feat(metrics): 投递失败与死信可观测指标，修 flushSchd 首次失败原因丢失（ACM2-98）

- PipelineCounters 按 target 累计发送失败；Dispatcher.retryOrDead 记账（OPS-2）
- 新增 deadByTarget 查询与 DeliveryDeadSnapshotProvider（TTL 缓存），死信存量按 target 暴露
- /metrics 注册 send_failures.total{target} 与 dead{target}，登记进 reference.md「指标与健康」
- flushSchd 失败改记真实异常原因，不再退回固定 send-failed

src/main/kotlin/com/gzzn/omms/msgexchange/delivery/Dispatcher.kt

@@ -45,6 +45,7 @@ class Dispatcher(
     private val port: DeliveryPort,
     private val props: PipelineProps,
     private val scheduler: FailureScheduler,
+    private val counters: com.gzzn.omms.msgexchange.infra.metrics.PipelineCounters,
 ) {
     private val log = org.slf4j.LoggerFactory.getLogger(Dispatcher::class.java)
 
@@ -162,7 +163,7 @@ class Dispatcher(
             lastFlush = scheduler.now()
             return
         }
-        val failures = mutableListOf<MsgEvent>()
+        val failures = mutableListOf<Pair<MsgEvent, String>>()
         for (e in batch) {
             try {
                 when (e.eventType) {
@@ -172,15 +173,17 @@ class Dispatcher(
                 // 条件确认：读取时刻的代次（EVENT_ID + STATE_VERSION）被新写入覆盖时不标记，留待下一轮重发。
                 e.eventId?.let { msgEvents.markSentIfVersion(it, e.stateVersion, scheduler.now()) }
             } catch (ex: Exception) {
-                failures.add(e)
+                // 首次失败就带上真实原因；不能退回字面量把根因抹掉
+                failures += e to (ex.message ?: ex.javaClass.simpleName)
             }
         }
-        failures.forEach { retryOrDead(it, it.lastError ?: "send-failed") }
+        failures.forEach { (e, error) -> retryOrDead(e, error) }
         lastFlush = scheduler.now()
     }
 
     /** 一条事件发失败之后怎么走：重试次数加一，到上限就标成 DEAD(EXHAUSTED) 留作死信（次数落库便于追查），否则按退避推到下次再发。 */
     private fun retryOrDead(e: MsgEvent, lastError: String) {
+        counters.sendFailedAdd(e.target)   // OPS-2：发送失败按 target 计数，供 /metrics 告警
         val eventId = e.eventId ?: return
         val attempts = e.attempts + 1
         if (scheduler.exhausted(attempts)) {