fix: do not release a claim while its output is still writing

Two windows where the claim was dropped early, both found in review.

A failed obs_output_start() does not mean the output is idle:
obs_output_can_begin_data_capture() refuses while active(output), so calling
start() twice lands in the failure path with the first muxer still writing.
Only release when the output is genuinely not active.

Both recording Destroy handlers deregistered the object before deleting it, and
only the destructor called DeleteOutput(), which can wait up to 20s for the
muxer to drain. For that whole window the file was still being written while its
claim was invisible to a concurrent Start. Stop first, then deregister;
DeleteOutput() is idempotent so the destructor's call becomes a no-op.

The on-disk check covers most of both windows -- the file exists by then -- but
not when overwrite is set, which suppresses exactly that check.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: summeroff

obs-studio-server/source/osn-advanced-recording.cpp

@@ -93,6 +93,11 @@ void osn::IAdvancedRecording::Destroy(void *data, const int64_t id, const std::v
 		PRETTY_ERROR_RETURN(ErrorCode::InvalidReference, "Recording reference is not valid.");
 	}
 
+	// Stop before deregistering. The destructor would do this anyway, but by then the object is out
+	// of the manager, and DeleteOutput() can wait up to 20s for the muxer to drain -- a window where
+	// the file is still being written but its claim is invisible to a concurrent Start.
+	recording->DeleteOutput();
+
 	osn::IAdvancedRecording::Manager::GetInstance().free(recording);
 	delete recording;
 

obs-studio-server/source/osn-output.cpp

@@ -179,8 +179,12 @@ void osn::Output::StartOutput()
 		code = OBS_OUTPUT_ERROR;
 	}
 
-	// libobs emits no "stop" here, so OnStopped() never runs for a failed start.
-	OnOutputStopped();
+	// libobs emits no "stop" here, so OnStopped() never runs for a failed start. But a failed start
+	// does not imply an idle output: obs_output_can_begin_data_capture() refuses while the output is
+	// already active, so starting twice lands here with the first muxer still writing. Releasing then
+	// would unclaim a file that is still being written to.
+	if (!obs_output_active(m_output))
+		OnOutputStopped();
 
 	PushReceivedSignal("stop", code, errorMessage);
 }

obs-studio-server/source/osn-simple-recording.cpp

@@ -88,6 +88,11 @@ void osn::ISimpleRecording::Destroy(void *data, const int64_t id, const std::vec
 		PRETTY_ERROR_RETURN(ErrorCode::InvalidReference, "Recording reference is not valid.");
 	}
 
+	// Stop before deregistering. The destructor would do this anyway, but by then the object is out
+	// of the manager, and DeleteOutput() can wait up to 20s for the muxer to drain -- a window where
+	// the file is still being written but its claim is invisible to a concurrent Start.
+	recording->DeleteOutput();
+
 	osn::ISimpleRecording::Manager::GetInstance().free(recording);
 	delete recording;