chore(autofix): Switch to claude for root cause step

[roaga]

Model switch from GPT 4o to Claude 3.5 Haiku results in better latency, better coding, and better root cause results.

(see row 2 below):

[trillville]

how much higher is the error rate? hard to tell from these metrics

[roaga]

how much higher is the error rate? hard to tell from these metrics

Good runs: 98
Errored runs: 5
Error rate: 0.04
Errored in root cause: 0
Errored in plan: 3
Error rate in root cause: 0.00
Error rate in plan: 0.02
Error rate in something after plan: 0.02
Runs with unapplied changes: 24
Missing change rate: 0.19

These are the correct numbers, the eval script is bugged @trillville

So actually error rate is unaffected

[jennmueng]

Bug in our eval, but the notebook error rate is correct

[roaga]

actually going for haiku now instead of sonnet, evals are better