ltm-finding: remove hardcoded 'main' model assumption

[bpowers]

Summary

discover_loops() is hardwired to analyze only a model named main.

Why this is a problem

The strongest-path algorithm itself is model-agnostic, but current API behavior assumes a specific model name. This limits reuse for projects that do not use main as the analysis root and makes discovery less composable with tooling that selects models dynamically.

Literature / algorithm alignment

LTM discovery is defined over a model’s causal graph and stock set; algorithmically it is not tied to a fixed model identifier.

Current implementation behavior

• find_main_model() fetches canonicalized "main" only.
• get_stock_variables() and downstream discovery depend on this.

Location:

• src/simlin-engine/src/ltm_finding.rs

Impact

• Discovery cannot be directly applied to non-main top-level models.
• Hidden coupling to naming convention in library behavior.

Success criteria

1. Add an API that allows specifying the target model ident for discovery.
2. Keep backward-compatible convenience behavior for main where useful.
3. Ensure stock selection and causal-graph construction use the caller-specified model.
4. Add tests for a project whose analysis model is not named main.
5. Document API behavior and migration path.

Notes

This is a usability/architecture fidelity issue and independent of the known flow-to-stock timing choice.