Security hardening pass: CSP docs, dependency audit workflow, abuse heuristics

[PAMulligan]

Summary

A focused sweep of the worker and widget for security baseline gaps: CSP guidance for host sites, automated dependency auditing, rate-limit tuning, and basic abuse-detection heuristics on top of the existing per-IP limits.

Motivation

The SECURITY.md file exists and KV-based rate limiting is in place, but there's no proactive defense-in-depth posture. Closing the obvious gaps now is much cheaper than after a public incident.

Acceptance Criteria

• CSP recipe in docs (script-src, frame-ancestors, connect-src to worker domain) for sites embedding Claudius
• pnpm audit runs in CI on a weekly cron and on every PR; high/critical vulns fail the build
• Dependabot configured for npm + GitHub Actions ecosystems with grouped PRs
• Rate-limit configs reviewed: per-IP, per-conversation, per-tenant, plus a global circuit breaker on Anthropic spend
• Abuse heuristics: prompt-injection detector (cheap classifier), profanity spike alarm, sudden conversation length anomaly
• Threat model document covering top 10 risks and current mitigations