feat: turn on graph capture for MUSA

supports_cuda_graph was False for MUSA, so every run there was pinned to
timer=direct. That was correct when it was written -- the capture path named
torch.cuda and could not have worked -- but it is not correct now, and the two
timers are not interchangeable: direct timing walks the full Python launch path
per iteration, so on small kernels it measures mostly launch overhead. Measured
on an S5000, capture against direct: batch_norm 16x16x8x48 0.0758ms -> 0.0089ms,
tanh and gelu at 4096x4096 about 3.3x, GEMMs unchanged at 1.00x. Leaving it off
discards that on every operator that can capture, and it also puts MUSA on a
different caliber from the CUDA numbers the harness is meant to compare against.

torch_musa supplies Stream, Event, synchronize, graph and MUSAGraph, so capture
works through the device interface. Verified on an S5000 over five operators,
one per path:

  tanh      captured, no marker
  dropout   15/15 rows fall back: MUSAGeneratorImpl::current_seed is not
            callable during capture
  rand      same
  perf_sort 25/25 rows fall back: operation not permitted when stream is
            capturing
  unique2   denied up front; rc=0 with 10 rows

MUSA gets its own error wording; none of the CUDA markers match it. "MUSA
error: unknown error" during capture is classified as poisoning rather than
recoverable, because the fallback runs and the process then dies in teardown:
before the deny entry was keyed correctly, unique2 reproduced exactly that,
ending at rc=1 with no rows, but now under an explicit BENCHMARK_CAPTURE_POISONED
marker instead of a bare ExceptionGroup.

capture_deny_ops is keyed by FLAGGEMS_PERF_CURRENT_OP, which holds the pytest
FUNCTION name minus `test_` -- unique2, perf_sort -- not the benchmark file
name. Getting that wrong fails open silently, which is how the first attempt
still hit rc=1 with the entry written as "unique"; a test now pins the key
space. The plugin-plan test is renamed because its assertion that
_cudagraph_plugin is absent from the plan holds on CUDA too and says nothing
about capture support.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backends/musa.py

@@ -8,9 +8,40 @@ class MusaBackend(DeviceBackend):
     visibility_env_keys = ("MUSA_VISIBLE_DEVICES",)
     torch_namespace = "musa"
     flaggems_vendor = "musa"
-    supports_cuda_graph = False
+    # torch_musa supplies Stream, Event, synchronize, graph and MUSAGraph, so
+    # capture works once the timing path takes them from Triton's device
+    # interface instead of torch.cuda. Measured on an S5000 over 15 operators:
+    # 11 captured, 3 fell back cleanly with a marker (dropout/rand on the RNG
+    # seed, sort on a capture-illegal stream op), 1 poisoned the context (see
+    # capture_deny_ops). Capture is worth having because it removes launch
+    # overhead the direct timer cannot: batch_norm 16x16x8x48 went 0.0758ms ->
+    # 0.0089ms, tanh/gelu 4096x4096 about 3.3x, while GEMMs stayed at 1.00x.
+    supports_cuda_graph = True
     supports_mm_cluster_fix = False
     supports_triton_hooks = True
+    # MUSA reports its own wording; none of the CUDA markers match it.
+    fatal_error_markers = (
+        "musa error: an illegal memory access",
+        "out of memory",
+        "device-side assert triggered",
+        "misaligned address",
+        "uncorrectable ecc",
+    )
+    # "unknown error" during capture is not recoverable on this runtime: the
+    # fallback runs, and the process then dies in teardown with zero rows. Treat
+    # it as poisoning so the operator is reported as capture-incompatible rather
+    # than silently losing its measurement.
+    poisoning_error_markers = ("musa error: unknown error",)
+    # Keyed by FLAGGEMS_PERF_CURRENT_OP, which is the pytest FUNCTION name minus
+    # `test_` -- `unique2`, `perf_sort` -- not the benchmark file name. Getting
+    # that wrong is silent: the entry simply never matches.
+    #
+    # Observed on an S5000: capturing test_unique2 takes two recoverable
+    # fallbacks and then raises "MUSA error: unknown error", ending the run at
+    # rc=1 with no measured rows. Skipping the attempt costs that operator's
+    # timing precision instead of its measurement. Revisit if a newer torch_musa
+    # changes the behaviour.
+    capture_deny_ops = ("unique2",)
     # Device count comes from torch.musa or MUSA_VISIBLE_DEVICES. Do not guess
     # a vendor CLI here; confirm the real listing tool before adding it.
     list_command = None