Detect routing collisions in skill library

Proactive skill routing audit before routing causes user confusion; prevents silent shadowing

• · SKILL.md files from .claude/skills/*/
• · skill name, description, whenToUse, trigger keywords, file-globs
• · reciprocal 'not when' cross-references

• · confirmed collision pairs: {skillA, skillB, sharedTrigger, severity}
• · unresolved pairs (no reciprocal disambiguation)
• · recommended fixes (add cross-ref OR reword trigger)

• · agent discovery (enumerate + extract triggers)
• · agent clustering (intra-cluster pairing, bound O(n^2))
• · agent detect (emit candidate pairs)
• · agent verify (default-reject: confirm real contention)

• · .claude/skills/*/SKILL.md files readable
• · ~82 skills (manageable cluster count)

• · clustering misses cross-domain collision (trigger vocab overlap but different domains)
• · trigger extraction too loose (over-reports collisions)
• · reciprocal 'not when' cross-ref not recognized
• · verify agent hallucinates a collision not in the actual code

• · 5-phase workflow: Discover, Cluster, Detect, Verify, Dedup & Report
• · Collision definition explicit: shared trigger + no reciprocal deflect
• · Clustering bounds O(n^2) to O(n*m) where m = cluster size
• · Default-reject verify: re-read both SKILL.md before confirming
• · Unordered pair dedup (A,B == B,A)
• · Blocking collisions flagged for immediate fix