Add SWE-bench Verified patch environment

[poofeth]

Summary

Adds a swe-bench-verified environment for the Prime SWE-bench Verified bounty.

This environment:

• loads princeton-nlp/SWE-bench_Verified from Hugging Face
• formats each instance as a repository repair prompt
• asks the model to return a unified diff inside <patch>...</patch> tags
• preserves SWE-bench metadata (test_patch, FAIL_TO_PASS, PASS_TO_PASS, base commit, difficulty, setup commit) in task["info"]
• provides deterministic local rewards/metrics for gold-patch exact match, patch similarity, submitted patch length, and gold patch length
• documents that this is a patch-generation/SFT/RL sanity-check environment and not a replacement for the official execution harness

/claim https://algora.io/PrimeIntellect-ai/bounties/znaiNF1eEZzF62SD

Validation

uv run pytest tests/test_swe_bench_verified_environment.py -q
# 6 passed

uv run python - <<'PY'
import sys
from pathlib import Path
sys.path.insert(0, str(Path('environments/swe_bench_verified').resolve()))
import swe_bench_verified as env
loaded = env.load_environment(train_limit=1, eval_limit=1)
print(type(loaded).__name__, loaded.taskset.taskset_id)
PY
# Env swe-bench/verified

uv run ruff check environments/swe_bench_verified tests/test_swe_bench_verified_environment.py
# All checks passed

uv run ruff format --check environments/swe_bench_verified tests/test_swe_bench_verified_environment.py
# 2 files already formatted

git diff --check
# no output

uv run --with ./environments/swe_bench_verified python - <<'PY'
from swe_bench_verified import load_environment
loaded = load_environment(train_limit=1, eval_limit=1)
print(loaded.taskset.taskset_id)
PY
# swe-bench/verified

[poofeth]

/claim https://algora.io/PrimeIntellect-ai/bounties/znaiNF1eEZzF62SD

[poofeth]

Follow-up commit b0ea80a strengthens the environment handoff and metrics:

• added official_submission(task, patch) for the official SWE-bench JSONL row shape (instance_id, model_patch)
• added changed-file overlap as a deterministic metric alongside patch similarity and line counts
• added tests for diff path extraction, official submission generation, and the new metric

Fresh validation:

uv run pytest tests/test_swe_bench_verified_environment.py -q
# 8 passed

uv run python - <<'PY'
import sys
from pathlib import Path
sys.path.insert(0, str(Path('environments/swe_bench_verified').resolve()))
import swe_bench_verified as env
loaded = env.load_environment(train_limit=1, eval_limit=1)
print(type(loaded).__name__, loaded.taskset.taskset_id)
PY
# Env swe-bench/verified

uv run ruff check environments/swe_bench_verified tests/test_swe_bench_verified_environment.py
# All checks passed

uv run ruff format --check environments/swe_bench_verified tests/test_swe_bench_verified_environment.py
# 2 files already formatted

git diff --check
# no output

[poofeth]

Additional package-install smoke passed for the environment package metadata:

uv run --with ./environments/swe_bench_verified python - <<'PY'
from swe_bench_verified import load_environment, official_submission
loaded = load_environment(train_limit=1, eval_limit=1)
row = list(loaded.taskset.source())[0]
print(loaded.taskset.taskset_id, official_submission(row, row['answer'])['instance_id'])
PY
# swe-bench/verified astropy__astropy-12907