Med3x sub-2-bit NIAH rescue test on Yi-6B-Chat

[jagmarques]

What

New Kaggle kernel testing whether Med3x outlier extraction rescues
sub-2-bit NIAH collapse on Yi-6B-Chat. Prior run: K2V1=6/40, K1V2=0/40,
K1V1=1/40. Hypothesis: magnitude outlier token-rows (norm > 3x median)
stored as fp16 restore recall.

Configs (nominal bpe before runtime overhead)

• FP16 baseline (must be non-degenerate at 4K)
• E8-K2V1 (1.8125 bpe) -- reproduces sub-2-bit collapse
• E8-K2V1 + Med3x -- rescue test
• E8-K1V2 (1.8125 bpe) -- worst collapse in prior run
• E8-K1V2 + Med3x -- rescue test
• E8-K3V2 (1.9375 bpe) -- near-lossless reference

Med3x bpe charged honestly: b+ = (1-p)b_E8 + p16 + 0.25 (p measured at runtime).

Protocol

• NIAH: 4K context, chat-template, n=20 trials per depth, verbatim value match
• PPL: paired-chunk wikitext-2 n>=40 segments, NF4 weights

Status

GPU result pending. No result claimed.

Proof of work

smoke test: exit 0
(a) Med3x outlier detection ... OK (row 22 flagged; fp16 rel_err=2.75e-04)
(b) Sign-flip invertible ... OK (max_diff=0.00e+00)
(c) NIAH needle recall (verbatim substring, not key trap) ... OK
(d) bpe formulas for K2V1, K1V2, K3V2 ... OK (K2V1=1.8125, K1V2=1.8125, K3V2=1.9375 bpe)
(e) E8 round-trip ... SKIP (nexusquant not installed locally)
All smoke tests PASSED.

kaggle kernels push: Kernel version 1 successfully pushed.
URL: https://www.kaggle.com/code/jooandrgomesmarques/nq-e8-med3x-yi

git diff --stat: 3 files changed, 852 insertions(+)
branch: company/e8-med3x-yi sha: 23dc020