feat(provider): add Scalattice

[romulushill]

Add Scalattice

Scalattice is an OpenAI-compatible inference marketplace (https://api.scalattice.cloud/v1, Authorization: Bearer slt_...) that serves open-weight models on a GPU fleet.

What's included

• providers/scalattice/provider.toml: @ai-sdk/openai-compatible, SCALATTICE_API_KEY
• providers/scalattice/logo.svg: mark, currentColor
• 18 models matching live IDs from GET https://scalattice.cloud/api/v1/public/models
• Lab metadata for models that did not yet have a models/ file, so host entries can use base_model

Llama 4 Scout, Ministral, Mistral Small, Qwen3.6 and Qwen3.8 are text-only on this host (attachment = false). VL models inherit vision from the lab files. [limit] is only set when it differs from the lab entry.

USD/MTok and served context windows are a snapshot of the public catalog on 2026-08-28. Rates are live marketplace prices and will move with GPU supply.

Docs: https://scalattice.cloud/docs/developers

[github-actions]

Action items

• [high] [possible mistake] models/alibaba/qwen3-vl-30b-a3b.toml:1 - Check: Lab base_model identity must name the real underlying model and stay consistent with established open-weight / first-party facts. Why: This lab id is bare qwen3-vl-30b-a3b with reasoning = false, release_date/last_updated = 2025-04, and weights at Qwen/Qwen3-VL-30B-A3B-Instruct, while sibling VL labs use an -instruct id (qwen3-vl-8b-instruct, qwen3-vl-235b-a22b-instruct), OpenRouter/Kilo/SiliconFlow Instruct peers use ~2025-10 dates and non-reasoning, and first-party providers/alibaba/models/qwen3-vl-30b-a3b.toml treats the same id as reasoning = true. Scalattice then inherits that lab via base_model = "alibaba/qwen3-vl-30b-a3b", so catalog consumers get a conflicted identity. Action: Author lab metadata as the Instruct open-weight model (e.g. models/alibaba/qwen3-vl-30b-a3b-instruct.toml) with dates/capabilities aligned to that checkpoint; point Scalattice base_model at it. Do not publish a bare qwen3-vl-30b-a3b lab that contradicts the Alibaba hybrid/thinking SKU of the same id.
• [medium] [possible mistake] models/mistral/ministral-8b-instruct-2512.toml:15 - Check: New lab [limit] values should match the model’s documented open-weight / peer defaults. Why: New lab sets output = 32_768 (context 262_144), but established hosts for this Ministral 3 8B 2512 checkpoint use much larger output caps (e.g. OpenRouter 209_715, Cortecs 256_000). Understating lab defaults will under-limit every override-only host that inherits them. Action: Verify max output (and context) against Mistral/HF docs for Ministral-3-8B-Instruct-2512 and align the lab [limit] (or document a deliberate lab default if 32k is intentional).