fix(eval): drop temperature for claude-opus-4-7

[ghost]

What this PR adds

Bug fix to the prompt-caching eval harness: removes a parameter that the Anthropic API now rejects, which currently prevents the script from running.

Skill checklist (for new skills)

N/A — this is a bug fix to eval/, not a skill change.

Lint output

N/A — no files under skills/ are touched by this PR.

Index regeneration

N/A — no skill files changed, so README.md and the index are unaffected.

Notes for review

run_eval.py passes temperature=0 to claude-opus-4-7, which the Anthropic API rejects:

{"type":"error","error":{"type":"invalid_request_error","message":"`temperature` is deprecated for this model."}}

The script fails before completing a single call. Removing the parameter fixes it.

Worth flagging: temperature=0 was presumably there for reproducible eval runs. Opus 4.7 doesn't accept any value other than 1, so determinism via temperature isn't available on this model. I left a comment in the code explaining why the param is gone rather than removing it silently.

For reference, I tested the sampling params across the current Anthropic lineup. Rejected on opus-5, sonnet-5, fable-5, opus-4-8 and opus-4-7. Still accepted on sonnet-4-6, opus-4-6, opus-4-5, haiku-4-5 and sonnet-4-5. So this only affects the 4-7-and-newer calls.