Summary

Introduce moderation v2 backend foundation in ClawHub: normalized verdict/reason/evidence model, deterministic static scanning, publish-time moderation derivation, and backfill support.

Why

Trust decisions need to be consistent and explainable across static, VT, and LLM signals while preserving compatibility with existing moderation fields.

Focused scope

This PR is scoped to one theme: core moderation v2 backend pipeline.

What changed

• Added normalized moderation fields in convex/schema.ts.
• Added canonical reason code + verdict utilities in convex/lib/moderationReasonCodes.ts.
• Added moderation engine in convex/lib/moderationEngine.ts.
• Integrated deterministic static scan in publish/backfill paths (convex/lib/skillPublish.ts, convex/skills.ts, convex/vt.ts).
• Updated moderation/public safety logic (convex/lib/moderation.ts, convex/lib/public.ts, convex/lib/skillSafety.ts).
• Follow-up fixes included:
  • escalateByVtInternal moderation flag overwrite bug
  • backfill cursor skip edge case
  • child_process false-positive fallback in scanner
  • rule name alignment to suspicious.nonstandard_network

Local validation

• bun run lint:oxlint
• bunx vitest run convex/lib/moderationEngine.test.ts convex/skills.rateLimit.test.ts

AI assistance transparency

• AI-assisted: Yes (implemented with Codex assistance)
• Testing level: Targeted local validation on touched modules
• I reviewed the final diffs and understand the behavior changes.