Add exclude parameter to S3 pull functionality

[corbt]

Summary

• Added typesafe exclude parameter to S3 pull functionality to filter out directories during sync
• Updated pull_model_trajectories helper to automatically exclude checkpoints and logs
• Enables pulling only trajectories for local analysis without downloading large model weights

Motivation

Often when analyzing trajectory data, we only need the trajectory files themselves, not the model checkpoints or logs. Model checkpoints can be very large (gigabytes), and downloading them consumes significant bandwidth and storage space. This is especially important when working on mobile connections or with limited bandwidth.

Changes

1. Added ExcludableOption type with valid options: "checkpoints", "logs", "trajectories"
2. Modified s3_sync() to accept and use exclude patterns with AWS CLI's --exclude flag
3. Updated pull_model_from_s3() and _experimental_pull_from_s3() to accept and pass through exclude parameter
4. Modified pull_model_trajectories() helper to automatically exclude checkpoints and logs

Test plan

Created and ran test_s3_exclude.py to verify:

• Helper function correctly excludes checkpoints and logs
• Direct backend call with exclude parameter works
• Only trajectory files are downloaded, excluded directories remain empty
• Tested with model email-agent-216-4 in project email_agent

🤖 Generated with Claude Code