diff --git a/fpga-imx6-flow/SKILL.md b/fpga-imx6-flow/SKILL.md
index 94053db..4dfb148 100644
--- a/fpga-imx6-flow/SKILL.md
+++ b/fpga-imx6-flow/SKILL.md
@@ -4,7 +4,7 @@ description: >
   Run the fpga-imx6-flow LangGraph workflow at E:\LangGraph_Design\FPGA-imx6-flow —
   the closed-loop SMOD (FPGA + i.MX6) hardware-in-the-loop design/debug pipeline:
   power on the board, connect JTAG, program bitstream/flash, restart the imx6 app,
-  analyze (ILA + COM6 serial), judge against design requirements, and iterate by
+  analyze (ILA + imx6 serial console), judge against design requirements, and iterate by
   modifying RTL/imx6 code. It drives the graph in-process (no langgraph dev) and
   hands the 5 interrupts (analyze / meets-requirement / propose / modify-FPGA /
   modify-imx6) back to this Claude session to execute.
@@ -43,8 +43,8 @@ USB-Blast 处于连接状态时给板卡上电或断电（保护 JTAG 适配器
 标准时序（图拓扑已按此实现，手动操作或写 `tools_*` / 真机脚本时也必须遵守）：
 
 ```
-断电(COM8 power_off) → 确认已断电 → 断开USB-Blast(COM4 blast_off) → 确认已断开
-  → 上电(COM8 power_on) → 确认已上电 → 连接USB-Blast(COM4 blast_on) → 连接JTAG/烧写/抓ILA
+断电(power_off) → 确认已断电 → 断开USB-Blast(blast_off) → 确认已断开
+  → 上电(power_on) → 确认已上电 → 连接USB-Blast(blast_on) → 连接JTAG/烧写/抓ILA
 ```
 
 **烧写 flash 的完整闭环**（固化后必须断电重启从 flash 加载再验证）：
@@ -59,7 +59,12 @@ USB-Blast 处于连接状态时给板卡上电或断电（保护 JTAG 适配器
   `power_off` 做整机断电重启，再次走完上述上电时序后才 `analyze`(抓ILA验证)。
 - **强化确认**：`power_on` 节点在上电前调用 `_ensure_blast_off_before_power()` 兜底确认
   USB-Blast 已断开；若发现仍连接会强制断开并在 `history` 标记 `rule_violation=True`。
-- vivado-mcp 工具：`power_on`/`power_off`(COM8 继电器)、`USB_Blast_on`/`USB_Blast_off`(COM4)。
+- 两个继电器的**串口号不在本工作流里绑定**：唯一来源是 `~/vivado_mcp_usb.conf`（逐机而异，
+  不入库）。`app/tools_hw.py` 通过 `vivado_mcp.serial_config.load_serial_ports()` 读取它，
+  串口收发/组帧/CRC 也复用 `vivado_mcp.serial_io`——协议只有一份实现。
+  文件不存在时调 vivado-mcp 的 **`detect_serial_relays`** 工具自动探测生成。
+  （对应的 MCP 工具是 `power_on`/`power_off`/`USB_Blast_on`/`USB_Blast_off`，和这里的
+  `tools_hw` 走同一套底层。）
 
 > **Run via the env python directly** (not bare `conda run`): on this Windows GBK
 > console, `conda run` re-prints the child's UTF-8 stdout through a GBK codec and
@@ -90,9 +95,17 @@ if (-not (Test-Path $py)) { Write-Host "[MISSING] langgraph env" } else {
   `langgraph` + `langgraph-checkpoint-sqlite` and the Vivado link). Do not proceed
   until it succeeds.
 
-**For a real-hardware run** (`--real`), additionally confirm: `env_config.json`'s
-`VIVADO_BIN_PATH` is valid; the vivado-mcp `power_on`/`USB_Blast_on` tools are
-available; COM4/COM6/COM8 are present (`[System.IO.Ports.SerialPort]::GetPortNames()`).
+**For a real-hardware run** (`--real`), additionally confirm the machine's two config
+files — **both live in `~`, neither is in the repo** (they hold COM numbers, VM/imx6
+passwords, and the Vivado path, all of which differ per machine):
+
+| 文件 | 装什么 | 缺了怎么办 |
+|---|---|---|
+| `~/vivado_mcp_usb.conf` | 电源继电器 + USB-Blast 继电器的 COM 号 | 调 vivado-mcp 的 **`detect_serial_relays`** 工具自动探测生成 |
+| `~/fpga_imx6_flow.json` | imx6 控制台 COM 号 (`ports.imx6`)、`VIVADO_BIN_PATH`、虚拟机/imx6 口令 | 照工程根的 `fpga_imx6_flow.example.json` 手工创建 |
+
+List the machine's ports with `[System.IO.Ports.SerialPort]::GetPortNames()`.
+工程根目录下**不该**再有 `env_config.json`——它已废弃，`app/config.py` 见到就报错。
 **Default runs are MOCK** (dry-run of the design loop with no hardware), because the
 workflow's `tools_*` real-hardware paths are still TODO — see "Mock vs real" below.
 
@@ -122,7 +135,7 @@ it, then resume the **same `--thread-id`** with a JSON payload. Repeat until exi
 
 | `action` | excalidraw node | What you (the session) do | Resume JSON |
 |---|---|---|---|
-| `analyze` | 通过ila/串口分析 | Grab COM6 serial log + pull ILA waveform (vivado-mcp); summarize | `{"analysis_log":"<摘要/路径>"}` |
+| `analyze` | 通过ila/串口分析 | Grab imx6 serial log + pull ILA waveform (vivado-mcp); summarize | `{"analysis_log":"<摘要/路径>"}` |
 | `meets_requirement` | 符合设计要求? | Judge the analysis against the design requirement | `{"meets": true}` or `{"meets": false}` |
 | `propose` | 分析原因，提出修改方案 | Diagnose root cause; decide what to change + next program mode | `{"reason":"...","need_fpga": true/false,"need_imx6": true/false,"program_mode":"bitstream"或"flash"}` |
 | `modify_fpga` | 改 RTL→综合→出 bit/flash | Edit RTL, run Vivado synth/impl, produce bit + flash files | `{"bit_path":"...","flash_path":"...","program_mode":"flash"}` |
@@ -160,17 +173,75 @@ To abandon a paused run and start clean, delete `<project>\runs\skill_checkpoint
 - **MOCK (default)** — `FPGA_IMX6_MOCK=1`; hardware/check nodes return success
   immediately. Use this to dry-run the *design-loop logic* and the 5 interrupts with
   no board attached. Great for demonstrating/validating the flow.
-- **Real (`--real`)** — drives actual hardware. Requires the project's
-  `app/tools_hw.py` / `tools_vivado.py` / `tools_imx6.py` real paths to be implemented
-  (they currently raise `NotImplementedError` as TODO — they will reuse the proven
-  vivado-mcp serial/Vivado logic). Until then, `--real` fails at the first hardware
-  node; tell the user to wire the tools first (see `Notes/Task00.md` §7).
+- **Real (`--real`)** — drives actual hardware. The `tools_hw` (power + USB-Blast relays,
+  ports from `~/vivado_mcp_usb.conf`), `tools_imx6` (console restart/kill/log, port from
+  `~/fpga_imx6_flow.json`'s `ports.imx6`), and `tools_vivado` (JTAG program/flash) real paths are
+  **now implemented and validated end-to-end (2026-07-05)** on the SMOD_ATSC board. Getting there
+  took **7 fixes** — see memory `fpga-imx6-flow-tx-safe`: power_on waits `sys init over` before
+  JTAG (else TX dies); `_PROJ`→SMOD_ATSC; `deploy_app`→no-op (deploy out-of-band, see above);
+  driver multi-interrupt resume (map by `action`); `power_off` blast-off-first + check retry guard
+  (else infinite loop); serial open-retry on transient err5 (now in `vivado_mcp.serial_io`);
+  `program_flash`→proven
+  `/fpga_flash` flasher + Mfg-ID verify (old tcl was a false-positive). Verify a flash landed with
+  JTAG `get_hw_ilas` (iter0 = **1 ILA / 0 VIO**; golden = **0 ILA / 2 VIO**).
+
+## imx6 access + file transfer (modify_imx6 / deploy step) — PROVEN 2026-07-05
+
+The `deploy_app` graph node is a no-op (deploy is done **out-of-band by the session**).
+Here is the working recipe for the `modify_imx6` interrupt (cross-compile HEAD's app and
+push it to the imx6). All of this is **network-based and fast** — no slow serial file push.
+
+**Hosts / creds**
+- **imx6 mgmt: `192.168.1.143`** (Windows host is on `192.168.1.125`, same subnet — `ping` works).
+  The old memory "imx6 not reachable over the network" is **outdated**; the mgmt port is reachable.
+  Login: **`root` / `anywavecom`**. (The serial console still works too but its login auto-logs-out and
+  is flaky; prefer network SSH for imx6 shell/log reads.)
+- **Cross-compile VM: `bowei@192.168.223.128`** (Ubuntu, password **`lr7402`** — ask the user, don't
+  guess/brute-force). Toolchain env-setup:
+  `/opt/fsl-imx-x11/4.1.15-2.0.0/environment-setup-cortexa7hf-neon-poky-linux-gnueabi`.
+  If VM SSH gives `kex_exchange_identification: Connection closed`, the VM's sshd is down/hung —
+  ask the user to `sudo systemctl restart sshd`.
+
+**SSH library quirk (critical)**
+- The imx6's SSH server is **old (legacy ciphers + `ssh-rsa` host key)**. paramiko **5.0 fails**
+  (`unknown cipher` / `no acceptable host key`). Install **paramiko 2.12** to a temp dir and use it
+  for the imx6 (it has TripleDES/aes-cbc/ssh-rsa):
+  `python -m pip install --target <dir> paramiko==2.12.0` then `PYTHONPATH=<dir> python ... -W ignore`.
+  The **VM** SSH is modern, so the env's paramiko 5.0 works for the VM (or use 2.12 for both).
+- The imx6 SSH has **no SFTP subsystem** (`open_sftp()` → `EOF during negotiation`). Transfer files
+  via **`cat > file` over an exec channel**:
+  ```python
+  ch = imx6_client.get_transport().open_session()
+  ch.exec_command("cat > /aw/anywave/app/imx_atsc_app_head")
+  with open(local_bin, "rb") as f: ch.sendall(f.read())
+  ch.shutdown_write(); ch.recv_exit_status()      # 0 = ok
+  ```
+  (The VM has SFTP; `open_sftp().get(...)` works there.)
+
+**Full modify_imx6 flow (validated)**
+1. Windows: `git archive --format=tar -o head_atsc.tar HEAD:imx6/src/atsc_app` (13 MB, self-contained).
+2. paramiko→VM (pw `lr7402`): SFTP the tar up; extract to a **FRESH dir** (`~/atsc_head_build`, never
+   touch `atsc_tsoip_build/` etc.); build:
+   `bash -lc 'source <envsetup>; cd ~/atsc_head_build && make'` → output `imx_atsc_app` (~2.92 MB ELF).
+3. paramiko→VM SFTP the built `imx_atsc_app` down to Windows.
+4. paramiko-2.12→imx6 (pw `anywavecom`): `cat >`-push it to `/aw/anywave/app/imx_atsc_app_head`,
+   `chmod 755`, `cp -a imx_atsc_app_head imx_atsc_app` (swap active; golden stays as
+   `imx_atsc_app.bak_pre_rx1`), md5-verify.
+5. Cold-boot (tools_hw blast_off→power_off→power_on, TX-safe) so rc.local runs the new app.
+6. Verify: network SSH grep `/var/module_log/imx_atsc_app.log` for `TALISE_radioOn` / `fb init state ok`
+   / `fpga[0xBB]=0x03` (sync=1/tvalid=1); spectrum for TX; (ILA needs an `.ltx` → skip unless re-impl).
+
+> The `git HEAD` app (b07d73b) is the known-good **RX1-on-framerA (M=4/lane0x03/RX1RX2) + correct TX**
+> build. On-device `imx_atsc_app.rx1` and any `M=2/lane0x04` variant **break TX** — do not use them.
 
 ## Troubleshooting
 
 | Symptom | Fix |
 |---|---|
 | `UnicodeEncodeError: 'gbk' ...` | Don't use bare `conda run`; use the env `python.exe` or `conda run --no-capture-output`. |
+| imx6 SSH `unknown cipher` / `no acceptable host key` (paramiko 5) | Use **paramiko 2.12** (temp `--target` dir) for the imx6 legacy SSH; see "imx6 access" above. |
+| imx6 `open_sftp` → `EOF during negotiation` | imx6 has no SFTP subsystem; push files via `cat > file` over an exec channel. |
+| VM SSH `kex_exchange_identification: Connection closed` | VM sshd down/hung — user `sudo systemctl restart sshd`. |
 | `import failed ... Run /langgraph-vivado-install` | The `langgraph` env is missing/incomplete → run `/langgraph-vivado-install`. |
 | Exit 1 `NotImplementedError ... tools_*` | A `--real` run hit an unwired hardware tool. Run without `--real` (mock), or implement `tools_*`. |
 | `resume` seems to start over | You passed a different `--thread-id` than `start`. Use the same one. |
diff --git a/fpga-imx6-flow/scripts/run_imx6_flow.py b/fpga-imx6-flow/scripts/run_imx6_flow.py
index ef3da1e..7d9ab29 100644
--- a/fpga-imx6-flow/scripts/run_imx6_flow.py
+++ b/fpga-imx6-flow/scripts/run_imx6_flow.py
@@ -95,13 +95,32 @@ def main() -> None:
                 graph.invoke(init, cfg)
             elif args.command == "resume":
                 payload = json.loads(args.resume_json) if args.resume_json else {}
-                graph.invoke(Command(resume=payload), cfg)
+                snap0 = graph.get_state(cfg)
+                pend = [intr for task in snap0.tasks for intr in task.interrupts]
+                if len(pend) <= 1:
+                    graph.invoke(Command(resume=payload), cfg)
+                else:
+                    # Multiple interrupts pending (parallel fan-out, e.g. modify_fpga +
+                    # modify_imx6). payload must be a map keyed by each interrupt's "action"
+                    # (or interrupt id); resume them all at once with an id->value map.
