opt: cumsum/chunked attention kernels, memory-flat CE, block checkpointing, v2 trainer (proven equivalent, 46 tests)

docs/TRAINING_LOG_1B.md

@@ -182,5 +182,24 @@ OUTPUT:  commands commands commands commands commands commands ...
 
 ---
 
+## 2026-08-22/23 — Optimization session (training PAUSED)
+
+No token digestion in this window: code-surgery session instead, on
+**AFKmoney/fractus-opt**.
+
+- Phase-2 status: started 2026-08-17T23:26 UTC; paused since ~2026-08-20
+  with <1% of the pass consumed (~0.7M tok/GPU at last sync).
+  **Remaining ≈ 425–430M tok/GPU → one full pass ≈ 4.5–5.5 days at the
+  measured baseline 900–1100 tok/s/GPU** (see HOW_FRACTUS_IS_TRAINED.md §8
+  for the math and post-opt scenarios).
+- Kernels: attention cumsum + chunked proven equivalent to the einsum
+  reference (fwd/grad/carry); chunked measured ×15–25 at 1B shapes,
+  memory-flat where the reference crashes.
+- Head: memory-flat chunked CE (loss identical within fp32 rounding).
+- Data: zero-copy int32 fetch (~27 GB RAM/pod saved).
+- Tests: 44/44 (28 repo + 14 equivalence proofs + 2 v2-loop smoke).
+
+---
+
 *Fractus CTE — continuous thought, phase-routed experts, progressive growth.*
 *Training log only. Not a claim of finished language competence.*