frames.fetch_add(1, Ordering::Relaxed);
                         // 端到端延迟 = 接收时刻 - 时钟偏移 - 采集时刻（EMA 平滑）
+                        let mut est = None;
                         if stats.offset_valid.load(Ordering::Relaxed) {
-                            let est =
+                            let e =
                                 now_ms() as i64 - stats.offset.load(Ordering::Relaxed)
                                     - hdr.ts_ms as i64;
-                            if (0..5000).contains(&est) {
+                            if (0..5000).contains(&e) {
                                 let old = stats.e2e.load(Ordering::Relaxed) as i64;
                                 let sm = if stats.e2e_valid.load(Ordering::Relaxed) {
-                                    (old * 4 + est) / 5
+                                    (old * 4 + e) / 5
                                 } else {
-                                    est
+                                    e
                                 };
                                 stats.e2e.store(sm as u64, Ordering::Relaxed);
                                 stats.e2e_valid.store(true, Ordering::Relaxed);
+                                est = Some(sm);
                             }
                         }
+                        stats.on_frame_arrival(est);
                         match decoder.decode(&payload) {
                             Ok(Some(yuv)) => {
                                 let (dw, dh) = yuv.dimensions();
-                                let mut rgb = vec![0u8; dw * dh * 3];
+                                let (dw, dh) = (dw as u32, dh as u32);
+                                if !size_logged {
+                                    size_logged = true;
+                                    eprintln!(
+                                        "[viewer] decoded frame {dw}x{dh} (window {}x{})",
+                                        hdr.width, hdr.height
+                                    );
+                                }
+                                let mut rgb = vec![0u8; dw as usize * dh as usize * 3];
                                 yuv.write_rgb8(&mut rgb);
-                                if frame_tx.send(rgb).is_err() {
+                                // latest-wins：直接覆盖旧帧，无积压
-                                    break;
+                                if let Ok(mut slot) = latest.lock() {
+                                    *slot = Some((dw, dh, rgb));
                                 }
                             }
                             Ok(None) => {}
                             Err(e) => {
                                 // 解码异常：请求被控端强制关键帧
+                                stats.dec_fail.fetch_add(1, Ordering::Relaxed);
                                 eprintln!("[viewer] decode failed: {e}, request keyframe");
-                                let _ = frame_tx.send(Vec::new()); // 空帧占位唤醒渲染循环
+                                if let Ok(mut slot) = latest.lock() {
+                                    *slot = Some((0, 0, Vec::new()));
+                                }
                             }
                         }
                     }
@@ -220,9 +323,11 @@ fn session(
         tokio::spawn(async move {
             loop {
                 match read_control(&mut irecv).await {
-                    Ok(ControlMsg::Pong { t, agent_ts }) => {
+                    Ok(ControlMsg::Pong { t, agent_ts, lost, sent }) => {
                         let r = now_ms();
                         stats.rtt.store(r.saturating_sub(t), Ordering::Relaxed);
+                        stats.agent_lost.store(lost, Ordering::Relaxed);
+                        stats.agent_sent.store(sent, Ordering::Relaxed);
                         // agent_ts ≈ (t+r)/2 时刻的 agent 时钟读数 → 偏移 EMA
                         let est = agent_ts as i64 - ((t + r) / 2) as i64;
                         let old = stats.offset.load(Ordering::Relaxed);
@@ -278,6 +383,8 @@ fn session(
     let mut hud_bytes = 0u64;
     let mut hud_fps = 0u64;
     let mut hud_mbps = 0f64;
+    let mut hud_lost = 0u64;
+    let mut hud_sent = 0u64;
 
     loop {
         if !window.is_open() || window.is_key_down(Key::Escape) {
@@ -290,18 +397,15 @@ fn session(
             push(&msg_tx, ControlMsg::Ping(now_ms()));
         }
 
-        // 1) 渲染最新帧（只保留最新，丢弃积压）；空帧=解码异常，请求关键帧
+        // 1) 渲染最新帧（单槽 latest-wins：每迭代只取最新一帧，天然防积压）
-        loop {
+        if let Ok(mut slot) = latest.lock() {
-            match frame_rx.try_recv() {
+            if let Some((sw, sh, raw)) = slot.take() {
-                Ok(raw) => {
+                if raw.is_empty() {
-                    if raw.is_empty() {
+                    push(&msg_tx, ControlMsg::KeyFrameReq);
-                        push(&msg_tx, ControlMsg::KeyFrameReq);
+                } else if sw > 0 && sh > 0 {
-                    } else if raw.len() == w * h * 3 {
+                    blit(&mut buffer, w, h, &raw, sw as usize, sh as usize);
-                        blit(&mut buffer, &raw);
+                    dirty = true;
-                        dirty = true;
-                    }
                 }
-                Err(_) => break,
             }
         }
 
