Studio: harden video/diffusion cache, attention, and CFG-parallel fault paths

- diffusion_attention: arch-gate FlashAttention 2 to Ampere (SM80)+ in both the
  primary selector and the heterogeneous-replica guard (it crashed on pre-Ampere).
- diffusion_cfg_parallel: convert boolean attn masks to additive bias before the direct
  cuDNN op so partial masks match F.scaled_dot_product_attention; make proxy disable_cache
  transactional (clean both branches, mark broken, surface a reload-required error).
- diffusion_cache: fail closed when a magcache step-count resize or below-threshold
  disable cannot remove the old cache; surface a failed enable+cleanup instead of a false
  uncached None.
- video: roll back earlier experts when a later expert raises in the all-or-none step-cache
  loop; fail the load when the primary-only cache cannot be re-engaged through the
  CFG-parallel proxy; validate transformer_cache_quality and cfg_parallel before the worker.
- scripts: place the fp8 ablation pipeline on CUDA; fail closed on a failed magcache resize
  in the speedmem bench; label OOM distinctly in the SDPA mask probe.
- tests: regressions for the FA2 arch gate, transactional proxy disable, all-or-none
  exception rollback, magcache fail-closed transitions, and enable+cleanup failure.

studio/backend/tests/test_diffusion_cache.py

@@ -372,6 +372,49 @@ def test_toggle_reengages_after_a_disable(monkeypatch):
     assert mode == TC_FBCACHE and t.enables == 2
 
 
+def test_toggle_magcache_step_change_hard_errors_when_disable_fails(monkeypatch):
+    # A failed removal during a magcache step-count change used to fall through and report
+    # "magcache" while the OLD #sN curve stayed armed (wrong ratio schedule). Fail closed.
+    import core.inference.diffusion_cache as dc_mod
+
+    _stub_diffusers(monkeypatch)
+    t = _ToggleTransformer()
+    t._unsloth_step_cache = "magcache@0.06#s50"
+    monkeypatch.setattr(dc_mod, "_disengage_step_cache", lambda *a, **k: False)
+    with pytest.raises(RuntimeError, match = "reload the video model"):
+        maybe_toggle_step_cache(
+            _pipe(t), steps = 30, mode = dc_mod.TC_MAGCACHE, family = "hunyuanvideo-1.5"
+        )
+
+
+def test_toggle_below_bar_hard_errors_when_disable_fails(monkeypatch):
+    # Below the cache threshold we want uncached; a failed disable used to report the stale
+    # mode instead of surfacing that the (wrong-step) cache is still armed.
+    import core.inference.diffusion_cache as dc_mod
+
+    _stub_diffusers(monkeypatch)
+    t = _ToggleTransformer()
+    t._unsloth_step_cache = "magcache@0.06#s50"
+    monkeypatch.setattr(dc_mod, "_disengage_step_cache", lambda *a, **k: False)
+    with pytest.raises(RuntimeError, match = "reload the video model"):
+        maybe_toggle_step_cache(
+            _pipe(t),
+            steps = FBCACHE_MIN_STEPS - 1,
+            mode = dc_mod.TC_MAGCACHE,
+            family = "hunyuanvideo-1.5",
+        )
+
+
+def test_apply_step_cache_enable_and_cleanup_failure_requires_reload(monkeypatch):
+    # enable_cache fails AFTER partially hooking and the cleanup disable_cache ALSO fails:
+    # the transformer may keep partial hooks, so surface it instead of a clean uncached None.
+    _stub_diffusers(monkeypatch)
+    t = _MixinTransformer(fail = True)
+    t.disable_cache = lambda: (_ for _ in ()).throw(RuntimeError("cleanup failed"))
+    with pytest.raises(RuntimeError, match = "partially cached and must be reloaded"):
+        apply_step_cache(_pipe(t), mode = "fbcache")
+
+
 def test_toggle_noop_without_cache_support(monkeypatch):
     _stub_diffusers(monkeypatch)
     t = _NonCacheMixinTransformer()