Counts host fallbacks so a silent one becomes a test failure.

Every op this backend cannot run natively transfers to Nx.BinaryBackend, computes, and transfers back. That keeps results correct — the fallback is the reference implementation — which is exactly why no assertion on values can ever detect that an op left the GPU. A fallback is bit-identical to the GPU path by construction, so correctness tests are structurally blind to it, and a performance cliff shows up as nothing at all.

That blindness is not hypothetical. Conv's backward pass ran entirely on the CPU for the whole life of the conv shaders: Nx.Defn.Grad (hidden, so not linked) emits convolutions with the first two axes swapped, those failed the identity-permutation gate, and every gradient conv fell back. The suite stayed green, the doctests stayed green, and a CNN training step took 30 seconds.

This module makes the invisible thing countable:

{result, counts} = Nx.Vulkan.Fallback.count(fn -> Nx.Defn.grad(...) end)
assert counts == %{}

Cost

Recording is per-process and off by default. When off, the instrumentation is a single Process.get/2 on a path that is already doing a device→host copy — unmeasurable. Attribution (which op fell back) reads the current stacktrace, which only happens while recording.

Scope

Instrumented at host_result/2 in Nx.Vulkan.VulkanoBackend, the common exit point of every fallback path. Counting is process-local, so a fallback that happens inside another process (e.g. work funnelled through Nx.Vulkan.Node) is not counted by the caller's count/1.

The count is a lower bound

Only ops that reach this backend can be counted. Once a fallback puts a tensor on Nx.BinaryBackend, Nx dispatches every subsequent op on that tensor straight to Nx.BinaryBackend — this module's callbacks are never invoked, so the downstream host work is invisible here.

That is not theoretical. A LeNet training step reported no window_max at all while dot/7 was still falling back: the dot dropped the pooling gradient onto the host, and everything after it ran there unseen. Fixing dot kept the tensor resident, and window_scatter_max/6 promptly appeared in the census.

So a count going up after a fix can mean the fix worked and exposed something that was already happening. Read the composition, not just the total.