harness-similarity

Run npx ruflo metaharness similarity with --a and --b pointing to two fingerprint JSON files, or use --a-key and --b-key for records persisted in memory. The command returns an overall score in [0,1] plus cosine, categorical, and jaccard components.

It uses a weighted composite: 0.60 times cosine similarity over a 9-dimension numeric vector, 0.25 times categorical agreement over 4 enum fields, and 0.15 times jaccard over the agent_topology array. This design prevents high cosine scores from masking unrelated verticals.

Use the --alert-below flag with a threshold such as 0.5 to emit an alert when two harnesses are too dissimilar. The --per-dimension flag shows which components (cosine, categorical, jaccard) drive the difference.

Malformed inputs never throw an exception. The function returns an object with degraded set to true and a reason field, and the CLI exits with code 2, preserving the graceful-degradation constraint from ADR-150.

No. It is a pure TypeScript/JavaScript module with zero static @metaharness/* imports and adds no new entries to package.json, keeping it removable and optional per the ADR-150 architectural constraints.