[zhouaihui]

Outcome

The reference-plan test expected draft_attention_execution to be omitted even though its Qwen3 DSpark example explicitly selects cudnn_flex_attention. The planner correctly records that backend as required with runtime evidence still pending, causing the CPU CI run to fail on an outdated expectation.

Update the reference-plan contract to assert the requested/resolved cuDNN backend, required capability, and unknown effective execution. Retain the exact omission-set check for the remaining capabilities, and add an explicit SDPA/CP1 case that checks optional, unresolved, omitted attention execution. Production planning behavior and the example configuration are unchanged.

Verification

• Existing failure reproduced by upstream CI run 42184209945: 535 passed, one reference-plan omission assertion failed; quality checks passed.
• PR CI run 42185276730 passed for head c7bc9f2bf9829bf186995863083a767610ad609d: lint, compilation, documentation links, 9 source contracts, 537 offline CPU tests, and 16 asset/patch tests. Tests ran on the hosted Linux runner with Python 3.11 and an isolated CPU venv using the unchanged CI dependency setup.
• git diff --check passed.
• GPU/runtime evidence: not applicable to this test-only correction. No GPU or tag-triggered wheel-build acceptance is claimed.

Compatibility and provenance

• No public configuration, checkpoint, export, runtime, dependency, or workflow changes.
• Updates ASpec planning contracts in the existing test module; retained source notice is unchanged and no upstream implementation is copied.