Add VK_ERROR_DEVICE_LOST exception handling

[mgajda]

Summary

When the GPU is lost during a compute dispatch (TDR, thermal shutdown, OOM, driver crash), Vulkan-Hpp throws vk::DeviceLostError from vkQueueSubmit or vkWaitForFences. Kompute's Sequence::evalAsync / evalAwait set mIsRunning = true just before these calls but never reset it on an exception path — so any follow-up call (even just to destroy and retry) hits the "already running" guard and deadlocks.

Fix

Catch-and-rethrow around the two Vulkan calls so mIsRunning is cleared on any throw, while letting the exception (including vk::DeviceLostError) propagate unchanged to the caller.

References

• Affects: Sequence::evalAsync(), Sequence::evalAwait() in src/Sequence.cpp
• PR: Sequence: reset mIsRunning on exception from submit/waitForFences #461

[mgajda]

Implementation Status: ✅ COMPLETED

PR: #460 - Add DeviceLostError exception for VK_ERROR_DEVICE_LOST handling

Changes:

• Created kp::DeviceLostError exception class (Exceptions.hpp)
• Catch vk::DeviceLostError in Sequence::evalAsync() (vkQueueSubmit)
• Catch vk::DeviceLostError in Sequence::evalAwait() (vkWaitForFences)
• Throw kp::DeviceLostError with context for callers to handle
• Documented @throws in public API headers

Testing: All 16 unit tests pass

Related: Issue #458 (queue-family selection fix) - companion fix also submitted

[axsaucedo]

Thank you for the attempt. Seems like a potentially valid topic, but not a fan of this contribution to be honest, it's adding an entire new file and complexity for a single exception; I appreciate and accept the use of AI for contributions, but these need to be aligned to the codebase understanding the architecture, framework, etc - seems to me this was just created and submitted without any of these. Closing the PR so first we align on the decision, or potentailly we can fix directly instead.

[mgajda]

@axsaucedo Not sure how to align with the codebase without constructive feedback. Can you please tell me what is wrong beside length?

[mgajda]

Opened a simplified replacement: #461 — 14 lines, no new header or exception class. Just resets the running flag so a throw from Vulkan doesn't deadlock the next call.