+                    resume_map = {}
+                    for intr in pend:
+                        val = intr.value if isinstance(intr.value, dict) else {}
+                        action = val.get("action")
+                        if action in payload:
+                            resume_map[intr.id] = payload[action]
+                        elif intr.id in payload:
+                            resume_map[intr.id] = payload[intr.id]
+                        else:
+                            resume_map[intr.id] = {}
+                    graph.invoke(Command(resume=resume_map), cfg)
             # status: no invoke, just snapshot
             snap = graph.get_state(cfg)
         except Exception as exc:  # noqa: BLE001
             _emit({"status": "error", "error": str(exc)}, 1)
 
         if snap.next:
+            all_intr = [intr.value for task in snap.tasks for intr in task.interrupts]
             _emit({
                 "status": "interrupt",
                 "thread_id": args.thread_id,
@@ -109,6 +128,8 @@ def main() -> None:
                 "next": list(snap.next),
                 "iteration": snap.values.get("iteration"),
                 "interrupt": _pending_interrupt(snap),
+                "pending_interrupts": all_intr,
+                "pending_count": len(all_intr),
             }, 2)
         else:
             _emit({
diff --git a/fpga_update/SKILL.md b/fpga_update/SKILL.md
new file mode 100644
index 0000000..15fbcd8
--- /dev/null
+++ b/fpga_update/SKILL.md
@@ -0,0 +1,119 @@
+---
+name: fpga_update
+description: 通过网络远程升级 SMOD 板（192.168.1.143）上 ATSC 激励器工程的 FPGA 镜像——先用 Vivado + perl 生成 quickboot 的 _update.mcs，再用 TCP 1234 的私有协议把它流式烧进 SPI flash 的 update 槽。只要用户提到"远程升级"、"升级 mcs"、"把新 bit 升上去"、"update mcs"、"远程更新 FPGA"、"Xfpga_ru"、"生成 quickboot mcs"、"quickboot"、"RU_golden 之后怎么升级"、"192.168.1.143 升级"，或者在 my_exciter 工程里跑完综合实现后想把结果固化到板子上，就必须用这个 skill——哪怕他们没说"skill"、没提 mcs，只说了"升上去"、"更新一下 FPGA"、"重新出个 mcs"。注意：这条路和网页上的 Remote Update 完全无关（那个只认 imx6 的 tar.gz 固件包）。JTAG 直烧打底镜像请改用 fpga_flash skill。
+---
+
+# SMOD FPGA 远程升级
+
+把 `my_exciter` 工程实现出的 `.bit` 变成 quickboot 更新镜像，再通过网络烧进板子的 SPI flash。
+
+板子必须**事先用 JTAG 烧过 `RU_golden`** 才支持远程升级 —— golden 提供上电时的兜底镜像和
+IPROG 跳转头。没烧过 golden 就先走 `/fpga_flash`。
+
+## 先搞清楚你在走哪条路
+
+这是最容易出错的地方，两条路互不相干：
+
+| | 这个 skill | 网页 Remote Update |
+|---|---|---|
+| 升什么 | **FPGA 镜像**（`.mcs`） | **imx6 固件包**（`.tar.gz`） |
+| 走哪 | TCP **1234**，私有帧协议 | HTTP `upload.cgi` → `check_package.sh` |
+
+把 `.mcs` 传到网页上不会报错，但 `check_package.sh` 找不到 `*.tar.gz` 就会
+`rm -rf /aw/upgrade/*` 把它删掉，什么都不会发生。原厂的 GUI 工具是 `tools/Xfpga_ru.exe`，
+没有命令行接口；`scripts/fpga_ru.py` 是它的等价替代，而且会告诉你失败在第几帧。
+
+## 流程
+
+### 1. 生成 `_update.mcs`
+
+```bash
+python scripts/gen_update_mcs.py                 # → bitstream/atsc_exciter_update.mcs
+python scripts/gen_update_mcs.py --tag smod_dpd  # → bitstream/smod_dpd_update.mcs
+```
+
+脚本按顺序跑项目里一直在用的那三个脚本（`post_bitstream.tcl` → `runScript_GenRu.tcl` →
+`MakeSpiFlashProgrammerMcsFiles_backup.pl -imagesize 64`），然后校验产物的起始地址是
+`0x800000`（这等价于"确实是用 `-imagesize 64` 生成的"）。
+
+它还替你挡了一个坑：`runScript_GenRu.tcl` 里是 `glob *.bit`，而 `bitstream/` 目录里躺着二十几个
+历史 `.bit`。脚本会把旧的临时挪走、跑完再放回去。**别绕过它去手动跑 tcl**，除非你已经清空了目录。
+
+`-imagesize 64` 不要改。它和 `SpiFlashProgrammer.vhd` 里的 `cAddrUpdateStart/End` 是一对，
+理由见 `references/quickboot.md`。
+
+### 2. 确认 ADPC 为 OFF（保险起见）
+
+DPD 的 sample&load 和升级数据流共用同一条 SPI1，而 FPGA 在升级期间会把 SPI1 上任何字都当成镜像数据。
+这条路径没被验证过（三次连续升级成功那次 `adpc=0`），所以稳妥起见先关掉：
+
+```bash
+# 登录后读 sys_cfg，第 3 个字段就是 adpc（0=OFF）
+curl -s -b "session=$SESS" "http://192.168.1.143/atsc_status_cfg.cgi?sys_cfg=&"
+```
+
+网页账号是 **`anywavecom` / `anywavecom`**（不是 `root`；`root/anywavecom` 是 SSH 的）。
+
+### 3. 烧进去
+
+```bash
+python scripts/fpga_ru.py --mcs bitstream/atsc_exciter_update.mcs
+python scripts/fpga_ru.py --mcs ... --dry-run     # 只校验文件，不碰硬件
+```
+
+先跑 `--dry-run`。它会检查行长、整行切分、EOF 记录、起始地址 —— 这些错误在真机上要等几分钟
+才会以一个没有上下文的 `ABORT_ERROR` 暴露出来。
+
+烧写要几分钟（8 MB，约 47600 帧）。**中途不要断电、不要重试。** 真断了也不会变砖：升级一开始就把
+切换字擦掉了，冷启动会退回 golden。
+
+成功后 app 会自动重启 imx6、复位 FPGA。然后**冷启动板子**（遵守 SMOD 上电时序）让 FPGA 从 flash
+重新加载新镜像。
+
+### 4. 验证
+
+抓 ILA、看 app 日志（`/var/module_log/imx_atsc_app.log`）、或者看频谱。
+闭环调试交给 `fpga-imx6-flow` skill。
+
+## 排错
+
+`fpga_ru.py` 会打印失败在哪一步、哪一帧、收到什么 ack。对着这张表看：
+
+- **擦除永不结束**（最容易误判的一个）→ 十有八九 programmer 已经报错退出了。
+  imx6 的 `POLL_ERASER` **不看 ERROR 位**，把"已失败"当成"还在擦"，所以客户端只会一直等。
+  别信进度条，直接读设备日志：
+
+  ```bash
+  ssh -o HostKeyAlgorithms=+ssh-rsa root@192.168.1.143 \
+    "grep 'ing|end' /var/module_log/imx_atsc_app.log | tail -3"
+  ```
+
+  `rstate=0x7`（ERROR=1, DONE=1, EraseOK=0）= programmer 中止；`rstate=0x0` 才是真的在擦，
+  擦完会看到 `rstate=0xc` + `UPSTATE_EARSE_END`。
+  历史上这里的元凶是 `smod_top.v` 没把 flash 的 SPI 引脚接出去，见 `references/protocol.md`。
+- **`ABORT_ERROR` 出现在第一帧** → mcs 格式不对，先跑 `--dry-run`。
+- **`ABORT_ERROR` 出现在数据流中后段** → 怀疑 SPI1 上有别人（DPD）在写，把 ADPC 关掉重试。
+- **没有应答 / 超时** → app 的 `sys_ready` 还是 0，或者 app 没在跑。
+  `ssh root@192.168.1.143 'ps | grep atsc'`（SSH 密码 `anywavecom`，dropbear 老得很，
+  命令行 ssh 要加 `-o HostKeyAlgorithms=+ssh-rsa`）。
+
+> 别用 `python ssh.py "tail -f ..."` 之类一次性读到 EOF 的封装去跟日志 —— 它永远不返回，
+> 什么也抓不到。升级完再 grep 日志文件即可，`log_printf` 已经落盘。
+
+## 不碰硬件地改协议客户端
+
+`scripts/mock_device.py` 是照着 `tcp_task.c` 的状态机实现的假设备，包括那三个硬限制和
+`':'` 重切分行为。改完 `fpga_ru.py` 先拿它跑一遍：
+
+```bash
+python scripts/mock_device.py --port 12340 &
+python scripts/fpga_ru.py --mcs <某个>_update.mcs --host 127.0.0.1 --port 12340 --yes
+```
+
+正确的一次会让 mock 收到**恰好 8388608 字节（8.000 MiB）**—— 正是 `cAddrUpdateEnd - cAddrUpdateStart`。
+对不上就是客户端的分帧错了。`--not-ready-every N` / `--abort-at-frame N` 可以逼出重传和中止路径。
+
+## 参考
+
+- `references/protocol.md` —— TCP 1234 帧格式、时序、三个硬限制、DPD 污染缺陷的完整证据链
+- `references/quickboot.md` —— flash 布局、切换字兜底、为什么是 `-imagesize 64`、两条烧写路径的区别
diff --git a/fpga_update/references/protocol.md b/fpga_update/references/protocol.md
new file mode 100644
index 0000000..9fcd6ba
--- /dev/null
+++ b/fpga_update/references/protocol.md
@@ -0,0 +1,137 @@
+# FPGA 远程升级协议（TCP 1234）
+
+## 首先：这条路和网页升级毫无关系
+
+设备网页 `sys_set.html` 的 **Remote Update** 框是给 **imx6 固件包**用的，不是给 FPGA 用的。
+把 `.mcs` 传上去等于扔进垃圾桶：
+
+```
+upload.cgi  →  存到 /aw/upgrade/  →  通知 app（WEB_UPGRADE）
+app: start_upgrade() → set_upgrade_flg(1) → imx_up_thr → check_package.sh
+check_package.sh:  find /aw/upgrade -name '*.tar.gz'   ← 只认 tar.gz
+                   找不到 → ret=3 → rm -rf /aw/upgrade/*  ← 你的 mcs 被删了
+```
+
+（`start_upgrade()` 还会在 `ctrl==CTRL_LOCAL` 时直接 return 0 —— 本地模式下网页升级本就不生效。）
+
+FPGA 走的是**另一条独立的路**：app 里的 `fpga_upgrade_task()` 在 **TCP 1234** 上起了个 server
+（`inc/config.h: UPGRADE_LISTEN_PORT 1234`），跑私有帧协议。原厂用 `tools/Xfpga_ru.exe` 驱动它
+（一个 16 KB 的 .NET GUI，没有命令行）。`scripts/fpga_ru.py` 就是它的替代品。
+
+服务端权威实现：`my_exciter/imx6/src/atsc_app/task/tcp_task.c` :: `fpga_upgrade_process()`。
+
+## 帧格式
+
+```
+请求   AA | cmd | datalen:u16 小端 | payload | sum | BB
+应答   AA | ack | 01 | 00          | ack     | sum | BB      （固定 7 字节）
+sum    从帧头到 sum 之前所有字节的 8-bit 累加和（get_check_sum，见 bsp/general_api.c）
+```
+
+服务端校验 `rxdata[0]==0xAA && rxdata[datalen+5]==0xBB`，再比对 `get_check_sum(rxdata, datalen+4)`。
+应答里 ack 同时出现在 `[1]` 和 `[4]`；校验和错时 `[1]` 被改写，所以**读 `[4]`**。
+
+| cmd | 值 | ack | 值 |
+|---|---|---|---|
+| `UPGRADE_START` | 0x01 | `START_OK` | 0x01 |
+| `POLL_ERASER` | 0x02 | `ERASER_ING` / `ERASER_END` | 0x02 / 0x03 |
+| `W_DATA_BLOCK` | 0x03 | `READY` / `NOT_READY` | 0x04 / 0x05 |
+| `W_DATA_BLOCK_LAST` | 0x04 | `ABORT_SUCCESS` / `ABORT_ERROR` | 0x06 / 0x07 |
+
+## 时序
+
+1. `UPGRADE_START`（空负载）→ `START_OK`。
+   服务端往 FPGA 寄存器 `0xE0` 写 `RESET_EN(bit0)=1` 再写 0，放开 `SpiFlashProgrammer`。
+2. 轮询 `POLL_ERASER` 直到 `ERASER_END`。擦 8 MB 要几十秒。
+3. 逐帧 `W_DATA_BLOCK`，负载是 **mcs 的原始 ASCII 文本行**。
+   - `READY` → 发下一帧
+   - `NOT_READY` → 等 1 ms **重发同一帧**
+4. 负载里出现 EOF 记录（`:00000001FF`）后，改发 `W_DATA_BLOCK_LAST`（空负载）轮询，
+   直到 `ABORT_SUCCESS` 或 `ABORT_ERROR`。服务端最多让你轮 100 次。
+5. `ABORT_SUCCESS` 后 app 调 `reset_fpga_system()` 并置 `fpga_up_sys_reboot=1` —— **imx6 会重启，
+   连接会断**。这是正常的，不是错误。
+
+## 三个会把你坑死的硬限制
+
+都在 `recv_upgrade_data_parse()` / `inc/tcp_task.h` 里，超了就 `ABORT_ERROR`：
+
+1. **负载 ≤ 538 字节**。`rxdata[MAX_DATA_LEN=544]` 装的是**整帧**，帧头尾占 6 字节。
+2. **每行长度必须在 `[11, 50]` 字节之间**（`UPDATA_RAW_FRAME_MIN/MAX_LEN`），**含行尾 CRLF**。
+   实际 mcs 只有三种行长：17（ELA）、45（数据）、13（EOF），都合法。
+3. **一帧里必须是整行**。服务端不看换行符，它用 `strchr(payload, ':')` 重新切分，
