`extension_dom_snapshot`, `extension_inspect`, `extension_eval`,
  `extension_storage`, `extension_open`, `extension_assert`,
  `extension_reload`, `extension_wait`, `extension_doctor`,
  `extension_list_extensions`) now fence what it wrote between
  `<untrusted-data-ID>` and `</untrusted-data-ID>`, with a fresh random
  `ID` per call named in `untrusted.boundary`, so text on an inspected
  page that says to publish or revoke reads as data, not as an
  instruction. The fence covers `value`, `error`, `warnings` and `hint`,
  because page text also reaches error messages and hints. A tag the
  page writes, even one that guesses the ID, is escaped in the raw text
  and parses back to the same characters, so every answer is still one
  JSON envelope with the same fields. Build, analyze and dev output stay
  unfenced: they carry the project's own source.