All 2,125 previously unattempted run2/run3 anchor draws submitted exactly once under the fixed no-retry protocol (t=1.0, top_p=0.95, top_k=20, pp=0, prompt<=2048, completion<=32768, thinking on, empty-response retries 0). 2,094 of 2,125 rows are train-ready; 31 returned an empty response and are preserved but NOT train-ready, so these caches are not complete. See run23.oneshot.provenance.json.
Publish Qwen3-4B subset20k anchor data and generation guide
Add regenerated v2 Qwen3-4B reasoning anchor cache (t1.0/p0.95/k20/pp0.0, 8192-token budget, no empty-response retries). 93882 rows: 82513 complete, 76691 hygiene-clean. Differs from the base v2 file in that the 21994 rows failing the hygiene validator were deleted and redrawn once (repetition loops -60%/-68%, near-empty answers largely persistent).
