Add Claude 4.x support and harden writeup pipeline

[SeoknamKo]

Summary

• Register the Claude 4.x model family (claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001) in AVAILABLE_LLMS so they are accepted by the CLI. Routing already works through the existing startswith("claude-") dispatcher in create_client / get_response_from_llm.
• Force English-only output during LaTeX writeup. When Claude produced Korean comments inside template.tex, Aider's SEARCH/REPLACE blocks failed to match on subsequent edits, triggering Only 3 reflections allowed, stopping and aborting the run. A new LANGUAGE_RULE is injected into the abstract / per-section / Related Work prompts, and a corresponding check is added to error_list.
• Fix a latent bug in the writeup Coder.create() dispatcher: it had no claude-* branch, so Claude models were passed to Aider without the required anthropic/ prefix. Experiments dispatcher had the branch; writeup did not. Both now match.
• Bump coder.max_reflections from Aider's default 3 to 6 after each Coder.create(), so transient SEARCH/REPLACE mismatches don't abort an otherwise healthy writeup.
• Guard import torch with a try/except. launch_scientist.py only uses torch for GPU enumeration (torch.cuda.device_count()), but a broken torch install (common on Apple Silicon conda setups with mixed pip/conda torch) caused immediate ImportError at module load, even for sklearn-only templates. get_available_gpus() now returns [] when torch is missing or CUDA is unavailable.

Why

This surfaced while running AI Scientist against a custom sklearn-only template with claude-opus-4-7 on an M4 Mac:

1. CLI rejected claude-opus-4-7 because it was not in AVAILABLE_LLMS.
2. After registering the model, the writeup Coder mis-routed the Claude call because the anthropic/ prefix was only applied in the experiments dispatcher.
3. Even after fixing the routing, writeup failed repeatedly: Claude emitted Korean LaTeX comments, Aider's SEARCH blocks included those comments verbatim, the target file did not contain them, diff edits failed, and Aider hit its 3-reflection cap.
4. Separately, a broken torch install in the conda base env hard-blocked launch for templates that don't need torch at all.

Each change is small and isolated, and keeps the existing behavior for non-Claude / torch-available setups intact.

Test plan

• python launch_scientist.py --model claude-opus-4-7 --experiment <sklearn_template> --writeup latex --num-ideas 1 completes through PDF generation on macOS M-series.
• python launch_scientist.py --model claude-sonnet-4-6 ... same.
• Existing gpt-4o / claude-3-5-sonnet-20241022 runs are unaffected (model dispatch unchanged for them).
• import launch_scientist succeeds in an env where torch is broken; get_available_gpus() returns [].
• Grep the resulting template.tex for any non-ASCII CJK ranges — should be empty.

🤖 Generated with Claude Code