docs: add OpenCode guide, split Goose out, fix LibreChat accuracy

[sergeiest]

Adds a dedicated OpenCode integration guide, separates Goose onto its own page, and corrects stale facts found while verifying both against the live gateway.

OpenCode — new page

cloud/guides/integrations/opencode.mdx, following the section contract in .github/INTEGRATION_GUIDE_TEMPLATE.md.

The thing that shaped the page: models.dev now carries the nearai provider, so OpenCode needs no provider block at all. The minimum working config is a single model line. The guide then covers the three cases that still need config — models newer than the catalog, trimming the 50+ model picker, and direct completions.

Verified end to end against OpenCode 1.18.31:

Check	Result
Zero-config path (model line + NEARAI_API_KEY)	opencode run → near-ai-ok
Manual provider.nearai.models entry	near-ai-ok on z-ai/glm-5.3-flash
Direct-completions baseURL override	near-ai-ok via glm-5-3-flash.completions.near.ai/v1
opencode models nearai	resolves 32 catalog models

Goose — split onto its own page

cloud/guides/opencode-goose.mdx is now Goose-only. Rewritten rather than trimmed, because verification turned up two errors:

• Config format was stale. The page showed flat GOOSE_PROVIDER / GOOSE_MODEL keys. Current Goose uses nested active_provider / providers. The old form is read for compatibility and migrated on update, so it wasn't broken — just legacy. Both are now documented, since those names remain valid as environment variables.
• The refresh story was backwards. The page treated Goose like OpenCode — stale picker, manual workarounds. Goose's NEAR AI Cloud provider does dynamic model discovery, so new models appear on their own. Manual steps are now the fallback.

Also removed https://glm-5-2.completions.near.ai/v1, which no longer resolves (TLS connect failure).

LibreChat — accuracy fixes

• The smoke test didn't work. max_tokens: 20 against a reasoning model returns content: null with finish_reason: length — 61 of 66 completion tokens went to reasoning_content. The guide then told readers to go fix their key or network path. Raised to 256.
• Reasoning output was silently dropped. Added customParams.reasoningFormat / reasoningKey: reasoning_content; the gateway streams the trace in that field.
• titleModel was a reasoning model, paying for a reasoning pass on every conversation title. Moved to google/gemini-2.5-flash-lite.
• Config version: 1.3.13 → 1.3.16, matching upstream librechat.example.yaml.
• Softened the blanket "do not add a provider field" — accurate for the OpenAI-compatible path, but it forecloses the native Anthropic one.

Across all three

z-ai/glm-5.2 → z-ai/glm-5.3-flash. The old ID survives only as an alias onto its successor, which is now noted inline on each page.

Reviewer notes

• Goose steps are not locally verified. Goose isn't installed on the authoring machine, so those steps come from Goose's published docs and block/goose's own providers.md, stated as such in Sources Checked. Worth a real Goose run before merge. One loose end: NEAR AI Cloud appears in Goose's docs and provider table, but there is no near* file under crates/goose/src/providers/ on main, so the config key nearai is inferred from the NEARAI_API_KEY convention rather than confirmed in source.
• The Goose page still lives at /cloud/guides/opencode-goose. Renaming it to cloud/guides/integrations/goose.mdx would match its siblings but needs a redirects entry. Left out of this PR deliberately.
• /v1/messages is undocumented. The gateway serves a native Anthropic Messages endpoint plus count_tokens — verified working with streaming and tool use — but neither appears in api-reference/openapi.json, so they are absent from the API Reference tab. Upstream spec gap; it also blocks documenting LibreChat's provider: anthropic path.
• The alias <Note> is now duplicated across pages. If it keeps spreading it probably belongs in model-discovery.mdx.

mint validate and mint broken-links both pass.

🤖 Generated with Claude Code