Add Parameter Golf submission: Vocab768_LinearPhaseInit_GatedXSA_EMA_…

[shram86]

Adds a new track_10min_16mb submission record based on a custom sp768 tokenizer export.

Result:

• final_int6_roundtrip val_loss: 1.87181421
• final_int6_roundtrip val_bpb: 1.21149167
• total submission size int6+zstd: 15082805 bytes

Included:

• record README
• submission.json
• native train_gpt.py
• run log

Main ingredients:

• vocab_size=768
• phase-mix init
• gated XSA on last 2 layers
• EMA
• late matrix-only QAT
• FlashAttention 3

Tokenizer/data export:

• the sp768 tokenizer artifacts and retokenized dataset export were published to the Hugging Face dataset repo shramoff/golf768
• locally, I used a small extension of cached_challenge_fineweb.py to support a custom repo id and root-level remote layout for sp768