chore(cerebras): replace Gemma 4 with Qwen 3.8

[sebastiand-cerebras]

Summary

• remove Gemma 4 31B from the Cerebras provider catalog
• add qwen-3.8-27b using the canonical Alibaba model metadata
• override the Cerebras 65,536-token context, 32,768-token output, and text/image input contract
• add $0.99/$1.49 per million token pricing and none/low/medium/high reasoning controls

Sources: Cerebras launch catalog contract and https://huggingface.co/Qwen/Qwen3.8-27B

Validation

• bun validate
• built the web catalog and inspected the normalized Cerebras entry

[github-actions]

The redundant identical limit.output is a clear override-only violation; I'll also flag the missing Cerebras-side citations for the material pricing/limits/reasoning claims.

Action items

• [medium] [violation] providers/cerebras/models/qwen-3.8-27b.toml:13 - Check: Provider base_model files must be override-only (no duplicated identical fields). Why: limit.output = 32_768 matches models/alibaba/qwen3.8-27b.toml exactly, so deep-merge already supplies it; restating it breaks the override-only rule. Action: Drop output from [limit] and keep only the Cerebras-specific context = 65_536 override (and only modality keys that actually differ, e.g. input without restating identical output = ["text"]).
• [low] [possible mistake] providers/cerebras/models/qwen-3.8-27b.toml:5 - Check: Data-changing PRs should cite direct provider pricing/docs/API mapped to claims. Why: Cost (0.99/1.49), context 65_536, and reasoning_effort none|low|medium|high are material host facts, but the PR only cites an unnamed “Cerebras launch catalog contract” plus Hugging Face—unlike existing Cerebras peers that link inference-docs.cerebras.ai model/reasoning pages. Action: Add a leading comment and/or PR-body URLs for the Cerebras model page (or launch contract) that cover pricing, limits, and accepted reasoning_effort values for qwen-3.8-27b.

[github-actions]

Action items

• [high] [violation] providers/cerebras/models/qwen-3.8-27b.toml:13 - Check: After base_model, provider files must be override-only and must not restate identical base values. Why: limit.output = 32_768 matches models/alibaba/qwen3.8-27b.toml exactly; plain objects deep-merge, so restating it is a redundant override. Action: Remove output from [limit] and keep only the real delta (context = 65_536).
• [high] [violation] providers/cerebras/models/qwen-3.8-27b.toml:17 - Check: After base_model, do not restate identical modality values. Why: modalities.output = ["text"] is identical to the lab entry; only the input list differs (video dropped). Action: Drop output under [modalities] and keep only input = ["text", "image"].
• [medium] [possible mistake] providers/cerebras/models/qwen-3.8-27b.toml:5 - Check: Relay reasoning_options must follow this host’s API using the lab/same-model peer baseline, not an invented effort set. Why: First-party Qwen3.8 effort is low/medium/xhigh (high is only an Alibaba alias → xhigh). Multiple qwen3.8-27b peers author none/low/medium/xhigh, and at least one live probe rejects high. This file uses none/low/medium/high with only an unverified “launch contract” note and no public Cerebras model/reasoning URL. Action: Confirm Cerebras accepts exactly these wire values (especially high vs xhigh); if the API uses xhigh, switch to that set; if high is correct, cite the Cerebras docs or contract snippet that lists the accepted reasoning_effort values.
• [low] [possible mistake] providers/cerebras/models/qwen-3.8-27b.toml:8 - Check: Material pricing/limit claims should point to direct provider sources. Why: Cost $0.99/$1.49 and 65,536 context are host-specific facts, but the PR only cites an unnamed “Cerebras launch catalog contract” plus Hugging Face (lab metadata), with no Cerebras pricing/model page URL. Action: Add a direct Cerebras pricing or model-doc citation (PR body and/or leading TOML comment) that supports cost, context, and modalities.