Backlog: README→website content migration + site gaps after the domain rename

[xiaguan]

Tracking list from the 2026-08-14 README refresh (pegainfer#868) and the GLM-5.2 page work (#19, #20). Ordered roughly by priority.

README → website migration

• API field reference page — pegainfer#868 trimmed the README API table down to a pointer at this site, so the target must exist: completions/chat request fields, per-model sampling limits, and per-model boundaries (e.g. GLM5.2: no logprobs, no n > 1, no penalties).
• Benchmark pages — move the README's RTX 5090 Qwen3 / Qwen3.5 vs vLLM sections (including the KV-offload tiering ladder) into the model pages or a dedicated benchmarks page; shrink the README to a summary + link.
• Kimi-K2 model page — in the README supported-models table but has no page here.
• DeepSeek-V2-Lite model page — same.
• Architecture page — migrate the README mermaid diagram and design-decision notes; redraw the diagram while at it.

Site gaps

• Canonical domain migration — SITE in astro.config.mjs / src/starlightRouteData.ts and the robots.txt sitemap still say open-infer.org. Set up a 301 from open-infer.org → pegainfer.org on the Cloudflare side first, then flip SITE. (The og-card image itself was fixed in fix: og-card shows pegainfer.org after the domain rename #20.)
• Stray old-domain link — blog/speculative-decoding.md links open-infer.org/models/qwen3-4b/... in prose; clean up together with the canonical switch.
• CI for tests — Cloudflare Workers Builds runs on PRs, but nothing runs npm test (build + static + Playwright e2e). Add a GitHub Action.
• GLM-5.2 page follow-ups — wider-fleet numbers when hardware allows: EP8/EP16 decode fleets, 2× prefill scaling, session-affinity routing effects on multi-turn TTFT (current page covers 1P + 1D EP4 only).

🤖 Generated with Claude Code