apu: guard the XMA loop-end frame test against degenerate loop metadata

The loop-end test runs before UpdateLoopStatus, so the degenerate-loop guard there never reaches it. With loop_count > 0 and loop_start >= loop_end, the clamped loop end equals the start offset of the stream at packet 0, arming the frame output limit and silently truncating 384 samples from its first frame. Cvar audio_xma_loop_end_guard, default on.

thirdparty/rexglue-sdk/src/native/audio/audio_system.cpp

@@ -84,6 +84,19 @@ REXCVAR_DEFINE_BOOL(audio_xma_header_straddle_fix, true, "Audio",
                     "combs on the stereo fold - the long-reported doubled/"
                     "robotic cutscene dialogue. Off restores the legacy "
                     "walk for A/B.");
+REXCVAR_DEFINE_BOOL(audio_xma_loop_end_guard, true, "Audio",
+                    "Apply the degenerate-loop guard (see "
+                    "audio_xma_loop_guard) to the loop-end frame test as "
+                    "well. That test is evaluated before the loop-status "
+                    "update, so the guard there never reaches it: with "
+                    "loop_count > 0 and loop_start >= loop_end (metadata "
+                    "AC6's streamed voices carry), the clamped loop_end "
+                    "equals a voice's starting read offset for the stream "
+                    "that begins at packet 0, arming the frame output limit "
+                    "and silently discarding 384 samples of that stream's "
+                    "first frame - permanently offsetting one context of a "
+                    "multi-stream voice from its siblings. Off restores the "
+                    "legacy behavior for A/B.");
 REXCVAR_DEFINE_BOOL(audio_xma_preserve_timeline, true, "Audio",
                     "When the XMA decoder backend rejects a damaged frame, emit that "
                     "frame's worth of silence instead of dropping it. The input read "

thirdparty/rexglue-sdk/src/native/audio/xma/context.cpp

@@ -20,6 +20,7 @@ REXCVAR_DECLARE(bool, audio_deep_trace);
 REXCVAR_DECLARE(bool, audio_xma_loop_guard);
 REXCVAR_DECLARE(bool, audio_xma_preserve_timeline);
 REXCVAR_DECLARE(bool, audio_xma_header_straddle_fix);
+REXCVAR_DECLARE(bool, audio_xma_loop_end_guard);
 
 namespace rex::audio {
 
@@ -726,6 +727,22 @@ void XmaContext::Decode(XMA_CONTEXT_DATA* data) {
   if (data->loop_count > 0) {
     const uint32_t loop_end = std::max(kBitsPerPacketHeader, data->loop_end);
     is_loop_end_frame = (data->input_buffer_read_offset == loop_end);
+    if (is_loop_end_frame && REXCVAR_GET(audio_xma_loop_end_guard) &&
+        data->loop_start >= data->loop_end) {
+      // This test is evaluated before UpdateLoopStatus, so the degenerate-loop
+      // guard there (audio_xma_loop_guard) never reaches it. With
+      // loop_count > 0 and loop_start >= loop_end the clamp above turns
+      // loop_end into kBitsPerPacketHeader - which is exactly a voice's
+      // starting read offset for the stream that begins at packet 0, and
+      // exactly what SwapInputBuffer resets the read offset to - so the
+      // frame output limit armed on the first frame of that stream (and on
+      // any post-swap pass that lands at the raw offset), and Consume()
+      // silently discarded the tail of the frame: 384 samples lost on ONE
+      // context of a multi-stream voice, which permanently offsets it from
+      // its siblings. Apply the same real-window requirement here that
+      // UpdateLoopStatus applies.
+      is_loop_end_frame = false;
+    }
   }
 
   UpdateLoopStatus(data);