Two numpy worlds on one Raspberry Pi
The dependency fix was a process boundary, not more pin juggling.
OUR OBSERVATIONLast verified: 2026-09-27 · Build 1.9.37
Picamera2/simplejpeg in this build needs numpy 1.x; TensorFlow training uses numpy 2.1.3. Installing numpy 2 into the runtime caused simplejpeg: numpy.dtype size changed.
~/akida-env
numpy 1.26.4→ job~/m3-train-env
numpy 2.1.3→ .fbzruntime loads
numpy 1.26.4→ job~/m3-train-env
numpy 2.1.3→ .fbzruntime loads
Two environments and one subprocess boundary turned the most common rebuild failure into explicit architecture.