fix(demo): stabilize visualizer bbox and mask rendering

Align bbox coordinate handling across primary and fallback paths, normalize Both-mode raw mask rendering, and tighten demo result typing to reduce runtime/display inconsistencies.

opengait/demo/input.py

@@ -7,7 +7,8 @@ Provides generator-based interfaces for video sources:
 """
 
 from collections.abc import AsyncIterator, Generator, Iterable
-from typing import TYPE_CHECKING, Protocol, cast
+from typing import Protocol, cast
+
 import logging
 
 import numpy as np
@@ -17,15 +18,15 @@ logger = logging.getLogger(__name__)
 # Type alias for frame stream: (frame_array, metadata_dict)
 FrameStream = Iterable[tuple[np.ndarray, dict[str, object]]]
 
-if TYPE_CHECKING:
-    # Protocol for cv-mmap metadata to avoid direct import
-    class _FrameMetadata(Protocol):
-        frame_count: int
-        timestamp_ns: int
+# Protocol for cv-mmap metadata (needed at runtime for nested function annotation)
+class _FrameMetadata(Protocol):
+    frame_count: int
+    timestamp_ns: int
 
-    # Protocol for cv-mmap client
-    class _CvMmapClient(Protocol):
-        def __aiter__(self) -> AsyncIterator[tuple[np.ndarray, _FrameMetadata]]: ...
+
+# Protocol for cv-mmap client (needed at runtime for cast)
+class _CvMmapClient(Protocol):
+    def __aiter__(self) -> AsyncIterator[tuple[np.ndarray, _FrameMetadata]]: ...
 
 
 def opencv_source(