diff --git a/docs/ci.md b/docs/ci.md
index c59ce836ff..14a916b836 100644
--- a/docs/ci.md
+++ b/docs/ci.md
@@ -28,8 +28,9 @@ The complete test-type × hardware-tier matrix. Empty cells have no tests yet; o
 PullRequest
   ├── pre-commit             (ubuntu-latest)
   ├── packaging-matrix       (ubuntu + macOS)        — [needs !examples_only]
+  ├── profiling-flags-smoke  (ubuntu-latest)         — (a2a3_changed || a5_changed) && !examples_only
   ├── ut                     (ubuntu + macOS)        — Python + C++ UT, no hardware [needs ut_affected]
-  ├── detect-changes         (ubuntu-latest)         — outputs non_code_only, a{2a3,5}_changed, {st,ut}_affected
+  ├── detect-changes         (ubuntu-latest)         — outputs non_code_only, a{2a3,5}_changed, {st,ut}_affected, examples_only
   ├── st-sim-a2a3            (ubuntu + macOS)        — a2a3_changed && st_affected
   ├── st-sim-a5              (ubuntu + macOS)        — a5_changed && st_affected
   ├── ut-a2a3                (a2a3 self-hosted)      — Python + C++ UT, a2a3 hardware [needs ut_affected]
@@ -44,7 +45,7 @@ PullRequest
 | `st-sim-a2a3` | `ubuntu-latest`, `macos-latest` | `pytest examples tests/st --platform a2a3sim` |
 | `st-sim-a5` | `ubuntu-latest`, `macos-latest` | `pytest examples tests/st --platform a5sim` |
 | `ut-a2a3` | a2a3 self-hosted | `pytest tests/ut --platform a2a3` + `ctest -L "^requires_hardware(_a2a3)?$" --resource-spec-file ...` + build `tools/cann-examples/query` and run `query version` (no device) + build `tools/cann-examples/aicpu-device-query` and `tools/cann-examples/aicpu-kernel-launch` (host + cross-compiled device SO, link smoke only) |
-| `st-onboard-a2a3` | a2a3 self-hosted | `pytest examples tests/st --platform a2a3 --device ...` |
+| `st-onboard-a2a3` | a2a3 self-hosted | `pytest examples tests/st -m "not sdma" --platform a2a3 --device ...`, then a separate `-m sdma` step, then the DFX per-feature smokes |
 | `ut-a5` | a5 self-hosted | `pytest tests/ut --platform a5` + `ctest -L "^requires_hardware(_a5)?$"` + build `tools/cann-examples/query` and run `query version` (no device) + build `tools/cann-examples/aicpu-device-query` and `tools/cann-examples/aicpu-kernel-launch` (link smoke only) |
 | `st-onboard-a5` | a5 self-hosted | `pytest examples tests/st --platform a5 --device ...` |
 
@@ -128,15 +129,17 @@ not need `--max-parallel` manually.
 ### Scheduling constraints
 
