Prepare LabCraft-Eval v0.1.2 integrity release

[jang1563]

What changed

• hardens benchmark scoring, task contracts, model identity, and runtime provenance across the 16 integrity commits
• removes application-only positioning, personal HPC paths, and exact scheduler identifiers from public-facing files
• generalizes the Slurm environment defaults while preserving runnable HPC workflows
• aligns package, citation, Zenodo, lockfile, and changelog metadata on v0.1.2
• adds a public-surface hygiene regression test
• includes the model registry in the Docker image and adds a packaging regression test
• documents v0.1.2 as a metadata-only Hugging Face release while preserving the frozen v0.1.1 scored snapshot

Why

The public main branch predates the current scorer, provenance, and model-identity fixes. This PR promotes the verified integrity branch and makes the repository suitable for a reader-facing public release without rewriting the frozen historical result artifacts.

Validation

• uv lock --check
• uv sync --frozen --extra dev --extra analysis --extra providers
• uv run --frozen python scripts/model_matrix.py validate
• uv run --frozen ruff check .
• uv run --frozen pytest -q (486 passed)
• CFF 1.2.0 validation
• shell syntax checks for Slurm and portfolio runners
• wheel build and isolated package smoke (labcraft 0.1.2)
• metadata-only HF schema 0.3.0 export, validator, and upload dry-run
• git diff --check

Local Docker validation was not available because the desktop Docker daemon was not running; the PR CI Docker job is the authoritative check.

CodeQL visibility note

CodeQL completed extraction and analysis of 84/84 Python files and 2/2 workflow files, but SARIF upload is unavailable while this repository is private. After merge, the repository will be made public and the main CodeQL push run must pass before tagging or announcing v0.1.2.