v3 trainer + English docs + production x8 results

docs/TRAINING_OPTIMIZATION.md

@@ -3,6 +3,17 @@
 **Date:** 2026-08-06
 **Measured on:** Ryzen 5 5500U, 12 threads
 
+> **Status update 2026-08-26:** this was the analysis phase. The implemented
+> follow-up (attention kernels cumsum/chunked proven equivalent, memory-flat
+> chunked CE, zero-copy data pipeline, v2/v3 trainers) lives in
+> **AFKmoney/fractus-opt** — see `OPTIMIZATION_2026-08-22.md` there for
+> proofs, deployment and measured results, including the live production
+> run on 8×RTX 5090 (~1,570 tok/s/GPU sustained, phase-2 finish ≈3 days).
+> Levers below resolved:
+> *Fused CE kernel* → done (`fractus/nn/ce.py`); *torch.compile* → flag in v2
+> trainer; *vocabulary reduction* → still open (breaks open-heart checkpoint
+> compatibility → new phase).
+
 ## Profile: where does the time go?
 
 ### Per-iteration breakdown (d=512, 4 blocks, 16 experts, chunk_len=32)
@@ -107,7 +118,7 @@ The CTE supports batch_size > 1 in tick_chunk_train.
 | 500M | ~4 days | Competent model |
 | 1.76B (Chinchilla) | ~14 days | Full Chinchilla |
 
-With progressive growth (warm start from palier 3), the model converges
+With progressive growth (warm start from stage 3), the model converges
 in ~1/4 of Chinchilla → **~3-4 days for a usable 1B model**.
 
 ## Remaining optimization levers
@@ -121,12 +132,12 @@ in ~1/4 of Chinchilla → **~3-4 days for a usable 1B model**.
 
 ## Training commands
 
-### CPU progressive growth
+### CPU progressive growth (stage list; `--paliers` is the literal CLI flag)
 ```bash
 python scripts/train_progressive.py --paliers 0,1,2,3 --accumulation-steps 8
 ```
 
-### GPU 1B training (from palier 3 checkpoint)
+### GPU 1B training (from stage-3 checkpoint)
 ```bash
 python scripts/train_1b_gpu.py \
     --checkpoint checkpoints/fractus_palier3.pt \