This is the DR-9 seam. The C++ exposes AtomicRequest::native_handle()
returning a raw drmModeAtomicReq*, because NVIDIA’s
eglStreamConsumerAcquireAttribEXT takes one through
EGL_DRM_ATOMIC_REQUEST_NV and submits the commit itself. That pointer
cannot exist here: drm-ffi builds the kernel struct from flat slices at
commit time and never materializes a libdrm request object.
Exporting the writes instead lets drmkit-scene-streams replay them into a
libdrm request it owns, so libdrm stays out of drmkit-core and out of
GPU-less builds. The same ordered view is what T7’s property-write trace
differ consumes, so the seam has a consumer in the minimal cut rather than
lying dormant until the streams tier lands.