Phase 0 docs 2/3: content accuracy fixes (defaults, GPU tables, curl snippets, regions)

[kylegani]

Context

Part of Phase 0 of the Agentic Cerebrium initiative: making Cerebrium fully usable by coding agents (Claude Code, Cursor, Codex, Copilot). Agents consume these docs through the Mintlify MCP server, llms.txt, and .md exports, and today the entry points are broken: the MCP handshake 406s at cerebrium.ai/docs/mcp, several proxy rewrites silently 404, and key config defaults documented here contradict the backend, which steers agent-authored cerebrium.toml files wrong. Every content claim in this stack was verified against dashboard-backend source (file:line refs below) on 2026-08-04.

This PR is layer 2 of a 3-PR stack (one reviewable unit per layer, merge bottom-up; merging a lower layer auto-retargets the ones above):

1. Phase 0 docs 1/3: MCP proxy wiring and contextual agent config #308 MCP proxy wiring + contextual agent config
2. Phase 0 docs 2/3: content accuracy fixes (defaults, GPU tables, curl snippets, regions) #309 content accuracy fixes (defaults, GPU tables, curl snippets, regions)
3. Phase 0 docs 3/3: delete intentionally orphaned pages #310 delete intentionally orphaned pages (+ redirects)

Related: the CloudFront half of the MCP fix (cerebrium.ai apex routing) is CerebriumAI/eks#431; this repo's vercel.json covers the docs Vercel project.

Middle layer of the Phase 0 docs stack: content-accuracy fixes only, all verified against dashboard-backend source. Builds on the proxy-wiring layer (#308).

Changes

• Cooldown default corrected 1800 -> 10 in the toml-reference table and sample (create_app/api/api.go:63 and the github/partner siblings; 1800 appears nowhere in backend code).
• replica_concurrency contradiction resolved with backend truth: default is 1 for GPU compute types, 100 for CPU (create_app/api/api.go:134-140); both toml-reference.mdx and batching-concurrency.mdx now say so.
• response_grace_period corrected 3600 -> 900 and max_replicas 2 -> 1 (same backend defaults file).
• GPU tables reconciled against the canonical 13-value enum (go-build-service/src/libs/apps/inputs.go:28): BLACKWELL_B300 removed everywhere (backend rejects it), CPU and INF2 rows added, AMPERE_A100 rows versioned, plan-availability section added (hobby 6 / standard 12 / enterprise 13 per plans.go).
• eu-north1 vs eu-north-1 "duplicate" was REFUTED: both are real distinct regions (Nebius Finland vs AWS Stockholm, region.go:76-78); rows in multi-region-deployment.mdx are now annotated with locations instead of one being deleted.
• Fixed five broken curl snippets (stray or unterminated quotes, angle-bracket URL, unquoted &) in endpoints/async.mdx, webhook.mdx and streaming.mdx (curl block only; the streaming falcon-link change lives in the layer above, Phase 0 docs 3/3: delete intentionally orphaned pages #310).
• localhost link in defining-container-images.mdx replaced with the relative page path.

How to test this layer

1. Extract each bash code fence from endpoints/async.mdx, webhook.mdx and streaming.mdx and run bash -n on it; all parse (they previously failed shell parsing). Or copy each fixed curl snippet and run it against a real app with a valid API key.
2. Spot-check content claims against backend: create_app/api/api.go:63 (cooldown 10), :73 (grace 900), :64 (max_replicas 1), :134-140 (concurrency 1 GPU / 100 CPU), inputs.go:28 (13 GPU types).
3. JSON sanity unchanged by this layer: python3 -m json.tool docs.json > /dev/null && python3 -m json.tool vercel.json > /dev/null.

Stack position: 2 of 3 (proxy wiring + contextual config -> content accuracy -> orphan cleanup)

🤖 Generated with Claude Code