perf(agent): 60fps上限/零拷贝转换/多线程编码/发送缓冲512KB + encbench基准

- capture: min_frame 33→16ms，RC 60fps；with_mapped 映射窗口内直接 BGRA→YUV（行距连续零拷贝）
- capture: num_threads(4) + 每120帧输出 convert/encode 耗时（实测：convert 2ms，encode 41ms/帧真实工况）
- agent: send_window 2MB→512KB，E2E 排队天花板 ~0.15s（修复 E2E 633ms 红）
- encbench: 1080p 编码基准（合成 5ms/帧，真实工况 41ms/帧 → 软编 1080p 上限 ~24fps，60fps 需 NVENC）
- tools/bench.md

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);
+}