feat: Import Repair

[0xCUB3]

Adds PythonImportRepair requirement for detecting and suggesting fixes for missing/incorrect imports in LLM-generated Python code.

Detection modes:

• Static analysis (default): undefined name detection + module availability checks
• Execution-based (when allow_unsafe_execution=True): runs code and parses runtime errors. Matches PythonExecutionReq interface (i.e. when code execution is enabled, import validation uses execution too)

Features:

• Handles ModuleNotFoundError, ImportError, NameError, AttributeError. Any I'm missing?
• 155 common aliases (np, pd, plt, Path, Optional, etc.)
• Module relocations for sklearn, tensorflow, torch, scipy
• Fuzzy matching for misspelled modules via rapidfuzz
• Works with RepairTemplateStrategy via ValidationResult.reason

Usage:

from mellea_contribs.reqlib import PythonImportRepair
from mellea.stdlib.sampling import RepairTemplateStrategy

result = session.instruct(
    "Write code using pandas",
    requirements=[PythonImportRepair()],
    strategy=RepairTemplateStrategy(loop_budget=3)
)

Files:

• mellea_contribs/reqlib/common_aliases.py - Alias database
• mellea_contribs/reqlib/import_resolution.py - Error parsing and resolution
• mellea_contribs/reqlib/import_repair.py - PythonImportRepair class
• test/test_import_repair.py - new tests

Dependencies:
rapidfuzz - Fast fuzzy matching for misspelled modules. Chosen over fuzzywuzzy (10x faster, no GPL licensing issues via python-Levenshtein, though way better name :) ) and difflib (faster, better ranking algorithm).

My tests

• 51 new unit tests passing. I tried to put in as many edge cases as possible.
• Ruff lint passing
• Mypy type check passing
• Tested with mellea + Ollama end-to-end on

[0xCUB3]

can remove if not necessary. I know mellea proper has this but it doesn't exist here.

[nrfulton]

Thanks for adding.

[nrfulton]

Thanks for adding.