Non-record: SP8192 D 5-seed base and R-series evidence package

[amrayach]

Summary

This PR adds a non-record submission package for the SP8192 D base and the 2026-04-09 R-series sweep.

The package is intended as a self-contained evidence folder under records/track_non_record_16mb/ and does not claim a new SOTA or a submission-valid lead result.

Key points

• Canonical base: D (SP8192 #1413 family with parallel residual + loop adjustment)
• Canonical result: 1.08128837 score-first TTT BPB (5-seed mean, sigma = 0.00058943)
• Best measured single-seed follow-up: R1_e_baseline = 1.08078562
• Important caveat: R1_e_baseline ran at 605s, so it is not a clean lead submission number
• Main measured finding: on this stack, OWC/CDQuant improve raw BPB but create a fixed-Brotli compression-entropy penalty that breaks the 16 MB cap

What is included

• README.md
• REPORT.md
• submission.json
• requirements.txt
• single-file package-local train_gpt.py derived from the archived seed-0 script and helper chain
• canonical D train logs for seeds 0, 42, 1234, 1337, 2025
• d_submission_summary.tsv
• r1_e_baseline.log
• r_series_combined_summary.tsv
• ARTIFACT_MAP.md

Important framing

• This is a non-record submission.
• The package’s main evidence claim is anchored to the completed RunPod D 5-seed bundle.
• Pegasus rerun failures are documented as operational context only, not as a missing primary-evidence requirement for the package.
• The package-local train_gpt.py is a single counted code file derived from the archived seed-0 runtime chain for submission review.

Notes

• Relative to the submission base, this branch only adds one new folder under records/track_non_record_16mb/.
• The package includes a minimal requirements.txt, but reproducing the measured runs still assumes the matching CUDA/Hopper environment with flash_attn_interface available.