fix(kb): term:tier told every AI session a typo grants shell

[cuttlefisch]

term:tier is shipped guidance about a security control, delivered to AI
sessions as authoritative. It said:

(lowercase values only — an unrecognised value currently falls back to shell)

Both halves are false, and the first is false in the least safe direction.

The dangerous half

config.rs:754 errors with unknown AI permission tier on an unrecognised
value — MAE refuses to start rather than resolving it to anything. There is
even a test at config.rs:1822 asserting an unknown tier "must be rejected,
never resolved".

So the node was describing the former bug as current behaviour. An agent
reading it would believe a typo in auto_approve_tier silently grants shell
— exactly backwards from a control that fails closed.

The other half

PermissionTier::parse (crates/ai/src/types.rs:66) does
s.trim().to_ascii_lowercase(), so values are case-insensitive. Harmless next to
the first, but the same drift, and it would have someone "fixing" a config that
already worked.

Bound to tests so it cannot drift back

Three tests, asserting the dangerous claim rather than the current wording —
the node can be rephrased freely, and only reintroducing the falsehood fails:

1. the body must not say an unknown value falls back to a tier;
2. it must positively state the refusal — a node that simply omitted the
   question would satisfy (1) while still leaving a reader to assume some silent
   default;
3. it must not claim lowercase-only.

All three were verified failing against the old text before the fix, not
merely passing after it:

tier_node_does_not_claim_an_unknown_value_falls_back_to_a_tier ... FAILED
tier_node_states_that_an_unknown_value_is_refused ... FAILED
tier_node_does_not_claim_lowercase_only ... FAILED

Why this is worth a PR of its own

This is the hazard CLAUDE.md principle #17 and SECURITY.md's own note describe: a
false claim in shipped guidance primes every future session, and is worse than no
claim. The same file already carries an honest NOTE about ai_tier being
badge-only — that part was accurate and is untouched.

No overlap with any PR currently open.

🤖 Generated with Claude Code