[FEATURE] Opt-out for the mid-turn "untrusted external data" wrapper on <channel> messages (using env var or per-source trust flag)

Ask: an opt-out for the mid-turn <channel> framing.

Please see example in next comment.

Possible Fix:

• an env var (e.g. CLAUDE_CHANNEL_TRUSTED_SOURCES=plugin:world:comms)
• or a trusted: true field on the channel capability / notifications/claude/channel meta.

Why it is wrong for us: the trust of a channel does not change with timing. Our plugin allowlists senders and conversations before a message is ever notified, so the operator's own DM is the only thing that can arrive — mid-turn or not.

The wrapper costs ~105 tokens per message and has made agents refuse or "defend against" their operator - see previous bug report #78399, closed as duplicate of #71069, which the stale bot closed - this is still an issue over a month later.

Scope: opt-in per source only; default behaviour unchanged.