Scoring & Survivorship Quality Upgrade — MST splitting, quality-weighted golden records, field provenance

[benzsevern]

Coming in the next release

PR #30 adds major improvements to cluster quality and golden record generation.

Cluster Quality

• MST-based auto-splitting — oversized clusters are automatically split by removing the weakest edge in the minimum spanning tree. Guaranteed to disconnect. Recursive until all sub-clusters are within max_cluster_size.
• Cluster quality labels — every cluster gets cluster_quality: "strong", "weak", or "split". Weak clusters (large gap between min and avg edge weight) get confidence downgraded by 0.7x.
• Configurable — auto_split, quality_weighting, and weak_cluster_threshold in GoldenRulesConfig.

Quality-Weighted Survivorship

All 5 merge strategies now accept optional quality weights from GoldenCheck:

• most_complete — ties broken by source quality
• majority_vote — votes weighted by quality
• source_priority — quality as tiebreaker within same priority level
• most_recent — confidence scales by quality
• first_non_null — picks highest-quality source first

When GoldenCheck is not installed, everything works exactly as before (optional-but-rewarded).

Field-Level Provenance

build_golden_record_with_provenance() tracks, per cluster, which source row contributed each field value, what strategy was used, and what candidates were available. Provenance is serialized in the lineage JSON sidecar as a golden_records section.

Backward Compatible

All new parameters have defaults. Existing callers unchanged. 85 tests pass.

PR: #30