Replace Step 5 CI sampling claim with canonical + fix label conflation

[PunchTheDev]

Summary

Step 5 of the Quickstart Guide had a factually-wrong CI sampling claim and conflated two distinct PR labels. This PR aligns the section with the canonical sources (eval.yml, score.yml) and surfaces both labels accurately.

Motivation

Two correctness issues found by canonical-source check:

1. CI sampling: copy said CI runs a quick check (1 easy problem per category). Per .github/workflows/eval.yml L150 + scripts/run_eval_pool.py, CI actually runs 3 specs — one random per round (any tier), 2× determinism on the first. Same shape as the anti-gaming sampling already explained in the Anti-gaming section.

2. PR label conflation: copy said the optimization label is applied if your agent passes all three categories. Per eval.yml L338–354, that's the passed label's behavior. The optimization label is actually applied when the PR beats SOTA in at least one category. The two were collapsed into one — and the passed label was missing entirely.

3. "Full scoring runs automatically" timing: ambiguous about whether per-PR or post-merge. Per .github/workflows/score.yml L3, the all-45-spec eval runs only after PR merges to main.

Changes

• QuickstartGuide.tsx L721–757: lead rewritten around canonical CI sampling shape; routed #anti-gaming ↓ link added; 199-char CI tooltip + 216-char 3-spec pool sample tooltip.
• Two-label <ul> replaces the single misleading sentence: passed (194-char tooltip) + optimization (186-char tooltip) — each accurate per eval.yml L338–339.
• Closing paragraph adds the score.yml post-merge framing (221-char tooltip) with routed /rankings → overall_score link.
• BACKLOG.md: bundled Step 1 → Step 5 row split into 5 individual rows (matching prior PRs Frame Guide Step 2 with CLI vs API roles and filter hints #295/Frame Guide Step 1 by canonical Docker vs Native paths #296/Surface fork CTA and sandbox limits in Guide Step 3 #297/Replace Step 4 generic copy with canonical 4-stage pipeline #298 split shape); Step 5 flipped ○ ○ ○ → ● ● ● with rationale.

Screenshots

Puppeteer-verified at 1440×900 on /guide#submit: all 5 dotted-underline tooltips render (199/216/194/186/221 chars), 3 <code> chips (passed, optimization, score.yml), routed #anti-gaming + /rankings anchors present, old 1 easy problem per category and passes all three categories phrases gone. 0 new console errors.