Experiment: SmearGate BOS Fix + train-only logit calibration

[someone114514]

Summary

Experimental variant of #1868 / #1851 SmearGate BOS Fix that adds a fixed post-GPTQ logit calibration pass.

The added calibration is deliberately small and train-only:

• global logit temperature
• coarse token-group bias buckets: byte length, starts-with-space, newline, digit, punctuation, alpha/case
• no validation-derived fitting state
• no frequency buckets by default
• frozen after fitting, then applied before softmax in quantized diagnostic eval and phased score-first TTT

This is intended as a direct test of whether the post-GPTQ calibration signal observed locally transfers to the stronger #1868 stack.

Controls

Defaults added in this branch:

LOGIT_CALIB_ENABLED=1
LOGIT_CALIB_TOKENS=100000
LOGIT_CALIB_STRIDE=64
LOGIT_CALIB_BATCH_SEQS=8
LOGIT_CALIB_LR=0.003
LOGIT_CALIB_L2=0.01
LOGIT_CALIB_EPOCHS=1
LOGIT_CALIB_APPLY_TTT_UPDATE=1

Set LOGIT_CALIB_ENABLED=0 to recover the original #1868 behavior.

Legality / causality

Calibration is fitted only from training-token shards after GPTQ. It does not read validation targets or build validation-time state. At validation time the correction is a fixed affine transformation of logits before normal softmax, so the distribution remains normalized.

Status

No new 8xH100 score yet. This branch is prepared for a direct single-seed run against the #1868 reproduction command.