Fix grader exploits, sandbox agent SQL, add tests/CI, and rewrite docs

[rajdeepchatale]

Summary

Repo-wide cleanup. Running the original code turned up several grading and safety bugs, which are fixed and covered by tests here. The docs have been rewritten to match what the code actually does.

Bugs fixed

Problem	Before	After
Writes were penalised but still executed	DELETE FROM employees then a wrong query scored 0.97 and ended the episode	Read-only DB (SQLite authorizer); destructive queries rejected unexecuted, score 0
DROP TABLE broke the episode	Every later step() raised	Blocked by the read-only DB
Correct answers didn't end the episode	4 reference queries scored 0.92–0.94, below the 0.95 threshold	Episode ends on an exact result match (is_correct)
Duplicate rows ignored	Join fan-out got full result credit	Multiset comparison
Reward farming	A pointless COALESCE added points	Bonus only when the task needs it
Runaway queries	A recursive CTE blocked the worker indefinitely	2 s / 10,000-row budget, RESOURCE_LIMIT diagnostic
Baseline ignored task_id	inference.py ran random tasks logged under the wrong name	Targets each task; covers all 14
Hidden task requirements	company_medium_2 filtered active employees without saying so	Questions state every filter

Tooling

• 71 pytest tests, covering the grader, the sandbox, episodes, and manifest/registry/baseline consistency
• GitHub Actions workflow: ruff, pytest and openenv validate on Python 3.10 and 3.12
• Committed uv.lock, which openenv validate failed without. Declared openai, added project metadata, bumped to 0.2.0.
• Removed the duplicate server/Dockerfile and the unused server/requirements.txt

Docs

• README renamed to SQL Query and rewritten, covering episodes, scoring, diagnostics, sandboxing, quickstart and limitations
• Removed the per-task baseline table because inference.py could not have produced those numbers. Re-run the fixed script to publish real results.
• CONTRIBUTING rewritten; LICENSE holder updated

Verification

• ruff check ., pytest (71 passed) and openenv validate . pass locally and in a clean Python 3.10 checkout
• End-to-end over WebSocket: reset(task_id=...) works, DELETE is blocked, the reference query ends the episode as correct
• Not verified: docker build, because Docker wasn't running locally

Behaviour changes

Scores aren't comparable with runs made before this change, because the grader now penalises duplicate rows, no longer rewards irrelevant COALESCE, and ends episodes on correctness.

🤖 Generated with Claude Code