feat: Hermite resampling, sprite shift registers, controller open bus

#3 audio.rs: replace linear interpolation with Catmull-Rom Hermite cubic.
  Stores prev_sample as p0 control point; m1=(p2-p0)/2, m2=(p2-p1)/2
  tangents give continuous first derivative across batch boundaries.

#4 ppu: add per-slot sprite shift registers (spr_shift_lo/hi, spr_x_counter,
  spr_attr_latch). load_sprite_shifters fetches pattern bytes with h-flip at
  dot 1 of each visible scanline. sprite_pixel_from_shifters replaces the
  per-pixel OAM scan; sprite-0 hit detection integrated into the shifter path.

#5 joypad.rs: format_controller_read now preserves bits 1-5,7 as open bus
  (!0x41 mask) instead of zeroing bits 1-4, matching NES hardware behaviour.

src/native_core/bus/joypad.rs

@@ -67,8 +67,11 @@ impl NativeBus {
     }
 
     fn format_controller_read(&self, bit: u8) -> u8 {
-        // Controller reads expose serial data in bit0, keep bit6 high, and
-        // preserve open-bus upper bits.
-        (self.cpu_open_bus & 0xE0) | 0x40 | (bit & 1)
+        // The NES controller port drives only bit 0 (serial data); bit 6 is
+        // held high by a pull-up on the expansion connector.  All other bits
+        // (1-5, 7) float and retain whatever is currently on the CPU data bus
+        // (open bus).  !0x41 clears bits 6 and 0 so we can OR in their
+        // canonical values without corrupting any open-bus bits.
+        (self.cpu_open_bus & !0x41u8) | 0x40 | (bit & 1)
     }
 }

src/native_core/ppu/api.rs

@@ -44,6 +44,10 @@ impl Ppu {
             sprite_count: 0,
             next_sprite_indices: [0; 8],
             next_sprite_count: 0,
+            spr_shift_lo: [0; 8],
+            spr_shift_hi: [0; 8],
+            spr_x_counter: [0; 8],
+            spr_attr_latch: [0; 8],
         }
     }
 
@@ -90,17 +94,33 @@ impl Ppu {
                 (self.read_palette(0), false)
             };
 
-            if !self.sprite0_hit_set() && self.sprite0_hit_at(mapper, y, dot) && bg_opaque {
-                self.set_sprite0_hit(true);
-            }
+            // Advance sprite shift registers for every visible pixel
+            // (unconditional so x-counters stay in sync even when the sprite
+            // layer is clipped on the left).
+            let spr_result = if self.sprites_enabled() {
+                self.sprite_pixel_from_shifters()
+            } else {
+                None
+            };
 
             let mut final_color = bg_color_index & 0x3F;
             let sprite_layer_enabled = show_spr && (x >= 8 || show_spr_left);
-            if sprite_layer_enabled
-                && let Some((spr_color_index, behind_bg)) = self.sprite_pixel(mapper, x, y)
-                && !(behind_bg && bg_opaque)
-            {
-                final_color = spr_color_index & 0x3F;
+            if sprite_layer_enabled {
+                if let Some((spr_color_index, behind_bg, is_sprite0)) = spr_result {
+                    // Sprite-0 hit: set when a non-transparent sprite-0 pixel
+                    // overlaps a non-transparent background pixel.  Suppressed
+                    // in the left 8 pixels when either clip bit is clear.
+                    if is_sprite0
+                        && bg_opaque
+                        && !self.sprite0_hit_set()
+                        && (x >= 8 || (show_bg_left && show_spr_left))
+                    {
+                        self.set_sprite0_hit(true);
+                    }
+                    if !(behind_bg && bg_opaque) {
+                        final_color = spr_color_index & 0x3F;
+                    }
+                }
             }
 
             let (r, g, b) = apply_color_emphasis(nes_rgb(final_color), self.mask);
@@ -160,11 +180,13 @@ impl Ppu {
         }
 
         if rendering_active {
-            // Transfer pre-evaluated sprite list at the start of each visible scanline,
-            // so dots 1-256 render with the correct sprites for *this* scanline.
+            // Transfer pre-evaluated sprite list at the start of each visible
+            // scanline, then immediately load the per-slot shift registers so
+            // that dots 1-256 render with the correct sprites for *this* line.
             if scanline < 240 && dot == 1 && self.sprites_enabled() {
                 self.sprite_count = self.next_sprite_count;
                 self.sprite_indices = self.next_sprite_indices;
+                self.load_sprite_shifters(mapper, scanline);
             }
 
             if dot == 256 {
@@ -284,6 +306,93 @@ impl Ppu {
         (count, indices, overflow)
     }
 
