Add HumanEval+/MBPP+ (EvalPlus) as native sandboxed LightEval tasks

[ezosa]

What

Adds humaneval_plus and mbpp_plus (EvalPlus) as native inspect-ai LightEval tasks, migrating these code-generation benchmarks off lm-eval-harness so they use LightEval's wiring, dashboard sample logging, and reasoning-trace separation.

New package src/lighteval/tasks/tasks/evalplus/ (auto-discovered like bfcl/):

• main.py — BFCL inspect-ai shape (record_to_sample + solver=[generate(cache=True)] + @scorer). Instruct-style prompt; extracts code from the model's post-reasoning final answer (with reasoning_parser=qwen3 the completion is the answer, not the <think> trace); grades by executing against EvalPlus's augmented unit tests. pass@1 estimated over N samples via the wrapper's epochs (mean reducer) — no custom metric.
• execution.py — check_correctness(sample, generation, timeout) / run_test(sample, test, timeout) grader that deliberately mirrors lcb's signature + multiprocessing.Process pattern, so the TensorWave network sandbox (sandbox_grading.py) wraps it via a single generic wrapper with no per-task code. The sample carries a one-element input_output purely so the wrapper's timeout math (which indexes sample["input_output"]["inputs"]) works unchanged. reliability_guard is vendored from OpenAI human-eval (defensive only; real isolation is the sandbox's CLONE_NEWUSER|CLONE_NEWNET).
• build_data.py — normalizes evalplus/humanevalplus + evalplus/mbppplus into a uniform multi-config HF dataset (ezosa/evalplus, configs humaneval_plus/mbpp_plus), deriving MBPP's entry point from its reference code.
• tests/unit/tasks/test_evalplus.py — pure-stdlib grading tests: correct→pass, wrong/raise/timeout/no-function→fail, and code extraction (incl. post-reasoning answers).

Why

The lm-eval *_instruct code tasks force an empty <think></think> block into the prompt, so the model never reasons. This migration keeps reasoning ON and gives proper LightEval logging.

Safety

Execution is unsafe and must run under eval_lighteval_sandboxed.sbatch. The consuming TensorWave repo registers the task prefixes (humaneval, mbpp) in sandboxed_evals.SANDBOXED_TASK_PREFIXES and the grader in GRADER_PATCHES (separate PR).

Verification

• Unit tests pass (grader validated locally under fork).
• ruff check + format clean.
• Registry auto-discovers evalplus/main.py; humaneval_plus/mbpp_plus resolve.

🤖 Generated with Claude Code