AriaStreamLab 0.2.4

• Lossless processing (max-speed replay, non-realtime mock) is now lossless and
  reproducible for processors with several inputs: secondary inputs use back-pressure
  instead of dropping their oldest samples, time-based joins wait for the data (not for a
  host-time timeout) and are never trimmed, and the join buffers are trimmed by time behind
  the primary. A raw secondary stream with nothing at a frame's time (a gap, or a stream
  that ended) stops holding the frame once the primary is 1 s of device time ahead.
• A processor returning something that is not an Emit/tuple or a list of them (a bare
  payload, a float) is counted as an error instead of silently ending its worker — which
  also hung a lossless replay for good; any unexpected error in a worker's loop is counted
  and the worker carries on.
• Quitting the app (closing the window, Ctrl-C in ariastreamlab app --browser or
  ariastreamlab design) waits for a session started from the browser to finalise; before,
  the process exited first and left its manifest at status: running.
• The experiment's logging section (level, format) is applied; it was ignored because
  the import-time defaults were never replaced. Reconfiguring keeps the session's
  logs/session.log handler.
• Live sessions: the device monitor survives unexpected errors (thermal mitigation reported
  without a temperature, an unparseable quality score, …) — before, one such error ended the
  thread and with it the stall detection that stops a session the glasses have stopped.
  The quality-stats parser no longer raises on a non-numeric score.
• Tests: the example-project tests no longer assume a non-realtime mock session is still
  running 0.3 s after it starts.
• Dashboard, sessions it did not start: Sessions → delete refuses the session attached to
  the dashboard whatever started it (ariastreamlab run, a notebook, a protocol script) —
  before, only a browser-started run was protected — and a session another process is still
  writing (manifest running, a file changed in the last 2 min). Run on Design and
  replay on Sessions refuse to start while any session is attached, instead of starting a
  second session against the same glasses and taking the dashboard over.
• Dashboard on the network (host: "0.0.0.0"): Design, Sessions and Setup — which
  read and write files, delete sessions, install packages and start sessions — answer only
  the computer running the experiment; other machines get the monitor. A warning is logged
  when the server listens beyond loopback, and the tutorial says what is exposed.
• Replay controls act at once: pause, seek, speed and stop interrupt the pacing wait
  instead of waiting for the next sample — across a multi-second hole in a recording they
  were dead for the whole gap (30 s at 0.25×) and stop() timed out. Seeking a paused replay
  keeps it paused at the target (it used to resume), and the position shows the target.
• Markers and intervals added during a replay (dashboard buttons, mark()/begin()/end()
  without at_ns) are stamped at the replay position; the host-clock estimate they used
  was wrong after a pause, a seek or at any speed other than 1×.
• VRS replay after a seek or with --start: sequence numbers continue from each stream's
  record index in the file instead of restarting at 0 (the statistics saw duplicates and
  the recorded sequence_number was wrong).
• Experiment.open() cleans up when it fails half-way (an unknown processor type, a
  storage or dashboard error): the statistics thread and whatever else had started are
  stopped, and a session directory that already exists is finalised as crashed instead of
  staying running.
• Experiment.wait() raises the error that ended a session started with start() (no
  glasses found, a source failure…); before, it returned True and a protocol loop went on
  annotating a session that had stopped.
• Ctrl-C ends a session normally (completed, with an "interrupted" warning) whether it was
  pressed during run() or in the protocol loop of a start()ed session; the latter used
  to mark it crashed.
• An Experiment is one session: open()/run() after close() raise instead of writing
  an empty "completed" session that carried the first one's statistics.
• Release workflow: a tag is published only after the whole CI suite (lint, types, import
  contracts, tests on Ubuntu and macOS, wheel build) passes on that tag; the tag must be
  vX.Y.Z and reaches the shell through an environment variable (a crafted tag name given
  to the manual re-publish could run commands); the workflow token is read-only except for
  the publishing job, and every third-party action is pinned to a commit.