Add NV-Reason-CT inference skill

[wyli]

Summary

Add skills/nv-reason-ct/ as the Medical AI Skills source for NV-Reason-CT inference on caller-provided 3D NIfTI chest or abdominal CT volumes. This source PR is the prerequisite for the catalog registration in NVIDIA/skills#608.

• Wrap the documented upstream Transformers inference path for a Hub model or complete local model-and-processor export.
• Require explicit --trust-model-code consent before executing custom model code; keep setup inspection read-only.
• Preserve structured JSON output, model/input provenance, and incomplete-generation handling.
• Document end-to-end use of the public upstream examples, isolated caches, and upstream-owned dependency installation. No maintained dependency pins or runtime synthetic-data generator are added.
• Include self-contained offline contract tests and opt-in real-volume/GPU parity tests. No scans, weights, generated medical responses, or local run artifacts are included.

Engineering and research workflows only; not for diagnosis, treatment, or clinical reporting.

Validation

• Focused offline suite: 125 passed, 5 skipped. The five skipped checks explicitly require separately staged upstream volumes or authorized live CUDA inference. Regression coverage includes repository-root collection without registered opt-in flags.
• Skill frontmatter validation passed.
• Repository structural lint: 0 errors, 1 advisory (new skill is not yet listed in the root README).
• Copyright checks passed for all five applicable source files.
• Full repository test suite (make test): 621 passed, 5 skipped; 64 existing pydicom deprecation warnings.
• Ruff, Black, isort, and whitespace checks passed for the skill source.
• GitHub lint/tests and compliance passed on f61264d7b30c492095466b3d4b8a816df229c0b4; markdownlint also passed locally with the repository's configured version. The initial table-delimiter failure was corrected.
• Requested managed evaluation and signing with /nvskills-ci after those checks passed. Two attempts on the current source failed during evaluator model-service preflight with HTTP 503, before skill trials ran. Managed validation has not passed, and no current-source signature has been generated.
• No new live GPU inference or comparative agent-evaluation result is claimed by this PR. The successful environment documented in the skill is historical evidence, not a dependency/security recommendation.

Approved unsigned source checkpoint

The maintainer explicitly approved merging this source into dev while managed validation and signing remain incomplete, so the implementation is available for engineering use. This is a one-off exception to the normal merge-after-validation workflow, not a claim that all CI is green or that the skill is signed. No checks or signing gates are disabled or changed.

Follow-up #68 carries current-source evaluation and CI-generated signing artifacts. It and the NVIDIA/skills#608 catalog registration are open for review, with validation/signing pending; neither must be treated as a verified publication.

Follow-up / catalog publication

• Complete current-source evaluation and generate public-safe BENCHMARK.md, skill-card.md, and skill.oms.sig; verify the final signature. Previous generated artifacts are deliberately not imported into this source checkpoint. The existing evaluation-evidence hold is not waived.
• Record the reviewed immutable public model revision required by the manifest's publication notes.
• Review the follow-up evidence and confirm its repository and managed CI checks pass before merging the signed revision.

Source changes are limited to the skill directory; no evaluator, shared infrastructure, or catalog policy is changed. The recorded environment's known security finding remains disclosed in the skill and is not resolved by this merge.

AI-assisted: Created with Codex/GPT at the user's request.

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci