rsscript: declare Tensor sin/trunc/floordiv/floormod + metal_* in tensor.rssi

The ml/tensor-alu and ml/metal-ffi slices registered these intrinsics + added
the Rust kernels, but the embedded .rssi interface declarations were left
uncommitted, so rss code calling them failed typecheck (RS0206 does not
resolve). Declares all six native methods. (.rssi is include_str!-embedded, so
callers also need a rebuilt rss binary.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: olwang

packages/ml/interface/tensor.rssi

@@ -37,6 +37,25 @@ pub native fn Tensor.matmul(
 ) -> Result<fresh Tensor, TensorError>
     effects(native)
 
+// Whether a Metal GPU is available for the native compute path (false off macOS or
+// when no device is present).
+pub native fn Tensor.metal_available() -> Bool
+    effects(native)
+
+// The Metal device name (e.g. "Apple M2 Max"), or an empty string when no GPU is
+// available.
+pub native fn Tensor.metal_device_name() -> fresh String
+    effects(native)
+
+// GPU matrix multiply: same contract as `matmul`, dispatched to Metal. NOT bit-
+// identical to the CPU kernel (GPU reductions reorder), so it matches to f32
+// tolerance. Errors on non-rank-2 operands, an inner-dim mismatch, or no GPU.
+pub native fn Tensor.matmul_metal(
+    a: read Tensor,
+    b: read Tensor,
+) -> Result<fresh Tensor, TensorError>
+    effects(native)
+
 // Elementwise binary ops over two same-shape tensors. No broadcasting — the
 // shapes must match exactly (broadcasting is deferred to a later slice). Errors
 // on a shape mismatch.
@@ -257,6 +276,15 @@ pub native fn Tensor.log2(t: read Tensor) -> fresh Tensor
 pub native fn Tensor.rsqrt(t: read Tensor) -> fresh Tensor
     effects(native)
 
+// Elementwise sine (radians). Output dtype F32.
+pub native fn Tensor.sin(t: read Tensor) -> fresh Tensor
+    effects(native)
+
+// Elementwise truncation toward zero (drops the fractional part, keeps the sign);
+// a float-result op that preserves dtype, unlike cast_i32 which retags to I32.
+pub native fn Tensor.trunc(t: read Tensor) -> fresh Tensor
+    effects(native)
+
 // Elementwise power a^b with broadcasting. Output dtype F32. Error if not
 // broadcast-compatible.
 pub native fn Tensor.pow(a: read Tensor, b: read Tensor) -> Result<fresh Tensor, TensorError>
@@ -283,6 +311,17 @@ pub native fn Tensor.idiv(a: read Tensor, b: read Tensor) -> Result<fresh Tensor
 pub native fn Tensor.modulo(a: read Tensor, b: read Tensor) -> Result<fresh Tensor, TensorError>
     effects(native)
 
+// Floor division (Python //): rounds toward negative infinity, so the quotient
+// sign follows the divisor (distinct from idiv's truncate-toward-zero). Div-by-zero
+// -> 0. Broadcasts; output dtype I32.
+pub native fn Tensor.floordiv(a: read Tensor, b: read Tensor) -> Result<fresh Tensor, TensorError>
+    effects(native)
+
+// Floor modulo (Python %): the remainder sign follows the DIVISOR (distinct from
+// modulo, whose sign follows the dividend). Mod-by-zero -> 0. Broadcasts; dtype I32.
+pub native fn Tensor.floormod(a: read Tensor, b: read Tensor) -> Result<fresh Tensor, TensorError>
+    effects(native)
+
 // Left / right shift on i64 values, shift count masked to 0..63.
 pub native fn Tensor.shl(a: read Tensor, b: read Tensor) -> Result<fresh Tensor, TensorError>
     effects(native)