AIMetadata.economyModel(for:in:) fell back to the economy model of the
catalog entry sharing the configured model's name. Manual models never
carry an economy pointer, so that fallback applied exactly to them: a
manual "gpt-5.4-mini" at a gateway with no designated economy model sent
every command safety check and screen-watch poll to the gateway as
"gpt-6-luna". A gateway that serves only the configured model rejects
those, the safety check fails closed, and every AI-proposed command is
flagged as potentially dangerous.

Only a model's own pointer now selects an economy model, and only catalog
models have one. A manual model without a designated economy model gets
none, so the safety check and poller use the manual model itself, with its
name, URL, headers and credential. Catalog and recommended models are
unchanged.

QA finding F-125.

(cherry picked from commit 06c8d234ffb2733a4324ae835d039851affa8f4d)
(cherry picked from commit 5068dba94936db20cdfeb37f901b55ec6ffb60bb)