+   最后一行按 `strlen` 量到负载末尾。**一行跨两帧 = 两行畸形行 = 直接中止。**
+
+另外服务端一次 `lwip_recv()` 读一帧，所以客户端要一次 `send()` 发完整帧，并关掉 Nagle。
+
+数据行 `:10 AAAA 00 <32 hex> CC` 里只有那 16 个数据字节被取走；地址和 ELA 记录全被忽略
+（详见 `references/quickboot.md`）。8 MB 镜像 = 524288 行 × 16 B，按 11 行/帧 → 约 47663 帧。
+
+## 曾经的"第二次远程升级必失败" —— 已定位并修复
+
+**症状**：JTAG 烧完 RU_golden 后第一次远程升级成功；冷启动后再升级第二次，卡在擦除阶段永不结束。
+
+**根因**：`smod_top.v` 里 flash 的三根 SPI 线**根本没接出去**。
+
+```verilog
+// 修复前 —— 只是内部 wire，MISO 还被钉死成常量
+wire outSpiCsB, outSpiMosi;
+wire inSpiMiso = 1'b1;
+
+// RU_golden_top.v 一直是对的
+output outSpiCsB ;
+output outSpiMosi ;
+input  inSpiMiso ;
+```
+
+`smod_ioplanning.xdc` 和 `timing.xdc` 早就把 `outSpiCsB`(P18) / `outSpiMosi`(R14) / `inSpiMiso`(R15)
+约束好了，但顶层没有这几个端口，于是 Vivado 只是丢了三条 CRITICAL WARNING 就继续：
+
+```
+CRITICAL WARNING: [Vivado 12-4739] set_output_delay:No valid object(s) found for '[get_ports outSpiCsB]'
+```
+
+后果：my_exciter 里的 `SpiFlashProgrammer` 片选和 MOSI 悬空、MISO 恒为 1，读 flash ID 得到全 `0xFF`，
+**在擦除之前就报错退出**。golden 的引脚是接好的，所以由 golden 执行的那一次升级永远成功 ——
+两次升级执行擦写的是**不同 bitstream 里的 programmer**，这正是"第一次成、第二次败"的由来。
+
+**修复**：把三个信号提升为 `smod_top` 的顶层端口（已改）。实测 `rstate` 从
+`0x7`（ERROR=1, DONE=1, EraseOK=0）变为 `0x0` → `0xc`（Ready=1, EraseOK=1），擦除正常结束。
+
+## ⚠ imx6 缺陷：`POLL_ERASER` 不看 ERROR 位
+
+上面那个 FPGA bug 之所以表现为"卡死"而不是"报错"，是因为 `tcp_task.c` 的擦除轮询只判断 `EARSE_OK`：
+
+```c
+rstate = fpga_byte_read(0x01, FPGA_UPGRADE_CTRL_ADDR);
+if (rstate & FPGA_UP_STATUS_EARSE_OK_MASK) ret_data = ERASER_END;
+else                                       ret_data = ERASER_ING;   // ERROR 位被无视
+```
+
+于是"已经失败"和"还在擦"返回同一个 `ERASER_ING`，客户端只能一直等到超时。原厂那个
+`Xfpga_ru.exe` 就是这样毫无提示地转圈。
+
+**这个 bug 还在。** 排查时不要信客户端的"erasing..."，直接读设备日志：
+
+```bash
+ssh -o HostKeyAlgorithms=+ssh-rsa root@192.168.1.143 \
+  "grep 'ing|end' /var/module_log/imx_atsc_app.log | tail -3"
+```
+
+`rstate=0x7` = programmer 已中止；`rstate=0x0` 才是真的在擦。
+
+**根治**：在 `POLL_ERASER` 分支里加 `if (rstate & FPGA_UPGRADE_STATUS_ERROR_MASK) ret_data = ABORT_ERROR;`。
+改完要重新交叉编译并部署 app。
+
+## 尚未证实的风险：DPD 与升级数据流共用 SPI1
+
+这一条**不是**上面那个 bug 的原因，但结构上确实存在，写在这里以免以后踩：
+
+- FPGA 侧升级期间 `inReset_EnableB==0`，`spi_slave32` 常开，SPI1 上任何 32-bit 字都会被当成镜像数据。
+  `smod_top.v` 自己的注释：`// when doing dpd S&L, MISO is spi read pin`。
+- imx6 侧 DPD 的 `ispi_tx_big_block()`（`bsp/dpd_proc.c`）和升级的 `ispi_fpga_send_buf_4Byte()`
+  **用同一个 `fpga_spi_fd`**（`bsp/isys_config.c`）。
+- 互斥量 `g_now_only_upgrade_fpga` 只在 `lcd_task.c` 一处被读，且在 `ERASER_END` 才置位。
+  另一个标志 `get_upgrade_flg()` 只有网页升级会置位。
+
+实测三次连续升级成功时 `adpc=0`（ADPC_OFF），所以 DPD 的 S&L 没有在跑 —— **这条路径从未被真正验证过**。
+稳妥起见，升级前把 ADPC 设为 OFF。
diff --git a/fpga_update/references/quickboot.md b/fpga_update/references/quickboot.md
new file mode 100644
index 0000000..e62c7e5
--- /dev/null
+++ b/fpga_update/references/quickboot.md
@@ -0,0 +1,77 @@
+# Flash 布局与 QuickBoot 兜底
+
+板上 flash：Winbond **w25q128bv**，128 Mbit = 16 MB，SPIx1，**24-bit 地址**，64 KB sector / 4 KB subsector。
+
+## 三个地址，一个都不能记错
+
+来自 `my_exciter.srcs/sources_1/quickboot/SpiFlashProgrammer.vhd`（**RU_golden 和 my_exciter 引用的是同一份文件**，
+见 `RU_golden.xpr` 里的 `$PPRDIR/../my_exciter/.../SpiFlashProgrammer.vhd`，所以两个工程的常数天然一致）：
+
+```vhdl
+cAddrUpdateStart     = X"00400000"   -- update 镜像起点
+cAddrUpdateEnd       = X"00C00000"   -- update 镜像终点+1  →  槽宽 8 MB
+cAddrSwitchWordN25Q  = X"00000FFC"   -- 切换字
+cAddrWidth           = 24            -- 和 w25q128bv 匹配
+```
+
+flash 实际布局（从 `golden_design_initial.mcs` 解码得到，不是推测）：
+
+```
+0x000000..0x000FFB  FF FF ...                  空
+0x000FFC            00 99 55 66                切换字（AA 被写成 00 = 无效）
+0x001000            20000000                   NOOP
+                    30020001 00400000          写 WBSTAR = 0x400000
+                    30008001 0000000F          写 CMD = IPROG   → 跳到 update 镜像
+0x001050            AA995566                   golden bitstream 自己的同步字
+        ..0x3FFFFF  golden bitstream（RU_golden_top.bit，3.4 MB）
+0x400000..0xBFFFFF  update 镜像（8 MB，日常远程升级写这里）
+0xC00000..0xFFFFFF  未用
+```
+
+**兜底机制**：上电时 FPGA 从地址 0 扫描同步字。
+- 切换字 = `AA995566` → 在 0xFFC 同步 → 执行 0x1000 处的 IPROG → 加载 update 镜像。
+- 切换字 ≠ `AA995566`（擦除后是 `FFFFFFFF`）→ 0xFFC 不同步 → 继续扫 → 在 0x1050 撞上 golden 的同步字 → **加载 golden**。
+
+远程升级一开始就把切换字所在的 subsector 擦掉（`sProgrammerEraseSwitchWord`，SSE 命令，
+擦 `0x000000..0x000FFF`），写完校验通过后再把切换字写回去。所以**升级中途断电只会退回 golden，不会变砖**。
+
+> subsector 是 4 KB，擦 0xFFC 只波及 `0x000000..0x000FFF`。IPROG 头在 `0x001000`，属于下一个
+> subsector，**不会被擦到**。golden 是安全的 —— 如果怀疑"升级把 golden 弄坏了"，可以排除这条。
+
+## 为什么 `-imagesize 64`
+
+perl 脚本的模型是"golden 和 update 一样大、update 紧跟 golden"：
+`imageSizeBytes = imagesize(Mbit) × 131072`，`updateStartAddr = imageSizeBytes`。
+
+- **golden**（`MakeSpiFlashProgrammerOnlyInitMcs.pl -imagesize 32`）→ 4 MB → IPROG 里的
+  WBSTAR = `0x400000`，和 `cAddrUpdateStart` 对上。日志见 `RU_golden/bitstream/gen_initial.log`。
+- **update**（`MakeSpiFlashProgrammerMcsFiles_backup.pl -imagesize 64`）→ 8 MB → 负载
+  `pad(bitstream, 0x7FFFFC) + crc32` 正好 8 MB，和槽宽 `0xC00000-0x400000` 对上。
+
+两个 imagesize 不一样是**故意的**（VHDL 注释里的 "Plan B: golden enlarged to 4MB"），别去"统一"它们。
+
+`_update.mcs` 里的 Intel-HEX 地址写的是 `0x800000`，看起来和 `cAddrUpdateStart=0x400000` 矛盾 ——
+**它不矛盾，因为地址被忽略**。imx6 的 `recv_upgrade_data_parse()` 只从 `:10xxxx00<32 hex>` 记录里抠出
+16 个数据字节，扩展线性地址记录 `:02000004` 直接当"中间控制字段"丢掉。FPGA 从 `cAddrUpdateStart`
+顺序写。所以对远程升级而言，`_update.mcs` 只有**负载长度**和**CRC**是有意义的。
+
+校验 `_update.mcs` 的起始地址是 `0x800000`，本质上是在校验"这个文件是用 `-imagesize 64` 生成的"。
+
+## 两条烧写路径，别搞混
+
+| | JTAG 直烧 | 网络远程升级 |
+|---|---|---|
+| 文件 | `RU_golden/bitstream/golden_design_initial.mcs` | `bitstream/<tag>_update.mcs` |
+| 工具 | `/fpga_flash` skill（flash 型号 `w25q128bv`） | `scripts/fpga_ru.py` |
+| 写到 | `0x000000..0x3FFFFF` | `0x400000..0xBFFFFF` |
+| 何时用 | 首次装机、板子变砖、换打底镜像 | 日常迭代 |
+
+`_update.mcs` **不能**拿去 JTAG 烧写：它只是 8 MB 裸数据 + CRC，没有 bitstream 头。
+反过来 `_initial.mcs` 也不能拿去远程升级。
+
+严守上电时序：断电 → 断 USB-Blast → 上电 → 连 USB-Blast → 烧写 → 断电 → 断 USB-Blast → 上电 → 连 USB-Blast。
+
+## 调试期的第三条路
+
+`program_device` 直接把 `.bit` 下到 FPGA，掉电即失，不碰 flash，配 `.ltx` 还能抓 ILA。
+迭代验证用它最快，功能确认后再走远程升级固化。固化用的 bit 要**去掉 ILA 重新实现**。
diff --git a/fpga_update/scripts/fpga_ru.py b/fpga_update/scripts/fpga_ru.py
new file mode 100644
index 0000000..4f0eec3
--- /dev/null
+++ b/fpga_update/scripts/fpga_ru.py
@@ -0,0 +1,307 @@
+#!/usr/bin/env python3
+"""Remote-update the SMOD FPGA by streaming a quickboot _update.mcs to the imx6.
+
+This is a Python reimplementation of what ``tools/Xfpga_ru.exe`` (a .NET GUI, no CLI)
+does. The device side is ``my_exciter/imx6/src/atsc_app/task/tcp_task.c`` ::
+``fpga_upgrade_task``, listening on TCP 1234. It parses the MCS text we send and
+pokes the bytes into the FPGA's SpiFlashProgrammer, which erases and writes the
+update half of the SPI flash by itself.
+
+This has nothing to do with the web page's "Remote Update" box -- that one runs
+check_package.sh, which only accepts *.tar.gz imx6 firmware bundles and deletes
+anything else it finds. See references/protocol.md.
+
+Wire format (all values from tcp_task.c / inc/tcp_task.h)::
+
+    request   AA | cmd | datalen:u16le | payload | sum | BB
+    response  AA | ack | 01 | 00       | ack     | sum | BB     (7 bytes)
+    sum       8-bit sum of every byte from the head up to (not incl.) sum
+
+Usage:
+    python fpga_ru.py --mcs .../bitstream/atsc_exciter_update.mcs
+    python fpga_ru.py --mcs ... --host 192.168.1.143 --yes
+    python fpga_ru.py --mcs ... --dry-run          # parse + frame, touch no hardware
+"""
+
+import argparse
+import os
+import re
+import socket
+import sys
+import time
+
+# ------------------------------------------------------------------ protocol constants
+
+HEAD, TAIL = 0xAA, 0xBB
+
+UPGRADE_START, POLL_ERASER, W_DATA_BLOCK, W_DATA_BLOCK_LAST = 0x01, 0x02, 0x03, 0x04
+
+ACK = {
+    0x01: "START_OK",
+    0x02: "ERASER_ING",
+    0x03: "ERASER_END",
+    0x04: "READY",
+    0x05: "NOT_READY",
+    0x06: "ABORT_SUCCESS",
+    0x07: "ABORT_ERROR",
+}
+START_OK, ERASER_ING, ERASER_END = 0x01, 0x02, 0x03
+READY, NOT_READY, ABORT_SUCCESS, ABORT_ERROR = 0x04, 0x05, 0x06, 0x07
+
+# rxdata[] on the device is MAX_DATA_LEN = 544 bytes and holds the *whole* frame,
+# so the payload can never exceed 544 - 6 bytes of framing. Going over this silently
+# truncates the frame and the tail check fails.
