[David Fan (jiafatom)]

Summary

Isolated follow-up to #33394, intentionally on a separate branch without modifying its head.

• Guard throwing ByteHelpers assertions for ORT_NO_EXCEPTIONS while preserving successful cases.
• Reject contradictory concrete scale/bias/zero-point row metadata when FC1 rows are symbolic; add regression mutations.
• Explain unpacked zero-point code 4 versus packed byte values.
• Add standalone reproducible padding benchmark, two corrected full runs and detailed report at docs/design/int3-qmoe-padding-experiment.md.

Experiment

A100-SXM4-80GB, CUDA13, actual Qwen FC1/FC2 dimensions, block32/64/128, compact versus 16/32-byte-aligned blocks, fixed8 and rotating128 experts, dequantization and one/32-token projections. Two runs of 108 cases each; seven trials per case; prefill16calls covers all128experts. Layout outputs match and independent scalar samples pass.

Block64 padding24->32bytes increases weight payload33.33%; rotating128expert decode is about11% slower and32token projection about10% slower in this scalar-load implementation. Byte-only padding adds nothing for these block sizes. This does not rule out future vectorized private layouts and is not a whole-model benchmark or pure bandwidth attribution.

Validation

Normal current-source schema/format tests17/17 passed. Format translation unit and real helper.cc compiled with ORT_NO_EXCEPTIONS:6/6 passed. Uses cached dependencies/isolated objects, not a clean full or globally exception-disabled build. After formatting, rebuilt benchmark quick108checks pass. Both full JSON protocols and report table verified.

Original #33394 code threads remain open until these changes enter its branch; this follow-up does not imply PR2-PR4 completion.