Add RLVR adapter and verifier environment

[yannickvilleneuve-hash]

Summary

• Add rlvr_adapter.py to export IFBench prompts as chat-message JSONL tasks and score response JSONL files with scalar instruction-following rewards.
• Add ifbench_verifiers.py, an optional Prime Intellect Verifiers environment that reuses the existing IFBench checkers as rollout rewards.
• Make evaluation response lookup tolerant of harmless prompt whitespace drift and avoid mutating input kwargs during strict scoring.
• Document the RLVR export, scoring, and Verifiers loading flow.

Motivation

This is intended to make IFBench directly usable for RLVR-style training/evaluation workflows while keeping the existing benchmark verifiers as the source of truth.

Validation

uv run python -m pytest rlvr_adapter_test.py -q
uv run python -m py_compile evaluation_lib.py rlvr_adapter.py ifbench_verifiers.py
uv run python -m run_eval --input_data=data/IFBench_test.jsonl --input_response_data=data/sample_output.jsonl --output_dir=/tmp/ifbench-eval-check
uv run --extra rlvr python - <<'PY'
from ifbench_verifiers import IFBenchEnvConfig, IFBenchTasksetConfig, load_environment
config = IFBenchEnvConfig(taskset=IFBenchTasksetConfig(limit=2, eval_limit=1))
env = load_environment(config)
print(len(env.taskset.get_dataset()))
print(len(env.taskset.get_eval_dataset()))
PY

Also verified the Verifiers reward path returns 1.0 for a positive single-example completion.

Bounty reference: https://algora.io/PrimeIntellect-ai/bounties/dderbjHtPwTiGVY4