+MAX_FRAME = 544
+FRAME_OVERHEAD = 6
+MAX_PAYLOAD = MAX_FRAME - FRAME_OVERHEAD  # 538
+
+# recv_upgrade_data_parse() rejects any line outside this range with -2 -> ABORT_ERROR.
+LINE_MIN, LINE_MAX = 11, 50
+
+UPDATE_START_ADDR = 0x800000
+ELA_RE = re.compile(rb"^:02000004([0-9A-Fa-f]{4})")
+EOF_RECORD = b":00000001FF"
+
+
+# ---------------------------------------------------------------------------- framing
+
+def build(cmd, payload=b""):
+    body = bytes([HEAD, cmd]) + len(payload).to_bytes(2, "little") + payload
+    return body + bytes([sum(body) & 0xFF, TAIL])
+
+
+def parse_ack(resp):
+    if len(resp) != 7 or resp[0] != HEAD or resp[6] != TAIL:
+        raise ProtocolError(f"malformed response: {resp.hex(' ')}")
+    return resp[4]
+
+
+class ProtocolError(RuntimeError):
+    pass
+
+
+class Device:
+    def __init__(self, host, port, timeout):
+        self.sock = socket.create_connection((host, port), timeout=timeout)
+        # Every frame is a request/response round trip; Nagle would just add latency.
+        self.sock.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)
+
+    def xfer(self, cmd, payload=b"", timeout=None):
+        if timeout is not None:
+            self.sock.settimeout(timeout)
+        self.sock.sendall(build(cmd, payload))
+        try:
+            resp = self.sock.recv(16)
+        except socket.timeout:
+            raise ProtocolError(
+                f"no response to cmd 0x{cmd:02X} within {self.sock.gettimeout()}s.\n"
+                "  The app returns an empty reply (and we block) when sys_ready==0 or the\n"
+                "  frame was malformed. Check the app is up: ssh root@<host> 'ps | grep atsc'"
+            )
+        if not resp:
+            raise ProtocolError("device closed the connection")
+        return parse_ack(resp)
+
+    def close(self):
+        try:
+            self.sock.close()
+        except OSError:
+            pass
+
+
+# ------------------------------------------------------------------------ mcs handling
+
+def load_chunks(path, max_payload, allow_any_image):
+    """Split the MCS into whole-line payloads.
+
+    Lines must stay whole: the device re-splits our payload on ':' and measures each
+    line's length itself, so a line straddling two frames is parsed as two malformed
+    lines and aborts the upgrade.
+    """
+    with open(path, "rb") as fh:
+        lines = fh.readlines()
+    if not lines:
+        raise SystemExit(f"[FAIL] {path} is empty")
+
+    first_ela = None
+    for line in lines:
+        m = ELA_RE.match(line)
+        if m:
+            first_ela = int(m.group(1), 16) << 16
+            break
+    if first_ela != UPDATE_START_ADDR and not allow_any_image:
+        got = "none" if first_ela is None else f"0x{first_ela:X}"
+        raise SystemExit(
+            f"[FAIL] {os.path.basename(path)} starts at {got}, expected 0x{UPDATE_START_ADDR:X}.\n"
+            "       That is not a quickboot _update.mcs. Uploading a base or _initial mcs\n"
+            "       writes garbage into the update slot; the board then falls back to golden\n"
+            "       on the next cold boot. Regenerate with scripts/gen_update_mcs.py.\n"
+            "       (Override with --allow-any-image if you really mean it.)"
+        )
+
+    for n, line in enumerate(lines, 1):
+        if not line.startswith(b":"):
+            raise SystemExit(f"[FAIL] line {n} does not start with ':' -- not an Intel-HEX mcs")
+        if not (LINE_MIN <= len(line) <= LINE_MAX):
+            raise SystemExit(
+                f"[FAIL] line {n} is {len(line)} bytes (incl EOL); the device only accepts "
+                f"{LINE_MIN}..{LINE_MAX} and aborts otherwise"
+            )
+    if not lines[-1].startswith(EOF_RECORD):
+        raise SystemExit(f"[FAIL] last line is {lines[-1]!r}, expected the EOF record {EOF_RECORD.decode()}")
+
+    chunks, cur = [], []
+    size = 0
+    for line in lines:
+        if size + len(line) > max_payload and cur:
+            chunks.append(b"".join(cur))
+            cur, size = [], 0
+        cur.append(line)
+        size += len(line)
+    if cur:
+        chunks.append(b"".join(cur))
+    return chunks, len(lines)
+
+
+# ------------------------------------------------------------------------------ stages
+
+def do_erase(dev, timeout):
+    print("[2/4] erasing the update image region (this is the slow part)")
+    ack = dev.xfer(UPGRADE_START, timeout=30)
+    if ack != START_OK:
+        raise ProtocolError(f"UPGRADE_START -> {ACK.get(ack, hex(ack))}, expected START_OK")
+
+    deadline = time.monotonic() + timeout
+    dots = 0
+    while time.monotonic() < deadline:
+        ack = dev.xfer(POLL_ERASER, timeout=30)
+        if ack == ERASER_END:
+            print("\n      erase complete")
+            return
+        if ack != ERASER_ING:
+            raise ProtocolError(f"POLL_ERASER -> {ACK.get(ack, hex(ack))}")
+        dots += 1
+        print(f"\r      erasing... {dots * 0.5:.0f}s", end="", flush=True)
+        time.sleep(0.5)
+    raise ProtocolError(
+        f"flash still erasing after {timeout}s.\n"
+        "  POLL_ERASER only ever reports ERASER_ING/ERASER_END -- the app never checks the\n"
+        "  ERROR bit, so a programmer that aborted looks identical to one that is still busy.\n"
+        "  Read the real status on the device:\n"
+        "    ssh root@<host> \"grep 'ing|end' /var/module_log/imx_atsc_app.log | tail -3\"\n"
+        "  rstate=0x7 (ERROR|DONE, EraseOK=0) means SpiFlashProgrammer aborted immediately --\n"
+        "  the running design cannot reach the flash. See references/protocol.md.")
+
+
+def do_write(dev, chunks, total_lines, retries):
+    print(f"[3/4] streaming {total_lines} mcs lines in {len(chunks)} frames")
+    sent_lines = 0
+    t0 = time.monotonic()
+    for i, chunk in enumerate(chunks):
+        for attempt in range(retries + 1):
+            ack = dev.xfer(W_DATA_BLOCK, chunk, timeout=60)
+            if ack == READY:
+                break
+            if ack == NOT_READY:
+                # Device asks us to re-send this exact block after a short pause.
+                time.sleep(0.001)
+                continue
+            if ack in (ABORT_SUCCESS, ABORT_ERROR):
+                return ack
+            raise ProtocolError(f"W_DATA_BLOCK -> {ACK.get(ack, hex(ack))}")
+        else:
+            raise ProtocolError(f"frame {i} still NOT_READY after {retries} retries")
+
+        sent_lines += chunk.count(b"\n")
+        if i % 2000 == 0 or i == len(chunks) - 1:
+            pct = 100.0 * sent_lines / total_lines
+            rate = sent_lines / max(time.monotonic() - t0, 1e-6)
+            eta = (total_lines - sent_lines) / max(rate, 1e-6)
+            print(f"\r      {pct:5.1f}%  {sent_lines}/{total_lines} lines  ETA {eta/60:.1f} min",
+                  end="", flush=True)
+    print()
+    return None
+
+
+def do_finish(dev, timeout):
+    print("[4/4] waiting for the FPGA to report the write is done")
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        try:
+            ack = dev.xfer(W_DATA_BLOCK_LAST, timeout=30)
+        except (ProtocolError, ConnectionResetError, OSError) as exc:
+            # On success the app calls reset_fpga_system() and reboots the imx6, so the
+            # socket dying right here is expected -- but we can no longer read the ack.
+            raise ProtocolError(f"link dropped while waiting for the final ack: {exc}")
+        if ack in (ABORT_SUCCESS, ABORT_ERROR):
+            return ack
+        if ack != READY:
+            raise ProtocolError(f"W_DATA_BLOCK_LAST -> {ACK.get(ack, hex(ack))}")
+        time.sleep(0.2)
+    raise ProtocolError(f"no final verdict after {timeout}s")
+
+
+# -------------------------------------------------------------------------------- main
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--mcs", required=True, help="path to the quickboot <tag>_update.mcs")
+    ap.add_argument("--host", default="192.168.1.143")
+    ap.add_argument("--port", type=int, default=1234)
+    ap.add_argument("--yes", action="store_true", help="skip the confirmation prompt")
+    ap.add_argument("--dry-run", action="store_true", help="validate and frame the mcs, contact nothing")
+    ap.add_argument("--allow-any-image", action="store_true",
+                    help="skip the 0x800000 start-address check (you almost never want this)")
+    ap.add_argument("--max-payload", type=int, default=MAX_PAYLOAD)
+    ap.add_argument("--erase-timeout", type=float, default=600)
+    ap.add_argument("--finish-timeout", type=float, default=120)
+    ap.add_argument("--retries", type=int, default=200)
+    args = ap.parse_args()
+
+    if not 0 < args.max_payload <= MAX_PAYLOAD:
+        raise SystemExit(f"[FAIL] --max-payload must be 1..{MAX_PAYLOAD} (device rxdata is {MAX_FRAME} bytes)")
+
+    print(f"[1/4] validating {args.mcs}")
+    chunks, total_lines = load_chunks(args.mcs, args.max_payload, args.allow_any_image)
+    mb = os.path.getsize(args.mcs) / 1e6
+    print(f"      {mb:.1f} MB, {total_lines} lines, {len(chunks)} frames, starts at 0x{UPDATE_START_ADDR:X}")
+
+    if args.dry_run:
+        print("[dry-run] mcs is well-formed and would stream cleanly. Nothing was sent.")
+        return
+
+    if not args.yes:
+        print(f"\nAbout to erase and rewrite the update image on {args.host}.")
+        print("The golden image is untouched, so a failure here is recoverable by cold boot.")
+        print("Do NOT power the board off while this runs.")
+        print("\nMake sure DPD is OFF. The exciter's DPD sample&load writes to the same SPI1 bus\n"
+              "we stream the image over, and the FPGA's spi_slave32 will happily fold those words\n"
+              "into the flash image. See references/protocol.md.")
+        if input("Continue? [y/N] ").strip().lower() not in ("y", "yes"):
+            raise SystemExit("aborted")
+
+    dev = Device(args.host, args.port, timeout=30)
+    try:
+        do_erase(dev, args.erase_timeout)
+        verdict = do_write(dev, chunks, total_lines, args.retries)
+        if verdict is None:
+            verdict = do_finish(dev, args.finish_timeout)
+    finally:
+        dev.close()
+
+    if verdict == ABORT_SUCCESS:
+        print("\n[OK] FPGA reports the update image was written and verified.")
+        print("The app has reset the FPGA and is rebooting the imx6.")
+        print("Now cold-boot the board (observe the SMOD power-up sequence) to load the new image.")
+        return
+    raise SystemExit(
+        f"\n[FAIL] device returned {ACK.get(verdict, hex(verdict))}.\n"
+        "The update slot is now half-written, but the switch word was erased at the start,\n"
+        "so the board still cold-boots golden. Nothing is bricked.\n\n"
+        "If this was ABORT_ERROR late in the stream, suspect SPI1 contention: the exciter's\n"
+        "DPD sample&load shares fpga_spi_fd with this upload, and the FPGA folds any SPI1 word\n"
+        "into the image. Turn DPD off and retry. references/protocol.md has the full story.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/fpga_update/scripts/gen_update_mcs.py b/fpga_update/scripts/gen_update_mcs.py
new file mode 100644
index 0000000..7ac6609
--- /dev/null
+++ b/fpga_update/scripts/gen_update_mcs.py
@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+"""Build the quickboot _update.mcs for the SMOD ATSC exciter.
+
+Runs exactly the three scripts the project has always used, in order:
+
+  1. tcl_1/post_bitstream.tcl                    impl_1/*.bit,.bin,.ltx -> bitstream/
+  2. tcl_1/runScript_GenRu.tcl                   write_cfgmem -size 64 -interface SPIx1
+                                                 -> bitstream/atsc_exciter.mcs
+  3. tcl_1/MakeSpiFlashProgrammerMcsFiles_backup.pl -imagesize 64 atsc_exciter.mcs
+                                                 -> bitstream/atsc_exciter_update.mcs
+
+Step 2 does ``glob *.bit`` inside bitstream/, so every stale .bit sitting there gets
+concatenated into one giant image. bitstream/ accumulates iteration bits, so this
+script moves the old ones aside for the duration and puts them back afterwards.
+
+The resulting _update.mcs is ONLY for the network remote update (scripts/fpga_ru.py).
+It cannot be JTAG-flashed -- see references/quickboot.md.
