Half the tie-stability device sweep is a duplicate: VT_MOE_ROUTER_WARP does not change the dispatch above E = 256

[localai-org-maint-bot]

Row: MODEL-MM-QWEN4-EXP

The gap

tests/vt/test_moe_router_tie_stability.cpp (landed by #2595) runs its device
sweep twice, once with VT_MOE_ROUTER_WARP pinned "1" and once pinned "0",
across E in {256, 512, 1024}. At E = 256 those are two different kernels. At
E = 512 and E = 1024 they are the same kernel launched twice:

• MoeRouterWarpValuesPerThread (src/vt/cuda/moe_router_warp.h:96-98) returns
  0 for any E outside {32, 64, 128, 256}.
• LaunchRouterWarp (src/vt/cuda/cuda_moe.cu:564) returns false on vpt == 0
  before touching the tensors, so LaunchRouter (:599-609) falls through to
  MoeRouterTopKKernel<Tin,false> whatever the env flag says.

What is and is not affected

The counts are honest. 108 rows, 4652 assertions and the mutant's 104
failures are all real assertions that really executed, and the closed-form
expectation is re-checked each time. Nothing is inflated in the sense of being
uncounted.

The COVERAGE at E > 256 is half what the doubled numbers imply. Of the 108
rows in case 1, 72 are at E in {512, 1024} and 36 of those are byte-identical
repeats of the other 36. In the mutation table the per-cell figure 12 is
2 arms x 3 h x 2 dtypes, so 6 of each 12 are the same disagreement counted a
second time. A reader who takes "both VT_MOE_ROUTER_WARP arms" as two
independent structures at the geometry qwen4_exp actually routes reads twice
the coverage that exists.

A comment saying so is added to the file by #2595. This issue owns the fix.

What closing this means

One of:

1. Restrict the arm loop to the geometries where the flag changes the dispatch
   (E = 256 here), and say in the file that E > 256 has one kernel; or
2. Keep both arms and assert the redundancy instead of implying independence —
   e.g. REQUIRE(MoeRouterWarpValuesPerThread(e) == 0) at E > 256, which
   turns the duplicate into a checked property of the dispatch rather than an
   unstated one; or
3. Widen MoeRouterWarpValuesPerThread so the warp kernel really does admit
   E = 512, which is a kernel change with its own derivation obligation
   (moe_router_warp.h excludes E > 256 on purpose, because the seed loops at
   cuda_moe.cu:71,83 accumulate several experts per thread in ascending order).

(3) is a product change and needs its own spec. (1) or (2) is a test repair and
needs GPU time to re-measure the counts every record cites, which is why it is
not folded into #2595.