[WIP] anneal swarm

[mayeechen]

Done:

• Create basic config for swarm with placeholder data sources
• Add swarm.py to construct swarm mixes and launch through cookbook

olmo-cookbook-swarm swarm -c src/cookbook/recipes/olmo3-midtraining/mayeec_swarms/basic_swarm.yaml --dry-run

• Dirichlet sampler with uniform prior
• Smoke test a standard flat swarm

To smoke test:

• Handling conditional superswarms

---

Note

^{Cursor Bugbot is generating a summary for commit f25a39d. Configure here.}

[cursor]

Bug: Cache Key Mismatch After Topic Expansion

Cache key is computed from original source configs (before topic expansion) but cache values are computed from expanded leaf configs (after topic expansion). The cache hash is calculated on line 52-57 using [(sc.name, sc.paths) for sc in source_configs], but then on lines 74-81, the source_configs are transformed to leaf_configs which have different names (e.g., "source:topic" format). This means the cache will be keyed by the unexpanded structure but will contain data using the expanded leaf names, causing cache mismatches where saved cache data cannot be properly retrieved or used. The cache hash should be recalculated after the leaf expansion, or the leaf expansion should happen before the hash calculation.