[Feature] Consider opt-in canonical frontier dedup for behavioral-frontier

[EdwardIrby]

Summary

Add an optional canonical deduplication mode to behavioral-frontier for larger explorations, while keeping the current trace-first behavior as the default.

Problem or proposal

behavioral-frontier currently deduplicates by full snapshot trace. That is useful for agent debugging because it preserves exactly how a generated behavioral spec reached a frontier or finding.

For larger generated skill/spec exploration, equivalent runtime frontiers may be reached through multiple event histories. An opt-in canonical dedup mode could reduce exploration size when the agent needs coverage or scale more than complete explanatory traces.

Use case

An agent iteratively generates skills, behavioral specs, and tool-call trigger sequences. During debugging and repair, trace-first output should remain the default because the selected-event history explains the failure. During broad search, the agent may want to collapse equivalent frontiers to keep exploration tractable.

Affected areas/files (if known)

• src/behavioral-frontier/explore-frontiers.ts
• src/behavioral-frontier/behavioral-frontier.schemas.ts
• src/behavioral-frontier/tests/*

Expected behavior

• Keep current trace-prefix deduplication as the default behavior.
• Add an explicit option such as dedupe: "trace" | "frontier" or similar.
• trace preserves current behavior and full replay evidence.
• frontier canonicalizes equivalent pending/frontier states and avoids exploring duplicate states.
• Output should remain useful for agent repair, ideally retaining at least one representative trace for each canonical frontier.

Acceptance criteria

• Existing trace-first behavior remains unchanged by default.
• New opt-in canonical mode has focused tests showing duplicate frontier collapse.
• Reports clearly identify the dedupe mode used.
• Representative traces are still available for findings.

Notes

This is intentionally a maybe-later enhancement. Current trace-based behavior is preferred for agent debugging and iterative spec repair.