Final CI evidence for candidate 38fb37ae748e5a7c6430012dc4610c1a7d7abe2a

All 18 jobs in CI run 41339383404 and the separate Docker run passed. Linux Python 3.11:
2,924 passed, 11 skipped, 2 deselected, 84.69% coverage. Full skip list and checkout
identity are retained in ci/linux-test-summary.txt. Quiet CI output does not
enumerate each passing test. The runtime recipe was actually installed and
executed on CPU. checks.boltz.checkpoint_loader_checked is true; cuda_tested
and models_downloaded are false. This proves the canonical preflight runs;
it does not establish publisher-checkpoint loading or full GPU inference.

GitHub artifact 13097056201 was downloaded and checked against its SHA256.
Both measured freezes and the separate CUDA11 inventory exactly match prior
candidate 493938376ffbbd554998c6b559a7c265827ffb73. Prior advisory response/query/input bytes are retained
with their original dates; no new vulnerability query was performed. DGL's
distributed-RPC advisory GHSA-3x5x-fw77-g54c remains applicable and explicitly
outside supported operation. Local/source-package exclusions and database
coverage limitations remain explicit. Unchanged versions do not prove
numerical equivalence under changed execution code.

The manifest hashes each file and source. CI paths alone are normalized to
logical roots. Full logs, signed URLs, model bytes, credentials and private
GPU identifiers are excluded. Earlier evidence is preserved unchanged.
