[fank]

The real H100 gate stopped because the unmodified baseline was not repeat-stable. This establishes no packed/unpacked mismatch.

Changes:

• Match both formats' tuning dictionaries rather than independently retuning them.
• Synchronize/snapshot first output before second launch; assert independent output pointers and no overwrite.
• Print compiled Stream-K setting, tuning, differing element count, max absolute difference and output pointers before exact repeat assertion.
• Add explicit --disable-stream-k controlled storage-kernel comparison. Verify the compiled control is honored and label the pass scope. This is NOT production-default qualification. No equality tolerance changes.

Source inspection: Humming runner allocates fresh zero outputs/locks for each launch; launcher returns supplied output. SM90 BF16 Stream-K may use native atomic reductions for more than3 slices, with barriers gating only the first slice before other reducers. This is a plausible repeat-order mechanism, not a confirmed diagnosis without the new runtime log.

Validation here: Python compilation, --help and20 repository CPU tests pass. No GPU launched locally; serving owns runtime tests. Real-model/graph/packed-scale serving qualification remains pending.