NEON kernel for the YUV420 superpixel convert (vld2q deinterleave, rounding-halving greens, mlal/rshrn luma, pairwise-add chroma block sums), bit-identical to the scalar path and proven so on a live frame via the FORGECTRL_NEON_CHECK one-shot memcmp. Convert 75 to 18 ms; the stream now runs at the OV5648 frame rate. Motion coexistence re-proven at 15 fps (clamped 0).
Demosaic the superpixels straight to planar YUV420 and encode on the CODA960 (mainline coda V4L2 mem2mem, node found by personality); libjpeg stays as the automatic fallback and the snapshot path. All camera paths now demosaic from a cached bounce copy of the frame: the V4L2 MMAP capture buffers are uncached, and reading them in-place costs ~340 ms/frame vs 43 ms memcpy + 75 ms cached convert. Per-frame stats logged every 100 frames; /cam/status reports the encoder.
A stream request for the other camera kicks current stream clients via a generation counter (their streams end cleanly; viewers freeze on the last frame) and switches. The index page retry consults /cam/status first so a preempted view does not steal the camera back.
Snapshots of the non-streaming camera no longer fail 409: the worker borrows the mux for one frame (stream viewers see a short freeze). Stream switches wait up to 3 s for the previous pin to drain. Arbitration compares the engine home camera, not the transiently borrowed pipeline camera. Index page: single toggled stream with retry plus a head-peek button.
ScottW514