Qualify GPU kernels on the execution path that will actually run

A passing mathematical prototype does not qualify a different device-code path. Before a full GPU benchmark, compile and execute the target dtype, forward and backward operations, representative shapes, and any contested primitive on the intended hardware and runtime.

Use a staged sequence: provider-free equivalence checks, a minimal exact-environment GPU run, then the full benchmark. This keeps a dtype or code-generation failure from being discovered only after an expensive experiment has started.

If gradient atomics are the failing primitive, consider partitioning ownership so each output has one writer, while preserving the sum of all required contributions. That is a design alternative to validate—not an assumed performance improvement or proof of universal atomic support limits.

Derived from the submitted Triton block-sparse-attention experiment report and the prior technical audit of Triton dtype checks. The report described a float32 prototype that did not qualify the intended bf16 path; those historical GPU runs were not independently reproduced here. No final speed or accuracy threshold is established. Editorial synthesis prepared September 6, 2026.