diff --git a/.claude/skills/test-all-device/SKILL.md b/.claude/skills/test-all-device/SKILL.md
index 52559a7ff2..dde20bd042 100644
--- a/.claude/skills/test-all-device/SKILL.md
+++ b/.claude/skills/test-all-device/SKILL.md
@@ -16,15 +16,34 @@ Detection / isolation procedures referenced below live in
    `--pto-session-timeout`.
 4. **Select a device range** (§C, range ≤4) — or, when wrapping in
    `task-submit`, let it pick via `--device auto --device-num <range size>`.
-5. **Run through `task-submit`** (§E). The underlying command:
+5. **Read the marker selector out of the same job** you took the timeout from.
+   `st-onboard-a2a3` is two pytest passes, not one: the sweep deselects
+   `-m "not sdma"` and a later step runs `-m sdma`. `st-onboard-a5` has no
+   marker filter. Take the expression from `ci.yml` rather than assuming, so
+   this skill cannot drift from the job it reproduces.
+6. **Run through `task-submit`** (§E). The underlying command, on a platform
+   whose job carries the filter:
 
    ```bash
-   pytest examples tests/st --platform <platform> --device <range-or-$TASK_DEVICE> \
-     --pto-session-timeout <timeout> -v
+   pytest examples tests/st -m "not sdma" --platform <platform> \
+     --device <range-or-$TASK_DEVICE> --pto-session-timeout <timeout> -v
    ```
 
+   Then the quarantined pass, after it — never at the same time, since the
+   point of the split is that no fault-injection case meets a device that has
+   already provisioned SDMA:
+
+   ```bash
+   pytest examples tests/st -m sdma --platform <platform> \
+     --device <2 devs or $TASK_DEVICE> --pto-session-timeout <timeout> -v
+   ```
+
+   Dropping either pass changes what you tested: a single flat sweep is not
+   what CI runs and reports failures CI never sees
+   ([`../testing/SKILL.md`](../testing/SKILL.md), issue #1425).
+
    Parallelism is auto-driven by `--device`: on hardware, one in-flight
    subprocess per device (`--max-parallel auto` = `len(--device)`); see
    `docs/testing.md` for the full reuse hierarchy.
-6. Report the results summary (pass/fail counts per task).
-7. If any tests fail, show the relevant error output and which device failed.
+7. Report the results summary (pass/fail counts per task), across both passes.
+8. If any tests fail, show the relevant error output and which device failed.
diff --git a/.claude/skills/test-runtime-device/SKILL.md b/.claude/skills/test-runtime-device/SKILL.md
index cd6bb55a06..6b3a026ca5 100644
--- a/.claude/skills/test-runtime-device/SKILL.md
+++ b/.claude/skills/test-runtime-device/SKILL.md
@@ -18,15 +18,30 @@ Detection / isolation procedures referenced below live in
    `--pto-session-timeout`.
 5. **Select a device range** (§C, range ≤4) — or, when wrapping in
    `task-submit`, let it pick via `--device auto --device-num <range size>`.
-6. **Run through `task-submit`** (§E). The underlying command:
+6. **Read the marker selector out of the same job** you took the timeout from.
+   `st-onboard-a2a3` is two pytest passes: the sweep deselects `-m "not sdma"`
+   and a later step runs `-m sdma`. `st-onboard-a5` has no marker filter. Both
+   quarantined tests are `tensormap_and_ringbuffer`, so the second pass is only
+   needed when `$ARGUMENTS` names that runtime.
+7. **Run through `task-submit`** (§E). The underlying command:
 
    ```bash
-   pytest examples tests/st --platform <platform> --runtime $ARGUMENTS \
+   pytest examples tests/st -m "not sdma" --platform <platform> --runtime $ARGUMENTS \
      --device <range-or-$TASK_DEVICE> \
      --pto-session-timeout <timeout> -v
    ```
 
+   Then, for `tensormap_and_ringbuffer` on a platform whose job carries the
+   filter, the quarantined pass afterwards — never concurrently, since the
+   split exists so that no fault-injection case meets a device that has already
+   provisioned SDMA ([`../testing/SKILL.md`](../testing/SKILL.md), issue #1425):
+
+   ```bash
+   pytest examples tests/st -m sdma --platform <platform> --runtime $ARGUMENTS \
+     --device <2 devs or $TASK_DEVICE> --pto-session-timeout <timeout> -v
+   ```
+
    Hardware parallelism is auto-driven by `--device` (one subprocess per
    device); no extra flag needed.
-7. Report the results summary (pass/fail counts per task).
-8. If any tests fail, show the relevant error output and which device failed.
+8. Report the results summary (pass/fail counts per task), across both passes.
+9. If any tests fail, show the relevant error output and which device failed.
diff --git a/docs/ci.md b/docs/ci.md
index 14a916b836..3149386401 100644
--- a/docs/ci.md
+++ b/docs/ci.md
@@ -82,10 +82,12 @@ benefit — device bin-packing for L3, xdist fanout for L2, and a shared
 `ChipWorker` per `(runtime, device)`:
 
 ```bash
-# Recommended CI invocation
-pytest examples tests/st --platform a2a3 --device 4-7 -x
+# Recommended CI invocation — a2a3 deselects the SDMA marker, as the job does,
+# and runs it as a second pass afterwards
+pytest examples tests/st -m "not sdma" --platform a2a3 --device 4-7 -x
+pytest examples tests/st -m sdma --platform a2a3 --device 4-5 -x
 
-# Same for a5
+# Same for a5, which has no marker filter
 pytest examples tests/st --platform a5 --device 0-7 -x
 ```