VALIDATION ONLY: discriminate PR132 ROCm versioned runtime selection

[nekomario28]

One-shot hosted A/B evidence carrier. Closed unmerged after discrimination. Base is the clean FlashML-org#132-derived version-selection candidate; head adds only the temporary workflow.

Pinned identities:

• upstream feat(rocm): add RDNA3 and RDNA4 runtime foundation FlashML-org/FreeToken#132 base: 45417560beeba896a9b0866e50154d2a7a320b8b
• clean candidate: 0c6c504b39a3cff39842209e4ef708e15f8e76e0
• carrier head: d7e96cbaf600233bd762ad549ff77965c9ee8435
• candidate diff: exactly _toolchain.py, setup.py, and tests/kernels/test_rocm_versioned_runtime_selection.py

Hosted A/B result

Run 32813417858, job 97697004925: SUCCESS on Ubuntu 24.04.

The same regression was applied to both trees:

#132 exact base:
ROCM_VERSIONED_RUNTIME_SELECTION_BASE_RC=1
ROCM_VERSIONED_RUNTIME_SELECTION_BASE=EXPECTED_FAIL

clean candidate:
ROCM_VERSIONED_RUNTIME_SELECTION=PASS_NUMERIC_7_14_OVER_7_9
ROCM_VERSIONED_RUNTIME_SELECTION_CANDIDATE=PASS

final:
ROCM_VERSIONED_RUNTIME_SELECTION_AB=BASE_RED_FIX_GREEN

The job also verified exact carrier/candidate/base ancestry, the old lexical sorted(library_dir.glob("libamdhip64.so.*")) pattern in the base, exact three-file candidate diff, and git diff --check.

This proves the source/build-selection defect and its correction. It does not claim a physical ROCm linker/runtime outcome or maintainer acceptance. Clean candidate remains fix/pr132-rocm-versioned-runtime-selection@0c6c504b39a3cff39842209e4ef708e15f8e76e0; carrier is intentionally not merged.