+"""
+
+import argparse
+import glob
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+
+REPO = r"E:\Xilinx_design\smod\SMOD_ATSC\my_exciter"
+TCL_DIR = os.path.join(REPO, "my_exciter.srcs", "tcl_1")
+BITSTREAM = os.path.join(REPO, "bitstream")
+IMPL = os.path.join(REPO, "my_exciter.runs", "impl_1")
+
+POST_BITSTREAM = os.path.join(TCL_DIR, "post_bitstream.tcl")
+GEN_RU = os.path.join(TCL_DIR, "runScript_GenRu.tcl")
+PERL_SCRIPT = os.path.join(TCL_DIR, "MakeSpiFlashProgrammerMcsFiles_backup.pl")
+DEFAULT_VIVADO = r"D:\Xilinx\Vivado\2018.1\bin\vivado.bat"
+
+# runScript_GenRu.tcl hardcodes this name.
+BASE_MCS = "atsc_exciter.mcs"
+UPDATE_MCS = "atsc_exciter_update.mcs"
+
+# The perl script emits these alongside the one we want.
+BYPRODUCTS = ("atsc_exciter_initial.mcs", "atsc_exciter_golden.mcs", "atsc_exciter_total.mcs",
+              "atsc_exciter_update2.mcs", "atsc_exciter_simple.mcs")
+
+# ":02000004HHHH" is the Intel-HEX extended-linear-address record.
+ELA_RE = re.compile(rb"^:02000004([0-9A-Fa-f]{4})")
+EXPECT_ELA = 0x800000
+
+
+def sh(cmd, cwd=None):
+    print(f"  $ {' '.join(cmd)}", flush=True)
+    if subprocess.run(cmd, cwd=cwd).returncode != 0:
+        sys.exit("[FAIL] command failed; see output above")
+
+
+def first_ela(path):
+    with open(path, "rb") as fh:
+        for line in fh:
+            m = ELA_RE.match(line)
+            if m:
+                return int(m.group(1), 16) << 16
+    return None
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--vivado", default=DEFAULT_VIVADO)
+    ap.add_argument("--tag", help="rename the result to <tag>_update.mcs so it survives the next run")
+    ap.add_argument("--keep-byproducts", action="store_true")
+    args = ap.parse_args()
+
+    for path, what in ((args.vivado, "vivado"), (POST_BITSTREAM, "post_bitstream.tcl"),
+                       (GEN_RU, "runScript_GenRu.tcl"), (PERL_SCRIPT, "perl script")):
+        if not os.path.exists(path):
+            sys.exit(f"[FAIL] {what} not found: {path}")
+
+    impl_bits = glob.glob(os.path.join(IMPL, "*.bit"))
+    if len(impl_bits) != 1:
+        sys.exit(f"[FAIL] expected exactly one .bit in {IMPL}, found {len(impl_bits)}: "
+                 f"{[os.path.basename(b) for b in impl_bits]}")
+    print(f"[1/4] implementation bit: {os.path.basename(impl_bits[0])}")
+
+    # Park every stale .bit so runScript_GenRu.tcl's `glob *.bit` can only see the fresh one.
+    stash = tempfile.mkdtemp(prefix="smod_bit_stash_")
+    parked = glob.glob(os.path.join(BITSTREAM, "*.bit"))
+    for b in parked:
+        shutil.move(b, os.path.join(stash, os.path.basename(b)))
+    if parked:
+        print(f"[2/4] parked {len(parked)} stale .bit files from bitstream/ (restored at the end)")
+
+    try:
+        # One vivado launch, both tcl scripts: each re-derives its own path from [info script].
+        wrapper = os.path.join(stash, "run_both.tcl")
+        with open(wrapper, "w") as fh:
+            fh.write(f'source {{{POST_BITSTREAM}}}\n')
+            fh.write(f'source {{{GEN_RU}}}\n')
+        print(f"[3/4] vivado: post_bitstream.tcl + runScript_GenRu.tcl -> {BASE_MCS}")
+        sh([args.vivado, "-mode", "batch", "-nojournal", "-notrace", "-source", wrapper], cwd=stash)
+
+        base = os.path.join(BITSTREAM, BASE_MCS)
+        if not os.path.exists(base):
+            sys.exit(f"[FAIL] {base} was not produced")
+
+        print(f"[4/4] perl MakeSpiFlashProgrammerMcsFiles_backup.pl -imagesize 64 {BASE_MCS}")
+        sh(["perl", PERL_SCRIPT, "-imagesize", "64", BASE_MCS], cwd=BITSTREAM)
+    finally:
+        for name in os.listdir(stash):
+            if not name.endswith(".bit"):
+                continue
+            dest = os.path.join(BITSTREAM, name)
+            # post_bitstream.tcl just copied the fresh impl_1 bit here under this same
+            # name; putting the parked (stale) one back would silently undo that.
+            if not os.path.exists(dest):
+                shutil.move(os.path.join(stash, name), dest)
+        shutil.rmtree(stash, ignore_errors=True)
+
+    update = os.path.join(BITSTREAM, UPDATE_MCS)
+    if not os.path.exists(update):
+        sys.exit(f"[FAIL] {UPDATE_MCS} was not produced")
+
+    # -imagesize 64 must place the image at 0x800000. If it doesn't, the payload is not
+    # 8 MB and the FPGA's CRC check over 0x400000..0xC00000 will fail after a long upload.
+    ela = first_ela(update)
+    if ela != EXPECT_ELA:
+        got = "none" if ela is None else f"0x{ela:X}"
+        sys.exit(f"[FAIL] {UPDATE_MCS} starts at {got}, expected 0x{EXPECT_ELA:X}. Do not upload it.")
+
+    if not args.keep_byproducts:
+        for name in BYPRODUCTS:
+            p = os.path.join(BITSTREAM, name)
+            if os.path.exists(p):
+                os.remove(p)
+
+    if args.tag:
+        tagged = os.path.join(BITSTREAM, f"{args.tag}_update.mcs")
+        shutil.move(update, tagged)
+        update = tagged
+
+    print(f"\n[OK] {update}  ({os.path.getsize(update)/1e6:.1f} MB)")
+    print(f"     base mcs kept at {os.path.join(BITSTREAM, BASE_MCS)}")
+    print("\nNext: python scripts/fpga_ru.py --mcs \"%s\"" % update)
+    print("      Turn DPD off first -- see references/protocol.md.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/fpga_update/scripts/mock_device.py b/fpga_update/scripts/mock_device.py
new file mode 100644
index 0000000..ef194f4
--- /dev/null
+++ b/fpga_update/scripts/mock_device.py
@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+"""A fake SMOD imx6 that speaks the FPGA remote-update protocol on TCP 1234.
+
+Mirrors ``fpga_upgrade_process()`` in tcp_task.c as literally as possible, including
+the parts that bite: the ':'-based re-splitting of our payload, the 11..50 byte line
+limit, and the READY-then-ABORT_SUCCESS handshake at the end.
+
+The point is to be able to exercise fpga_ru.py end-to-end without erasing a real
+flash. If a change to fpga_ru.py passes against this and against the real board,
+the two agree; if it passes here but not there, this file is the thing that's wrong.
+
+    python mock_device.py --port 12340
+    python fpga_ru.py --mcs some_update.mcs --host 127.0.0.1 --port 12340 --yes
+
+Fault injection, to prove the client's retry/abort paths actually work:
+    --not-ready-every N   answer NOT_READY to every Nth data frame
+    --erase-polls N       answer ERASER_ING to the first N POLL_ERASER frames
+    --abort-at-frame N    answer ABORT_ERROR on the Nth data frame
+"""
+
+import argparse
+import socket
+import sys
+
+HEAD, TAIL = 0xAA, 0xBB
+UPGRADE_START, POLL_ERASER, W_DATA_BLOCK, W_DATA_BLOCK_LAST = 0x01, 0x02, 0x03, 0x04
+START_OK, ERASER_ING, ERASER_END = 0x01, 0x02, 0x03
+READY, NOT_READY, ABORT_SUCCESS, ABORT_ERROR = 0x04, 0x05, 0x06, 0x07
+
+LINE_MIN, LINE_MAX = 11, 50
+MAX_DATA_LEN = 544
+
+
+def reply(ack):
+    tx = bytes([HEAD, ack, 1, 0, ack])
+    return tx + bytes([sum(tx) & 0xFF, TAIL])
+
+
+def split_lines_like_c(payload):
+    """Reproduce the strchr(':') walk in W_DATA_BLOCK.
+
+    The device never looks at newlines: it finds the first ':', then each following
+    ':' ends the previous line. The final line runs to the NUL terminator, which is
+    why a payload that ends mid-line silently mis-measures that line.
+    """
+    out = []
+    start = payload.find(b":")
+    if start < 0:
+        return None
+    while True:
+        nxt = payload.find(b":", start + 1)
+        out.append(payload[start:] if nxt < 0 else payload[start:nxt])
+        if nxt < 0:
+            return out
+        start = nxt
+
+
+def parse_line(line):
+    """recv_upgrade_data_parse(): -2 bad length, -1 EOF record, 16 data bytes, 0 control."""
+    if not (LINE_MIN <= len(line) <= LINE_MAX):
+        return -2
+    length = int(line[1:3], 16)
+    rectype = int(line[7:9], 16)
+    if length == 0x10 and rectype == 0x00:
+        return 16
+    if length == 0x00 and rectype == 0x01:
+        return -1
+    return 0
+
+
+class Mock:
+    def __init__(self, args):
+        self.args = args
+        self.bytes_written = 0
+        self.data_frames = 0
+        self.erase_polls = 0
+        self.file_end = False
+        self.done_polls = 0
+
+    def on_frame(self, cmd, payload):
+        if cmd == UPGRADE_START:
+            self.bytes_written = 0
+            self.erase_polls = 0
+            self.file_end = False
+            return START_OK
+
+        if cmd == POLL_ERASER:
+            self.erase_polls += 1
+            return ERASER_ING if self.erase_polls <= self.args.erase_polls else ERASER_END
+
+        if cmd == W_DATA_BLOCK:
+            self.data_frames += 1
+            if self.args.abort_at_frame and self.data_frames == self.args.abort_at_frame:
+                return ABORT_ERROR
+            if self.args.not_ready_every and self.data_frames % self.args.not_ready_every == 0:
+                return NOT_READY  # client must re-send this exact frame
+
+            lines = split_lines_like_c(payload)
+            if lines is None:
+                return ABORT_ERROR
+            ack = READY
+            for line in lines:
+                n = parse_line(line)
+                if n == -2:
+                    print(f"  [mock] ABORT: bad line ({len(line)} bytes) {line[:20]!r}", file=sys.stderr)
+                    return ABORT_ERROR
+                if n == -1:
+                    self.file_end = True
+                    ack = READY  # check_fpga_upgrade_done() is not DONE yet
+                elif n > 0:
+                    self.bytes_written += n
+            return ack
+
+        if cmd == W_DATA_BLOCK_LAST:
+            self.done_polls += 1
+            if self.done_polls < self.args.done_polls:
+                return READY
+            return ABORT_SUCCESS if self.file_end else ABORT_ERROR
+
+        return ABORT_ERROR
+
+    def serve_one(self, conn):
+        buf = b""
+        while True:
+            data = conn.recv(MAX_DATA_LEN)
+            if not data:
+                return
+            buf += data
+            # The real device does a single recv() per frame; we tolerate splits so a
+            # client bug shows up as a protocol error rather than a hang.
+            while len(buf) >= 6:
+                datalen = int.from_bytes(buf[2:4], "little")
+                total = datalen + 6
+                if len(buf) < total:
+                    break
+                frame, buf = buf[:total], buf[total:]
+                if frame[0] != HEAD or frame[-1] != TAIL:
+                    print(f"  [mock] bad framing {frame[:4].hex(' ')}", file=sys.stderr)
+                    return
+                if sum(frame[:datalen + 4]) & 0xFF != frame[datalen + 4]:
+                    conn.sendall(reply(NOT_READY))
+                    continue
+                conn.sendall(reply(self.on_frame(frame[1], frame[4:4 + datalen])))
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--port", type=int, default=12340)
+    ap.add_argument("--erase-polls", type=int, default=2)
+    ap.add_argument("--done-polls", type=int, default=1)
+    ap.add_argument("--not-ready-every", type=int, default=0)
+    ap.add_argument("--abort-at-frame", type=int, default=0)
+    args = ap.parse_args()
+
+    srv = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    srv.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+    srv.bind(("127.0.0.1", args.port))
+    srv.listen(1)
+    print(f"[mock] listening on 127.0.0.1:{args.port}", flush=True)
+
+    conn, _ = srv.accept()
+    mock = Mock(args)
+    try:
+        mock.serve_one(conn)
+    finally:
+        conn.close()
+    mib = mock.bytes_written / (1 << 20)
+    print(f"[mock] {mock.data_frames} data frames, {mock.bytes_written} bytes ({mib:.3f} MiB), "
+          f"file_end={mock.file_end}")
+    # A correct run streams exactly the 8 MiB update image.
