feat: add swegym entry (full SWE-Gym 2438 + official lite subset)

[recursix]

Adds the swegym registry entry — full SWE-Gym (2438 tasks) with the 230-task lite split as an official named subset (cube-harness#504, package swegym-cube).

Submitted as a same-repo branch (I have admin) rather than a fork, so quick-compliance can push the CI-derived fields back — the fork variant (#62) passes compliance but fails the write-back step (GITHUB_TOKEN can't push to a fork branch). #62 will be closed as superseded.

Validated locally + in CI on #62: ✅ Schema valid, ✅ swegym-cube==0.1.0 installed, ✅ SWEGymBenchmarkConfig, task_count 2438, has_debug_task True (2), ✅ Quick check PASSED.

Note: dev_install_url pins @feat/swegym-lite-cube until cube-harness#504 merges; drop the ref (→ default branch) after merge. The cube depends on cube-standard>=0.1.0rc9 (now on PyPI), which is what unblocked this.

🤖 Generated with Claude Code

[github-actions]

Entry review — swegym

Verdict: CONCERN

Check	Result
description_matches_package	⚠️ unverified
authors_consistent_with_git	⚠️ unverified
no_id_squat_vs_existing	✅ pass
no_brand_impersonation	✅ pass
wrapper_license_plausible	⚠️ unverified

Notes:

Core claims could not be verified. PyPI package swegym-cube appears unpublished: empty summary, empty license field, no long description. The dev_install_url points at a feature branch (cube-harness@feat/swegym-lite-cube#subdirectory=cubes/swegym-cube) and the entry's own NOTE says it is blocked until both a cube-harness PR merges and cube-standard rc9 lands on PyPI — so the package state is explicitly pre-merge/incomplete. No public README reachable for the cube wrapper repo, so description-vs-package consistency cannot be confirmed.

authors: only author is recursix (Alexandre Lacoste). The cube subdirectory (cubes/swegym-cube) is on a private/unreachable feature branch; could not confirm this handle appears in commit history for the wrapper. recursix is not in known-authors.yaml for this id. Unverified.

wrapper_license: Apache-2.0 is a valid SPDX id, but the cube-harness repo LICENSE was not reachable to confirm. Upstream SWE-Gym benchmark_license reported Apache-2.0 with source_url to SWE-Gym/SWE-Gym/blob/main/LICENSE — plausible but not fetched/confirmed here.

id swegym already exists in the registry; this PR modifies an existing entry (own-entry ownership presumed verified by CI), so not an id-squat. name/description faithfully describe the real SWE-Gym (arXiv 2412.21139), no brand impersonation.

Verdict CONCERN driven by the pattern of unverifiable core claims: unpublished package + unreachable repo block confirmation of description, authorship, and license. Recommend re-review once swegym-cube is published to PyPI and the cube-harness wrapper lands on a public default branch.

[github-actions]

✅ ownership-check + quick-compliance passed (slow-compliance is informational).

Auto-merge skipped because:

• LLM review flagged a CONCERN (see review comment above)

A maintainer will review and merge.

[github-actions]

Entry review — swegym

Verdict: CONCERN

Check	Result
description_matches_package	✅ pass
authors_consistent_with_git	⚠️ unverified
no_id_squat_vs_existing	✅ pass
no_brand_impersonation	✅ pass
wrapper_license_plausible	✅ pass

Notes:

Description: consistent with SWE-Gym (arxiv 2412.21139), a SWE-bench-methodology dataset of 2438 real-world Python SWE tasks across 11 repos with a 230-instance lite split. task_count 2438 matches. Docker-based fail_to_pass/pass_to_pass framing is correct for this benchmark family. Description plausible; cannot cross-check the swegym-cube subdir README directly but consistent with cube-harness repo structure.

ID squat: existing ids include swebench-live and swebench-verified; 'swegym' is a distinct benchmark (different dataset/authors than SWE-bench), not a confusable near-duplicate. Pass.

Brand impersonation: 'SWE-Gym' is the actual upstream name (SWE-Gym/SWE-Gym), faithful port — pass.

License: wrapper Apache-2.0 is a valid SPDX id; benchmark_license.reported Apache-2.0 with source_url to SWE-Gym/SWE-Gym/blob/main/LICENSE — plausible, verified_by_original_authors:false noted. Could not fetch the actual upstream LICENSE to confirm Apache-2.0, but claim is internally consistent. Pass on available evidence.

CONCERN drivers (not fails, but blocking):

1. PyPI metadata for swegym-cube is EMPTY / package appears unpublished (dev_install_url is a git+ subdirectory URL, which may be acceptable, but no PyPI long description to cross-check description_matches_package against).
2. authors_consistent_with_git UNVERIFIED: sole author 'recursix' (Alexandre Lacoste) is an AI Alliance / cube-harness maintainer, not a known SWE-Gym original author (known-authors.yaml lists none for this id). The linked repo README was readable but I could not inspect commit history for cubes/swegym-cube to confirm this handle contributed to that specific wrapper subdir. recursix is plausible as a cube-harness contributor (matches the harness authorship), so no impersonation evidence — but the original-author attribution for the SWE-Gym dataset itself is not represented.

Net: core claims look legitimate (faithful SWE-Gym port by a cube-harness maintainer), but empty PyPI page + inability to verify the wrapper author against the cube subdir commit history prevent a clean PASS. Recommend maintainer confirm (a) swegym-cube is actually published/installable and (b) recursix committed to cubes/swegym-cube.

[github-actions]

✅ ownership-check + quick-compliance passed (slow-compliance is informational).

Auto-merge skipped because:

• LLM review flagged a CONCERN (see review comment above)

A maintainer will review and merge.