chore(model): align qwen recommendations with current Ollama library

[OisinKyne]

Previous recommendations had drift: qwen3:0.6b appeared throughout
docs/, qwen3:8b was the laptop default in obol model pull, and
the qwen3.6:27b-coding-mxfp8 size claim (~13 GB) was wrong.

Cross-checked against ollama.com/library:

• qwen3.6: smallest is 27b at 17 GB; no small variants exist
• qwen3.5: 4b (3.4 GB) and 9b (6.6 GB) are the practical small/mid
  options. 9b is already our integration-test baseline.
• qwen3:0.6b still exists but is far too small to be a useful
  default in 2026 — left only as a rank.go regression guard.

New tier:

• Capable host (≥32GB): qwen3.6:27b (17 GB)
• Coding (capable host): qwen3.6:27b-coding-mxfp8 (31 GB, corrected)
• Validated baseline: qwen3.5:9b (6.6 GB)
• Low-RAM laptop: qwen3.5:4b (3.4 GB)
• Reasoning: deepseek-r1:8b (4.9 GB)
• Lightweight: gemma3:4b (3.3 GB)

Files swept:

• cmd/obol/model.go: obol model pull menu + no-models hint
• internal/openclaw/openclaw.go: no-models hint
• internal/embed/skills/{sell,monetize-guide}/SKILL.md: examples
• docs/getting-started.md, docs/guides/monetize-inference.md
• CLAUDE.md, flows/lib.sh, flows/flow-03-inference.sh
• internal/openclaw/monetize_integration_test.go (comment)

internal/model/rank.go and rank_test.go intentionally untouched;
the qwen3:0.6b references there are regression guards for users
who happen to have it pulled, not recommendations.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

[bussyjd]

Superseded by #386 — all commits from this branch are already present at the tip of integration/pr377-pr381 (verified by git log <branch> ^origin/integration/pr377-pr381 returning empty). Closing to keep the queue tidy. Track final landing on main via #386.