Reframe Guide Whitelisted models around harness-picks-model

[PunchTheDev]

Summary

Guide Whitelisted models (L604-637) was a list of 18 chips fronted by a generic "LLM agents receive a harness-injected LLMClient" lead and footnoted with the most important fact (the harness picks which model runs, agents cannot override). Rebuilt around that fact, deduped a 3rd llm.chat() code block already covered in Step 3 and Patterns, and tooltipped the two canonical sources first-timers would otherwise need to grep the repo for.

Motivation

A first-timer reading the prior version sees 18 chips and reasonably concludes "let me pick claude-opus." That's wrong — agents call llm.chat() once and the harness sets FORGE_MODEL per eval run (default anthropic/claude-haiku-4-5, overridable by secrets.FORGE_MODEL in CI). The footnote saying "model is fixed by harness" buried the only fact that prevents the misconception.

The section also re-documented the SDK call shape that Steps 3 and Patterns already cover post-PRs #297 and #300 — a 3rd duplicate. Per feedback_link_to_canonical_explainer.md, route to the canonical site, don't duplicate.

Changes

• QuickstartGuide.tsx L604-637 — lead rewritten with harness picks which model runs it foregrounded; FORGE_MODEL env var tooltip (326 chars) cites the exact CI workflow lines (eval.yml L157, score.yml L104/151) and SDK enforcement (forge/sdk/llm.py L23, L32-35 — whitelist check raises before chat() ever runs)
• config/model-whitelist.txt link gets a 200-char tooltip citing the file's own "CI reads it directly. Update this file to add or remove models" header — names it as the canonical authority, not just a reference
• Removed the 3rd duplicate llm.chat([...]) code block; replaced with a 1-sentence routed <a href="#write">Step 3 - Write your agent ↑</a> pointer carrying the canonical max_tokens=4096 SDK signature from forge/sdk/llm.py L41
• Anti-gaming rationale moved from buried footnote to a cross-link to #anti-gaming ↓ (carries the step 281/372/373/375 cross-link pattern — surfaces the enforcement chain instead of restating it inline)
• Chip count auto-derived from WHITELISTED_MODELS.length so the lead and grid never drift if the list is edited
• BACKLOG.md L111 flipped ● ● ● with rationale

Verification

Puppeteer 1440x900 on /guide#models:

• hasHarnessPicks=true hasForgeModelInline=true hasNoApiKey=true
• oldLlmChatBlockGone=true oldFootnoteGone=true (no The model is fixed by the harness via phrase)
• sdkRefersStep3=true maxTokens4096=true linkToWrite=true linkToAntiGaming=true linkToWhitelistFile=true
• chipCount=18 (matches canonical config/model-whitelist.txt 18 entries)
• 2 tooltips: FORGE_MODEL env var (326 chars), config/model-whitelist.txt (200 chars)
• 0 new console errors (1 pre-existing 404 on a tile asset unchanged)