Support multi-file experiment bundles

[michi774]

Some user requests require experimenting with multiple algorithms/recsys libraries (e.g. lenskit and recbole).
In the current agent loop, the model returns a single python source string, which is written to a single .py file and executed.
This strongly discourages modularity (seperate runner scripts), and it becomes painful to maintain or iterate on larger experiments.

This is especially relevant because LensKit and RecBole are typically used through our wrapper package "omnirec".

Current behavior (repo-specific)

Execution pipeline today:

• The LLM returns a single code string (see PlanAndCode in https://github.com/ISG-Siegen/AutoRecLab/blob/develop/treesearch/function_specs.py
• Type checking writes that string to runfile.py in the workspace and runs uv run ty check on that single file (see https://github.com/ISG-Siegen/AutoRecLab/blob/develop/treesearch/type_checker.py).
• Execution writes that string to runfile.py in the workspace and runs uv run runfile.py (see https://github.com/ISG-Siegen/AutoRecLab/blob/develop/treesearch/interpreter.py).
• Checkpointing stores a copy as code.py and moves generated artifacts from out/workspace/ and out/workspace/working/* into checkpoint/<node-id>/generated/ (see https://github.com/ISG-Siegen/AutoRecLab/blob/develop/treesearch/search.py).

Important nuance: the checkpoint stores code.py, but the interpreter actually executes runfile.py.

Why this is limiting

• Large monolithic scripts are harder to debug and review.
• No way to keep “experiment A” and “experiment B” as separate modules.

Goal

Enable the agent to produce and execute a multi-file experiment bundle:

• Multiple .py files created per node/run
• A clear entrypoint file to execute
• Type checking across the bundle (at least the entrypoint + its local imports)
• All files captured in the checkpoint for reproducibility

Type checking across multiple files

ty check can be invoked on:

• a directory, or
• multiple file paths

For bundles, we likely want to run ty check on the bundle root directory (workspace subdir) or the entrypoint plus all .py files in the bundle.