Initial commit: Lush vs Bash AI benchmarking framework

Benchmark harness that uses LLM agents to solve shell scripting tasks
in both Bash and Lush, then compares correctness and code quality.

- CLI with run, run-all, list-tasks, report, and export commands
- Agent loop with retry support via Anthropic Claude provider
- Test harness executing solutions in sandboxed subprocesses
- LLM-driven questionnaire for subjective code quality evaluation
- HTML report export with charts (matplotlib)
- 8 Category A tasks (write-from-scratch in both languages)
- 4 Category B tasks (verify provided Bash, convert to Lush)
- Lush language reference for agent context

.python-version

@@ -0,0 +1 @@
+3.11