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

tasks/category_a/reverse_string.toml

@@ -0,0 +1,21 @@
+name = "reverse_string"
+category = "a"
+description = """
+Read a single line from stdin and print it reversed to stdout.
+"""
+
+[[test_cases]]
+stdin = "hello"
+expected_stdout = "olleh"
+
+[[test_cases]]
+stdin = "abcdef"
+expected_stdout = "fedcba"
+
+[[test_cases]]
+stdin = "racecar"
+expected_stdout = "racecar"
+
+[[test_cases]]
+stdin = "a"
+expected_stdout = "a"