index row for #1954 wrongly claims #907 documents the same test and assertion

Row: BACKEND-ROCM

The #1954 row in
.agents/issue-index.md claims issue #907 documents "the same test name and
the same assertion" as #1954 on GB10. This is false.

#907 records test_cuda_ops's "CUDA silu_and_mul matches CPU" case failing
at 439 of 440 assertions on GB10. #1954 is about
test_backend_cross_device's "MoeSiluMul matches the CPU oracle within
NMSE <= 5e-4" case, at :2063, on gfx1200 ROCm. Different test, different
assertion. Both are last-digit bf16 numerics in a silu-and-mul kernel, so
they are the same defect family, but #907 is not a second recorded instance
of #1954's specific test and assertion.

#1954's own GitHub title and body have already been corrected to state this
plainly. .agents/environment.md and .agents/specs/rocm-kquant-nwarps-decode.md
## Owed have also been corrected on row/ROCM-KQUANT-NWARPS-DECODE. The
.agents/issue-index.md row cannot be corrected in place: the file is
append-only, and scripts/check-agent-record.py refuses a second row citing
an issue number already in the index. This issue exists so a corrective row
can cite it instead, following the same pattern as
#1339 superseding #1280 and
#1796 retracting #1456.

Owned by BACKEND-ROCM.

Start with .agents/issue-index.md and compare the superseding and retraction patterns in issues #1339 and #1796. Check scripts/check-agent-record.py to understand the append-only and duplicate-issue constraints. Done means adding a corrective row that cites issue #2013 without claiming #907 records the same test and assertion as #1954.