Summary

Closes #416.

This adds explicit agent-friendly shape metadata to load_data_source and load_data_source_async responses so MCP clients can reason about loaded data handles without guessing from raw column names and dtypes.

What changed

• add a shared metadata builder in Executor
• expose:
  • target_scitype
  • target_variates
  • has_exog
  • exog_variates
  • index_type
  • n_target_columns
  • n_exog_columns
• apply the same metadata enrichment to both sync and async load paths
• add focused tests for:
  • datetime-indexed target with multivariate exogenous data
  • range-indexed target without exogenous data

Why this matters

This makes loaded data self-describing for agents before they choose estimators or workflow branches. In particular, it reduces ambiguity around:

• whether a handle is univariate or multivariate
• whether exogenous variables are present
• whether the time index is datetime-like or integer/range based

Validation

• .venv/bin/ruff check src/sktime_mcp/runtime/executor.py tests/test_data_sources.py
• .venv/bin/ruff format --check src/sktime_mcp/runtime/executor.py tests/test_data_sources.py
• .venv/bin/pytest tests/test_data_sources.py -q
• .venv/bin/pytest -q