Skip tracklet stitching when no animals were detected - #3504

videos with zero
detections would fail after serialization on `np.stack([])`

Return an
empty DataFrame with the expected DLC column schema instead and add a
regression test for the empty-output case.

- add informative warning
- add guard in dataframe export: raise ValueError if no predictions

Ensure inference emits results for images whenever a prediction exists, even when no detections are produced and thus no queue item is generated. This avoids dropping frames with valid predictions but empty detection outputs.

Fixes a mismatch in `analyze_videos` where the DataFrame export step still used `predictions` instead of `ctd_predictions`. This ensures saved video analysis results use the corrected prediction set from the current processing path.

When a video produces no pose predictions, analysis now writes all-NaN placeholder predictions so downstream HDF5 export and filtering can continue instead of failing on an empty result set. The DataFrame export path was adjusted to handle these fallback frames correctly and raises clearer errors only for genuinely invalid empty input. FMPOSE 3D inference now also raises a clear RuntimeError when no 2D poses are detected for a video.

Move frame key zero-padding logic into `runners.shelving.frame_key_width` and use it from both video output generation and shelf writing. This keeps frame key formatting consistent across full-data exports and on-the-fly shelving, while preserving the default width when the frame count is unknown.

Refines `video_inference` logging so empty results distinguish unreadable videos from detector misses, and keeps re-encoding guidance for partial reads. Expands tests to cover zero-prediction, partial-prediction, and shelf-writer cases.

Simplifies `analyze_videos` by always passing the inference `predictions` directly to `create_df_from_prediction` when `save_as_df` is enabled, and removes the special-case branch that generated all-NaN placeholder rows when no predictions were produced. The unused `_nan_predictions` helper and its warning/error handling were also deleted.

Use the CTD prediction list when deciding whether to write the dataframe, so `.h5` export still happens based on the converted CTD results instead of an unrelated variable. The warning now also explains that the CTD pickle had no frames when export is skipped.

Add an explanatory comment in `analyze_videos` documenting why `ctd_predictions` can be empty, and why the code warns and skips dataframe creation instead of raising. This captures edge cases like zero-frame pickle output and `save_as_df=False`.

Refines the test docstring in `test_runners_inference.py` to describe the async behavior more clearly by comparing it to the sequential path without implying a regression pin.

ShelfWriter now always derives its frame key width through `frame_key_width()` instead of keeping a separate default. This removes duplicated fallback logic and keeps frame index formatting consistent with the shared helper.

Avoid running stitch/tracklet conversion when no animals were detected in a video. `analyze_videos` now detects empty assemblies, warns, and leaves the pose predictions available without creating tracking files. `TrackletStitcher` raises a dedicated `EmptyTrackletsError`, and the stitcher test now expects that case.

Cover the no-detection path in pose-estimation assembly generation and the header-only tracklet case in the stitcher. These tests pin the expected empty-input behavior and the error type for invalid tracklet files.