feat: 链路控制、延迟校准与长测工具，补全文档

- 固件：新增 link_control / stream_control，覆盖运行中 Wi-Fi 断线恢复、取帧节拍、
  温控与质量/帧率减载；UDP 发送增加有界重试预算与帧期限，PSRAM JPEG 直发减少拷贝。
- 协议：重组器严格校验分片布局，统计未收齐、超时与完全缺失的帧，处理帧号回绕与设备重启。
- 桌面端：视频改 latest-value 通道并前置帧时间标记，拒绝积压和慢解码的过期画面；新增
  延迟标记页与 --latency-report 校准，完善录制、服务端与退出流程。
- 文档与脚本：新增硬件、开发、延迟、功耗、实时性、稳定性文档和改动记录，重写根 README；
  新增构建、性能、延迟与串口长测脚本，以及长测客户端脚本。
- 构建：固定 esp-1.97.0.0 工具链，分区表随构建复制，更新 sdkconfig 与忽略规则。

scripts/perf_probe.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+"""Measure real UDP video throughput, complete-frame loss and receive jitter.
+
+Uses an ephemeral port by default so it can run beside the desktop client.
+The timings are receive-side metrics, not camera-to-display latency.
+"""
+import argparse
+from collections import Counter
+import json
+from pathlib import Path
+import socket
+import secrets
+import struct
+import time
+
+from udp_probe import decode
+
+PORT = 5600
+FRAGMENT = 1400
+
+
+def percentile(values, p):
+    if not values:
+        return None
+    values = sorted(values)
+    return round(values[int((len(values) - 1) * p)], 3)
+
+
+def summary(source, duration):
+    counts = source['counts']
+    frames = counts['complete_frames']
+    dropped = counts['superseded_frames'] + counts['missing_frames'] + counts['invalid_jpeg'] + counts['timed_out_frames']
+    span = source['last_complete'] - source['first_complete'] if frames > 1 else 0
+    return dict(
+        address=source['address'], counts=dict(counts),
+        window_fps=round(frames / duration, 3) if duration else 0,
+        active_fps=round((frames - 1) / span, 3) if span else 0,
+        frame_loss_percent=round(100 * dropped / max(1, frames + dropped), 3),
+        resolution=source.get('resolution'),
+        mean_jpeg_bytes=round(counts['jpeg_bytes'] / frames) if frames else 0,
+        inter_frame_p50_ms=percentile(source['intervals'], .5),
+        inter_frame_p95_ms=percentile(source['intervals'], .95),
+        inter_frame_max_ms=round(max(source['intervals']), 3) if source['intervals'] else None,
+        assembly_p95_ms=percentile(source['assembly'], .95),
+        pending_frames=int(source['pending'] is not None),
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('host', help='Board IPv4 address; avoids VPN limited-broadcast routing')
+    parser.add_argument('--seconds', type=float, default=30)
+    parser.add_argument('--local-port', type=int, default=0)
+    parser.add_argument('--local-ip', default='')
+    parser.add_argument('--receive-buffer', type=int, default=1024 * 1024, help='0 keeps the OS default')
+    parser.add_argument('--json', type=Path)
+    parser.add_argument('--save-frame', type=Path)
+    args = parser.parse_args()
+    client_id = secrets.randbelow(0xFFFFFFFF) + 1
+    if client_id == 0x0C110001:
+        client_id += 1
+    subscribe = struct.pack('<IBHIIHHIHHB', 0x4D564331, 2, 1, client_id, 0, 0, 0, 0, 0, 0, 2)
+    sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+    sock.setsockopt(socket.SOL_SOCKET, socket.SO_BROADCAST, 1)
+    if args.receive_buffer:
+        sock.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, args.receive_buffer)
+    sock.bind((args.local_ip, args.local_port))
+    local_port = sock.getsockname()[1]
+    buffer_bytes = sock.getsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF)
+    sock.settimeout(.05)
+    sources = {}
+    invalid = 0
+    saved = False
+    started = time.perf_counter()
+    renew = started
+    report_at = started + 5
+    previous_frames = 0
+    previous_report = started
+    try:
+        while time.perf_counter() - started < args.seconds:
+            now = time.perf_counter()
+            for source in sources.values():
+                if source['pending'] is not None and now-source['pending']['started'] > .5:
+                    source['counts']['timed_out_frames'] += 1
+                    source['pending'] = None
+            if now >= renew:
+                sock.sendto(subscribe, (args.host, PORT))
+                renew = now + .8
+            if now >= report_at:
+                frames = sum(s['counts']['complete_frames'] for s in sources.values())
+                print(f'{now-started:5.1f}s  {(frames-previous_frames)/(now-previous_report):5.1f} fps  complete={frames}', flush=True)
+                previous_frames, previous_report, report_at = frames, now, now + 5
