Decide the speech-to-text default for installs without a component picker

[gregjkal]

Part of #55. This is a product decision for the maintainers; the fallback if the default stays is the README sentence tracked in the docs issue.

Observation

On a fresh macOS install from the DMG (built from master 3121cdd, launched with a scratch HOME), the gateway writes its default gateway.toml with two speech-to-text models and a default profile that loads both:

item	size
whisper.cpp Metal runtime	8.8 MB
whisper-base-en (interim)	145 MB
whisper-small-en (final)	480 MB

Download and engine load took 42 s on a fast link. Nothing on screen announces the download before it starts, and the installer never asked.

The README describes STT as an installer component ("a configuration gate, since the runtime and models download on demand") that a user can decline. That component exists only in the Windows NSIS installer. A DMG has no component picker, and neither does the AppImage or the .deb, so every non-Windows install gets STT on.

Ask

Decide one of these and file the follow-up:

1. Off by default outside the Windows installer. The generated gateway.toml on macOS and Linux omits the [[stt_model]] entries; the user turns speech on from the Config UI, which then downloads on demand. This matches the README's "download on demand" wording.
2. Keep on by default, but prompt. First launch shows the download size and a Continue / Skip choice in the Workshop before the gateway starts pulling models.

Either way, the STT sentence in the README Downloads section (docs issue) gets updated to match.

Done means

• Decision recorded here.
• Follow-up PR filed for the chosen option, or this issue closed with the docs sentence as the resolution.