onnxruntime

e379df24 - [WebGPU] Fix deferred pipeline callback lifetime after failed waits (#33209)

[WebGPU] Fix deferred pipeline callback lifetime after failed waits (#33209) ### Description <!-- Describe your changes. --> Share `PipelineCallbackContext` ownership between `PendingPipelineBuild` and Dawn's one-shot pipeline callback instead of passing a raw pointer to state owned solely by the pending build. A failed wait can still discard deferred dispatches and return failure, but the callback retains its result state until completion or cancellation. The callback captures only the result state—not `WebGpuContext` or `CommandRecordingState`—and releases its ownership when invoked. Successful pipeline creation and caching behavior remain unchanged. Add regression tests covering: - Successful pipeline completion and callback-state cleanup. - A failed wait followed by late callback delivery after deferred dispatches are cleared. - Callback delivery and cleanup after the ORT WebGPU context is destroyed. Update the ownership documentation and document the remaining cancellation-test coverage gap. No public API/ABI, replay-operator, GenAI, or checked-completion behavior changes are included. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> `ProgramManager::Build` previously passed Dawn a raw `PipelineCallbackContext*` whose lifetime depended on a `unique_ptr` in the deferred dispatch state. If `Wait(build.future)` failed, `EncodeDeferredDispatches` could clear that state while Dawn still had an outstanding callback. A later callback could then write through the freed pointer. Shared ownership makes callback-state lifetime independent of dispatch cleanup without adding a permanent retention list or retaining ORT context objects. **Validation** Native Release provider tests and the shared WebGPU plugin both rebuilt successfully: ```bash cmake --build build/webgpu-pipeline-lifetime/Release --target onnxruntime_provider_test --parallel 32 cmake --build build/webgpu-plugin-20260929/Release --target onnxruntime_providers_webgpu --parallel 8 ``` Context, concurrent-context, and data-transfer regressions on Mesa lavapipe: **54 passed, 1 expected configuration skip**. ```bash cd build/webgpu-pipeline-lifetime/Release VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/lvp_icd.x86_64.json \ ./onnxruntime_provider_test \ --gtest_filter='WebGpuContextTest.*:WebGpuConcurrentContextTestStandalone.*:WebGpuDataTransferTest.*' ``` The three focused lifetime tests also passed across **20 repetitions (60 executions)**: ```bash VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/lvp_icd.x86_64.json \ ./onnxruntime_provider_test \ --gtest_filter='WebGpuContextTest.SuccessfulPipelineBuildReleasesCallbackOwnership:WebGpuContextTest.FailedPipelineWait*' \ --gtest_repeat=20 ``` The failed-wait tests deterministically use an instance without `TimedWaitAny` and a ready-but-undelivered validation-error pipeline callback. Actual native instance-shutdown cancellation is not injected because outstanding Dawn events can retain the instance; delivery after ORT context destruction and final callback-state reclamation are directly verified. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>