+    /// Fetch pattern bytes for the sprites evaluated for `scanline` and load
+    /// them into the per-slot shift registers.  Called once at dot 1 of each
+    /// visible scanline so that `sprite_pixel_from_shifters` can provide
+    /// cycle-accurate, shift-register-based sprite rendering for dots 1-256.
+    pub(super) fn load_sprite_shifters(&mut self, mapper: &dyn Mapper, scanline: u32) {
+        let sprite_height = if (self.ctrl & 0x20) != 0 { 16i16 } else { 8i16 };
+        for slot in 0..8usize {
+            // Default: inactive slot — shifters transparent, counter parked.
+            self.spr_shift_lo[slot] = 0;
+            self.spr_shift_hi[slot] = 0;
+            self.spr_x_counter[slot] = 0xFF;
+            self.spr_attr_latch[slot] = 0;
+            if slot >= self.sprite_count as usize {
+                continue;
+            }
+            let i = self.sprite_indices[slot] as usize;
+            let oam_idx = i * 4;
+            let attr = self.oam[oam_idx + 2];
+            let tile = self.oam[oam_idx + 1];
+            let sprite_y = self.oam[oam_idx] as i16 + 1;
+            let mut row = scanline as i16 - sprite_y;
+            if row < 0 || row >= sprite_height {
+                continue;
+            }
+            if (attr & 0x80) != 0 {
+                row = sprite_height - 1 - row; // vertical flip
+            }
+            let (lo_addr, hi_addr) = if sprite_height == 16 {
+                let table = ((tile & 1) as u16) << 12;
+                let tile_num = (tile & 0xFE).wrapping_add((row / 8) as u8) as u16;
+                let row_in_tile = (row & 7) as u16;
+                let lo = table + tile_num * 16 + row_in_tile;
+                (lo, lo + 8)
+            } else {
+                let table = if (self.ctrl & 0x08) != 0 { 0x1000u16 } else { 0u16 };
+                let lo = table + (tile as u16) * 16 + row as u16;
+                (lo, lo + 8)
+            };
+            let mut lo = mapper.ppu_read(lo_addr);
+            let mut hi = mapper.ppu_read(hi_addr);
+            if (attr & 0x40) != 0 {
+                // Horizontal flip: reverse bit order so MSB is always the
+                // leftmost pixel when we shift out from bit 7.
+                lo = lo.reverse_bits();
+                hi = hi.reverse_bits();
+            }
+            self.spr_shift_lo[slot] = lo;
+            self.spr_shift_hi[slot] = hi;
+            self.spr_x_counter[slot] = self.oam[oam_idx + 3];
+            self.spr_attr_latch[slot] = attr;
+        }
+    }
+
+    /// Advance all active sprite shift registers by one pixel and return the
+    /// colour and priority of the first non-transparent sprite pixel found.
+    /// The third element of the tuple is `true` when the winning sprite is
+    /// OAM sprite 0 (used for sprite-0 hit detection).
+    ///
+    /// Every active slot is always updated regardless of which slot wins, so
+    /// this must be called exactly once per visible pixel dot (1-256).
+    pub(super) fn sprite_pixel_from_shifters(&mut self) -> Option<(u8, bool, bool)> {
+        let mut result: Option<(u8, bool, bool)> = None;
+        for slot in 0..self.sprite_count as usize {
+            if self.spr_x_counter[slot] > 0 {
+                self.spr_x_counter[slot] -= 1;
+                continue;
+            }
+            // Extract the MSB from each pattern plane then advance the shifter.
+            let lo_bit = (self.spr_shift_lo[slot] >> 7) & 1;
+            let hi_bit = (self.spr_shift_hi[slot] >> 7) & 1;
+            self.spr_shift_lo[slot] <<= 1;
+            self.spr_shift_hi[slot] <<= 1;
+            if result.is_none() {
+                let pix = lo_bit | (hi_bit << 1);
+                if pix != 0 {
+                    let attr = self.spr_attr_latch[slot];
+                    let pal_idx = (((attr & 0x03) as u16) << 2) | pix as u16;
+                    let color = self.read_palette(0x10 | pal_idx);
+                    let behind_bg = (attr & 0x20) != 0;
+                    let is_sprite0 = self.sprite_indices[slot] == 0;
+                    result = Some((color, behind_bg, is_sprite0));
+                }
+            }
+        }
+        result
+    }
+
     pub fn note_scroll_register_write_legacy(&mut self, scanline: usize, dot: u32) {
         let mut target_scanline = scanline;
         let mut x_start = 0u8;