+            try:
+                data, address = sock.recvfrom(2048)
+            except socket.timeout:
+                continue
+            now = time.perf_counter()
+            parsed = decode(data)
+            if parsed is None:
+                invalid += 1
+                continue
+            kind, h, payload = parsed
+            if kind != 1:
+                continue
+            size = h['total_len']
+            index, count = h['frag_index'], h['frag_count']
+            offset = index * FRAGMENT
+            if h['version'] != 1 or size > 4 * 1024 * 1024 or count != max(1, (size + FRAGMENT - 1) // FRAGMENT) or index >= count or offset > size or len(payload) != min(FRAGMENT, size-offset):
+                invalid += 1
+                continue
+            sender, frame = h['sender_id'], h['frame_id']
+            if sender not in sources:
+                sources[sender] = dict(address=address[0], counts=Counter(), latest=None, last_seen=now, pending=None, intervals=[], assembly=[], first_complete=None, last_complete=None)
+            s = sources[sender]
+            counts = s['counts']
+            counts['video_datagrams'] += 1
+            latest = s['latest']
+            if latest is not None:
+                distance = (frame - latest) & 0xFFFFFFFF
+                if distance >= 0x80000000 and now-s['last_seen'] >= 1:
+                    counts['restarts'] += 1
+                    latest = s['latest'] = None
+                    s['pending'] = None
+                elif distance >= 0x80000000 or (distance == 0 and s['pending'] is None):
+                    counts['late_or_repeated_datagrams'] += 1
+                    continue
+                elif distance > 0:
+                    counts['missing_frames'] += distance - 1
+                    counts['superseded_frames'] += int(s['pending'] is not None)
+                    s['pending'] = None
+            s['latest'], s['last_seen'] = frame, now
+            if s['pending'] is None:
+                s['pending'] = dict(buffer=bytearray(size), seen=set(), count=count, size=size, width=h['width'], height=h['height'], started=now)
+            pending = s['pending']
+            if (size, count, h['width'], h['height']) != (pending['size'], pending['count'], pending['width'], pending['height']):
+                counts['inconsistent_metadata'] += 1
+                continue
+            if index in pending['seen']:
+                counts['duplicate_fragments'] += 1
+                continue
+            pending['seen'].add(index)
+            pending['buffer'][offset:offset+len(payload)] = payload
+            if len(pending['seen']) != count:
+                continue
+            jpeg = bytes(pending['buffer'])
+            if jpeg[:2] == b'\xff\xd8' and jpeg[-2:] == b'\xff\xd9':
+                completed = time.perf_counter()
+                counts['complete_frames'] += 1
+                counts['jpeg_bytes'] += size
+                if s['last_complete'] is not None:
+                    s['intervals'].append((completed-s['last_complete'])*1000)
+                s['last_complete'] = completed
+                if s['first_complete'] is None:
+                    s['first_complete'] = completed
+                s['assembly'].append((completed-pending['started'])*1000)
+                s['resolution'] = [h['width'], h['height']]
+                if args.save_frame and not saved:
+                    args.save_frame.write_bytes(jpeg)
+                    saved = True
+            else:
+                counts['invalid_jpeg'] += 1
+            s['pending'] = None
+    finally:
+        sock.close()
+    elapsed = time.perf_counter()-started
+    result = dict(seconds=round(elapsed, 3), host=args.host, local_port=local_port, receive_buffer_bytes=buffer_bytes, invalid_datagrams=invalid, sources={f'0x{sender:08X}': summary(s, elapsed) for sender, s in sources.items()}, timing_scope='receiver inter-frame spacing and first-to-last fragment assembly; not camera-to-display latency')
+    if args.json:
+        args.json.write_text(json.dumps(result, ensure_ascii=False, indent=2)+'\n', encoding='utf-8')
+    print(json.dumps(result, ensure_ascii=False, indent=2))
+    return 0 if any(s['counts']['complete_frames'] for s in sources.values()) else 1
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())