Release v2.9.0

Breaking Changes

• Replace multimodal-embedding's video.camera_key with video.camera_keys, accepting a
  nonempty list or all (the default). Results now include camera names in their keys by default;
  select one camera and set output.tag_camera_key: false to retain untagged field names.
  Earlier episode-only staging and checkpoints are not migrated.
• Move mp4-index metadata and frame indexes from Arrow sidecars into manifest.lance, with
  record schema version 7. Update consumers to read the nested metadata and frames fields
  and allow null box metadata that PyAV does not report. Create a fresh manifest dataset for
  existing outputs; existing moov/moof caches remain reusable.
• Rename video-transcode's execution.spans_per_checkpoint to execution.parquet_rows_per_file.
  Shared recovery changes transcode field metadata and publication receipts, and replaces both
  transcode and video-embed staging layouts. Existing transcode outputs and earlier recovery
  workspaces have no automatic migration; rebuild affected outputs with fresh staging.
• Default managed Slurm Ray head allocations to exclusive nodes and advertise their allocated
  resources for application work. Set slurm.head.exclusive: false to retain shared-node CPU
  and memory requests; cpus and memory are ignored for exclusive allocations.

Added

• Embed multiple cameras in one multimodal-embedding run through shared model actors, with
  per-camera checkpoints and one atomic publication of all selected camera results. Adding a
  camera reuses already-published results for other cameras.
• Extend fragmented MP4 indexing to multiple frames per fragment. Add custom UUID-box parsing
  hooks, optional camera timing fields, and validation of frame byte ranges against source size.
• Add input.index_manifest_uri to data-integrity and --index-manifest to its standalone
  CLIs to reuse mp4-index timelines without scanning each covered video. Record container
  decode-timestamp provenance when sensor timestamps are absent, and fall back to media reads
  for missing, unusable, or size-mismatched index records.
• Add HesaiLidarParser for calibrated Pandar128E3X UDP decoding into full-sweep point clouds,
  including intensity, ring IDs, and per-point sensor timestamps.
• Run managed Ray workloads on the Slurm head, including GPU work, and allow worker_lanes: 0
  for head-only runs that can add workers later. Head and worker resource requests are independent.
• Accept registry image references in managed Slurm Ray configs and add
  slurm.worker.max_lane_allocations to cap renewal arrays on clusters with job-count limits.
• Add opt-in OTLP metrics export around config-backed pipeline runs through
  COSMOS_CURATOR_OTLP_METRICS_PUSH, with a configurable push interval.

Fixed

• Select re-encoded robot-action-split frames by source presentation timestamps and verify
  the selected range, preventing shifted clips when seeking MPEG-TS sources or reordered frames.
• Resolve local file:// paths consistently in multimodal-split preflight, discovery, and
  sensor reads, including file://localhost paths.
• Validate GPS protobuf fields without the removed FieldDescriptor.label API, restoring
  compatibility with current protobuf descriptors.
• Clear inherited PIXI_WORKSPACE_* activation variables in Slurm launches so container
  commands do not use the launcher's host workspace.

Changed

• Adopt shared checkpointing and recovery for video-transcode and video-embed. Resume saved
  selections across fresh clusters, appended fragments, and unrelated enrichment commits;
  retry publication from durable results without repeating checkpointed encoding or inference.
• Publish MP4 metadata, frame indexes, source summaries, and provenance together in Lance,
  allowing consumers to query video summaries without loading full frame arrays.
• Build mp4-index frame indexes with PyAV from extracted moov/moof boxes, supporting
  additional fragmented layouts without rereading source media when the boxes are cached.
  Stream selection, timestamps, and durations follow PyAV; UUID extensions retain box-walker
  indexing, with distinct extraction contracts for the two indexers.
• Upgrade Ray to 2.59.0. Local clusters started with ray.init() now enable token authentication by default.
• Upgrade Lance (pylance) to 13.0.0 and Pixi in runtime images to 0.82.0.

Documentation

• Add a VITRA / EPIC-KITCHENS public sample guide, an annotation-to-LeRobot converter, and an
  example robot-action-split config for an end-to-end egocentric workflow.
• Document shared checkpoint recovery, multicamera embedding, PyAV frame indexing, MP4
  extensions and schemas, and managed Slurm Ray head execution.
• Refresh the quickstart around Curator Next recipes and canonical Lance outputs, and clarify
  pipeline execution and robot-action split configuration.