Repair published PR-diff tasks after the oracle isolation fix

[adithya-s-k]

The generator fix in #145 addresses #144 for new exports. Existing published tasks need a separate migration; changing the library cannot remove files baked into an old task image.

The current manifest for AdithyaSK/repo2rlenv-pr-diff lists 181 tasks. On 2026-09-22, the sampled ripgrep task Dockerfile at revision 971b463606b0ffa69ac6b6125d90c6868744be2b still embeds /verifier/oracle.patch in the agent image. This confirms the published dataset has not been migrated; it is not a claim that all 181 tasks were individually inspected.

Required follow-up:

• Audit all published PR-diff tasks and any copies for baked grading assets.
• Move oracle/verifier assets into tests/, update the evaluation script, and clear both reward formats before grading.
• Regenerate hashes/manifests where applicable, rebuild affected images, and invalidate old image caches.
• Check no-op and oracle behavior, and confirm the oracle is absent during the agent phase.
• Publish a new dataset revision with migration notes. Mark old revisions as affected; previous oracle passes do not establish resistance to answer leakage.

Keep this open until the repaired dataset revision and validation evidence are linked here.

[KNambiarDJsc]

Opened #157: repo2rlenv migrate pr-diff <dir> [--apply], covering the file-repair half of this issue's checklist (detect the baked-oracle Dockerfile, rewrite it + tests/ via the same builder functions generate uses today, keep content_hash unchanged since instruction.md/solution/patch.diff are never touched, refuse to touch anything whose content_hash doesn't match what's on disk). Full detail in the PR description.

Two things from this issue's checklist that PR doesn't cover, flagging here rather than silently leaving them:

1. Image rebuild + registry cache invalidation + publishing the repaired revision — needs your HF Hub push access to AdithyaSK/repo2rlenv-pr-diff, not something I can do from outside. The tool's output (a dataset dir with rewritten environment/Dockerfile + tests/) is meant to feed directly into your existing repo2rlenv push flow for that.
2. "Confirm the oracle is absent during the agent phase" — I verified the rewritten Dockerfile is byte-identical to what generate produces today (whose oracle-absence pr_diff: ship oracle + verifier via tests/, not baked in image #145 already verified end-to-end via Harbor), but I don't have Docker in the environment I built this in, so I have not independently rebuilt a migrated task's image and re-run the find / -name oracle.patch check from pr_diff: ship oracle + verifier via tests/, not baked in image #145's own test plan against it. Worth doing before trusting the migration output for the real dataset, even though the code path is shared with fresh generation.

Also: "any copies" — I searched the repo's own docs/RFCs/README for other referenced pr_diff datasets and only found AdithyaSK/repo2rlenv-pr-diff (181 tasks, the one already linked here). If there are other forks/copies elsewhere on the Hub, I have no visibility into that from the repo alone — let me know if there's a known list.