Clean SP8192 LegalTTT reproduction metadata

[teslaeco]

Motivation

• Final hygiene cleanup to ensure metadata files are plain UTF-8 / ASCII-safe and free of bidi/control characters without changing any reported metrics, logs, claims, or folder structure.

Description

• Rewrote submission.json as clean pretty-printed JSON (expanded the seeds array) and ensured README.md is plain UTF-8 text; preserved all numeric metrics, seeds, dates, hardware, compliance flags, logs, train_gpt.py, folder name, and the root leaderboard.

Testing

• Validated submission.json parses with python json.load, ran a byte-level scan for bidi/control characters on README.md and submission.json (no issues), verified required per-seed val_bpb values and dataset/tokenizer markers exist in each train_seed*.log, observed the upstream train_gpt.py reference was missing so no replacement was applied, and confirmed git diff --stat shows only submission.json changed (1 file changed, 5 insertions(+), 1 deletion(-)).

---

Codex Task