feat(talk): report which speech-to-text path talk.catalog uses

* feat(talk): report which speech-to-text path talk.catalog uses

talk.catalog's transcription group now carries an optional `source`:
`streaming` (Voice Call streaming provider), `voice-model`
(agents.defaults.voiceModel), or `media-understanding` (tools.media audio
models). buildTalkTranscriptionConfig, which already picks the path, sets it.
When Talk uses another path, optional `mediaProviders` reports the
tools.media audio providers' readiness, since voice notes still use them.

Settings -> Talk reads both: the status names the path Talk uses, and the
speech-to-text model row shows the picked model's readiness, checked
against voice-note readiness, with the key hint and Open Models on the same
row. The row also says whether the model drives Talk or only voice notes.
Older Gateways omit the fields; the page then keeps the stated precedence
and the existing check.

* test(gateway): add talk transcription source to the narrow-mutations mock

* fix(talk): describe streaming without its config location, drop source fallback

The streaming status no longer says the provider was set in the Voice Call
plugin: with nothing configured, the resolver can pick a provider
automatically. The media-understanding status names the configured models
rather than "the model below", since the Gateway can fall back past an
unready picked model.

Settings -> Talk consumes `source` directly: the picked model drives Talk
only on the media-understanding path, and a catalog without `source` gets
no readiness check against another path's providers. The precedence text
stays as the description until the catalog answers.

* feat(talk): split Talk and voice-notes speech-to-text rows, report streaming-auto

talk.catalog transcription source gains streaming-auto for the automatic
pick from configured keys, so streaming now means the Voice Call streaming
config names the provider. Settings → Talk shows a read-only Talk row that
links to the setting that chose its provider, and a Voice notes row with
the model picker.

* fix(talk): send Talk speech-to-text setup to Voice Call settings

Every streaming transcription provider reads its key from the Voice Call
streaming config, while only some (OpenAI, xAI) also accept a Models key;
a Models key cannot enable Deepgram streaming. The Talk row's automatic and
not-set-up cases now open Voice Call settings on its configuration tab
instead of Models.

* fix(talk): link voice-notes keys to Models only when Models can add them

Voice notes read a provider's key from Models config, sign-in profiles or
the Gateway environment, but Settings → Models lists model providers only,
so media-only providers such as Deepgram and ElevenLabs have no entry there.
The Talk page now reads the providers Models can connect (models.authStatus)
with its catalog, keeps Open Models for those, and otherwise says the key
goes in the Gateway's environment.

* fix(talk): publish the catalog without Models, report unnamed selection as automatic, name the fallback provider

- The Talk page read talk.catalog and models.authStatus together, so a
  slow Models answer held back every Talk fact. The catalog now publishes
  on its own and the Models key hint follows when Models answers.
- talk.catalog reported `source: "streaming"` whenever the Voice Call
  config held provider options, even with no provider named. The runtime
  picks automatically in that case, so it now reports `streaming-auto`.
- On the voice-notes path the Talk row said it used the model below even
  when another audio model was serving as fallback. It now names the
  provider in use when that is not the picked model's provider.

* refactor(talk): drop unused setUp key and one-use source helper

* fix(talk): type the Talk row state instead of casting its links

* ci: rerun CI after Maven download failure

* ci: rerun CI after hosted runner lost connection

* ci: rerun CI after flaky chat notices test