Add large_instance_bias for Cassandra model

[homatthew]

Summary

• Adds large_instance_bias parameter that penalizes selecting the largest instance in a family
• Reserves max-size instances for emergency scale-ups (e.g., avoid i3.4xlarge so it's available if we need to scale up quickly)
• Max instance size per family is computed dynamically from hardware catalog at runtime

Implementation

• Added large_instance_bias field to NflxCassandraArguments
• Added helper _compute_max_instance_size_per_family() in capacity_planner.py
• Applied rank penalty in plan_certain() sort key for largest instances
• Annotated plans with is_largest_in_family metadata in cluster_params
• Override regret() in Cassandra model to add regret for largest instances (supports both plan_certain() and plan() uncertainty planning paths)

Usage

planner.plan_certain(
    model_name="org.netflix.cassandra",
    region="us-east-1",
    desires=desires,
    extra_model_arguments={
        "large_instance_bias": 5.0,  # Adds rank penalty for largest instances
    },
)

Test plan

• Added unit tests for argument acceptance and default values
• Added test verifying largest instances get rank penalty
• Added test verifying metadata stored in cluster_params
• All 378 existing tests pass

🤖 Generated with Claude Code