fla-optimization-loop

Treat the op's test file and naive.py reference as frozen, and run python -m benchmarks.ops.verify --op <op> after every single change. Each iteration makes one change, checks the gate stays green, logs the result, and commits before the next attempt.

Run python -m benchmarks.ops.verify --op <op> --base main to compare against main in a single session. Use --gate-k only for fast iteration signal on a shape subset, and promote only on a full green gate with a repeatable measured win.

The loop supports hand-written kernels in Triton, Gluon, TileLang, and CuTe DSL. Torch operations are allowed only as glue around a kernel you wrote, not as a wholesale replacement that delegates the compute to a vendor library.

The conftest replaces torch.empty allocations with NaN-filled memory, so partially written outputs surface as NaN in assert_close. An ad-hoc script often sees coincidentally zeroed memory and hides the partial-write bug; trust the gate and fully initialize every output element.

Stop when the iteration cap is reached, when re-assessment shows hard evidence of a floor such as HBM bandwidth or launch-overhead limits, or after documenting three distinct tried directions. A no-go verdict requires a recorded baseline, a reasoned candidate attempt, gate status, bench evidence, and a named blocker.

No. Tolerance is part of the frozen contract, and widening it converts a correctness regression into a silent one. If you believe a test or tolerance is genuinely wrong, that must be a separate justified PR, never bundled into a performance change.