+    sys.exit(0 if mock.file_end else 1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/smod_atsc_dev/SKILL.md b/smod_atsc_dev/SKILL.md
new file mode 100644
index 0000000..0813494
--- /dev/null
+++ b/smod_atsc_dev/SKILL.md
@@ -0,0 +1,151 @@
+---
+name: smod_atsc_dev
+description: >
+  在 SMOD 硬件平台（Artix-7 xc7a200t + i.MX6 + ADRV9009）上开发、调试、固化 ATSC 激励器：
+  改 RTL 并用 vivado-mcp 综合实现、在 Ubuntu VM 上交叉编译 imx6 的 atsc_app 并部署、
+  用 ILA/串口/频谱做闭环验证。（生成 quickboot .mcs 和远程升级 FPGA 转交 fpga_update skill。）
+
+  只要用户提到 SMOD、my_exciter、RU_golden、smod_top、atsc_app、ATSC 激励器、8VSB 调制、
+  ADRV9009/ORx/DPD/TxAGC、tuner(Sony CXD2878)、TSoIP、quickboot、远程升级 mcs、192.168.1.143、
+  或在 E:\Xilinx_design\smod\SMOD_ATSC 下做任何 FPGA / imx6 改动，就必须使用这个 skill —— 即便
+  他们没点名 "skill"，也没说 "SMOD"，只是说 "把新 bit 升上去"、"改一下激励器的 RTL"、
+  "imx6 app 编译部署一下"、"抓个 ILA 看看"、"重新出个 mcs"。这个 skill 给出平台事实、
+  两条烧写路径的选择、上电时序铁律，并把闭环调试转交给 fpga-imx6-flow。
+---
+
+# SMOD ATSC 激励器开发
+
+SMOD 是一块 **Artix-7 (xc7a200tfbg676-2) + i.MX6 + ADRV9009** 的 ATSC 激励器板。
+FPGA 做 8VSB 调制 / DPD / AGC / TS 前处理，imx6 跑 `atsc_app`（配置 ADRV9009、tuner、
+网页、SNMP），两者通过一个 I2C/寄存器桥通信。
+
+这个 skill 是**总入口**：它告诉你平台的硬事实、该走哪条烧写路径、以及什么时候把活
+交给别的 skill。具体细节在 `references/`，可直接运行的脚本在 `scripts/`。
+
+## 平台事实速查
+
+| 项 | 值 |
+|---|---|
+| 激励器 Vivado 工程 | `E:\Xilinx_design\smod\SMOD_ATSC\my_exciter\my_exciter.xpr`（top = `smod_top`，也是 **git 仓库根**） |
+| 打底 golden 工程 | `E:\Xilinx_design\smod\SMOD_ATSC\RU_golden\RU_golden.xpr`（top = `RU_golden_top`） |
+| FPGA / Flash | `xc7a200tfbg676-2` / Winbond **w25q128bv**（128 Mbit = 16 MB，SPIx1） |
+| Vivado | `D:\Xilinx\Vivado\2018.1`（`bin\vivado.bat`） |
+| imx6 源码 | `my_exciter\imx6\src\atsc_app`（Makefile 在该目录根） |
+| imx6 设备 | `192.168.1.143`，`root` / `anywavecom`（网页 + SSH；Windows 主机 `192.168.1.125` 同网段） |
+| 交叉编译 VM | `bowei@192.168.223.128`，密码 `lr7402`，构建目录 `~/atsc_head_build` |
+| 工具链 | `/opt/fsl-imx-x11/4.1.15-2.0.0/environment-setup-cortexa7hf-neon-poky-linux-gnueabi` |
+| app 部署路径 | `/aw/anywave/app/imx_atsc_app`（先写 `imx_atsc_app_head`，再 `cp -a` 覆盖 active） |
+| app 日志 | `/var/module_log/imx_atsc_app.log` |
+| 串口 | COM 号**逐机而异，两处配置都在 `~` 下、都不入库**：两个继电器（电源 / USB-Blast）看 `~/vivado_mcp_usb.conf`（缺了就跑 vivado-mcp 的 `detect_serial_relays` 自动探测生成）；imx6 控制台看 `~/fpga_imx6_flow.json` 的 `ports.imx6`。**任何地方都不要写死 COM 号。** |
+| 原理图 | `my_exciter\Notes\SMOD_V1.0_SCH251010.PDF` |
+| 构建产物约定 | `my_exciter\bitstream\<tag>.{bit,ltx,mcs,prm}`，一次迭代一个 tag（`smod_p1`、`smod_dpd`…） |
+
+## ⚠️ 硬件铁律：上电时序
+
+**给 SMOD 上电/断电之前，USB-Blast 必须先断开。** 绝不在 USB-Blast 连接状态下切换板卡电源
+——这会打坏 JTAG 适配器和板卡。
+
+```
+断电(power_off) → 断开 USB-Blast(USB_Blast_off) → 上电(power_on) → 连接 USB-Blast(USB_Blast_on) → JTAG/ILA
+```
+
+用 vivado-mcp 的 `power_on` / `power_off` / `USB_Blast_on` / `USB_Blast_off` 工具执行。
+上电后要**等 imx6 串口控制台打印 `sys init over` 再连 JTAG**，否则 TX 链路会挂掉。
+
+## 两条烧写路径 —— 先选对，再动手
+
+这是本平台最容易搞错的地方。板上 flash 是 **quickboot 双镜像**布局：golden 镜像在
+`0x000000..0x3FFFFF`，update 镜像在 `0x400000..0xBFFFFF`，切换字 `0xAA995566` 在 `0x000FFC`。
+
+| | **A. JTAG 直烧** | **B. 网络远程升级**（PR07 主线） |
+|---|---|---|
+| 何时用 | 板子是砖 / 首次装机 / 要换 golden | 日常迭代，板子已有 golden 且网络能通 |
+| 烧什么 | `RU_golden\bitstream\golden_design_initial.mcs` | `my_exciter\bitstream\<tag>_update.mcs` |
+| 怎么做 | `/fpga_flash` skill（flash 型号 `w25q128bv`） | **`/fpga_update` skill** |
+| 需要 | USB-Blast + 上电时序 | 只要 `192.168.1.143` 能访问 |
+| 风险 | 高（要动电源和 JTAG） | 低（写坏了 golden 会兜底） |
+
+**调试期**还有第三种：`program_device` 直接把 `.bit` 下到 FPGA（掉电即失），配合 `.ltx` 抓 ILA。
+迭代阶段用它最快，验证通过后再走 A 或 B 固化。
+
+### B 走 `/fpga_update`
+
+**远程升级不走网页。** 网页的 Remote Update 只认 imx6 的 `*.tar.gz` 固件包，`.mcs` 传上去会被
+`check_package.sh` 删掉。FPGA 走的是 TCP **1234** 上的私有协议。生成 mcs、烧写、排错的完整流程
+都在 `fpga_update` skill 里，包括两个必须知道的事实：
+
+- **擦除卡住 ≠ 还在擦。** imx6 的 `POLL_ERASER` 不看 ERROR 位，programmer 报错和正常忙碌
+  返回同一个 `ERASER_ING`。读设备日志的 `rstate` 才知道真相（`0x7`=已中止，`0x0`=真在擦）。
+- `-imagesize 64` 和 `SpiFlashProgrammer.vhd` 里的 `cAddrUpdateStart/End` 是一对，别改。
+
+升级完成后**断电冷启动**，让 FPGA 从 flash 重新加载，再验证。
+
+## 三条典型工作流
+
+**只改 RTL** → 编辑 `my_exciter.srcs/sources_1/**` → vivado-mcp `run_synthesis` +
+`run_implementation` + `generate_bitstream` → `program_device` 下 bit → 抓 ILA 验证 →
+满意后走 `/fpga_update` 固化。
+
+**只改 imx6** → 编辑 `imx6/src/atsc_app/**` → 提交到 git（脚本默认打包 `HEAD`）→
+
+```bash
+python .../scripts/imx6_app.py build-deploy --tag mychange
+```
+
+→ 冷启动让 `rc.local` 拉起新 app → `imx6_app.py log` 看日志。
+
+**FPGA + imx6 联调** → 交给 **`/fpga-imx6-flow`**。那个 skill 驱动一张 LangGraph 图，
+把上电/JTAG/烧写/重启做成确定性节点，把「分析 → 判断是否达标 → 提方案 → 改 FPGA / 改 imx6」
+这四步作为中断交回当前会话执行，一轮一轮迭代。PR07 明确要求走这条路做联合调试。
+
+## 什么时候调别的 skill
+
+| 场景 | 用 |
+|---|---|
+| 上电 / JTAG / 综合实现 / 抓 ILA / 下 bit | **vivado-mcp** 工具（`power_on`、`program_device`、`run_tcl`…） |
+| 闭环设计-调试迭代（本 skill 的主推荐） | **`/fpga-imx6-flow`** |
+| JTAG 烧 flash（golden 或砖机恢复） | **`/fpga_flash`**（flash 型号选 `w25q128bv`） |
+| 单模块 RTL 仿真 / 出 testbench | **`/vivado-sim`** |
+| ILA 抓出来的 IQ CSV 做频谱 / OBW / 带肩 | **`/fft-analyse`** |
+| 网页操作（登录、升级、看频谱页） | **claude-in-chrome** |
+| 实测频谱（N9320B 频谱仪） | `mcp__mcp_n9320b__*` 工具 |
+
+## scripts/
+
+| 脚本 | 用途 |
+|---|---|
+| `imx6_app.py` | `build` / `deploy` / `build-deploy` / `log` —— VM 交叉编译 + 推到 imx6 + 读日志 |
+
+出 `.mcs` 和远程升级的脚本在 `/fpga_update` skill 里（`gen_update_mcs.py`、`fpga_ru.py`）。
+
+两个脚本都 `--help` 可查。`imx6_app.py deploy` 会自动装 paramiko 2.12 到缓存目录并 re-exec
+（imx6 的 dropbear 太老，paramiko 5 连不上，见下）。
+
+## references/
+
+按需读，别一次全读进来：
+
+- `references/fpga.md` —— RTL 目录结构、XDC、vivado-mcp 用法、ILA/ltx 约定、已知时序告警的性质
+- `references/imx6.md` —— app 源码结构、交叉编译细节、部署/日志/寄存器访问、网页
+- `references/upgrade.md` —— quickboot 双镜像布局、mcs 生成的确切参数、golden 恢复
+- `references/history.md` —— 历史 PR/Task 笔记索引（tuner、TSoIP、ORx、DPD、AGC 各自踩过什么坑）
+
+## 常见坑
+
+| 症状 | 原因 / 处理 |
+|---|---|
+| 远程升级后重新上电 FPGA 不跑 | 传错了文件 —— 必须是 `_update.mcs`（起始 `0x800000`），不是 `write_cfgmem` 出的基础 mcs |
+| 远程升级卡在 "erasing" 不动 | programmer 已报错退出（`rstate=0x7`），imx6 的轮询不看 ERROR 位。见 `/fpga_update` |
+| imx6 SSH 报 `unknown cipher` / `no acceptable host key` | dropbear 太老。用 paramiko **2.12**（`imx6_app.py` 已处理）；命令行 ssh 加 `-o HostKeyAlgorithms=+ssh-rsa` |
+| imx6 `open_sftp()` → `EOF during negotiation` | 设备没有 SFTP 子系统。用 `cat > file` 走 exec channel 传文件 |
+| VM SSH `kex_exchange_identification: Connection closed` | VM 的 sshd 挂了，请用户 `sudo systemctl restart sshd` |
+| 上电后 TX 链路死掉 | 上电后没等 `sys init over` 就连了 JTAG |
+| `power_off` 卡在循环里 | 先 `USB_Blast_off` 再 `power_off`，顺序反了 |
+| WNS ≈ -9 ~ -13 ns | JESD204 `init_seq_data` 的既有 false path，与功能无关，不要为它改设计 |
+| `i2cget` 读 FPGA 寄存器时好时坏 | 带调试 ILA 的 bit 时序更差；且会和 app 抢总线。以生产 bit 为准，或从 app 日志读 |
+
+## 写笔记
+
+这个工程的约定是 `my_exciter\Notes\PRxx.md` 记**需求**，`Taskxx.md` 记**方案与实现总结**
+（含验证数据和踩过的坑）。做完一轮有结论的工作，补一份 `Taskxx.md` —— 后来的人（和后来的你）
+靠它避免重复踩坑。`references/history.md` 是已有笔记的索引。
diff --git a/smod_atsc_dev/references/fpga.md b/smod_atsc_dev/references/fpga.md
new file mode 100644
index 0000000..4a091bd
--- /dev/null
+++ b/smod_atsc_dev/references/fpga.md
@@ -0,0 +1,81 @@
+# FPGA 侧（my_exciter / RU_golden）
+
+## 工程
+
+| | my_exciter | RU_golden |
+|---|---|---|
+| xpr | `E:\Xilinx_design\smod\SMOD_ATSC\my_exciter\my_exciter.xpr` | `...\RU_golden\RU_golden.xpr` |
+| top | `smod_top`（`my_exciter.srcs/sources_1/top/smod_top.v`，~2300 行） | `RU_golden_top` |
+| 作用 | ATSC 激励器全功能 | 打底镜像：网口 + quickboot 自烧，用来兜底和支撑远程升级 |
+| 器件 | xc7a200tfbg676-2 | 同左 |
+
+`my_exciter` 目录本身是 **git 仓库根**（`imx6/` 也在里面）。RU_golden 的 xpr 引用了
+my_exciter 里的少数源文件（`ts_lock_top.vhd`、`sdram_tsoip_top.vhd`），改动时留意。
+
+## RTL 目录（`my_exciter.srcs/sources_1/`）
+
+| 目录 | 内容 |
+|---|---|
+| `top/` | `smod_top.v` 顶层 + 时钟/AGC/DA 前处理/ORx 抽取等杂项 |
+| `atsc_mod_xip/` | 8VSB 调制核（加密 IP） |
+| `dpd/` | 数字预失真 |
+| `txagc/` | 发射 AGC |
+| `asi_pre/`, `dvb_asi/`, `ts/` | ASI / TS 输入与包同步（`src_preproc`、`spi_lock_det`） |
+| `tsoip/`, `tsoip_atsc/`, `udp/`, `sdram/` | TS over IP 收发与缓存 |
+| `tuner/` | Sony CXD2878 tuner 的 TS 接口（`TUN_D[7:0]/TUN_VLD/TUN_CLK`） |
+| `quickboot/` | `SpiFlashProgrammer.vhd` + `SpiSerDes.vhd` + `spi_slave32.v` —— 远程升级的 FPGA 侧 |
+| `i2c/` | imx6↔FPGA 的寄存器桥 |
+| `gps/`, `uart/`, `misc/`, `bd_cfg/`, `coef/`, `ip/` | 其余 |
+
+约束在 `constrs_1/new/`：`smod_ioplanning.xdc`（SMOD 板的引脚，**这块板用这个**）、
+`project_ioplanning.xdc`（旧板，别混用）、`timing.xdc`、`timing_cleanup_120m.xdc`、
+`bitstream.xdc`、`debug.xdc`（当前为空，ILA 用 IP 例化而非 XDC 标记）。
+
+## 用 vivado-mcp 干活
+
+典型顺序（工具名即 `mcp__vivado-mcp__*`）：
+
+```
+start_session(project=<xpr>)     # 打开工程
+  → verilog_compile_check        # 改完 RTL 先做语法检查，比跑综合快得多
