feat!: track MLflow in SQLite and harden the canonical gate

[fmind]

What

Moves MLflow off the deprecated local file store onto a SQLite backend, re-locks every dependency to latest stable, and brings the repository up to the current canonical stack (single mise run all gate, workflow and Dockerfile linting, real filesystem scanning, committed mise.lock).

Breaking: MlflowService now defaults to sqlite:///mlflow.db for tracking and registry. The MLFLOW_ALLOW_FILE_STORE opt-in is gone. Existing ./mlruns runs are not migrated.

Why

• MLflow 3 put the filesystem store in maintenance mode; SQLite is a real SQLAlchemy backend, the one the model registry is designed for, and one env var away from a production database.
• A clean checkout of v5.0.0 failed mise run check today with 25 known vulnerabilities across 4 transitive dependencies — with zero code changes. The lockfile had simply aged.
• check:scan was not scanning: trivy config . silently deferred to any TRIVY_CONFIG exported in the user's shell, and config runs only the misconfig scanner, so 3 of the 4 scanners declared in trivy.yaml never ran.
• Workflows and the Dockerfile were the only code in the repository nothing linted.

How

• MlflowService defaults to SQLite; tests copy a session-scoped, pre-migrated database instead of paying MLflow's Alembic migration per test (7s → 0.07s each).
• New all task composes format → check → test → build; CI runs that one task.
• New checks: check:actions (actionlint + zizmor, with .github/zizmor.yml encoding the tag-pinning policy) and check:dockerfile (hadolint).
• check:scan → trivy --config trivy.yaml fs .; check:vuln skips the editable project and caches.
• Coverage gate 80% → 100%, which is what the suite actually reaches.
• cryptography is overridden past MLflow's stale <50 cap so the PYSEC-2026-3552 fix is installed; the suite and every MLflow job are the proof.
• Weekly full-history security.yml; hardened ci.yml/cd.yml; grouped Dependabot with the chore(deps) prefix cliff.toml already skipped.
• Untracked an accidentally committed 790 KB mlflow.db; deleted the stale vendored copy of the Agent Skills under .gemini/skills/, which had drifted back to teaching just and pre-commit.

Test plan

• mise run all — green (format, 8 checks, 45 passed / 3 xfailed at 100% coverage, wheel + sdist).
• lefthook run pre-commit --all-files — green.
• mise run project — all six MLflow jobs run end to end on SQLite; model registered as version 1 and promoted to the Champion alias.
• pip-audit — no known vulnerabilities.