Replace Step 4 generic copy with canonical 4-stage pipeline

[PunchTheDev]

Summary

Step 4 of the Guide (/guide#eval) was generic: "Test your agent before submitting" + a code block + a 4-item passing list that did not match the actual benchmark/evaluate.py pipeline. This rewrites the section around the canonical 4 stages (agent → geometry → fea → similarity) and frames why a local run matters.

Motivation

• The "passing result means" list lumped 3 geometry checks into items 1-3 and missed Stage 1 (agent) and Stage 4 (similarity) entirely.
• "bolt hole clearance" was factually wrong — _check_bolt_holes in benchmark/geometry.py ray-shoots to verify hole presence at the bolt_pattern_mm centers, not clearance.
• Lead never named why you'd run locally before submitting (CI per-PR runs only 3 of 45 problems; iterate offline first).

Changes

• Lead rewritten: names the 3-of-45 CI sampling (routed to #anti-gaming), forge eval --docker parity with the forge-eval:latest GHCR image (209-char tooltip).
• Bottom paragraph replaced with a 4-stage <ol> from benchmark/evaluate.py:
  • agent (221-char tip — sandbox + STEP file return)
  • geometry (373-char tip — 4 sequential checks, each named with the canonical spec field)
  • fea (323-char tip — von Mises vs yield_stress_mpa / safety_factor)
  • similarity (261-char tip — check_source_similarity.py; routed to #anti-gaming)
• forge status command line gains --spec flag for clarity.
• BACKLOG: split the bundled "Step 1 → Step 5" line into 5 individual rows; flipped Step 4 to ● ● ●.

Verified live at localhost:8080/guide#eval: 4 stages render, all tooltips populated, both #anti-gaming links navigate, no new console errors.