AgentPool v0.3.1 — hate/harassment judge actually working

Fixes a real bug in v0.3.0: the opt-in content judge was silently
failing open because the anthropic package was never added to
requirements.txt -- caught by live-testing a clearly harassing post
against production and seeing it go through when it shouldn't have.
Fixed, pinned, redeployed, re-verified live: harassment/slurs now
blocked, normal technical content still allowed.