[Arkex-ai]

Summary

A failed vkBindBufferMemory could still be reported as a usable imported buffer. Both host-pointer and non-sparse DMA-buf factories now reject failed binding and roll back the resources they own before publishing an address.

Technical changes

• Check binding after successful allocation; free memory only when allocation succeeded, then reset failed host-import handles.
• Preserve main's destroy_import_buffer helper so associated texel views are released before raw buffer destruction.
• Keep the original binding result, range retry hold, DMA-to-host fallback and sparse-route delegation. No API change or extra GPU wait.

Validation

Head 27d7e7c, based on main 117eae7:

• 34 cases pass on Windows LLVM and Linux GCC, including ASan/UBSan. 11 mutations are rejected, including omission of the owner-release hook.
• Full local Windows suite: 298 passed / 2 failed / 21 skipped, versus 297 / 2 / 21 on the exact baseline. The same NUP-path and SDK-header tests fail; the aggregate is not green.
• The approved factory behavior is unchanged after normalizing main's owner helper. Current-head native gameplay has not been repeated.

Current-head fork check	Result
Linux AppImage	Passed
Windows build	Passed
Source suite	Passed

Earlier native evidence and limits

At 2ec6ae3 (API12), serial and parallel/batch checkpoint runs passed with fixed 2x interpolation, resize and six decoded 3440x1440 captures. Both exited normally and preserved the original checkpoint. This covers the healthy Windows host-pointer path; native DMA-buf failure, GPU OOM and campaign coverage are untested.

[Arkex-ai]

Rebased onto main d224dc06; resolved the runner conflict using per-test RUN_ALL_ARGS. The production patch ID is unchanged. Focused imported-buffer ownership and runner-declaration checks pass.

Head ca857d8fd249c8d03265b357bc6a70e1eedc0c21; fresh CIs pending:

• source-regressions.yml: https://github.com/Arkex-ai/Hookshot/actions/runs/42870250346
• windows-builder.yml: https://github.com/Arkex-ai/Hookshot/actions/runs/42870255120
• linux-appimage.yml: https://github.com/Arkex-ai/Hookshot/actions/runs/42870259765

Full-suite/current API13 native validation remains separate and pending.

[Arkex-ai]

Thanks, agreed. I rewrote this and the other four active PR descriptions: the plain-language problem comes first, followed by brief technical changes and one validation section with measured test results and current-head CI links. Rebase history and repeated status lists are removed; earlier native evidence is folded separately so it is not confused with the current head. No code changed.

[Arkex-ai]

Rebased onto main 117eae7 at 27d7e7c. The binding checks, rollback conditions, fallback and retry holds are unchanged; both rollback sites retain main's destroy_import_buffer owner-release boundary. The fixture checks that the release hook runs before destruction.

All 34 cases pass on Windows/Linux, including sanitizers; 11 mutations are rejected. Full local suite: 298 passed / 2 failed / 21 skipped, versus 297 / 2 / 21 on the exact baseline, with the same two local failures. Fresh exact-head CIs are linked in the description; current-head native gameplay remains untested.