diff --git a/docs/comm-domain.md b/docs/comm-domain.md
index b7db6f67c4..f2de05c3ae 100644
--- a/docs/comm-domain.md
+++ b/docs/comm-domain.md
@@ -237,6 +237,19 @@ the resident `KernelArgs`, and injects it into every run's kernel
 SDMA streams and its kernels read a zero workspace address. The workspace is
 released at Worker finalize by ordinary stream/manager teardown.
 
+**`enable_sdma` is a defect quarantine, not a capability switch, and it has an
+exit condition.** Selecting an engine already belongs to the kernel, which reads
+whichever addresses were injected; the runtime would otherwise provision every
+engine the platform supports. SDMA is the exception because
+`SdmaWorkspaceManager::Init()` is indivisible — the 16 KB workspace *is* the
+descriptor table for 48 CP-process STARS streams, so there is no way to have the
+address without holding the streams — and a Worker holding those streams gets a
+single device-reset attempt instead of three after an AICore fault. Defaulting it
+on would therefore put every ordinary Worker in that population, which is exactly
+the regression [the investigation](investigations/2026-07-a2a3-sdma-fault-teardown.md)
+records. The flag disappears once CANN bounds the final CP-process stream release,
+or once pto-isa offers an `Init()` that separates the workspace from the streams.
+
 Provisioning also warms the SDMA control path once, in the same call: a
 vector-only AICore ELF (`sdma_warmup_kernel.o`, staged per arch under
 `build/lib/<arch>/sdma_warmup/`) walks every channel so the first
@@ -260,7 +273,7 @@ and issue #1425. `enable_sdma` is currently honored only by the a2a3 onboard
 `tensormap_and_ringbuffer` runtime; host-build-graph, simulation, a5, and
 provider-disabled builds fail Worker init fast when it is set. A5 provisions
 its communication-context SDMA workspace by default; this is separate from
-the callable-declared workspace mechanism controlled by `enable_sdma`.
+the Worker-level workspace mechanism controlled by `enable_sdma`.
 
 ---
 
diff --git a/src/common/platform/include/aicpu/aicpu_device_config.h b/src/common/platform/include/aicpu/aicpu_device_config.h
index 88051e7013..4b9a65a658 100644
--- a/src/common/platform/include/aicpu/aicpu_device_config.h
+++ b/src/common/platform/include/aicpu/aicpu_device_config.h
@@ -55,9 +55,10 @@ int get_scheduler_timeout_ms();
  * kind (see DmaWorkspaceKind). Published by simpler_aicpu_init (from
  * InitArgs.dma_workspace_addr[]) into a resident-SO array; the scheduler
  * copies each slot into every core's GlobalContext, so kernels read it via
- * get_dma_workspace(args, kind). 0 means no callable has provisioned that
- * engine. A callable that declares the engine is rejected before launch if the
- * platform cannot provide a non-zero address. Out-of-range kinds are ignored.
+ * get_dma_workspace(args, kind). 0 means this Worker did not provision that
+ * engine — the platform does not support it, or the Worker declined it. A
+ * kernel that reads 0 is responsible for its own fallback; nothing rejects it
+ * before launch. Out-of-range kinds are ignored.
  */
 void set_dma_workspace_addr(int kind, unsigned long long addr);
 
diff --git a/src/common/platform/include/common/dma_workspace.h b/src/common/platform/include/common/dma_workspace.h
index de601d7904..f6bdbf5e15 100644
--- a/src/common/platform/include/common/dma_workspace.h
+++ b/src/common/platform/include/common/dma_workspace.h
@@ -11,22 +11,34 @@
 /**
  * Async-DMA engine workspace kinds.
  *
- * The runtime provisions a per-device scratch workspace for each async-DMA
- * engine it supports and injects the device addresses into every core's
- * GlobalContext, so kernels obtain them via get_dma_workspace(args, kind)
- * without threading them as user args. One slot per engine, indexed by this
- * enum; DMA_WORKSPACE_KIND_COUNT sizes the injected array end to end
- * (InitArgs, the resident AICPU config, and GlobalContext).
+ * The runtime provisions a per-device scratch workspace and injects the device
+ * addresses into every core's GlobalContext, so kernels obtain them via
+ * get_dma_workspace(args, kind) without threading them as user args. Selecting
+ * an engine is the kernel's decision, not the runtime's: every provisioned
+ * address is injected, and a kind the runtime did not provision reads back 0.
+ * One slot per engine, indexed by this enum; DMA_WORKSPACE_KIND_COUNT sizes the
+ * injected array end to end (InitArgs, the resident AICPU config, and
+ * GlobalContext).
  *
  * simpler-owned and deliberately independent of pto-isa's comm::DmaEngine —
  * the kernel maps this kind to the pto-isa engine tag at the call boundary.
  * Shared by host (provisioning + InitArgs) and device (scheduler + kernels),
  * so it carries no dependencies beyond the enum itself.
  *
+ * A Worker provisions an engine only when it both requested that engine and the
+ * device supports it (dma_workspace_supported_mask()), so an unrequested or
+ * unsupported kind stays 0. SDMA is the only engine anyone declines today, and
+ * the only one worth declining: provisioning it is inseparable from holding 48
+ * CP-process STARS streams whose post-fault release CANN bounds neither with a
+ * completion fence nor with a portable timeout — see
+ * docs/investigations/2026-07-a2a3-sdma-fault-teardown.md. A Worker declines it
+ * by leaving `enable_sdma` off, which is the default.
+ *
  * Current support matrix: SDMA is available only on a2a3 onboard with the
  * tensormap_and_ringbuffer runtime. URMA is reserved for the future a5
  * per-domain provider. Host-build-graph, simulation, a5, and builds without
- * the a2a3 PTO-SDMA provider reject non-empty requirements at registration.
+ * the a2a3 PTO-SDMA provider reject a request for SDMA during Worker
+ * initialization, when the workspace would be provisioned.
  */
 
 #ifndef PLATFORM_COMMON_DMA_WORKSPACE_H_