diff --git a/.github/workflows/pr-gate.yml b/.github/workflows/pr-gate.yml
index a1fbdc6..ec887e9 100644
--- a/.github/workflows/pr-gate.yml
+++ b/.github/workflows/pr-gate.yml
@@ -185,6 +185,18 @@ jobs:
             toolchain/output/linx_blockisa_llvm_musl
           key: llvm-${{ runner.os }}-tileop-required-gate-v2-${{ steps.llvm-source.outputs.sha }}
 
+      - name: Install candidate TileOP headers into the cached toolchain
+        run: |
+          set -euo pipefail
+          RESOURCE_DIR="$(${TOOLCHAIN_INSTALL_DIR}/bin/clang -print-resource-dir)"
+          mkdir -p "${RESOURCE_DIR}/include/tileop-api"
+          rsync -a --delete \
+            toolchain/src/Linx-TileOP-API/include/ \
+            "${RESOURCE_DIR}/include/tileop-api/"
+          cmp \
+            toolchain/src/Linx-TileOP-API/include/common/pto_tile.hpp \
+            "${RESOURCE_DIR}/include/tileop-api/common/pto_tile.hpp"
+
       - name: Run TileOP contract checks
         working-directory: toolchain/src/Linx-TileOP-API
         env:
diff --git a/docs/tileop-usage/tlsu/load-store-move/TLOAD.md b/docs/tileop-usage/tlsu/load-store-move/TLOAD.md
index 0c8fb88..d3294f1 100644
--- a/docs/tileop-usage/tlsu/load-store-move/TLOAD.md
+++ b/docs/tileop-usage/tlsu/load-store-move/TLOAD.md
@@ -26,6 +26,11 @@ PTO_SHARED_INLINE void TLOAD(SharedTile<shp> &dst, const gm_shape &src);
 
 除通用 Tile 约束外，必须满足 PTO-SPEC 对本操作规定的操作数角色、数据类型组合、形状、布局、有效区域、容量、存储位置、PE mask 以及 alias 规则。对于需要 Shared Tile、标量、索引、scale、bias 或选项对象的重载，只能使用接口声明的参数形式；不能通过省略参数来伪造另一种操作数组合。
 
+`B.IOR.RegSrc1` carries the GM row stride in **bytes**, not elements. The
+ordinary, Shared, and CUBE transport overloads therefore derive the encoded
+stride with `GetStrideBytes`; callers must not pre-scale or reinterpret that
+value in element units.
+
 ## 默认值
 
 未显式传入的可选参数使用该 C++ 重载和 PTO-SPEC contract 规定的默认值。默认选项、维度、布局、padding、scale mask 和属性字段可能与显式编码的零值不同；调用者不得把“省略”与“传入零值”自动等同。
diff --git a/include/jcore/template_asm.hpp b/include/jcore/template_asm.hpp
index f2d7c36..7ce87de 100644
--- a/include/jcore/template_asm.hpp
+++ b/include/jcore/template_asm.hpp
@@ -6945,24 +6945,29 @@ void TPRELU(tile_shape &dst, tile_shape &src0, tile_shape &src1) {
 }
 
 
-// TSEL: select between two tiles using mask
+// TSEL: select true_src where mask is set, otherwise preserve the prior dst.
+// PTO requires the false source to be bound explicitly.  Model the public
+// in-place API as two B.IOT bindings and keep dst read/write so the second
+// binding snapshots its old value before publishing the new destination.
 template <is_tile_data_v tile_shape>
-void TSEL(tile_shape &dst, tile_shape &src0, tile_shape &src1) {
+void TSEL(tile_shape &dst, tile_shape &mask, tile_shape &true_src) {
   asm volatile(
-    "BSTART.TEPL 26, %D1\n"
-    "B.DIM %2, 0, ->lb0\n"
-    "B.DIM %3, 0, ->lb1\n"
-    "B.DIM zero, %c4, ->lb2\n"
-    "B.IOT %5, %6, mask=1111, last, ->%0<%Z7>\n"
+    "BSTART.TEPL 26, %D[DataType]\n"
+    "B.DIM %[ValidCol], 0, ->lb0\n"
+    "B.DIM %[ValidRow], 0, ->lb1\n"
+    "B.DIM zero, %c[Cols], ->lb2\n"
+    "B.IOT %[Mask], %[True], mask=1111\n"
+    "B.IOT %[Prior], mask=1111, last, ->%[Dst]<%Z[TileSize]>\n"
     ""
-    : "=Tr"(dst.data())
-    : "i"(type_traits<typename tile_shape::DType>::TypeCode),
-      "r"(src0.GetValidCol()),
-      "r"(src0.GetValidRow()),
-      "i"(tile_shape::Cols),
-      "Tr"(src0.data()),
-      "Tr"(src1.data()),
-      "i"(tile_type_traits<typename tile_shape::TileDType>::TilesizeCode)
+    : [Dst] "=Tr"(dst.data())
+    : [Prior] "0"(dst.data()),
+      [DataType] "i"(type_traits<typename tile_shape::DType>::TypeCode),
+      [ValidCol] "r"(mask.GetValidCol()),
+      [ValidRow] "r"(mask.GetValidRow()),
+      [Cols] "i"(tile_shape::Cols),
+      [Mask] "Tr"(mask.data()),
+      [True] "Tr"(true_src.data()),
+      [TileSize] "i"(tile_type_traits<typename tile_shape::TileDType>::TilesizeCode)
   );
 }
 
