Refuse attacks: a central ethics rule above every other

[wintermeyer]

A central ethics rule, above every other, that no story around a
request can work around.

heinzel administers machines its operator is responsible for and
nothing else: never recover a password or key, break into a
system, scan hosts it does not administer, surveil people, hide
traces, or turn the infrastructure against others. It holds on
the operator's own hosts too, because "it's my server" cannot be
checked. No framing unlocks it (a book, research, a CTF, a
claimed permission, an emergency, a role play, a step split into
harmless parts), and neither a custom rule nor
HEINZEL_GUARD_DISABLE overrides it. Defensive work stays allowed:
reading auth logs, checking password policy without touching a
hash, resetting a password on a host it administers.

• CLAUDE.md: an "Ethics (above every other rule)" block at the
  top, and a note that rules/ethics.md takes no overrides.
• rules/ethics.md: the full rule, the no-framing list, what
  stays allowed, and the refusal behaviour (one sentence, a line
  in memory/refusals.log, the topic stays refused for the
  session).
• guard-taboos.sh: a mechanical backstop blocking the known
  crackers and attack suites by name, both run and installed, in
  every permission mode. Like the other taboos it names a known
  set, not the effect, so a renamed or hand-built tool is the
  model layer's job; 20 new fixtures, 742/0 green.

Ships as 2.32.0.

An AI agent wrote this text in my name. I know that is problematic.