cpu2tensor 0.5.1

cpu2tensor 0.5.1 adds a supported exact wire-frame observer for profiling and replay capture.

from cpu2tensor import Pool

with open("capture.bin", "wb") as capture:
    def record(frame):
        capture.write(frame.data)

    with Pool([endpoint], wire_observer=record) as pool:
        for batch in pool.read():
            train(batch)

Each callback receives the complete header and payload as a read-only borrowed memoryview, plus stable worker, endpoint, and header metadata. The view is released on return. Per-worker order is preserved; multiworker callbacks may overlap. With capture disabled, the normal path allocates no observer metadata or views. See docs/wire-observer.md.

The attached offline learner runtime contains cpu2tensor 0.5.1 built from commit aec51e7c415b4c6b2a59a136773d2ed727eeed13, CPU PyTorch 2.13.0, and dependencies for CPython 3.12 on Linux x86-64. Its builder boundary is Ubuntu 24.04 with glibc 2.39. QEMU, targets, credentials, and host orchestration remain operator-owned.