diff --git a/src/linker/mod.rs b/src/linker/mod.rs
index ad5ea76..5c9bbe2 100644
--- a/src/linker/mod.rs
+++ b/src/linker/mod.rs
@@ -396,13 +396,19 @@ impl Linker {
 
         let mut all_instructions = Vec::new();
 
-        // `__oam_used` marker from IR codegen — set whenever user
-        // code contains at least one `draw` statement. Gates the
-        // `$FE` OAM shadow fill inside `gen_init`, the OAM DMA
-        // inside `gen_nmi`, and (cascaded below) the default palette
-        // / smiley / rendering-enable machinery. Programs that don't
+        // Does this program need the OAM DMA plumbing? True when
+        // either of two markers the IR codegen drops is present:
+        //   - `__oam_used`: user code contains at least one `draw`.
+        //   - `__sprite_cycle_used`: user code calls `cycle_sprites`
+        //     (which rotates the DMA start offset each frame; it
+        //     presupposes the DMA is running).
+        // Gates the `$FE` OAM shadow fill inside `gen_init`, the
+        // OAM DMA inside `gen_nmi`, and — cascaded via the
+        // `has_visual_output` check below — the default palette /
+        // smiley / rendering-enable machinery. Programs that don't
         // draw save ~520 cycles per NMI plus a handful of bytes.
-        let has_oam = has_label(user_code, "__oam_used");
+        let has_oam =
+            has_label(user_code, "__oam_used") || has_label(user_code, "__sprite_cycle_used");
 
         // RESET entry point
         all_instructions.push(Instruction::new(NOP, AM::Label("__reset".into())));
diff --git a/src/linker/tests.rs b/src/linker/tests.rs
index 22614b6..66ce23d 100644
--- a/src/linker/tests.rs
+++ b/src/linker/tests.rs
@@ -97,6 +97,28 @@ fn link_rom_size_correct() {
     assert_eq!(rom_data.len(), 16 + 16384 + 8192);
 }
 
+#[test]
+fn sprite_cycle_marker_implies_oam_dma() {
+    // `cycle_sprites` rotates the OAM DMA start offset each frame;
+    // without the DMA itself the cycling is a no-op. The linker's
+    // `has_oam` gate therefore ORs `__sprite_cycle_used` with
+    // `__oam_used` so a program that only cycles still gets the
+    // DMA plumbing. The NMI of such a program writes the DMA
+    // trigger at `$4014`.
+    let linker = Linker::new(Mirroring::Horizontal);
+    let user_code = vec![
+        Instruction::new(NOP, AM::Label("__sprite_cycle_used".into())),
+        Instruction::implied(NOP),
+    ];
+    let rom = linker.link(&user_code);
+    // The OAM DMA write is `STA $4014` — `8D 14 40` in bytes.
+    let prg = &rom[16..16 + 16_384];
+    assert!(
+        prg.windows(3).any(|w| w == [0x8D, 0x14, 0x40]),
+        "cycle_sprites should force the OAM DMA even without a draw"
+    );
+}
+
 #[test]
 fn link_with_sprites_places_chr_data() {
     let linker = Linker::new(Mirroring::Horizontal);
diff --git a/src/runtime/mod.rs b/src/runtime/mod.rs
index 552ea0e..ebe2d57 100644
--- a/src/runtime/mod.rs
+++ b/src/runtime/mod.rs
@@ -388,29 +388,19 @@ pub struct NmiOptions {
     pub has_oam: bool,
     /// When false, drop the three instructions that shift `$4017`
     /// (JOY2) into `ZP_INPUT_P2` from the NMI's 8-iteration input
-    /// loop. Single-player programs save ~6 bytes of code and ~30
-    /// cycles per frame (a `LDA abs`, an `LSR A`, and a `ROL zp`
-    /// running 8 times).
+    /// loop. Single-player programs save ~6 bytes of code and ~88
+    /// cycles per frame (LDA abs + LSR A + ROL zp = 11 cycles × 8
+    /// iterations).
     pub has_p2_input: bool,
     /// When false, drop the three instructions that shift `$4016`
     /// (JOY1) into `ZP_INPUT_P1`. If both `has_p1_input` and
     /// `has_p2_input` are false the whole strobe-and-loop block
     /// disappears — programs that never touch `button.*` pay
-    /// zero cycles for input sampling.
+    /// zero cycles for input sampling (~100 cycles: the 88-cycle
+    /// body plus the ~12-cycle strobe and loop scaffold).
     pub has_p1_input: bool,
 }
 
-impl NmiOptions {
-    /// Whether the program reads any controller input — the
-    /// necessary condition for emitting the strobe write to
-    /// `$4016` and the 8-iteration shift loop. Skipped entirely
-    /// when both ports are unused.
-    #[must_use]
-    pub fn any_input(&self) -> bool {
-        self.has_p1_input || self.has_p2_input
-    }
-}
-
 #[must_use]
 pub fn gen_nmi(opts: NmiOptions) -> Vec<Instruction> {
     let NmiOptions {