[Improve] Align DeepSeek V4 Pro provider metadata

[zoomote]

​Opened on behalf of @taltas. Follow up by mentioning @roomote, in the web UI, or in Discord.

What changed

• Labels the stable deepseek-v4-pro alias as DeepSeek V4 Pro 0813 only for the first-party DeepSeek API and OpenCode Go proxy route.
• Adds Baseten's published deepseek-ai/DeepSeek-V4-Pro preview-weight model without applying the 0813 checkpoint label, and leaves Fireworks AI on its existing preview-weight metadata.
• Corrects first-party V4 Pro to text-only, fixes OpenCode Go pricing to $0.435 input, $0.87 output, and $0.003625 cache reads, and omits an unverified Baseten cache-write price.
• Adds regression coverage for the first-party-versus-self-hosted distinction and the corrected metadata.

Why this change was made

DeepSeek swapped the checkpoint behind its existing deepseek-v4-pro API alias to 0813 without introducing a new model ID. OpenCode Go proxies that first-party route, while Fireworks AI and Baseten self-host the previously published V4 Pro preview weights, so their labels must remain undated.

Impact

Users see accurate version labels and cost/capability metadata for each route without a fabricated dated model ID. DeepSeek warns that it plans a significant overall price increase, so these hardcoded rates may need a follow-up update.

Related PRs

• Documentation: [Docs] Explain DeepSeek V4 Pro provider versions Zoo-Code-Docs#27

Screenshots

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[zoomote]

Addressed in 498a89d: limited 0813 labeling to DeepSeek and OpenCode Go, restored Fireworks preview-weight metadata, kept Baseten undated, corrected V4 Pro image support and OpenCode Go pricing, removed the unverified Baseten cache-write price, and expanded regression coverage. The draft PR metadata and final proof now match the shipped distinction.