feat(anthropic): support Claude Opus 4.7 by omitting deprecated temperature

[Khurdhula-Harshavardhan]

Summary

• Opus 4.7 deprecated temperature, top_p, and top_k — passing any of them now returns a 400. The current Anthropic provider unconditionally sends temperature=config.temperature, so any run with --model-id claude-opus-4-7 fails immediately.
• Guard the kwarg by model id: temperature is still passed for every existing model (Opus 4.6, Sonnet 4.6, future Anthropic models without the opus-4-7 substring), and is omitted only when the id contains opus-4-7.
• Reasoning is unaffected — adaptive thinking is off by default on 4.7 unless the caller passes a thinking field, which this provider does not.

Ref: What's new in Claude Opus 4.7 — "Starting with Claude Opus 4.7, setting temperature, top_p, or top_k to any non-default value will return a 400 error. The safest migration path is to omit these parameters entirely."

Methodology note

Anthropic's docs also state: "If you were using temperature = 0 for determinism, note that it never guaranteed identical outputs." The within-Anthropic comparison stays roughly apples-to-apples; a one-line README footnote on Opus 4.7's row is recommended when results land.

Test plan

• SDK-level smoke test against claude-opus-4-7 confirms the model returns valid JSON without sampling params.
• Verified guard: temperature is still in kwargs for claude-opus-4-6 and claude-sonnet-4-6, and absent for claude-opus-4-7.
• End-to-end inference run: python -m sob.run --provider anthropic --modality text --model-id claude-opus-4-7 --sample-size 5
• Full benchmark runs across text / image / audio.
• Add "claude-opus-4-7": "Claude-Opus-4.7" to data/evaluation/display_names.json after eval summaries land (separate PR or follow-up commit).

🤖 Generated with Claude Code