polish: log atexit registration failure; trim comments

- Log a warning to stderr if std::atexit registration fails so the loss of
  the last-resort shm cleanup is visible; primary cleanup path is unaffected.
- Tighten inline comments to explain only intent/constraints, not mechanics.

Author: Vinya567

src/shm_manager.cc

@@ -61,7 +61,11 @@ RegisterParentShmRegion(const std::string& shm_region_name)
     parent_shm_regions.insert(shm_region_name);
   }
   if (!parent_shm_atexit_registered.exchange(true)) {
-    std::atexit(CleanupParentShmRegions);
+    if (std::atexit(CleanupParentShmRegions) != 0) {
+      std::cerr << "python_backend: failed to register atexit shm cleanup "
+                   "handler; relying on TerminateStub for cleanup"
+                << std::endl;
+    }
   }
 }
 

src/shm_manager.h

@@ -194,7 +194,7 @@ class SharedMemoryManager {
 
   void SetDeleteRegion(bool delete_region);
 
-  // Remove the parent-owned shared memory region from the filesystem.
+  // Idempotent: removes the parent-owned shm region and clears delete_region_.
   void RemoveShmRegion();
 
   std::unique_ptr<CUDAMemoryPoolManager>& GetCUDAMemoryPoolManager()

src/stub_launcher.cc

@@ -805,15 +805,12 @@ void
 StubLauncher::TerminateStub()
 {
   if (is_initialized_) {
-    // Enforce stub_timeout_seconds_ as a single total budget across the
-    // finalize wait and the subsequent process-exit wait; without this the
-    // healthy teardown path could take up to 2 * stub_timeout_seconds_.
+    // Single teardown budget: finalize + wait share stub_timeout_seconds_.
     bool force_kill = false;
     int64_t remaining_seconds = stub_timeout_seconds_;
     if (is_healthy_) {
       const int64_t total_timeout_ms = stub_timeout_seconds_ * 1000;
-      // MessageQueue::Pop takes `int const&`; clamp to INT_MAX so a large
-      // stub_timeout_seconds_ cannot silently narrow to a negative value.
+      // Clamp to INT_MAX; MessageQueue::Pop takes int.
       const int pop_timeout_ms = static_cast<int>(std::min<int64_t>(
           total_timeout_ms,
           static_cast<int64_t>(std::numeric_limits<int>::max())));
@@ -983,8 +980,7 @@ StubLauncher::WaitForStubProcessWithTimeout(int64_t timeout_seconds)
     sleep(1);
   }
 
-  // The process may have exited during the final sleep window; re-check once
-  // more so we don't force-kill an already-dead process.
+  // Stub may have exited during the last sleep(1); recheck before killing.
   {
     int status;
     pid_t ret = waitpid(stub_pid_, &status, WNOHANG);