Real-OSS-repo case study

[Connorrmcd6]

STATUS (2026-06-14): deferred to stretch / v3. Per the approved v2 plan, the real-OSS case study is the highest-effort, hardest-to-grade item and is out of the core published study — which lands multi-turn (#102–#106) + multi-provider (#107–#108) + diverse synthetic scenarios (#104) first. Revisit as an external-validity credibility anchor once the core result holds. Original scope below.

---

Real-OSS-repo case study.

The Connorrmcd6/surface#11 suite is synthetic for control; this adds external validity with at least one real project.

• pick a real repo + a genuine documented invariant; anchor it with a real hub
• inject a logic drift surf check catches; generate the genuine surf_report.json
• author a realistic task + a deterministic (or carefully-rubric'd) grader
• run + report; note the higher noise vs synthetic and how it's handled