Add entity resolution to deduplicate and merge cross-source observations

Read docs/_drafts/06-roadmap.md and inspect the existing Upsert behavior first, focusing on the Tier 1 exact-URN scope. Done means observations merge idempotently into an existing entity, provenance is tracked, and the merge behavior is defined; Tier 2 and Tier 3 are follow-ups.

Context

Compass stores entities from multiple sources as separate records. A Kafka topic ingested from two different systems creates two unrelated entities with different URNs. The graph is fragmented — context assembly, impact analysis, and search all operate on disconnected duplicates.

Entity resolution is the mechanism that matches incoming observations against existing entities, merges properties, and maintains unified identity. This is the prerequisite for a coherent knowledge graph.

Scope

Tier 1: Exact URN Match

• When an observation arrives with a URN that already exists, merge properties into the existing entity
• Track provenance: which source contributed which properties
• Idempotent — re-sending the same observation must not create duplicates or mutate state unexpectedly
• This is what Upsert partially does today, but without provenance tracking or merge strategy

Tier 2: Heuristic Matching

• Match observations where URN differs but type + name + source pattern suggests the same logical entity
• Configurable matching rules (e.g., "bigquery table names map to dbt model names via this pattern")
• Candidate scoring with a confidence threshold

Tier 3: Semantic Similarity (follow-up)

• Use embedding similarity to catch non-obvious matches
• Only viable after the embedding pipeline has indexed sufficient entities
• Should be a signal fed into Tier 2 scoring, not a standalone matcher

Merge Strategy

• When a match is found, merge properties from the new observation into the existing entity
• Default: last-write-wins per field
• Track which source contributed which properties (provenance)
• Resolution audit log: record what was matched, merged, and why

Design Considerations

• Resolution must be idempotent
• Meteor sends raw observations, Compass resolves — keep the interface simple
• Start with Tier 1 (exact URN match with provenance). Ship it. Tier 2 and 3 are follow-ups.
• Graph-aware ranking (#237) depends on a coherent, deduplicated graph — this should ship first

References

• Compass Roadmap — Entity Resolution