docs(sdk): 补 1.3.0 的新增能力与发版口径；main 路牌推到 1.4.0

用户指出：两个仓库的文档都没跟着更新。本仓的缺口：

1. `README.md` / `README_EN.md` 头部仍写「当前：SDK 1.2.0（需内核 1.2.0+）」，
   版本表停在 1.2.0 —— 1.3.0 的两项新增能力（`InjectOptions.Priority`/`PriorityL1`–`PriorityL4`、
   `UnregisterOutputChannel` 一族）在 README 里**一个字都没有**，而它们正是这一版
   插件作者最需要知道的东西。
2. 发版口径没写清"**SDK 仓不发 patch tag**"：核心 1.3.x 的后续 patch 不伴随 SDK 发版，
   patch 位恒为 `.0`（§七.1）。这条以前只在规范里，README 没提，结果我自己在
   2026-09-13 误发了 `v1.3.1`（已撤回）；`v1.2.1` 是同一类历史遗留。
3. `meta/meta.go` 的版本语义注释还停在「现为 1.2.0：核心的 1.2.x 线正在发布中」，
   与事实相反。1.3.0 既已随核心正式 tag 定版，该号归发布线所有 ⇒ main 推进到 **1.4.0**。

补写内容：1.3.0 能力小节（四级中断优先级、动态输出通道、通道名约束与那起
`device/<id>` 生产事故）、发版口径两段、版本表补 1.3.0 行、meta 注释与路牌。

README_EN.md

@@ -4,7 +4,7 @@ Plugin development SDK for building intelligent plugins that interact with the H
 
 ## Version and Compatibility
 
-Current: **SDK 1.2.0** (requires kernel **1.2.0+**).
+Current: **SDK 1.3.0** (requires kernel **1.3.0+**).
 
 **The version tracks the kernel's minor version, with the patch position pinned at `.0`**:
 
@@ -12,13 +12,33 @@ Current: **SDK 1.2.0** (requires kernel **1.2.0+**).
 |---|---|
 | 1.0.0 / 1.0.1 / … / 1.0.4 | 1.0.0 |
 | 1.1.0 / 1.1.1 / … / 1.1.N | **1.1.0** |
-| 1.2.0 onward | 1.2.0 |
+| 1.2.0 / 1.2.1 / … / 1.2.N | **1.2.0** |
+| 1.3.0 onward | **1.3.0** |
 
 The kernel's patch position is reserved for bugfixes and vulnerability fixes, which never touch the
 public interface, so the SDK version has no reason to move with it — otherwise you would either be
 forced to chase releases or suspect your version is stale, when not one character of the interface
 has changed.
 
+The SDK repository therefore publishes **exactly once per minor version** (`vX.Y.0`); kernel patches
+such as `v1.3.1` do not trigger an SDK release. (A `v1.3.1` tag was mistakenly cut on 2026-09-13 and
+has been withdrawn — any SDK tag with a non-zero patch position is wrong.)
+
+## New in 1.3.0: Injection Priority and Dynamic Output Channels
+
+- **`InjectOptions.Priority` / `PriorityL1`–`PriorityL4`** — a plugin declares the interrupt level of
+  its own injection; the kernel schedules L1–L4, where **L4 is reserved for the kernel and
+  kernel-level plugins**. The zero value is fully equivalent to the old three-argument call
+  (queued, never preempting), so existing plugins need neither a code change nor a rebuild.
+  Queued input has no level: anything can jump ahead of it.
+- **`UnregisterOutputChannel` / `OutputChannelUnregistrar` / `SetOutputChannelUnregistrar`** —
+  channels that die with their resource (one channel per remote device) can now be unregistered;
+  previously they lingered and the model kept "successfully" sending into a dead channel.
+- **Channel names must be legal and unique.** The name is spliced into the LLM function name
+  `output_send__<name>`, so it may only contain `[A-Za-z0-9_-]`. A real production incident
+  (2026-09-13): `device/<id>` made every LLM request fail with 403. Derive channel names from
+  external IDs — never use the raw ID.
+
 **Upgrading a 1.0.x plugin to 1.1.x: no code changes, no rebuild.** Everything added in 1.1.0 is
 in the "plugin calls, kernel implements" direction, so not calling it means not being affected
 (verified with an old `plugin.bin` built against SDK 0.9.2: it handshakes fine on the new kernel,