Run the cross-lineage overlap on real Gemini + open-weights models

[sebasmos]

Run stage 1 across a real roster (Gemini tiers + at least one open-weights family) and run experiments.run_holes_test: do same-lineage failure vectors overlap more than cross-lineage ones. Report the verdict.

Status: the harness is ready; this task runs it on real data / models (the contributor step).

Model backend. Agents use the Gemini API for now behind one gateway; extensible to other model APIs later. No fine-tuning.

[sebasmos]

#265 is 906 lines of code and tests with zero result artifacts, so under the real-data rule the cross-lineage claim is unmet, and the paper lists it as future work rather than a result. One GPU turn closes it.

Two things worth knowing before you spend that turn, both from reading the smoke test rather than the description. It reports cross_lineage, lineages, committee and n_turns from declared config rather than observed behaviour, so those four fields pass as hardcoded literals. And main() returns 0 unconditionally, so a run that resolved no image and made no model calls still prints "cross_lineage": true and exits 0. A skip is currently indistinguishable from a pass to anything reading the output.

Fixing those first would mean the GPU turn produces an artifact that actually demonstrates what it claims.

[sebasmos]

Closing as part of the pre-release cleanup. The paper is submitted and the repository is being prepared to go public, so the board is being reduced to the single item that still gates a published number: #393, the MIMIC-CXR same-lineage judge cell. Everything else is out of scope for this release. This is not a judgement on the work: unmerged branches were archived to git bundles before deletion, and anything here can be reopened if the lane is picked up again.