v0.5.0

Export several moments from one event into a single video, plus a defect
sweep across the export pipeline, the timeline and the build.

Highlights

• Multi-interval export — pick more than one range on the timeline and
  get one file. Drag the markers to a range, press Add interval to
  commit it, repeat. Committed ranges show as green bands on the timeline
  and as deletable chips, and the duration readout is the sum of what will
  actually be exported, not the span it covers. Intervals are joined with a
  hard cut, sorted and merged if they overlap.
• Each interval keeps its own burned-in wall clock, so the timestamp
  jumps to the correct time at every join instead of counting straight
  through the gaps. Telemetry HUD and GPS follow the same way.
• A camera missing from one of the selected ranges no longer freezes its
  tile on the last frame for the rest of the export — the gap is filled
  with black of exactly the right length so all tiles stay in sync.

Fixes

• Exports could silently lose up to a second off the end. The HUD overlay
  determines output length, and with non-whole-frame ranges it came out
  fractionally shorter than the video, so ffmpeg trimmed the video and
  still reported success.
• Telemetry HUD could show the wrong speed and GPS for the remainder of an
  export: if a segment failed to parse mid-selection, every later reading
  was placed early by that segment's duration. The output clock now cannot
  be skipped.
• Timeline markers no longer stick to the cursor. Releasing the mouse
  outside the timeline left the marker dragging, so the next hover — with
  no button held — moved it.
• The timeline slider now seeks when driven by the keyboard. Arrow/Home/
  End moved the thumb without moving the video, which then snapped back.
• Parsed telemetry is capped at 10 cached segments in the browser. Scrubbing
  a long Sentry event previously pinned every segment it touched (~1 MB
  each) for the lifetime of the page.
• Local and non-Windows builds no longer compile with the Windows platform
  define. On macOS this selected a Windows ffprobe.exe that cannot run, so
  a clip the native MP4 parser could not read was dropped from the index
  instead of falling back. linux-arm / linux-musl-arm publishes were
  affected the same way.

Internal

• Interval and frame-grid arithmetic extracted to ExportInterval /
  ExportTiming and covered by the repo's first unit tests (27).
• Dead code removed (ExtractSeiMessagesForTimeRange, an unread timeline
  field) and the ffprobe registration collapsed to a single Windows /
  non-Windows branch.
• The protobuf C# for the telemetry schema is now generated from
  SeiMetadata.proto at build time instead of being checked in, so editing
  the schema can no longer leave the generated code silently stale. The
  generated output is byte-identical to the file it replaced.
• The camera filter is compared by value rather than field by field, and
  the audit artifacts from the architecture sweep were removed now that
  every item in them has landed.
• Container images are built on every pull request. Previously the images
  were only ever built during a release, so a build break could not surface
  until after the tag existed — which is exactly how an arm64-only failure
  slipped through: it left macOS, the unit tests and the amd64 image all
  green.

Compatibility

No migration required. The export API keeps the existing
startTimeUtc/endTimeUtc fields — a request without the new intervals
list behaves exactly as before — so older clients keep working.