feat(ai-providers): add OpenCode Zen provider (#1400)

[datlechin]

Summary

Adds OpenCode Zen (https://opencode.ai/docs/zen) as a first-class AI provider, and fixes the base-URL handling that made it (and other gateways) fail as a Custom Provider. Closes #1400.

OpenCode Zen is a clean OpenAI-compatible gateway: POST /zen/v1/chat/completions (Bearer auth) serves every model family it offers, and GET /zen/v1/models returns the list. So it reuses the existing OpenAICompatibleProvider with no new transport.

• New AIProviderType.openCode — display name "OpenCode Zen", default endpoint https://opencode.ai/zen, API-key auth, sparkles icon.
• Registered through the existing OpenAI-compatible loop in AIProviderRegistration (same wiring as OpenRouter); the model picker auto-loads from /v1/models, so there's no hardcoded model list to go stale.
• Added to the "Add Provider" list in AI settings.

Why the reporter's Custom Provider failed (root cause)

OpenCode documents its base URL as https://opencode.ai/zen/v1, but OpenAICompatibleProvider appended /v1/... to the stored endpoint, so a pasted .../zen/v1 became .../zen/v1/v1/chat/completions and 404'd. This is a footgun for every gateway whose documented base ends in /v1.

Fix: a new String.openAIPath(_:) appends /v1/<resource> only when the endpoint doesn't already end in /v1. Used in the three URL builds (chat, test-connection, models). Backward-compatible: the built-in endpoints (https://api.openai.com, https://openrouter.ai/api) don't end in /v1, so their behavior is unchanged; and any custom gateway now tolerates a pasted /v1 base URL.

Verified against the live API

• GET https://opencode.ai/zen/v1/models returns a standard OpenAI list with bare ids (gpt-5.5, claude-opus-4-7, big-pickle, ...).
• POST .../v1/chat/completions with a Claude/GPT/Gemini id returns 401 Missing API key (accepted, just needs a key), and the free big-pickle model returns a full chat.completion with no key. So one OpenAI-compatible transport covers all of Zen's models.

Files

• Models/AI/AIModels.swift — new .openCode case + displayName / defaultEndpoint / symbolName.
• Core/AI/Registry/AIProviderRegistration.swift — .openCode added to the OpenAI-compatible loop.
• Views/Settings/AISettingsView.swift — added to the addable provider list.
• Core/AI/String+AIEndpoint.swift — new openAIPath(_:).
• Core/AI/OpenAICompatibleProvider.swift — chat / test / models URLs use openAIPath.
• TableProTests/Core/AI/StringAIEndpointTests.swift — new.
• docs/features/ai-assistant.mdx, CHANGELOG.md.

Notes

• The model picker lists all of Zen's models (including Claude/GPT/Gemini, which also work through Zen); Zen's real value here is its open models (DeepSeek, GLM, Qwen, Kimi, ...) under one key. Claude/OpenAI/Gemini still have their own first-class providers.
• "OpenCode Go" is a billing tier on the same endpoint/key, so this one preset covers it.
• "OpenCode Zen" is a brand name, so the display name is not localized (matches OpenRouter/Gemini).

Test plan

• Settings > AI > Add Provider lists "OpenCode Zen"; adding it pre-fills the endpoint and the model picker loads from /v1/models after pasting a key.
• A chat turn against an OpenCode model streams a response.
• StringAIEndpointTests passes: endpoint without /v1 gets one appended; endpoint ending in /v1 is not doubled; trailing slash normalized.
• Existing OpenRouter / Ollama / Custom providers still build the same URLs as before.
• A Custom Provider with base URL https://opencode.ai/zen/v1 now works (no doubled /v1).