Fix sampler_weights checkpoint loading

[Jah-yee]

Summary

Fixes Issue #25: sampler_weights not available while weights is available

Problem

When loading a sampler_weights checkpoint using a tinker path (e.g., tinker://run-id/sampler_weights/0001), the checkpoint loading fails with a "Path is invalid" error. This happens because the API call was being made with an incorrectly formatted checkpoint ID.

Root Cause

In ParsedCheckpointTinkerPath.from_tinker_path(), the checkpoint_id was being set to include the type prefix (e.g., sampler_weights/0001), but the API was receiving inconsistent checkpoint ID formats that didn't properly distinguish between training and sampler checkpoints.

Fix

1. Added a new api_checkpoint_id property to ParsedCheckpointTinkerPath that returns:

   • For training checkpoints: the pure checkpoint ID (e.g., 0001)
   • For sampler checkpoints: the prefixed ID for API calls (e.g., sampler_weights/0001)

2. Updated all methods that use tinker paths in rest_client.py to use api_checkpoint_id instead of checkpoint_id

Testing

Added new test file tests/test_sampler_weights_loading.py with tests for:

• Parsing weights checkpoint paths
• Parsing sampler_weights checkpoint paths
• Verifying the correct api_checkpoint_id format is returned for each type

How to Test

1. Save sampler weights: training_client.save_weights_for_sampler("sampler-001")
2. Get the returned tinker path (e.g., tinker://run-id/sampler_weights/sampler-001)
3. Try to download: rest_client.get_checkpoint_archive_url_from_tinker_path(tinker_path)
4. Verify the download URL is correctly generated

The fix ensures both weights and sampler_weights checkpoints can be loaded without errors.