[DjDeveloperr]

The G17 watchdog renews a published job's deadline, up to thirty seconds in all, while some command retired within the last five seconds: the GPU is then shared, not hung. A ray traversal dispatch beside another one gets no such renewal, because neither retires anything until it ends, even though both progress. Each gets only the ten seconds a long job has without it.

On Neo, dEQP-VK.ray_query.stress.geometry_shader.triangles takes about 6.3 s alone. Two copies side by side took 10 to 17 s each, and the watchdog failed 18 of 28 runs in three minutes with ETIMEDOUT. Their clients lost the device. The same timeout appears in 6-worker RT CTS runs whenever two traversal-heavy cases overlap.

This change also counts the GPU as shared while a long job (ray traversal or AS build) of another queue runs beside a long job. The thirty-second bound is unchanged, so a pair that never finishes still fails.

Testing

Neo (A18 Pro), Mesa agx3 at #62's head, this commit on integration/aurora at #423's merge:

• Two copies of ray_query.stress.geometry_shader.{triangles,aabbs} side by side for three minutes: 24 runs, all pass, no watchdog timeouts; each took 14 to 18 s. Without this change, the same leg failed 18 of 28 runs with ETIMEDOUT.
• Regression gate passes.
• Targeted CTS (5368 cases): 4687 pass, 681 not supported, 0 fail.
• Ray query and ray tracing pipeline CTS groups at 6 workers (57190 cases): 47100 pass, 10088 not supported, 0 fail. Two host-built 1M-primitive acceleration-structure cases were killed for memory exhaustion, as on the base.
• No asahi or G17 kernel errors.