When setting the AVC descriptor fails in MP4.Close(), attempt a fallback that constructs an AvcC/avc1 sample entry from available SPS/PPS NALUs. Adds github.com/Eyevinn/mp4ff/avc import and two helpers: addAVCDescriptorFallback (builds a visual sample entry, sets tkhd width/height if available, and inserts it into stsd) and buildAVCDecConfRecFromSPS (creates an avc.DecConfRec from SPS/PPS bytes by extracting profile/compat/level and filling defaults). Logs errors and warns when the fallback is used. This provides resilience against SPS parsing errors when writing the MP4 track descriptor.
Replace direct sanitizeParameterSets usage with normalizeH264ParameterSets and normalizeH265ParameterSets in mp4.Close. The new functions split Annex-B blobs, strip start codes, detect NALU types (SPS/PPS for AVC; VPS/SPS/PPS for HEVC), aggregate distinct parameter sets and fall back to sanitizeParameterSets if none are found. Added splitParamSetNALUs and formatNaluDebug helpers and debug logging to output concise parameter-set summaries before setting AVC/HEVC descriptors. These changes improve handling of concatenated Annex-B parameter set blobs and make debugging parameter extraction easier.
Fill missing SPS/PPS/VPS from camera config before closing recordings and warn when parameter sets are incomplete (for both continuous and motion-detection flows). Sanitize parameter sets (remove Annex-B start codes and drop empty NALUs) before writing AVC/HEVC descriptors. Prevent creation of empty MP4 files by flushing/closing and removing files when no audio/video samples were added, and only add an audio track when audio samples exist.
Add updateVideoParameterSetsFromAnnexB to parse Annex B NALUs and populate missing SPS/PPS/VPS for H.264/H.265 streams. Call this helper when adding video samples so in-band parameter sets can be recovered early. Also add error logging in Close() when setting AVC/HEVC descriptors fails. These changes improve robustness for streams that carry SPS/PPS/VPS inline.
Add --provenance=false to docker build invocations in .github/workflows/release-create.yml (both default and arm64 steps) to suppress Docker provenance metadata during CI builds.
Pass VERSION from CI into Docker builds and embed it into the Go binary via ldflags. Updated .github workflow to supply --build-arg VERSION for both architectures. Added ARG VERSION and logic in Dockerfile and Dockerfile.arm64 to derive the version from git (git describe --tags) or fall back to the provided build-arg, then set it with -X during go build. Changed VERSION in machinery/src/utils/main.go from a const to a var defaulting to "0.0.0" and documented that it is overridden at build time. This ensures released images contain the correct agent version while local/dev builds keep a sensible default.
Uncomment and explicitly set mdhd.Duration = 0 in machinery/src/video/mp4.go for relevant tracks (video H264/H265 and audio track). This ensures mdhd.Duration is zero for fragmented MP4 so players derive duration from fragments (avoiding QuickTime adding fragment durations and doubling the reported duration).
Compute actual video duration from SegmentDurations and ensure container headers reflect fragment durations. Set mvhd.Duration and mvex/mehd.FragmentDuration to the maximum of video (sum of segments) and audio durations so the overall mvhd matches the longest track. Use the summed segment duration for track tkhd.Duration and keep mdhd.Duration at 0 for fragmented MP4s (to avoid double-counting). Add a warning log when accumulated video duration differs from the recorded VideoTotalDuration. Harden fingerprint generation and private key handling with nil checks.
Add mp4_duration_test.go: unit test that creates a simulated H.264 fragmented MP4 (150 frames at 40ms), closes it, parses the output and verifies that mvhd/mehd and trun sample durations are consistent and that mdhd.Duration is zero.
Add MacEpochOffset constant and convert mp4.StartTime to Mac HFS time for QuickTime compatibility. Compute macTime = mp4.StartTime + MacEpochOffset and use it for mvhd CreationTime/ModificationTime, as well as track tkhd and mdhd creation/modification timestamps for video and audio tracks. Also set mvhd Rate, Volume and NextTrackID. These changes ensure generated MP4s use QuickTime-compatible epoch and include proper mvhd metadata.
Replace call to log.Log.Warn with log.Log.Warning in MP4.flushPendingVideoSample to match the logger API. This is a non-functional change that preserves the original message and behavior while using the correct logging method name.
Before flushing a segment when mp4.Start is true, add any pending VideoFullSample for the current video track to the current fragment. The change computes and updates LastVideoSampleDTS and VideoTotalDuration, adjusts the sample DecodeTime and Dur, calls AddFullSampleToTrack, logs errors, and clears VideoFullSample so the pending sample is included in the segment before starting a new one. This ensures segments contain all frames up to (but not including) the keyframe that triggered the flush.