+  → run_synthesis
+  → get_critical_warnings        # 综合后立刻看，别等到实现完
+  → run_implementation
+  → generate_bitstream
+  → get_timing_report / get_utilization_report
+  → power_off → USB_Blast_off → power_on → (等 sys init over) → USB_Blast_on
+  → program_device               # 下 bit（掉电即失）
+```
+
+其它常用：`run_tcl` / `safe_tcl`（任意 Tcl）、`get_run_progress`（长任务轮询）、
+`get_ip_status` / `inspect_ip_params`、`xdc_lint` / `xdc_auto_fix`、`get_io_report`。
+
+综合+实现在这个工程上要 **20~40 分钟**。改一行就重跑之前，先用 `verilog_compile_check`。
+
+## ILA
+
+`smod_top.v` 里 ILA 是**手工例化的 IP**（不是 `mark_debug`）。历史上用过 `ila_eth2`、
+`ila_2`、`ila_3`、`ila_adrv9009_fb`；当前活跃的是 `u_ila_adrv9009_rx`（复用
+`ila_adrv9009_fb` IP，探针宽度 16,16,1,8,1,1,1,8,8,8,8,5）。
+
+要换探针就直接改那个例化的 probe 连接。**探针宽度必须和 IP 里配的一致**，否则综合报错。
+需要新宽度就改 IP 或换一个 ILA IP。
+
+抓波形要 `.ltx`（`impl_1/smod_top.ltx`）。归档时 `.bit` / `.ltx` / `.mcs` 三件套用**同一个 tag**
+放进 `bitstream/`，否则下次想复现就对不上号了。ILA 导出的 CSV 交给 `/fft-analyse` 做频谱。
+
+**固化前把调试 ILA 去掉**，重新实现出干净的生产 bit：ILA 吃资源、恶化时序，还会和 imx6
+抢寄存器总线。
+
+## 时序
+
+`get_timing_report` 上 WNS 常年在 **-9 ~ -13 ns**，路径落在 JESD204 TX 的
+`init_seq_data`（`jesd204_phy_1lane_38j82_refclk` group）。这是基础设计里**既有的
+初始化序列 false path**，与 ATSC 数据通路无关 —— JESD 链路实测能起来，RF/调制/tuner 都正常。
+
+不要为了这个数字去改设计。真正要盯的是你改动那部分的时钟域和 CDC。
+
+## 出 mcs
+
+走 `/fpga_update` skill。一句话：远程升级要的是 `_update.mcs`，不是 `write_cfgmem` 的直接产物；
+而且它走 TCP 1234，不走网页。
diff --git a/smod_atsc_dev/references/history.md b/smod_atsc_dev/references/history.md
new file mode 100644
index 0000000..9fc2b17
--- /dev/null
+++ b/smod_atsc_dev/references/history.md
@@ -0,0 +1,35 @@
+# 历史笔记索引（`my_exciter\Notes\`）
+
+约定：**`PRxx.md` = 需求**（用户提出的要做什么），**`Taskxx.md` = 方案 / 实现总结**
+（怎么做的、踩了什么坑、验证结论）。`discussNN-*.md` 是过程讨论。
+遇到相关主题**先读对应的 Task 文件**——里面的结论往往能省掉一整轮硬件迭代。
+
+| 主题 | 需求 | 方案 / 总结 | 结论要点 |
+|---|---|---|---|
+| 从 EMV7.1 移植到 SMOD | `PR00.md`、`PR01.md` | `Task00.md`、`Task01.md` | 引脚重排到 `smod_ioplanning.xdc`；IO bank / 自由引脚勘查脚本也在 Notes 里（`task01_*.tcl`） |
+| ADRV9009 JESD + SPI 寄存器核对 | `PR02.md` | `Task02.md` | JESD204B TX/RX lock status 寄存器位置 |
+| TSoIP（网口 2 收 IP 流取 TS） | `PR03.md` | — | 根因见 memory `tsoip-root-cause-0x50` |
+| Tuner（473 MHz RF → TS → 调制核） | `PR04.md` | `Task04.md` | demod 是 **Sony CXD2878 + FREIA**，不是 MN88436；移植 EMV9.3 的 `sonytuner/`；ILA 实测 19.41 Mbps、`ts_lock_a=1`、0x47 间隔 188 |
+| ATSC 信号质量分析（RX1 路） | `PR06.md` | `Task06.md` | 用 RX1 接收自发信号，算带肩 + 8VSB 星座 + SNR，网页 `sys_spectrum.html` 显示 |
+| 改用 ORx→DPD 40 MHz 反馈 | `PR06-1.md` | `Task06-1.md` | 反馈通道换成 ORx |
+| ORx/RX1 双通道可切换 + DPD 内部 AGC | `PR06-2.md` | `Task06-2.md` | 共用滤波下采样；大量频谱截图见 `PR06-2_*.png`，索引在 `PR06-2_频谱截图索引.md` |
+| 本 skill 的由来 | `PR07.md` | （本 skill） | 要求：SMOD ATSC 开发总入口，联调走 `/fpga-imx6-flow` |
+
+## 其它散落资源
+
+- `Notes/*.png` —— 历次频谱 / 星座 / 波形截图，文件名带 iter 号，配合 Task 文件看
+- `Notes/create_ila_adrv9009_fb.tcl` —— 建 ILA IP 的 Tcl
+- `Notes/task01_*.tcl` —— IO / DSP / 未约束端口的勘查脚本
+- `Notes/workflow_state.db`、`workflow_threads.json` —— fpga-imx6-flow 的 LangGraph 状态
+- `RU_golden/skills/*.py` —— 早期 RMII/GMII 抓包分析脚本（网口调试用）
+- `bitstream/ila_*.csv`、`orx_*.csv` —— 历次 ILA 导出，交给 `/fft-analyse` 复现频谱
+
+## 已固化的经验（别重新踩）
+
+- 上电时序铁律；上电后等 `sys init over` 再连 JTAG，否则 TX 死。
+- WNS -9~-13 ns 是 JESD204 `init_seq_data` 的既有 false path，与功能无关。
+- imx6 dropbear 老：paramiko 2.12 + `cat >` 传文件（无 SFTP）。
+- 设备上 `imx_atsc_app.rx1` 和 `M=2/lane 0x04` 变体会打断 TX；git `HEAD` 的版本是好的。
+- 远程升级走 TCP 1234（不是网页！），只能传 `_update.mcs`，传错文件会导致冷启动后 FPGA 不跑。
+- 远程升级"卡在擦除"通常是 programmer 已报错（rstate=0x7），imx6 轮询不看 ERROR 位。见 `/fpga_update`。
+- 带调试 ILA 的 bit 下 `i2cget` 读 FPGA 寄存器不可靠。
diff --git a/smod_atsc_dev/references/imx6.md b/smod_atsc_dev/references/imx6.md
new file mode 100644
index 0000000..0e37b6f
--- /dev/null
+++ b/smod_atsc_dev/references/imx6.md
@@ -0,0 +1,84 @@
+# imx6 侧（atsc_app）
+
+## 源码
+
+`E:\Xilinx_design\smod\SMOD_ATSC\my_exciter\imx6\src\atsc_app`（Makefile 在这一层）：
+
+| 目录 | 内容 |
+|---|---|
+| `main.c` | 起线程：`fpga_upgrade_task`、`tcp_task`、`web`、`lcd`、`gps`… |
+| `bsp/` | 板级：`i2c_drv.c`、`fpga.c`（FPGA 寄存器读写）、`eeprom.c`（配置）、`dac38j82.c`、`dpd_proc.c`、`isys_config.c` |
+| `talise/`, `adrv9009/`, `adi_hal/` | ADI ADRV9009 驱动 |
+| `sonytuner/` | Sony CXD2878 + FREIA(CXD6821) 解调驱动（从 EMV9.3 移植，见 PR04） |
+| `task/` | `upgrade.c`（软件包升级）、`tcp_task.c`、`iweb_task.c`、`tsoip.c`、`iq_rx.c` |
+| `inc/` | 头文件；`fpga.h` 里 `FPGA_UPGRADE_I2C_REG_ADDRES = 0xE0` |
+
+同级还有 `imx6/src/web/`（cgi + www 页面，含 `upload.cgi` 远程升级入口）。
+
+## 交叉编译
+
+工具链在 VM `bowei@192.168.223.128`（密码 `lr7402`）：
+
+```
+/opt/fsl-imx-x11/4.1.15-2.0.0/environment-setup-cortexa7hf-neon-poky-linux-gnueabi
+```
+
+流程（`scripts/imx6_app.py build` 已封装）：
+
+1. Windows 侧从 git 打包源码 —— **用 `HEAD` 而不是工作区**，这样构建的东西可追溯：
+   `git -C my_exciter archive --format=tar -o app.tar HEAD:imx6/src/atsc_app`（约 13 MB）
+2. SFTP 上传到 VM，解到 **`~/atsc_head_build`**（一个干净目录；不要碰 `~/atsc_tsoip_build/` 之类的旧目录）
+3. `bash -lc 'source <envsetup>; cd ~/atsc_head_build && make'` → 产出 `imx_atsc_app`（~2.9 MB ELF）
+4. SFTP 取回 Windows
+
+VM 的 sshd 是现代版本，paramiko 5 / OpenSSH 都能连。若报
+`kex_exchange_identification: Connection closed`，是 VM 的 sshd 挂了，请用户 `sudo systemctl restart sshd`。
+
+## 部署到设备
+
+设备 `192.168.1.143`，`root` / `anywavecom`。**它的 SSH 是老 dropbear**，有两个坑：
+
+1. **算法太老**：paramiko 5 会报 `unknown cipher` / `no acceptable host key`。
+   要用 **paramiko 2.12**（有 TripleDES / aes-cbc / ssh-rsa）：
+   `python -m pip install --target <dir> paramiko==2.12.0`，然后 `PYTHONPATH=<dir>` 跑。
+   命令行 ssh 则加 `-o HostKeyAlgorithms=+ssh-rsa`。
+2. **没有 SFTP 子系统**：`open_sftp()` 直接 `EOF during negotiation`。传文件走 exec channel：
+
+   ```python
+   ch = client.get_transport().open_session()
+   ch.exec_command("cat > /aw/anywave/app/imx_atsc_app_head")
+   ch.sendall(open(local_bin, "rb").read())
+   ch.shutdown_write(); assert ch.recv_exit_status() == 0
+   ```
+
+   （另一条老路是 SSH 反向端口转发 + 设备侧 `wget`，见 Task04；现在没必要了。）
+
+部署顺序 —— **先落一个新名字，验证 md5，再覆盖 active**，这样传输失败不会毁掉能跑的 app：
+
+```
+cat > /aw/anywave/app/imx_atsc_app_head
+chmod 755 imx_atsc_app_head
+md5sum 对比
+cp -a imx_atsc_app_head imx_atsc_app     # 覆盖 active
+```
+
+`scripts/imx6_app.py deploy` 就是这个顺序。已知好用的备份是 `imx_atsc_app.bak_pre_rx1`。
+
+改完 app 要**冷启动**（断电重启，遵守上电时序）让 `rc.local` 拉起新进程。
+
+## 观察设备
+
+- 日志：`/var/module_log/imx_atsc_app.log`。健康标志：`TALISE_radioOn`、`fb init state ok`、
+  `fpga[0xBB]=0x03`（sync=1 / tvalid=1）。`imx6_app.py log` 直接拉。
+- 配置：`/aw/anywave/cfg/exc_cfg.conf`，`[BLK_SYS_PARA] reg50` 选输入源
+  —— `0`=Tuner，`1`=TS，`2`=EAS，`3`=TSoIP。
+- 网页：`http://192.168.1.143`（`sys_rf.html`、`sys_ts.html`、`sys_set.html`、
+  `status_info.html`、`log_output.html`、以及远程升级的 `upload.cgi`）。
+- 串口控制台（COM 号见 `~/fpga_imx6_flow.json` 的 `ports.imx6`，逐机而异、不入库）能用，但会自动登出、比较飘。**优先走网络 SSH**读日志和跑命令。
+- 直接读 FPGA 寄存器可以 `i2cget`，但它会和 app 抢总线；带调试 ILA 的 bit 下更不可靠。
+  以 app 日志为准，或者临时停掉 app。
+
+## 已知不能用的构建
+
+设备上的 `imx_atsc_app.rx1`、以及任何 `M=2 / lane 0x04` 的变体**会打断 TX**。
+git `HEAD` 的版本是 RX1-on-framerA（`M=4` / lane `0x03` / RX1RX2）+ TX 正常的已知好版本。
diff --git a/smod_atsc_dev/references/upgrade.md b/smod_atsc_dev/references/upgrade.md
new file mode 100644
index 0000000..91ce7df
--- /dev/null
+++ b/smod_atsc_dev/references/upgrade.md
@@ -0,0 +1,24 @@
+# 远程升级 / Flash 布局
+
+**这块内容的唯一真源是 `fpga_update` skill。** 要生成 `_update.mcs` 或者把它烧进板子，
+直接用那个 skill，别在这里另起炉灶。
+
+- `~/.claude/skills/fpga_update/SKILL.md` —— 流程
+- `~/.claude/skills/fpga_update/references/quickboot.md` —— flash 布局、切换字兜底、`-imagesize 64` 的由来
+- `~/.claude/skills/fpga_update/references/protocol.md` —— TCP 1234 协议、以及 DPD 污染升级数据流的缺陷
+
+三件容易记反的事：
+
+1. **FPGA 的远程升级不走网页。** 网页的 Remote Update 只认 imx6 的 `*.tar.gz` 固件包；
+   `.mcs` 传上去会被 `check_package.sh` 直接删掉。FPGA 走 TCP **1234** 的私有协议
+   （原厂工具 `tools/Xfpga_ru.exe`，替代品 `fpga_update/scripts/fpga_ru.py`）。
+2. **升级前必须关掉 DPD**，否则 DPD 的 SPI1 流量会混进镜像，CRC 校验失败。
+3. **`_update.mcs` 不能 JTAG 烧写**，`_initial.mcs` 也不能远程升级。首次装机 / 救砖走
+   `/fpga_flash`（flash 型号 `w25q128bv`，烧 `RU_golden/bitstream/golden_design_initial.mcs`）。
+
+## 调试期的第三条路
+
+`program_device` 直接把 `.bit` 下到 FPGA，掉电即失。迭代验证用它最快 —— 不碰 flash，
+配 `.ltx` 就能抓 ILA。功能确认后再走远程升级固化。
+
+固化用的 bit 要**去掉调试 ILA 重新实现**：ILA 吃资源、恶化时序，还会和 imx6 抢寄存器总线。
diff --git a/smod_atsc_dev/scripts/imx6_app.py b/smod_atsc_dev/scripts/imx6_app.py
new file mode 100644
index 0000000..81ba84a
--- /dev/null
+++ b/smod_atsc_dev/scripts/imx6_app.py
@@ -0,0 +1,263 @@
+#!/usr/bin/env python3
+"""Cross-compile the SMOD imx6 atsc_app on the Ubuntu VM and deploy it to the board.
