Add MergedTop3_v3 clean 8xH100 record-track submission

[hesong0222-dev]

Summary

This PR adds a clean track_10min_16mb submission folder:

• records/track_10min_16mb/2026-03-25_MergedTop3_v3_clean_h100

The submission is a merged top-stack recipe built from the public leaderboard lineage:

• 11 layers
• XSA on the last 4 layers
• EMA only
• 3x MLP
• SmearGate
• BigramHash with 2048 buckets
• mixed int6 quantization with zstd
• sequence length 2048
• Muon/AdamW weight decay 0.04
• sliding-window eval with stride 64
• Partial RoPE (ROPE_DIMS=16)
• layerwise LN scaling
• GPTQ-lite clip search
• WARMDOWN_ITERS=3500

This folder is the clean rerun package after removing recovery-only framing and adding strict runtime gates for a fresh uninterrupted 8x H100 execution.

Clean run result

Measured on 8x H100 with a fresh uninterrupted single-seed run:

• step_stop=5347
• train_time=580.213s
• final_int6_roundtrip_exact val_loss=1.96565872
• final_int6_roundtrip_exact val_bpb=1.16417381
• eval_time=44.398s
• bytes_model_int6_zstd=15,562,277
• bytes_code=72,924
• bytes_total=15,635,201

This run stayed under both required caps:

• training time < 600s
• evaluation time < 600s
• artifact size < 16,000,000

What is included

• train_gpt.py
• README.md
• submission.json
• requirements.txt
• helper scripts for remote bootstrap, strict preflight, one-shot run, and artifact collection

Verification

• local python3 -m py_compile train_gpt.py passed
• strict single-process and 8-process remote preflight passed before training
• clean 8x H100 one-shot run completed
• final exact metric and final byte accounting were captured from the successful run output and later cross-checked against the recovered remote log/artifacts

Limitations

• This is a clean single-seed result, not a new-SOTA statistical claim
• no multi-seed significance test is included in this PR
• this PR does not include multi-seed statistical evidence