docs(wiki): coverage read honestly, production caveats, per-process state The documentation site is large and complete, so the wiki collects the operator view spread across it: what each layer catches and misses, what to settle before production, and what changes with more than one worker. Threat coverage reads the published OWASP numbers with their own caveats: 100% is against the corpus rather than the problem, the 6% benign false-positive rate has a known tripping case that will affect security users, and the number deliberately excludes the AI escalation path. It also records that Presidio and tiktoken are not installed by default, so PII detection and budget accounting are narrower than they look. Before production leads with #108: an admin bearer can disable auth and the firewall in a single config/apply request, with non-blocking validation warnings, so an admin token is equivalent to the ability to disable the proxy until that lands. Scaling and state states the consequence of per-process rate limits plainly: with N workers a limit of R is an effective N*R, and one worker's circuit breaker does not stop the others. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>