fix(hf): verify empty-metadata kernels from exact Git

[Lutar, Stephen P. (stephenlutar2-hash)]

Outcome

Repairs the still-failing real controller path after #354. Post-merge run 30379450034 disproved the first hypothesis: Kernel metadata revision resolution succeeded, but kernels.get_kernel(...) raised ValueError: min() iterable argument is empty during exact-revision selfcheck.

Root cause

The exception occurs after [validated] kernel-card-publish and before kernel-verify. The pinned kernels==0.16.0 package asks the Kernel metadata API to enumerate build variants even when the already verified Git repository contains a valid build/torch-cpu tree. The #354 fallback surrounded the earlier metadata-revision call, not this loader call.

Change

• catch only the exact observed empty-build ValueError around kernels.get_kernel;
• materialize only build/** from authenticated Kernel Hub Git;
• require the fetched main revision to equal the already bound immutable 40-character revision;
• pass the repository root to pinned kernels.get_local_kernel, whose local resolver checks repo/build;
• run the same callable/selfcheck result validation and record the selected selfcheck transport;
• preserve every unrelated metadata, load, dependency, revision-drift, and selfcheck error as fail-closed.

Validation

• 11 finalizer tests passed locally, including the real finalize_kernel call path;
• 2 focused Git transport tests passed locally;
• Python compilation passed;
• exact-revision binding, moved-main rejection, read-only materialization, repository-root loading, no metadata-API call in the fallback, and unrelated-error rejection are covered;
• live public Kernel Git trees were observed at governed-inference-meter@dd133d351d5d77ff666ce16d452edfdfdc3f1a61 and szl-governed-norm@52ebdccd8d181e2b367387353ad87484e2d9d8b3, both with build/torch-cpu/** variants;
• commit 2670c692009764b74b7522383f921069b3e9cc1e has a verified SSH signature and DCO trailer.

Boundaries

This does not mutate model, dataset receipt, Kernel build, visibility, hardware, secrets, rulesets, or workflow policy. It does not convert a failed selfcheck into success; it gives the exact existing build a second supported local loading path and still requires its actual selfcheck() to pass.

Reopens/follows #301.

Signed-off-by: Lutar, Stephen P. stephenlutar2@gmail.com

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2670c69200

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Materialize Git LFS build objects before loading

When a Kernel variant contains LFS-tracked artifacts such as a compiled .so, this checkout runs with _auth_environment(), which sets GIT_LFS_SKIP_SMUDGE=1, and the sparse tree also omits .gitattributes; consequently build/** contains LFS pointer text rather than the binary that get_local_kernel() must import, so the new fallback still fails for compiled builds. This is reproducible with a local LFS-backed Kernel repository, and git lfs smudge --help confirms that GIT_LFS_SKIP_SMUDGE disables smudging. Enable LFS materialization for this read-only checkout (including the relevant attributes or an explicit authenticated LFS fetch/checkout) before invoking the loader.

Useful? React with 👍 / 👎.