[pdfinn]

Fixes the frozen recognition boxes in the vision demo. After echo place dla0 > /mnt/vision/N/ctl, wm/camera kept drawing the last boxes. They updated again only after the layer was unmounted and bound again.

Cause

1. A place change starts a new follower for the same feed (restart in visionfs's settings). So does a change of threshold, classes, model or rate.
2. It stopped the old follower with stopfeed. That marks the session ended and answers every waiting reader with end of file.
3. layers(2)'s reader takes end of file as the end of the layer: it keeps what it has and exits.
4. A new reader starts only when the file's stamp changes or the layer is bound again. So the boxes froze until the unmount and bind.

It isn't specific to the DLA: any of those settings does the same.

Fix

The restart only kills the old follower (killfeed). The sequence numbers continue, so a reader takes the new follower's next frame. src none still ends readers, as before.

Test

tests/host/vision_feed_test.sh, against a camera (camsrv -t): cat reads the session's detections in the background, which stops at end of file as a layer's reader does. The threshold changes from 0.3 to 0.45 while cat is waiting, and its output must keep growing.

Run on the Jetson with the yolox-nano model:

visionfs	result
unfixed	cat stopped at 6 then 6 lines (fails)
fixed	12 then 24 lines (passes)

My first two versions of the check didn't reproduce the bug. One used shell reads that weren't waiting at the restart. The other wrote thresh 0.3, which is the default, so nothing restarted.

Two notes on that run:

• The other failure in the run, frame 3's detection, comes from the picture I used: the test expects a dog, still.jpg shows a person. The live and camera checks pass.
• The test skips without a model, so it doesn't run in CI.