feat: native frozen-fit HDBSCAN deployment for saved Judges

This was generated by AI during triage.

Goal

Add density-supported frozen-fit HDBSCAN prediction through ClusterFit.classify, saved-Judge prediction and patient OOD. Related #91 explicitly excludes HDBSCAN; this issue supplies its native deployment mechanism.

Acceptance

Document density semantics before code. Retain sklearn fitting, PCA and fitted partition unless compatibility is explicitly resolved. Persist prediction state without refitting on load. Keep backend-local/refined-global mappings, distinguish density groups, promoted noise representatives, rejected queries and deployable counts. Cover empty, all-noise, single-cluster, batching and legacy-state cases; preserve selection and L/R/F plus KMeans/DECAF behavior.

Delivery and isolation

Baseline 2d5544ec; branch feat/native-hdbscan-deployment in isolated atom-hifi-hdbscan-deploy worktree. Preserve original checkout, shared environment, PIT-WATER study artifacts. Small behavioral fixtures and bounded reference-only canary only; no lifetime run. Deliver committed docs, runnable saved-Judge example and later-arm integration instructions. No merge without explicit approval.