src/native_core/ppu/types.rs

@@ -42,6 +42,15 @@ pub struct Ppu {
     pub(super) sprite_count: u8,
     pub(super) next_sprite_indices: [u8; 8],
     pub(super) next_sprite_count: u8,
+    // Per-slot sprite shift registers loaded at dot 1 of each visible scanline.
+    // spr_shift_lo/hi hold the 8-bit pattern row (h-flip already applied).
+    // spr_x_counter counts down the remaining pixels before a slot becomes
+    // active; when it reaches 0 the slot starts shifting out pixel bits.
+    // spr_attr_latch stores the OAM attribute byte (priority, palette, flips).
+    pub(super) spr_shift_lo: [u8; 8],
+    pub(super) spr_shift_hi: [u8; 8],
+    pub(super) spr_x_counter: [u8; 8],
+    pub(super) spr_attr_latch: [u8; 8],
 }
 
 impl Default for Ppu {

src/runtime/audio.rs

@@ -7,6 +7,13 @@ pub struct AudioMixer {
     samples_per_cpu_cycle: f64,
     sample_accumulator: f64,
     last_output_sample: f32,
+    // Previous output sample (two batches ago) used as the p0 control point
+    // for Catmull-Rom Hermite interpolation.  Storing p0 allows the tangent at
+    // the start of each interpolation interval to be computed as
+    //   m1 = (p2 - p0) / 2
+    // which produces a smooth, continuous first derivative across batch
+    // boundaries rather than the kink introduced by linear ramps.
+    prev_sample: f32,
     // One-pole IIR low-pass filter state (approximates NES ~14 kHz RC filter).
     // Coefficient: a = exp(-2π * fc / fs). At fc=14000, fs=48000: a ≈ 0.160
     lp_coeff: f32,
@@ -31,6 +38,7 @@ impl AudioMixer {
             samples_per_cpu_cycle: sample_rate as f64 / cpu_hz,
             sample_accumulator: 0.0,
             last_output_sample: 0.0,
+            prev_sample: 0.0,
             lp_coeff,
             lp_state: 0.0,
             hp_coeff,
@@ -46,6 +54,7 @@ impl AudioMixer {
     pub fn reset(&mut self) {
         self.sample_accumulator = 0.0;
         self.last_output_sample = 0.0;
+        self.prev_sample = 0.0;
         self.lp_state = 0.0;
         self.hp_prev_x = 0.0;
         self.hp_prev_y = 0.0;
@@ -82,29 +91,49 @@ impl AudioMixer {
             return;
         }
 
-        let start = self.last_output_sample;
+        // Catmull-Rom Hermite interpolation between the previous batch sample
+        // (p1 = last_output_sample) and the current batch sample (p2 = sample).
+        //
+        // The tangent at p1 uses the two-point central difference:
+        //   m1 = (p2 - p0) / 2,  where p0 = prev_sample (two batches ago).
+        // The tangent at p2 uses the forward difference (p3 approximated as p2,
+        // i.e. the signal stays flat beyond the current batch):
+        //   m2 = (p2 - p1) / 2.
+        //
+        // Hermite basis:
+        //   h00(t) = 2t³ - 3t² + 1
+        //   h10(t) = t³  - 2t² + t
+        //   h01(t) = -2t³ + 3t²
+        //   h11(t) = t³  -  t²
+        //   f(t)   = h00·p1 + h10·m1 + h01·p2 + h11·m2
+        //
+        // For t = 1 this collapses to p2, so the last output of each batch
+        // always lands exactly on the current APU sample value.
+        let p0 = self.prev_sample;
+        let p1 = self.last_output_sample;
+        let p2 = sample;
+        let m1 = (p2 - p0) * 0.5;
+        let m2 = (p2 - p1) * 0.5;
+
+        let denom = samples as f32;
         let a = self.lp_coeff;
         let b = 1.0 - a;
-        if samples == 1 {
-            let lp = a * self.lp_state + b * sample;
+        for idx in 0..samples {
+            let t = (idx + 1) as f32 / denom;
+            let t2 = t * t;
+            let t3 = t2 * t;
+            let interp = (2.0 * t3 - 3.0 * t2 + 1.0) * p1
+                + (t3 - 2.0 * t2 + t) * m1
+                + (-2.0 * t3 + 3.0 * t2) * p2
+                + (t3 - t2) * m2;
+            let lp = a * self.lp_state + b * interp;
             self.lp_state = lp;
             let hp = self.hp_coeff * self.hp_prev_y + lp - self.hp_prev_x;
             self.hp_prev_x = lp;
             self.hp_prev_y = hp;
             out.push(hp);
-        } else {
-            let denom = samples as f32;
-            for idx in 0..samples {
-                let t = (idx + 1) as f32 / denom;
-                let interp = start + (sample - start) * t;
-                let lp = a * self.lp_state + b * interp;
-                self.lp_state = lp;
-                let hp = self.hp_coeff * self.hp_prev_y + lp - self.hp_prev_x;
-                self.hp_prev_x = lp;
-                self.hp_prev_y = hp;
-                out.push(hp);
-            }
         }
+        self.prev_sample = p1;
         self.last_output_sample = sample;
     }
 }