A small studio sent us a frame capture with a symptom we recognised: textures flickered for one frame after a level load, only on one GPU vendor, only when the frame rate was above 50. Nothing crashed, nothing logged, and the bug report had been open since November.

The cause was a missing pipeline barrier between a compute pass that wrote a lightmap and the graphics pass that sampled it. Vulkan does not require the driver to order those accesses, and the driver that happened to work had been serialising them for its own reasons. Turning on VK_LAYER_KHRONOS_validation with sync validation enabled printed the hazard immediately, naming both the write and the read with line numbers from their command buffer builder.

Two things made this cheap to fix and expensive to find. Cheap, because the fix is one barrier. Expensive, because the symptom was intermittent and the team's test devices were all from one vendor. We now recommend running sync validation on one automated run per commit, even when it costs 30 percent more time: a hazard you cannot see is a hazard you will ship.