Improve game dictionaries with dual-judge validation

Extensively improve LexiClash game dictionaries: per language, LLM-generate common missing words, dual-judge them (two independent personas, both must agree), persist survivors to the committed candidate files that the backend verify->promote->heal pipeline consumes, then report. Args: {langs?:string[], lang?:string, limit?:number}.

Complex workflows requiring parallel agents, synthesized judgment, or multi-phase triage.

• · refined artifact

• · Claude Code agent runtime (parallel/fan-out)
• · subagent execution
• · JSON schema validation

• · Artifact must conform to specified schema
• · Claude Code workflow harness

• · File conflicts if fix groups overlap
• · Agent returns data not matching schema → structured rejection

• · Structured schema enforcement
• · Explicit concurrency isolation (disjoint files/resources)
• · Independent verification lens