Add W104 faithful SP8192 LegalTTT bad-seed probe

[teslaeco]

Motivation

• Reduce bad-seed variance by providing a faithful, evaluator-friendly replay candidate of the SP8192 + LegalTTT near‑SOTA stack with key defaults source-visible.
• Start reproducibility probe by running only seed=314 first and gate additional seeds (42, 999) on its pass to focus effort on the bad-seed failure mode.

Description

• Add a new working folder records/track_10min_16mb/2026-04-20_SP8192_LegalTTT_W104_FaithfulReplay/ containing the replay artifacts and guidance.
• Add train_gpt.py that surfaces source-visible defaults at the top (VOCAB_SIZE, TOKENIZER_PATH, DATA_PATH, TRAIN_SHARDS_OVERRIDE, QK_GAIN_INIT, TTT_ENABLED, TTT_LR, TTT_EPOCHS) and exports them via os.environ.setdefault(...), while preserving the original packed SP8192 LegalTTT payload and architecture/compression surface (3-layer recurrence, parallel residuals, QK gain 5.25, legal score-first TTT, quantized+brotli target under 16 MB).
• Add run_w104_seed314_probe.sh which bootstraps a venv and deps, caches the official FineWeb sp8192 using MATCHED_FINEWEB_REPO_ID="kevclark/parameter-golf", runs only SEED=314, writes logs to /workspace/w104_seed314.log, and prints the final quantized_ttt val_bpb.
• Add README.md explaining this is not a new submission, documenting intent and pass criteria for seed314 (must beat 1.08168719, strong pass <1.0812) and instructing to run seeds 42/999 only after seed314 passes.

Testing

• Ran python -m py_compile train_gpt.py in the new folder and it succeeded.
• Verified source‑visible defaults with grep for the expected lines in train_gpt.py and all expected defaults were found.
• Confirmed the change summary via a diff-stat showing the three new files and the expected insertions.

---

Codex Task