@@ -368,6 +472,7 @@ fn session(
         }
 
         // 6) HUD：每秒刷新一次指标，画在右上角
+        let (hud_loss, hud_jit, hud_p95, hud_dec);
         if hud_last.elapsed() >= Duration::from_secs(1) {
             let f = stats.frames.load(Ordering::Relaxed);
             let b = stats.bytes.load(Ordering::Relaxed);
@@ -375,12 +480,46 @@ fn session(
             hud_mbps = (b.saturating_sub(hud_bytes)) as f64 * 8.0 / 1e6;
             hud_frames = f;
             hud_bytes = b;
+            // 丢包率 = Δlost / Δsent（agent 发送方向，随 Pong 同步）
+            let l = stats.agent_lost.load(Ordering::Relaxed);
+            let s = stats.agent_sent.load(Ordering::Relaxed);
+            hud_loss = if s > hud_sent {
+                ((l - hud_lost) as f64 / (s - hud_sent) as f64 * 100.0).min(99.9)
+            } else {
+                -1.0
+            };
+            hud_lost = l;
+            hud_sent = s;
+            hud_jit = stats.jitter_x16.load(Ordering::Relaxed) / 16;
+            hud_p95 = stats.gap_p95();
+            hud_dec = stats.dec_fail.load(Ordering::Relaxed);
             hud_last = Instant::now();
+        } else {
+            hud_loss = f64::NAN;
+            hud_jit = 0;
+            hud_p95 = None;
+            hud_dec = 0;
         }
         if dirty {
             let rtt = stats.rtt.load(Ordering::Relaxed);
             let e2e = stats.e2e.load(Ordering::Relaxed);
             let e2e_ok = stats.e2e_valid.load(Ordering::Relaxed);
+            let loss_str = if hud_loss.is_nan() {
+                "--".to_string()
+            } else if hud_loss < 0.0 {
+                "--".to_string()
+            } else {
+                format!("{hud_loss:.1}%")
+            };
+            let loss_c = if hud_loss.is_nan() || hud_loss < 0.0 {
+                C_GRAY
+            } else {
+                lat_color(hud_loss < 1.0, hud_loss < 5.0)
+            };
+            let jit_c = lat_color(hud_jit < 10, hud_jit < 30);
+            let p95_c = hud_p95
+                .map(|p| lat_color(p < 60, p < 150))
+                .unwrap_or(C_GRAY);
             draw_hud(
                 &mut buffer,
                 w,
@@ -392,6 +531,10 @@ fn session(
                         &if e2e_ok { format!("{e2e}ms") } else { "--".into() },
                         lat_color(e2e < 100, e2e < 200),
                     ),
+                    ("JIT", &format!("{hud_jit}ms"), jit_c),
+                    ("P95", &hud_p95.map(|p| format!("{p}ms")).unwrap_or("--".into()), p95_c),
+                    ("LOSS", &loss_str, loss_c),
+                    ("DECFAIL", &format!("{hud_dec}"), if hud_dec == 0 { C_GRAY } else { C_RED }),
                     ("RATE", &format!("{:.0}MBP", hud_mbps), C_GRAY),
                     ("RES", &format!("{w}x{h}"), C_GRAY),
                 ],
@@ -473,6 +616,7 @@ fn glyph(c: u8) -> [u8; 5] {
         b':' => [0x00, 0x36, 0x36, 0x00, 0x00],
         b'.' => [0x00, 0x60, 0x60, 0x00, 0x00],
         b'-' => [0x08, 0x08, 0x08, 0x08, 0x08],
+        b'%' => [0x13, 0x18, 0x04, 0x0A, 0x11],
         b' ' => [0x00; 5],
         _ => [0x00; 5],
     }
@@ -546,9 +690,17 @@ fn draw_hud(buffer: &mut [u32], w: usize, lines: &[(&str, &str, u32)]) {
     }
 }
 
-/// RGB888 字节（openh264 解码输出）→ minifb 0RGB u32 缓冲
+/// RGB888 字节（openh264 解码输出）→ minifb 0RGB u32 缓冲。
-fn blit(buffer: &mut [u32], raw: &[u8]) {
+/// 支持源尺寸与窗口不同（解码器可能对齐到 16 倍数，如 1080→1088）：
-    for (px, chunk) in buffer.iter_mut().zip(raw.chunks_exact(3)) {
+/// 逐行拷贝重叠区域，源小于窗口时右侧/下方留黑。
-        *px = ((chunk[0] as u32) << 16) | ((chunk[1] as u32) << 8) | chunk[2] as u32;
+fn blit(buffer: &mut [u32], w: usize, h: usize, raw: &[u8], dw: usize, dh: usize) {
+    let cw = w.min(dw);
+    let ch = h.min(dh);
+    for row in 0..ch {
+        let src = &raw[row * dw * 3..row * dw * 3 + cw * 3];
+        let dst = &mut buffer[row * w..row * w + cw];
+        for (px, chunk) in dst.iter_mut().zip(src.chunks_exact(3)) {
+            *px = ((chunk[0] as u32) << 16) | ((chunk[1] as u32) << 8) | chunk[2] as u32;
+        }
     }
 }