EmptyPoseBody: zero-allocation reads + fix v0.1 frame count via BytesIOReader

[AmitMY]

Changes

1. unpack_empty_tensor no longer allocates

Returns np.broadcast_to(np.array(0, dtype=...), shape) — a zero-stride, read-only view backed by 4 bytes — instead of np.empty(shape). EmptyPoseBody keeps the full shape and dtype without shape-sized memory, and values are now deterministically 0 instead of uninitialized garbage.

2. Fix: v0.1 files read 0 frames through BytesIOReader

Adding a shape test for EmptyPoseBody exposed a pre-existing bug: read_v0_1 infers the frame count from reader.bytes_left(), but BytesIOReader.bytes_left() only counted already-buffered bytes — nearly zero right after the header. This broke every v0.1 read through the BytesIOReader path, not just EmptyPoseBody:

with open('tests/data/mediapipe.pose', 'rb') as f:
    Pose.read(f, start_frame=5)
# ValueError: Start frame 5 is greater than the number of frames 0

bytes_left() now reports the stream's remaining bytes (matching BufferReader semantics), while expect_to_read uses a new buffered_bytes_left() to decide when to fetch more from the file — the two callers need different quantities.

3. Tests

• test_unpack_empty_tensor: shape, dtype, and buffer advancement.
• test_read_empty_pose_body_shape_matches_numpy_pose_body: reads every sample .pose file with both NumPyPoseBody and EmptyPoseBody and asserts identical data shape, dtype, confidence shape, and fps. Skips the legacy v0.0 openpose.pose (EmptyPoseBody has never supported v0.0, which stores variable people per frame).

Verification

• EmptyPoseBody on mediapipe.pose: shape (170, 1, 178, 3), strides (0, 0, 0, 0), 4-byte backing buffer (previously read as 0 frames).
• Pose.read(f, start_frame=5) on a v0.1 file now matches a bytes-based read exactly (shape and values).
• Full suite: 197 passed; only pre-existing failures (TensorFlow not installed) remain.

🤖 Generated with Claude Code