Add MLflow tracking and harden NV-Segment-CT fine-tuning

[wyli]

Summary

Add opt-in MLflow tracking to nv-segment-ct-finetune while keeping the upstream MONAI bundle as the sole training implementation. The change also tightens runtime reproducibility, makes bundle staging explicit, and refreshes the skill's documentation and publication evidence.

What changed

• Add mlflow_experiment_name, mlflow_tracking_uri, and mlflow_run_name inputs.
  • Setting an experiment name enables MONAI 1.4.0's built-in --tracking mlflow path.
  • Omitting the tracking URI uses a local file store under <output-dir>/mlruns.
  • Caller-configured remote stores, including Databricks, remain supported.
  • Non-secret tracking metadata is recorded in output.json; executed-config uploads are disabled.
• Preserve the existing training behavior: no custom MLflow run lifecycle, trainer replacement, DataLoader override, or training-config patching.
• Require the exact behavior-bearing monai==1.4.0 version and bound the optional MLflow dependency to >=2.10,<4.
• Replace runtime downloads, environment creation, package installation, and fallback repair logic with one explicit local staging path.
• Verify all seven pinned bundle assets by SHA-256 before MONAI starts, and load checkpoint comparisons with weights_only=True.
• Document required permissions, network endpoints, filesystem writes, caller-managed credentials, setup steps, and failure behavior.
• Bump the skill manifest to 0.4.2 and refresh the skill index, reference material, benchmark report, skill card, tests, and OMS signature.

Validation

• make test — 424 passed.
• make verify-skills — 21/21 real targets passed completeness and reproducibility audits with no advisories.
• make verify-negative-fixtures — 2/2 fixtures failed at their expected gates.
• make verify — repository lint, hooks, canonical pack replay, and drift checks passed.
• Focused nv-segment-ct-finetune tests — 17 passed.
• Pinned local staging and integrity checks passed for the declared configs, label dictionary, metadata, and model weights.
• The refreshed NVSkills evaluation and signature verification passed.
• Current GitHub lint/test, compliance, and NVSkills CI checks are green.

Scope and limitations

• A live GPU fine-tuning run was not repeated for the final implementation-only simplification.
• MLflow remains optional. Tracking failures propagate from the upstream MONAI run.
• This skill is for development workflows and is not for clinical deployment, clinical interpretation, autonomous diagnosis, or regulatory submission.

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

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci

[wyli]

/nvskills-ci