Refactor README and streaming_turn_model.py to clarify fixed-grid frame ownership and update documentation. Remove merge conflict markers and enhance explanations regarding session launching conditions and audio slicing behavior.

README.md

@@ -497,8 +497,8 @@ python streaming_turn_model.py \
   --start 0 --end 4 \
   --video-fps 2 \
   --strict-fixed-video-grid \
-  --verify-every 5 \
-  --verify-continuous-every 5 \
+  --verify-every 1 \
+  --verify-continuous-every 1 \
   --flash-attention-2 \
   --output features/Session_1_open_first_4s.pt
 ```
@@ -554,10 +554,12 @@ The sidecar selects the matching fixed-FPS MP4 and sets `N` to its exact
 `--strict-fixed-video-grid` is enabled.
 
 In strict A+V mode the extractor does not delegate temporal frame selection to
-the generic video utility. It reads the already-fixed MP4 grid explicitly,
-uses only frames with timestamps `<= t`, and duplicates the most recent past
-frame when the visual tower needs an even frame count. Audio is loaded once at
-16 kHz and sliced exactly at each causal boundary. The processor still receives
+the generic video utility. It reads the already-fixed MP4 grid explicitly and
+assigns both video frames and audio samples with half-open intervals
+`[start, end)`. Thus a frame exactly on a two-second boundary belongs only to
+the following block. It duplicates the most recent available frame when the
+visual tower needs an even frame count. Audio is loaded once at 16 kHz and
+sliced at the same causal boundary. The processor still receives
 `use_audio_in_video=True`, so Qwen's audio/video token interleaving is retained.
 Long runs use `tqdm` and update the progress bar after every completed state,
 showing percentage, media time, elapsed time, ETA, and throughput. Pass
@@ -576,15 +578,17 @@ python streaming_turn_model.py \
   --start 0 --end 4 \
   --video-fps 2 \
   --strict-fixed-video-grid \
-  --verify-every 5 \
-  --verify-continuous-every 5 \
+  --verify-every 1 \
+  --verify-continuous-every 1 \
   --flash-attention-2 \
   --output features/Session_1_first_4s_check.pt
 ```
 
-Do not launch all sessions unless `cache_replay` cosine is nearly 1. The
-`continuous_window_semantics` cosine answers a different question and may be
-lower because blockwise media preprocessing changes the representation.
+Do not launch all sessions unless `cache_replay` cosine is nearly 1. For
+`open-stream-replay`, a materially lower `continuous_window_semantics` cosine
+means that the custom payload stitching is not equivalent to one continuous
+Qwen input; it must not be dismissed as a media-turn-boundary effect because
+this mode intentionally keeps a single media turn open.
 
 Video can be sampled at 2 FPS while predictions remain at 5 Hz. Audio and the
 causal boundary update every 200 ms; consecutive prediction times may use the
@@ -667,6 +671,7 @@ Each `.pt` file contains:
 - `config`: step, native block, memory, FPS, and pixel settings;
 - `verification`: cached-versus-exact numerical checks;
 - `input_video_fps` and `video_sampling_grid`: sampling provenance;
+- `media_interval_policy`: fixed-grid AV uses half-open `[start, end)` ownership;
 - `logits`: optional `[N, 10, 3]` output if a trained head is supplied;
 - `probabilities`: optional softmax probabilities.
 

streaming_turn_model.py

@@ -243,7 +243,13 @@ class FixedGridAV:
             raise RuntimeError(f"Decoded audio is empty: {self.path}")
 
     def frames(self, start: float, end: float) -> list[Any]:
-        """Return fixed-grid frames whose timestamps lie in [start, end)."""
+        """Return fixed-grid frames whose timestamps lie in ``[start, end)``.
+
+        Half-open ownership is essential at native block boundaries. For
+        example, the frame at exactly 2.0 s belongs to ``[2.0, 4.0)`` and must
+        not also be committed in ``[0.0, 2.0)``. Audio slicing below follows
+        the same convention.
+        """
 
         import numpy as np
         from PIL import Image
@@ -1112,11 +1118,32 @@ class OpenStreamReplayExtractor:
             manual_audios=manual_audios,
             video_fps=self.config.video_fps,
         )
-        return _open_media_pieces(
+        pieces = _open_media_pieces(
             self.model,
             inputs,
             use_audio_in_video=use_audio_in_video,
         )
+        # A standalone tail is at most one native 2 s block. Its temporal
+        # TMRoPE coordinates must therefore stay inside one 50-position chunk.
+        # This catches endpoint-inclusive frame selection (or another
+        # preprocessing regression) before overlapping coordinates are
+        # committed to the next block.
+        duration = interval[1] - interval[0]
+        if duration <= self.config.native_block_seconds + 1e-6:
+            temporal_positions = pieces.payload_position_ids[0]
+            temporal_span = int(
+                temporal_positions.max().item()
+                - temporal_positions.min().item()
+            )
+            positions_per_chunk = _tmrope_positions_per_chunk(self.model)
+            if temporal_span >= positions_per_chunk:
+                raise RuntimeError(
+                    "One open-stream AV interval spans multiple native TMRoPE "
+                    f"chunks: interval={interval}, temporal_span={temporal_span}, "
+                    f"chunk_size={positions_per_chunk}. Check frame-boundary "
+                    "ownership and audio preprocessing."
+                )
+        return pieces
 
     def state_from_scratch(
         self,
@@ -1701,6 +1728,11 @@ def main() -> None:
             if fixed_grid_av is not None
             else "qwen_utility_adaptive_sampling"
         ),
+        "media_interval_policy": (
+            "half_open_[start,end)"
+            if fixed_grid_av is not None
+            else "qwen_utility_defined"
+        ),
         "open_stream_policy": (
             {
                 "commit": "completed_2s_av_payload_only",