[cornelislouisa]

First of five stacked PRs that bring the k-fold work from guille/kfold_relaunch (the code behind the October k-fold results) into main.

Changes

• STRING threshold. adjacency_threshold is 0.4 in every dataset config. WGCNA ignores it since Keep a target number of WGCNA edges instead of a frozen cutoff #138 (it uses adjacency_target_connectivity: 0.10), so the stale per-dataset WGCNA cutoffs and their "UPDATED" headers are removed. The searches already forced STRING to 0.4; this makes direct runs match. wgcna_binarization: target_connectivity is now stated explicitly (it was already the default).
• STRING cache safety. Alias, bulk and interaction caches are locked per artifact, JSON is written atomically, truncated gzip downloads are re-downloaded, and corrupt cache JSON is rebuilt. Copied from kfold_relaunch.
• Pre-transform cache safety. PreProcessor builds its cache under a directory lock and saves through a .partial file, so concurrent runs never read a half-written cache.
• SAGN guard. gene_identity is rejected for SAGN, which trains on precomputed aggregates and would silently ignore the identity embedding.
• Tests. The config consistency test now requires the same adjacency_threshold everywhere. test_string_cache_concurrency.py (from Add production-ready fold-mean Optuna search #139) covers the locks and atomic writes. A new gene-identity test covers the SAGN guard.
• Docs. README graph-construction bullets give the WGCNA and STRING values, and ogbench/data/adjacency/README.md describes per-fold binarization.

Checks

• pytest -m "not slow": 207 passed
• pre-commit clean
• 1-epoch runs: MotrPAC fixed and k-fold, BRCA STRING (graph built at 0.4), BRCA SAGN with gene_identity=learnable fails with the new error

[cornelislouisa]

The whole stack (#141 → #145) passed the reproducibility check against guille/kfold_relaunch. Full results are in #145 (comment) (the comment on #145). In brief: the pipeline code is identical to relaunch, the data caches are byte-identical, the sklearn baselines match exactly, an MLP best trial reproduces October bit for bit, and graph models differ only by CUDA nondeterminism.