diff --git a/test/test_v058_engine_contract.py b/test/test_v058_engine_contract.py
index 1ba9fc2..c43c8d1 100644
--- a/test/test_v058_engine_contract.py
+++ b/test/test_v058_engine_contract.py
@@ -200,6 +200,21 @@ def test_tlsu_load_store_stride_is_expressed_in_bytes(self) -> None:
         tlsu_doc = (ROOT / "docs" / "tileop-usage" / "tlsu" / "load-store-move" / "TLOAD.md").read_text(encoding="utf-8")
         self.assertIn("row stride in **bytes**", tlsu_doc)
 
+    def test_tsel_binds_the_prior_destination_as_explicit_false_source(self) -> None:
+        start = self.header.index("void TSEL(")
+        end = self.header.index("// TABS:", start)
+        tsel = self.header[start:end]
+        self.assertIn('"B.IOT %[Mask], %[True], mask=1111\\n"', tsel)
+        self.assertIn(
+            '"B.IOT %[Prior], mask=1111, last, ->%[Dst]<%Z[TileSize]>\\n"',
+            tsel,
+        )
+        self.assertIn(': [Dst] "=Tr"(dst.data())', tsel)
+        self.assertIn('[Prior] "0"(dst.data())', tsel)
+        self.assertNotIn(
+            '"B.IOT %5, %6, mask=1111, last, ->%0<%Z7>\\n"', tsel
+        )
+
     def test_fpatr_carries_shared_transpose_controls(self) -> None:
         tile = PTO_TILE.read_text(encoding="utf-8")
         self.assertIn("bool TransA = false", tile)
diff --git a/test/tileop_api/src/TSELCanonical.cpp b/test/tileop_api/src/TSELCanonical.cpp
new file mode 100644
index 0000000..6102d1c
--- /dev/null
+++ b/test/tileop_api/src/TSELCanonical.cpp
@@ -0,0 +1,11 @@
+#include <common/pto_tileop.hpp>
+
+using namespace pto;
+
+using SelectTile = Tile<Location::Vec, uint16_t, 16, 16,
+                        BLayout::RowMajor>;
+
+void tsel_canonical(SelectTile &dst, SelectTile &mask,
+                    SelectTile &true_src) {
+  TSEL(dst, mask, true_src);
+}
diff --git a/test/tileop_api/verify_target_cxx_frontend.sh b/test/tileop_api/verify_target_cxx_frontend.sh
index 35f7616..a5321c2 100755
--- a/test/tileop_api/verify_target_cxx_frontend.sh
+++ b/test/tileop_api/verify_target_cxx_frontend.sh
@@ -8,9 +8,26 @@ ROOT=$(cd "$(dirname "$0")/../.." && pwd)
 OUT=$(mktemp -d "${TMPDIR:-/tmp}/tileop-target-cxx.XXXXXX")
 trap 'rm -rf "$OUT"' EXIT
 
+# -mlxbc injects the TileOP headers installed in Clang's resource directory
+# ahead of ordinary -I paths.  Build a temporary resource overlay that keeps
+# the toolchain's builtin headers but replaces only tileop-api with this
+# checkout, otherwise this gate silently validates the packaged API instead of
+# the proposed source tree.
+CLANG_RESOURCE_DIR=$("$TC_DIR/clang++" -print-resource-dir)
+CLANG_RESOURCE_OVERLAY="$OUT/clang-resource"
+mkdir -p "$CLANG_RESOURCE_OVERLAY/include"
+for resource_path in "$CLANG_RESOURCE_DIR"/include/*; do
+  resource_name=${resource_path##*/}
+  if [[ "$resource_name" != tileop-api ]]; then
+    ln -s "$resource_path" "$CLANG_RESOURCE_OVERLAY/include/$resource_name"
+  fi
+done
+ln -s "$ROOT/include" "$CLANG_RESOURCE_OVERLAY/include/tileop-api"
+
 FLAGS=(
   --target="$LINX_TARGET"
   -mlxbc
+  -resource-dir="$CLANG_RESOURCE_OVERLAY"
   --sysroot="$LINX_SYSROOT"
   -nostdinc++
   -isystem "$LINX_SYSROOT/include/c++/v1"
@@ -22,7 +39,7 @@ FLAGS=(
   -I"$ROOT/include"
 )
 
