Bump prompt-testing reviewer to Opus 4.7

[mattgodbolt-molty]

Summary

• Bump the default reviewer model from claude-opus-4-6 to claude-opus-4-7 in prompt_testing/reviewer.py and both spots in prompt_testing/cli.py (run --review-model and the standalone review --model). Same $5/$25 price tier, stronger reasoning.
• Drop the hard-coded temperature=0.0 from the reviewer's messages.create call — Opus 4.7 rejects the parameter (temperature is deprecated for this model).
• Replace the hard-coded $15/$75 cost calc in _run_reviews with app.model_costs.get_model_cost(model). The old calc would have over-reported review cost by 3× against the new Opus pricing and would silently drift again on any future model bump.

Test plan

• uv run pytest — 92 passing

• uv run pre-commit run --all-files

• Live smoke test against the Anthropic API (3 cases × Sonnet 4.6 explainer + Opus 4.7 reviewer):

  Running 3 test cases with prompt: current
    [1/3] ✓ square_cpp_o1 (in=895 out=381)
    [2/3] ✓ basic_inline_001 (in=931 out=469)
    [3/3] ✓ factorial_beginner_assembly (in=1293 out=611)
  
  Reviewing 3 results with claude-opus-4-7...
    [1/3] ✗ square_cpp_o1 (2 issues, $0.0155)
    [2/3] ✓ basic_inline_001 (0 issues, $0.0081)
    [3/3] ✓ factorial_beginner_assembly (1 issues, $0.0150)
  
  3/3 succeeded, total cost: $0.0698
  Correctness: 2/3 passed
  Review cost: $0.0385 (claude-opus-4-7)
  

  Cost arithmetic checks out: (1120+1254+1491) × $5/M + (394+74+300) × $25/M = $0.0385 ✓

  Notable side-effect: Opus 4.7 caught a real factual error in the existing Sonnet 4.6 explanation for square_cpp_o1 — the explanation claimed imul eax, edi, edi was a valid three-operand alternative, when in fact the third operand of three-operand imul must be an immediate. Useful signal that the upgraded reviewer is paying off.

🤖 Generated with Claude Code