Every image thread took the socket's mutex to add its message, and the thread that cut a batch then
serialised and sent the whole period's images while holding it, so all image threads waited for the
encode. Now the message is copied outside the lock without spots, reflections and image (the metadata
message never carries them), the lock only appends and swaps the batch out, and the batch is serialised
into its own buffer and sent with the lock released. Two batches cut close together can leave in either
order, which the documentation already allows; END stays last.
Two latent faults go with it: a disabled socket (no period) kept every image of the run and sent them
all at END - now it keeps nothing; and the batch buffer was a fixed 20 kB per image, too small for a 2D
azimuthal integration profile - it is now sized from the profile and ADU histogram lengths.
(ported from rc184-2 4147934e6)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HVeDQsU5a8AJUdJPgWevAK
leonarski_fandClaude Opus 5.5
79fe568795