Group Guide whitelisted models by provider and clarify selection

[PunchTheDev]

Summary

Guide Whitelisted models section rebuilt against the 3-pillar rubric (Beautiful · Seamless · Understandable instantly). Closes the corresponding BACKLOG.md row.

Changes

• QuickstartGuide.tsx L604–676 — lead rewritten to explain the whitelist↔FORGE_MODEL relationship up front (was buried in trailing footnote): the harness picks one model per eval via FORGE_MODEL, agent can't read or override it, prompts must hold across the list
• Chips regrouped by provider with uppercase headers + count (ANTHROPIC · 7, OPENAI · 2, DEEPSEEK · 3, META · 2, GOOGLE · 2, MISTRAL · 1, QWEN · 1); provider prefix stripped from chip text (anthropic/claude-haiku-4-5 → claude-haiku-4-5) since the group header already names it
• Footnote replaced with dotted-underline cursor-help tooltip on FORGE_MODEL_WHITELIST (245 chars) naming the env var, where it's read (forge.sdk.llm at chat time), and the rejection point (before OpenRouter)
• BACKLOG.md L111 flipped ○ ○ ○ → ● ● ● with the rationale + line range

Verification

Puppeteer 1440×900 on /guide#models:

• 7 provider headers in order, 18 chips total (matches config/model-whitelist.txt)
• Lead text exact match: "Whitelist below is the candidate set; the harness picks one model per eval via the FORGE_MODEL env var"
• FORGE_MODEL_WHITELIST span carries cursor-help + 245-char tooltip
• 0 new console errors