fix(train): drop selection_metric from --model-config before model construction

[jayhesselberth]

Summary

• leech model optimize writes selection_metric into best_params.json for provenance, but handle_train had no consumer for that key — it fell through extra_kwargs into get_model()'s constructor kwargs and raised TypeError: model got unexpected keyword argument(s): selection_metric, breaking the documented optimize -> train --model-config best_params.json workflow.
• Add "selection_metric" to the _explicit_keys set in src/leech/commands/train.py::handle_train, matching the existing checkpoint_metric precedent — a metric-selection value that is pure bookkeeping/provenance, not an architecture param.
• best_params.json still records selection_metric after this change; only --model-config consumption strips it before get_model() is called.

Closes #324

Test plan

• uv run pytest tests/test_gridsearch.py -v — 35 passed, including two new regression tests (TestBestParamsJsonRoundTrip):
  • test_best_params_json_records_selection_metric — confirms best_params.json still contains selection_metric (provenance not regressed)
  • test_best_params_json_does_not_crash_train — confirms handle_train(..., model_config=best_params.json) no longer raises TypeError
• uv run pytest tests/test_config_propagation.py -v — 31 passed (existing _explicit_keys/handle_train coverage, no regression)
• uv run ruff format --check src/leech/commands/train.py tests/test_gridsearch.py — clean
• uv run ruff check src/leech/commands/train.py tests/test_gridsearch.py — clean
• uv run ty check src/leech/ — clean

🤖 Generated with Claude Code