docs(agents): require line-by-line audit of AI-generated destructive PRs

Add two rules learned from earlier PR review pain.

AI-generated PRs touching find / mass-delete / sandbox-container / bundle-id
matchers must be reviewed per branch (fallback branches regress to broad
globs), per protected-path coverage, and per user-confirmation step. Ask the
contributor to narrow matchers to exact constants instead of approving on
plausibility.

Test-orphan grep must cover the whole repo including top-level entry scripts,
not only lib / bin / scripts / tests. Trace string-built call sites, eval,
declare -f, compgen, and write-only helpers' data flow before declaring a
function dead.