+
+    build         package app source -> VM (~/atsc_head_build) -> make -> pull imx_atsc_app back
+    deploy        push a local imx_atsc_app to 192.168.1.143, md5-verify, then swap it in
+    build-deploy  both
+    log           tail /var/module_log/imx_atsc_app.log from the device
+
+Two device quirks drive the design here:
+
+  * The imx6 runs an old dropbear. Modern paramiko can't negotiate with it
+    ("unknown cipher" / "no acceptable host key"), so this script vendors
+    paramiko 2.12 into ~/.cache/smod_atsc_dev and re-execs itself under it.
+  * That dropbear has no SFTP subsystem, so files go over an exec channel with
+    ``cat > file`` rather than sftp.put().
+
+Deploy writes imx_atsc_app_head first and only overwrites the active imx_atsc_app once
+the md5 matches -- a half-transferred binary should never become the thing rc.local runs.
+
+After deploying you still need a cold boot (power_off -> USB_Blast_off -> power_on ->
+USB_Blast_on) for rc.local to pick up the new app. This script deliberately does not
+touch the relays; that sequencing lives in the vivado-mcp tools.
+"""
+
+import argparse
+import hashlib
+import os
+import subprocess
+import sys
+import tarfile
+import tempfile
+
+REPO = r"E:\Xilinx_design\smod\SMOD_ATSC\my_exciter"
+APP_SUBDIR = "imx6/src/atsc_app"
+
+VM_HOST, VM_USER, VM_PASS = "192.168.223.128", "bowei", "lr7402"
+VM_BUILD_DIR = "atsc_head_build"  # relative to the VM home dir
+TOOLCHAIN = "/opt/fsl-imx-x11/4.1.15-2.0.0/environment-setup-cortexa7hf-neon-poky-linux-gnueabi"
+
+IMX_HOST, IMX_USER, IMX_PASS = "192.168.1.143", "root", "anywavecom"
+IMX_APP_DIR = "/aw/anywave/app"
+IMX_STAGING = f"{IMX_APP_DIR}/imx_atsc_app_head"
+IMX_ACTIVE = f"{IMX_APP_DIR}/imx_atsc_app"
+IMX_LOG = "/var/module_log/imx_atsc_app.log"
+
+BINARY = "imx_atsc_app"
+VENDOR_MARKER = "SMOD_PARAMIKO_VENDORED"
+# paramiko 2.12 reaches for algorithms.TripleDES, which cryptography >= 43 removed.
+VENDOR_PKGS = ["paramiko==2.12.0", "cryptography<43"]
+
+
+# --------------------------------------------------------------------------- vendoring
+
+def ensure_legacy_paramiko():
+    """Re-exec under a vendored paramiko 2.12 unless we're already running under it."""
+    if os.environ.get(VENDOR_MARKER):
+        return
+    cache = os.path.join(os.path.expanduser("~"), ".cache", "smod_atsc_dev", "paramiko212")
+    if not os.path.isdir(os.path.join(cache, "paramiko")):
+        print(f"[setup] installing legacy paramiko into {cache} (one time)", flush=True)
+        os.makedirs(cache, exist_ok=True)
+        # --no-warn-conflicts: pip diffs --target installs against the ambient env and
+        # reports conflicts that don't apply here, which reads like a failure.
+        subprocess.run([sys.executable, "-m", "pip", "install", "--quiet", "--no-warn-conflicts",
+                        "--target", cache] + VENDOR_PKGS, check=True)
+    env = dict(os.environ, **{VENDOR_MARKER: "1"})
+    env["PYTHONPATH"] = cache + os.pathsep + env.get("PYTHONPATH", "")
+    env["PYTHONWARNINGS"] = "ignore"
+    sys.exit(subprocess.run([sys.executable, os.path.abspath(__file__)] + sys.argv[1:], env=env).returncode)
+
+
+# ------------------------------------------------------------------------------- ssh
+
+def connect(host, user, password):
+    import paramiko
+    cli = paramiko.SSHClient()
+    cli.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+    cli.connect(host, username=user, password=password, timeout=20, look_for_keys=False, allow_agent=False)
+    return cli
+
+
+def run(cli, cmd, check=True, quiet=False):
+    if not quiet:
+        print(f"  [ssh] {cmd}")
+    _, out, err = cli.exec_command(cmd, timeout=1800)
+    rc = out.channel.recv_exit_status()
+    stdout, stderr = out.read().decode(errors="replace"), err.read().decode(errors="replace")
+    if check and rc != 0:
+        sys.exit(f"[FAIL] rc={rc}\n{stdout}\n{stderr}")
+    return rc, stdout, stderr
+
+
+def md5_of(path):
+    h = hashlib.md5()
+    with open(path, "rb") as fh:
+        for chunk in iter(lambda: fh.read(1 << 20), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+# ----------------------------------------------------------------------------- build
+
+def package_source(repo, source_mode):
+    tar_path = os.path.join(tempfile.mkdtemp(prefix="smod_app_"), "app.tar")
+    if source_mode == "git":
+        subprocess.run(["git", "-C", repo, "archive", "--format=tar", "-o", tar_path, f"HEAD:{APP_SUBDIR}"],
+                       check=True)
+        rev = subprocess.run(["git", "-C", repo, "rev-parse", "--short", "HEAD"],
+                             capture_output=True, text=True, check=True).stdout.strip()
+        print(f"[build] packaged {APP_SUBDIR} from git HEAD ({rev})")
+    else:
+        src = os.path.join(repo, *APP_SUBDIR.split("/"))
+        skip = {".o", ".bak", ".zip", ".tar"}
+        with tarfile.open(tar_path, "w") as tf:
+            for root, _, files in os.walk(src):
+                for name in files:
+                    if os.path.splitext(name)[1] in skip or name == BINARY:
+                        continue
+                    full = os.path.join(root, name)
+                    tf.add(full, arcname=os.path.relpath(full, src))
+        print(f"[build] packaged {APP_SUBDIR} from the working tree (stale .o excluded)")
+    return tar_path
+
+
+def cmd_build(args):
+    tar_path = package_source(args.repo, args.source)
+    out_path = os.path.abspath(args.out)
+
+    vm = connect(VM_HOST, VM_USER, VM_PASS)
+    try:
+        sftp = vm.open_sftp()
+        print(f"[build] uploading {os.path.getsize(tar_path)/1e6:.1f} MB to VM")
+        sftp.put(tar_path, f"/home/{VM_USER}/app_src.tar")
+
+        run(vm, f"rm -rf ~/{VM_BUILD_DIR} && mkdir -p ~/{VM_BUILD_DIR} && "
+                f"tar xf ~/app_src.tar -C ~/{VM_BUILD_DIR}")
+
+        print("[build] cross-compiling (this takes a few minutes)")
+        rc, out, err = run(vm, f"bash -lc 'source {TOOLCHAIN} && cd ~/{VM_BUILD_DIR} && make {args.make_args}'",
+                           check=False)
+        tail = "\n".join((out + err).splitlines()[-40:])
+        if rc != 0:
+            sys.exit(f"[FAIL] make exited {rc}. Last output:\n{tail}")
+        print(tail)
+
+        remote_bin = f"/home/{VM_USER}/{VM_BUILD_DIR}/{BINARY}"
+        rc, _, _ = run(vm, f"test -f {remote_bin}", check=False)
+        if rc != 0:
+            sys.exit(f"[FAIL] make succeeded but {remote_bin} is missing")
+
+        os.makedirs(os.path.dirname(out_path) or ".", exist_ok=True)
+        sftp.get(remote_bin, out_path)
+    finally:
+        vm.close()
+
+    print(f"\n[OK] {out_path}  ({os.path.getsize(out_path)/1e6:.2f} MB, md5 {md5_of(out_path)})")
+    return out_path
+
+
+# ---------------------------------------------------------------------------- deploy
+
+def push_file(cli, local, remote):
+    """Send a file over an exec channel: the device's dropbear has no SFTP subsystem."""
+    stdin, stdout, _ = cli.exec_command(f"cat > {remote}", timeout=600)
+    with open(local, "rb") as fh:
+        for chunk in iter(lambda: fh.read(1 << 16), b""):
+            stdin.write(chunk)
+    stdin.channel.shutdown_write()
+    rc = stdout.channel.recv_exit_status()
+    if rc != 0:
+        sys.exit(f"[FAIL] `cat > {remote}` exited {rc}")
+
+
+def cmd_deploy(args):
+    local = os.path.abspath(args.binary)
+    if not os.path.exists(local):
+        sys.exit(f"[FAIL] not found: {local}")
+    want = md5_of(local)
+    print(f"[deploy] {local} (md5 {want})")
+
+    imx = connect(IMX_HOST, IMX_USER, IMX_PASS)
+    try:
+        print(f"[deploy] staging -> {IMX_STAGING}")
+        push_file(imx, local, IMX_STAGING)
+        run(imx, f"chmod 755 {IMX_STAGING}")
+
+        _, out, _ = run(imx, f"md5sum {IMX_STAGING}")
+        got = out.split()[0]
+        if got != want:
+            sys.exit(f"[FAIL] md5 mismatch after transfer: device {got} != local {want}. "
+                     f"Active app untouched.")
+        print("[deploy] md5 verified")
+
+        run(imx, f"cp -a {IMX_STAGING} {IMX_ACTIVE}")
+        _, out, _ = run(imx, f"md5sum {IMX_ACTIVE}")
+        if out.split()[0] != want:
+            sys.exit("[FAIL] active binary md5 mismatch after swap")
+    finally:
+        imx.close()
+
+    print(f"\n[OK] deployed to {IMX_ACTIVE}")
+    print("Now cold-boot the board so rc.local starts the new app:")
+    print("  vivado-mcp: power_off -> USB_Blast_off -> power_on -> (wait 'sys init over') -> USB_Blast_on")
+    print(f"Then check the log:  python {os.path.basename(__file__)} log")
+
+
+# ------------------------------------------------------------------------------- log
+
+def cmd_log(args):
+    imx = connect(IMX_HOST, IMX_USER, IMX_PASS)
+    try:
+        cmd = f"tail -n {args.lines} {IMX_LOG}"
+        if args.grep:
+            cmd += f" | grep -E '{args.grep}'"
+        _, out, _ = run(imx, cmd, check=False, quiet=True)
+        print(out)
+    finally:
+        imx.close()
+    print("\nHealthy markers: TALISE_radioOn / 'fb init state ok' / fpga[0xBB]=0x03")
+
+
+# ------------------------------------------------------------------------------ main
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    sub = ap.add_subparsers(dest="cmd", required=True)
+
+    def add_build_args(p):
+        p.add_argument("--repo", default=REPO)
+        p.add_argument("--source", choices=["git", "worktree"], default="git",
+                       help="git (default, reproducible) packages HEAD; worktree packages uncommitted edits")
+        p.add_argument("--make-args", default="", help="extra args passed to make on the VM")
+        p.add_argument("--out", default=os.path.join(tempfile.gettempdir(), BINARY))
+
+    b = sub.add_parser("build", help="cross-compile on the VM and pull the binary back")
+    add_build_args(b)
+
+    d = sub.add_parser("deploy", help="push a local binary to the imx6 and swap it in")
+    d.add_argument("binary", nargs="?", default=os.path.join(tempfile.gettempdir(), BINARY))
+
+    bd = sub.add_parser("build-deploy", help="build then deploy")
+    add_build_args(bd)
+
+    lg = sub.add_parser("log", help="tail the app log on the device")
+    lg.add_argument("-n", "--lines", type=int, default=80)
+    lg.add_argument("-g", "--grep", help="extended regex to filter with")
+
+    args = ap.parse_args()
+    ensure_legacy_paramiko()  # after parsing, so --help never triggers a pip install
+
+    if args.cmd == "build":
+        cmd_build(args)
+    elif args.cmd == "deploy":
+        cmd_deploy(args)
+    elif args.cmd == "build-deploy":
+        args.binary = cmd_build(args)
+        cmd_deploy(args)
+    elif args.cmd == "log":
+        cmd_log(args)
+
+
+if __name__ == "__main__":
+    main()