feat(health): EventStore latency + llama-server liveness probes

- EventStoreLatencyProbe: times a side-effect-free lastGlobalSequence read,
  wired into HealthMonitor; HealthConfig.eventStoreWarnLatencyMs (BACKLOG A §4)
- LlamaServerHealthProbe: liveness + tokens/sec degradation trend, injectable
  client (HttpLlamaLivenessClient over Ktor); deterministic unit tests
- llama probe left unregistered (TODO) — HttpClient not in monitor scope yet

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt

@@ -23,7 +23,9 @@ data class CorrexConfig(
  * Continuous health watch (observability-spec §4). When [enabled], a background monitor polls
  * the on-disk footprint every [intervalMs] and records a degraded/restored event on the edge.
  * [diskWarnBytes] is the absolute event-log + CAS size that trips a warning; [diskWarnGrowthBytesPerMin]
- * is the growth rate that warns log compaction is becoming non-theoretical. Defaults suit a homelab.
+ * is the growth rate that warns log compaction is becoming non-theoretical. [eventStoreWarnLatencyMs]
+ * is the event-store latency watermark: a single side-effect-free store read slower than this trips a
+ * warning that the append/fsync path is contended or the disk is failing. Defaults suit a homelab.
  */
 @Serializable
 data class HealthConfig(
@@ -31,6 +33,7 @@ data class HealthConfig(
     val intervalMs: Long = 30_000,
     val diskWarnBytes: Long = 5_000_000_000,
     val diskWarnGrowthBytesPerMin: Long = 50_000_000,
+    val eventStoreWarnLatencyMs: Long = 250,
 )
 
 /**