[BUG] Example 93 GQA initialization overwrites active remote contributions

Which component has the problem?

CUTLASS C++

Bug Report

Describe the bug

Example 93 initializes active split mailboxes after cluster startup synchronization. A receiving CTA can erase a contribution already written by a remote producer. Incorrect attention output was observed under Compute Sanitizer racecheck.

Steps/Code to reproduce bug

Use revision 246e9206: CUTLASS 59e3a3338d516ca6ce0e073af8da65289678a35c plus the separate scratch-reduction barrier. Keeping that barrier removes scratch reuse as a cause.

curl -fL https://raw.githubusercontent.com/1sgtpepper/cutlass/923e9400d6901f90f30d18510320d08d838fbcac/test/regression/gqa_max_scratch.cu -o gqa_max_scratch.cu
nvcc -std=c++17 -O3 -lineinfo --expt-relaxed-constexpr \
  -gencode=arch=compute_100a,code=sm_100a \
  -Iinclude -Itools/util/include -Iexamples/93_blackwell_low_latency_gqa \
  gqa_max_scratch.cu -o gqa_max_scratch
compute-sanitizer --tool racecheck --error-exitcode 3 \
  ./gqa_max_scratch --mailboxes --repetitions 3

Across seven eight-CTA cases, output contained 473 wrong values, despite zero reported racecheck hazards. Cases include all-active producers and inactive producers with a partial tail. The reference is double-precision softmax with 1/256 absolute tolerance.

Expected behavior

Active producers exclusively write their mailbox contributions. Inactive entries contain the neutral values needed by the final reduction.

Environment details (please complete the following information):

• CUTLASS revisions above; main as of September 8, 2026: 59e3a3338d516ca6ce0e073af8da65289678a35c. The affected mailbox initialization is unchanged in v4.8.0dev and v4.7.1; these releases were not executed.
• B200/SM100, driver 580.95.05.
• CUDA 13.1.1 and bundled Compute Sanitizer.
• Modal cloud GPU, Ubuntu 24.04 container, Linux x86_64.

Additional context

The local epilog barrier cannot order remote stores against initialization. Initializing only inactive entries changes the wrong-value count from 473 to 0; restoring the old initialization gives 196. Counts vary with scheduling. Ordinary launches passed; these numerical failures occurred under racecheck.

Start by reproducing the reported command with test/regression/gqa_max_scratch.cu, the referenced CUTLASS revision, and Compute Sanitizer racecheck. Inspect mailbox initialization in examples/93_blackwell_low_latency_gqa and compare initialization of active and inactive entries. Done means the reproduced cases produce no wrong values while inactive entries retain the neutral values required by the final reduction.