llama-router: async polling for process termination after SIGKILL

Poll process_running() at ROUTER_POLL_INTERVAL_MS until exit confirmed.
Ensures VRAM freed before hot-swap. Update docs to use constant names

Author: ServeurpersoCom

tools/router/ARCHITECTURE.md

@@ -85,15 +85,15 @@ This aligns with the project philosophy: **everything configurable at runtime, z
 3. If not running, or if a conflicting group is active:
    - Terminate conflicting backends
    - Spawn new llama-server with assigned port
-   - Poll `/health` until ready (10s timeout)
+   - Poll `/health` until ready (`ROUTER_BACKEND_READY_TIMEOUT_MS` timeout)
 4. Forward request to backend, streaming response back to client
 5. Backend remains running for subsequent requests
 
 ### Process Lifecycle
 
 - **Spawn**: `fork()`/`CreateProcess()` with stdout/stderr capture
-- **Health polling**: 200ms intervals, 10s timeout
-- **Graceful shutdown**: SIGTERM → 1s wait → SIGKILL
+- **Health polling**: `ROUTER_BACKEND_HEALTH_POLL_MS` intervals, `ROUTER_BACKEND_READY_TIMEOUT_MS` timeout
+- **Graceful shutdown**: SIGTERM → wait `ROUTER_PROCESS_SHUTDOWN_TIMEOUT_MS` → SIGKILL → poll every `ROUTER_POLL_INTERVAL_MS` until exit
 - **Cleanup**: File descriptors closed, waitpid() called
 
 ---

tools/router/router-process.cpp

@@ -132,7 +132,9 @@ void terminate_process(ProcessHandle & handle) {
     if (!wait_for_process_exit(handle, ROUTER_PROCESS_SHUTDOWN_TIMEOUT_MS)) {
         LOG_ERR("Process pid=%d did not terminate, sending SIGKILL\n", static_cast<int>(handle.pid));
         kill(handle.pid, SIGKILL);
-        wait_for_process_exit(handle, 1000);
+        while (process_running(handle)) {
+            std::this_thread::sleep_for(std::chrono::milliseconds(ROUTER_POLL_INTERVAL_MS));
+        }
     }
     close_process(handle);
 #endif