 - Sim scene tests and no-hardware unit tests run on github-hosted runners (no hardware).
-- `detect-changes` computes three flags (`a2a3_changed`, `a5_changed`, `non_code_only`) from the PR diff, **all three derived from one `NON_CODE` set**: `docs/`, `.docs/`, `.claude/`, `mkdocs.yml`, `.github/workflows/docs.yml`, `.gitignore`, `.pre-commit-config.yaml`, and any `*.md` file anywhere. Membership follows a file's *effect*, not its path — `mkdocs.yml` and `docs.yml` are docs tooling that happens to live outside `docs/`. An arch flag is `false` only when every changed file is in the opposite platform's tree (`src/{arch}/`, `examples/{arch}/`, `tests/{st,ut/cpp}/{arch}/`) or in `NON_CODE`. Anything else — shared C++ (`src/common/`), Python (`python/`, `simpler_setup/`), build files (`CMakeLists.txt`, `pyproject.toml`), shared test infra (`tests/ut/py/`, `tests/lint/`), tooling (`tools/`), or **`.github/workflows/ci.yml` itself** — flips both flags to `true`. `ci.yml` is deliberately excluded from `NON_CODE`: a change to the gates must run everything, including whatever it just switched off.
+- `detect-changes` computes four axes from the PR diff — non-code (`non_code_only`), architecture (`a2a3_changed` / `a5_changed`), test category (`st_affected` / `ut_affected`), and example corpus (`examples_only`) — **all of them derived from one `NON_CODE` set**: `docs/`, `.docs/`, `.claude/`, `mkdocs.yml`, `.github/workflows/docs.yml`, `.gitignore`, `.pre-commit-config.yaml`, and any `*.md` file anywhere. Membership follows a file's *effect*, not its path — `mkdocs.yml` and `docs.yml` are docs tooling that happens to live outside `docs/`. An arch flag is `false` only when every changed file is in the opposite platform's tree (`src/{arch}/`, `examples/{arch}/`, `tests/{st,ut/cpp}/{arch}/`) or in `NON_CODE`. Anything else — shared C++ (`src/common/`), Python (`python/`, `simpler_setup/`), build files (`CMakeLists.txt`, `pyproject.toml`), shared test infra (`tests/ut/py/`, `tests/lint/`), tooling (`tools/`), or **`.github/workflows/ci.yml` itself** — flips both flags to `true`. `ci.yml` is deliberately excluded from `NON_CODE`: a change to the gates must run everything, including whatever it just switched off.
 - **Test-category axis:** `ST_ONLY='^(tests/st/|examples/)'` and `UT_ONLY='^tests/ut/'`, applied in the same shape as the arch patterns — a category is unaffected only when *every* changed file is exclusively the other's. `st_affected` gates the four scene-test jobs; `ut_affected` gates `ut`, `ut-a2a3`, `ut-a5`. Anything belonging to neither (root `conftest.py`, `pyproject.toml`, `simpler_setup/`, `tests/lint/`) flips both, so shared infrastructure always runs both suites.
 - **Example-corpus axis:** `EXAMPLES_ONLY='^examples/'`, same shape again. `packaging-matrix` and `profiling-flags-smoke` build and install the product and then exercise it with a fixed, tiny payload — one entry-point script and one `vector_example` — so they read neither suite as a corpus and nothing under `examples/` can reach them: it is absent from both `wheel.packages` and the sdist `include`. `tests/` deliberately stays in, because `tools/verify_packaging.sh` runs a `tests/st/` file as its entry-point smoke.
 - **Gated jobs (scene tests):** `st-sim-{a2a3,a5}`, `st-onboard-{a2a3,a5}` run iff their platform's flag **and** `st_affected` are `true`.
 - **Platform-independent jobs (all UT + packaging):** `ut`, `ut-a2a3`, `ut-a5`, `packaging-matrix` ignore the *platform* flags — unit tests exercise shared contracts (nanobind bindings, RuntimeBuilder, ring buffers, etc.) and the risk of silently skipping a regression outweighs the CI minutes saved. The `tests/ut/cpp/{arch}/` entry in the gating regex only *attributes* an arch-specific C++ UT change to that platform (so it does not spuriously flip the other arch's scene-test flag); it does not gate the UT jobs themselves. The three UT jobs do respect `ut_affected`, which is a statement about test category rather than silicon. `packaging-matrix` respects `examples_only`, which is a statement about which corpus a job reads: `examples/` is in neither `wheel.packages` nor the sdist `include`, and `tools/verify_packaging.sh` never opens it.
-- **`non_code_only` is the same `NON_CODE` set, not a narrower one.** It is `true` when no changed file falls outside it. Nothing in the set can change what the code does, and no workflow consumes any of it beyond its own gate: `pre-commit` is ungated so it always exercises `.pre-commit-config.yaml`, `docs.yml` is unconditional on every PR so it always exercises `mkdocs.yml` / `docs/`, and **no workflow invokes anything under `.claude/`** (`grep -rn '\.claude' .github/workflows/` finds only comments). An **empty diff short-circuits the whole step**: attribution is impossible, so a single guard sets `non_code_only=false` and every arch and category flag `true`, then returns — running the full matrix. That guard is deliberately one place; testing emptiness per flag is what previously left `non_code_only` false while both arch flags also came out false, running UT and packaging but skipping every scene test.
+- **`non_code_only` is the same `NON_CODE` set, not a narrower one.** It is `true` when no changed file falls outside it. Nothing in the set can change what the code does, and no workflow consumes any of it beyond its own gate: `pre-commit` is ungated so it always exercises `.pre-commit-config.yaml`, `docs.yml` is unconditional on every PR so it always exercises `mkdocs.yml` / `docs/`, and **no workflow invokes anything under `.claude/`** (`grep -rn '\.claude' .github/workflows/` finds only comments). An **empty diff short-circuits the whole step**: attribution is impossible, so a single guard sets `non_code_only=false`, every arch and category flag `true`, and `examples_only=false`, then returns — running the full matrix, packaging and the profiling smoke included. That guard is deliberately one place; testing emptiness per flag is what previously left `non_code_only` false while both arch flags also came out false, running UT and packaging but skipping every scene test.
 
   The arch flags subtract `NON_CODE` before deciding, so a non-code-only change already makes both `false`. An arch-gated job therefore needs no separate non-code check. See [`.claude/rules/ci-change-detection.md`](../.claude/rules/ci-change-detection.md) for the invariants these gates must keep.
 
+- **SDMA tests run as their own step inside `st-onboard-a2a3`.** The sweep deselects them with `-m "not sdma"` and a later step runs `-m sdma`. Ordering is what the two paths share: the SDMA step is always second, so no fault-injection case can land on a device that has already provisioned SDMA. Device acquisition differs by host arch — on aarch64 the SDMA step takes its own `task-submit --device auto --device-num 2`, so the two steps are disjoint in devices as well; on x86_64 there is no `task-submit` and both steps use the same `${DEVICE_RANGE}`, leaving ordering as the only separation. Provisioning the SDMA workspace creates device-only STARS streams that live in the device fault domain, so an AICore fault on a device that has provisioned SDMA costs minutes instead of milliseconds — the sweep's `aicore_op_timeout` fault injection must therefore never share a device with them ([#1425](https://github.com/hw-native-sys/simpler/issues/1425)). Selection is by marker on both sides, so the two cannot drift apart; the split can be dropped once #1425 is fixed. Nothing outside `st-onboard-a2a3` filters on the marker, so a local `pytest examples tests/st` still runs everything.
+
 ## Hardware Classification
 
 Three hardware tiers, applied to all test categories. See [testing.md](testing.md#hardware-classification) for the full table including per-category mechanisms (pytest markers, ctest labels, folder structure).