[non-record] 1xH100 screening: compression + eval strategy

[numb3r33]

Summary

This PR adds a non-record 1xH100 screening bundle documenting the March 26 log-backed experiment matrix we used to prioritize further compute.

What’s Included

• B0, Q1, Q3, C1, and C2 raw screening logs
• submission.json anchored on the checked B0 baseline result
• a README summarizing the matrix and the main finding
• surviving script-family snapshots for:
  • dense baseline
  • fp16-embedding family
  • 10-layer mixed-precision family

Main Finding

The main result from this screen is that pre-quant and post-quant quality diverge sharply once we push capacity or compression too hard.

• C1 improved pre-quant BPB over the baseline, but lost after quantization
• Q1 stayed near the baseline while producing a meaningfully smaller artifact
• Q3 and C2 show that making the model smaller or easier to compress is not enough if the post-quant weight distribution becomes too fragile

This points the next round of compute toward:

• evaluation strategy, especially rerunning the missing-log sliding-window candidate
• compression-aware training and quantization-friendly schedules
• only then 8xH100 validation on the top candidates

Notes

• This is explicitly a non-record submission and not an 8xH100 leaderboard claim.
• A later 1xH100 run summary suggested stronger sliding-window results, but those logs were not preserved, so they are intentionally excluded from this public bundle.