Add Telnyx as a model provider

[a692570]

What this changes

Adds Telnyx as a model provider, so OpenLive can talk to open-source LLMs hosted on Telnyx Inference GPU infrastructure.

Telnyx Inference is an OpenAI-compatible endpoint:

• Base URL: https://api.telnyx.com/v2/ai
• Auth: Bearer $TELNYX_API_KEY
• Lists models live at /v2/ai/models (OpenAI { object: "list", data: [...] } shape)
• Speaks the Chat Completions wire (/chat/completions)

Because it speaks the universal /chat/completions dialect with Bearer auth, the existing openai-chat adapter works unchanged. This PR is a single BUILTIN_PROVIDERS entry plus a seed MODEL_SNAPSHOT — the same shape as the DeepSeek / Together / Fireworks / Perplexity entries added in v0.1.9.

Seed models (verified current against the live /v2/ai/models endpoint, 2026-07-11)

Model ID	Params	Context	Notes
moonshotai/Kimi-K2.6	1.0T	256K	Zero-click default. Recommended for voice AI (thinking disabled). Fast + vision-capable.
zai-org/GLM-5.2	753.9B	1M	Coding, reasoning, 1M context window.
MiniMaxAI/MiniMax-M3-MXFP8	428B	1M	Cheapest while maintaining high intelligence.

These are seeds so the picker has a default before the live /v2/ai/models fetch loads; the fetch overrides them the moment the key is set. catalogId: "telnyx" is set for forward-compat if Telnyx ever lands in the models.dev catalog (today it isn't, and the code path falls back to the live fetch / snapshot gracefully).

How I tested it

• pnpm typecheck passes clean across all workspace packages.
• Verified the live /v2/ai/models endpoint returns OpenAI-compatible shape ({ object: "list", data: [...] }) with 21 models, so the fetchOpenAICompat path in packages/harness/src/models.ts picks it up without adapter changes.

Checklist

• pnpm typecheck passes
• One focused change, not several bundled together
• Screenshots added for UI changes (N/A — no UI changes; the provider shows up in the existing Settings picker automatically)

Known limitations

The vision badge in the picker reads from models.dev modalities.input metadata or the OpenRouter architecture.input_modalities shape. Telnyx's /v2/ai/models returns vision capability as a boolean is_vision_supported field instead, so the badge may not light up for Telnyx models even when they support images. The models still work for voice + camera; only the picker badge is affected. I kept this PR scoped to provider registration only — vendor-specific metadata parsing is a separate concern.