Experiment: Hybrid Semantic + Structural Grouping

[monadic]

Summary

We should test a hybrid semantic + structural grouping model that combines meaning signals with repo/deployer structure for higher-trust grouping.

Context from Browse code by meaning: semantic browse works best when labels are short, comparative, and explainable.

Experiment Scope (Large)

• Build a prototype grouping engine that combines:
  • semantic-ish signals (resource/name/label tokenization)
  • structural signals (owner chain, deployer type, trace path, repo path fragments when available)
• Emit group evidence payloads suitable for both CLI display and JSON debugging.
• Add snapshot/fixture coverage for grouping determinism.

Deliverables

• Experimental grouping API (internal) with versioned output schema.
• CLI integration in one experimental command path.
• Goldens for representative mixed-tool fixtures.
• Short design note documenting algorithm and tradeoffs.

Success Criteria

• Grouping quality better than structure-only baseline in fixture review.
• Deterministic outputs on fixed inputs.
• Explainability: each grouped resource has attributable evidence.

Risks

• Over-grouping unrelated resources.
• Label noise if token hygiene is weak.
• Performance degradation on large fleet snapshots.

Mitigations

• Hard limits and fallback to baseline grouping.
• Keep algorithm deterministic and bounded.
• Ship as opt-in experiment until confidence is high.