refactor: decomposition WIP (orchestrator/server/execution-plan)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DhFXmKe4WisSSPf9LrmmTg

infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/DefaultModelManager.kt

@@ -154,9 +154,6 @@ class DefaultModelManager(
     private suspend fun waitForHealthy(process: LlamaProcess): Boolean {
         val endTime = Clock.System.now().toEpochMilliseconds() + healthTimeoutMs
         while (Clock.System.now().toEpochMilliseconds() < endTime) {
-            // A crashed process (bad --model, port clash) never gets healthy; polling the full
-            // timeout is pure dead time. Bail as soon as it's gone.
-            if (!process.isAlive) return false
             try {
                 val response = httpClient.get("http://$host:$port/health").body<String>()
                 if (response.contains("\"status\":\"healthy\"") || response.contains("ok")) {
@@ -165,6 +162,10 @@ class DefaultModelManager(
             } catch (_: Exception) {
                 // Continue polling
             }
+            // A crashed process (bad --model, port clash) never gets healthy; polling the full
+            // timeout is pure dead time. Bail as soon as it's gone — but only after probing health,
+            // so a server that's already answering isn't missed on a lost startup/exit race.
+            if (!process.isAlive) return false
             delay(POLL_INTERVAL)
         }
         return false