-for source in MXScaleVariants.cpp SharedTransposeNonSquare.cpp; do
+for source in MXScaleVariants.cpp SharedTransposeNonSquare.cpp TSELCanonical.cpp; do
   "$TC_DIR/clang++" "${FLAGS[@]}" -fsyntax-only \
     "$ROOT/test/tileop_api/src/$source"
 done
@@ -33,7 +50,10 @@ done
   "$ROOT/test/tileop_api/src/MXScaleVariants.cpp" -o "$OUT/MXScaleVariants.ll"
 "$TC_DIR/clang++" "${FLAGS[@]}" -c \
   "$ROOT/test/tileop_api/src/MXScaleVariants.cpp" -o "$OUT/MXScaleVariants.o"
+"$TC_DIR/clang++" "${FLAGS[@]}" -c \
+  "$ROOT/test/tileop_api/src/TSELCanonical.cpp" -o "$OUT/TSELCanonical.o"
 "$TC_DIR/llvm-objdump" -d "$OUT/MXScaleVariants.o" > "$OUT/MXScaleVariants.dis"
+"$TC_DIR/llvm-objdump" -d "$OUT/TSELCanonical.o" > "$OUT/TSELCanonical.dis"
 grep -q '<1024 x i32> asm sideeffect' "$OUT/SharedTLoad.ll"
 grep -q 'i64 asm sideeffect.*=@2Sr' "$OUT/SharedTLoad.ll"
 grep -q '<1024 x i32> asm sideeffect.*@2Sr' "$OUT/SharedTLoad.ll"
@@ -93,4 +113,15 @@ grep -Eiq 'BSTART\.CUBE[[:space:]]+TMATMULMX\.ACC,[[:space:]]+E4M3' "$OUT/MXScal
 grep -Eiq 'BSTART\.CUBE[[:space:]]+TGEMVMX\.BIAS,[[:space:]]+BF16' "$OUT/MXScaleVariants.dis"
 grep -Eq 'B\.IOT[[:space:]]+t#[1-8], t#[1-8], mask=1111' "$OUT/MXScaleVariants.dis"
 
+grep -Eq 'B\.IOT[[:space:]]+[tumn]#[1-8], [tumn]#[1-8], mask=1111[[:space:]]*$' \
+  "$OUT/TSELCanonical.dis"
+grep -Eq 'B\.IOT[[:space:]]+[tumn]#[1-8], mask=1111, last,.*->[tumn]<[0-9]+(B|KB)>' \
+  "$OUT/TSELCanonical.dis"
+if grep -Eq 'B\.IOT[[:space:]]+[tumn]#[1-8], [tumn]#[1-8], mask=1111, last,.*->' \
+    "$OUT/TSELCanonical.dis"; then
+  echo "FAIL: TSEL collapsed to the forbidden single-B.IOT in-place form" >&2
+  sed -n '1,120p' "$OUT/TSELCanonical.dis" >&2
+  exit 1
+fi
+
 echo "Linx target C++ frontend MX/effective-shape contract: PASS"
diff --git a/tools/generate_engine_docs.py b/tools/generate_engine_docs.py
index 79f8cb4..dc0c030 100644
--- a/tools/generate_engine_docs.py
+++ b/tools/generate_engine_docs.py
@@ -84,6 +84,25 @@ def render() -> str:
             "The `BSTART.VEC` / `BSTART.SFU` spellings are canonical aliases of",
             "the unique `BSTART.TEPL` compiled carrier (ADR 0057).",
             "",
+            "## 使用要求",
+            "",
+            "优先根据操作页面中的 C++ API 调用 TileOP；本页仅用于按执行引擎查找操作。不要仅凭分类名称推断执行引擎，具体 engine 和支持的 Tile location 以操作页面为准。",
+            "",
+            "## 约束、默认值、异常和边界行为",
+            "",
+            "每个 engine 都有自己的 Tile、PE participation、容量和 bundle 约束。省略的 bundle 字段采用具体操作 contract 的默认值；非法 engine、binder、Tile location、mask 或属性组合会在执行前被拒绝。操作的边界、padding、异常和结果语义不由本索引页定义。",
+            "",
+            "## 使用示例",
+            "",
+            "```cpp",
+            "// 例如从 VEC 操作页面调用 C++ wrapper。",
+            "TADD(dst, lhs, rhs);",
+            "```",
+            "",
+            "## 完整语义",
+            "",
+            "各操作的完整语义请参阅其操作页面末尾的 PTO-SPEC 链接，规范总目录见 [PTO-SPEC tile 文档](https://github.com/PTO-ISA/pto-spec/tree/v0.58.4.1/docs/tile)。",
+            "",
         ]
     )