romdevtools 0.27.0 → 0.28.0

package/examples/gb/templates/puzzle.c

@@ -1,21 +1,34 @@
 /* ── puzzle.c — Game Boy match-3 falling-block scaffold ─────────────
  *
- * Mirrors the NES/Genesis/SNES puzzle scaffolds. 6-wide × 12-tall
- * grid drawn via BG tilemap (each cell = 1 BG tile). 1×3 vertical
- * active piece; LEFT/RIGHT shifts, A rotates colour order, DOWN
- * soft-drops, START hard-drops. Horizontal triples clear and score.
+ * 8-wide × 14-tall well drawn via BG tilemap (each cell = 1 BG tile).
+ * 1×3 vertical active piece; LEFT/RIGHT shifts, A/B cycles the colour
+ * order, DOWN soft-drops, START hard-drops. Matches of 3+ in a row —
+ * horizontal, vertical, or either diagonal — clear, survivors fall
+ * (gravity), and cascades chain with rising score.
  *
- * On the Game Boy we don't have a built-in font, so we render the
- * grid as coloured tile cells using three BG tile shapes (R/G/B
- * stripes). Score is kept in WRAM; rendering a numeric HUD requires
- * a digit-tile blob — left as an extension.
+ * On DMG we differentiate the three block kinds by SHAPE (2bpp stripe
+ * patterns), not colour. The GBC template is the full-colour version.
+ *
+ * RENDERING CONTRACT (the "pieces flash / don't render" fix): this
+ * core silently DROPS VRAM writes during active display — and can
+ * even drop one early in vblank. So (mirroring the GBC reference
+ * puzzle):
+ *  - The FALLING piece is OAM sprites 0-2 (one OAM DMA per frame —
+ *    no BG writes at all to move it, no erase artifacts).
+ *  - The LOCKED well is BG tiles, written ONLY right after
+ *    wait_vblank(): a budgeted diff (grid vs shadow) plus a rolling
+ *    SCRUB that continuously repaints the well from grid[], so any
+ *    dropped write self-heals within ~half a second.
+ *  - enable_vblank_irq() at boot → wait_vblank HALTs to the real
+ *    vblank leading edge (also ~30x faster on the WASM core than
+ *    the LY-polling fallback).
  */
 
 #include "gb_hardware.h"
 #include "gb_runtime.h"
 
-#define COLS 6
-#define ROWS 12
+#define COLS 8
+#define ROWS 14
 
 #define T_BLANK 0
 #define T_R     1
@@ -23,6 +36,10 @@
 #define T_B     3
 #define T_WALL  4
 
+/* Map placement: centre the 8-col well → BG col offset +6, row offset +1. */
+#define WELL_MX 6
+#define WELL_MY 1
+
 /* tile_blank is the EMPTY-cell / backdrop tile. It is NOT all-zero: a
  * subtle dither (colour 0 + faint colour 1) so the empty playfield and the
  * area around the well read as a textured surface, never one flat colour
@@ -38,8 +55,7 @@ static const uint8_t tile_wall[16] = {
     0xFF,0xFF, 0xFF,0xFF, 0xFF,0xFF, 0xFF,0xFF,
 };
 /* Three distinct tile shapes (since GB BG is 2bpp, we differentiate
- * by *shape*, not colour-on-CGB). The CGB palette path could give us
- * real colours; for DMG-compatibility we use shape. */
+ * by *shape*, not colour-on-CGB). */
 static const uint8_t tile_r[16] = {
     0xFF,0x00, 0xFF,0x00, 0xFF,0x00, 0xFF,0x00,
     0xFF,0x00, 0xFF,0x00, 0xFF,0x00, 0xFF,0x00,
@@ -55,18 +71,30 @@ static const uint8_t tile_b[16] = {
 
 static const uint16_t bg_palette[4]  = { 0x7FFF, 0x5294, 0x294A, 0x0000 };
 
-static uint8_t grid[ROWS][COLS];
+#define NCELL (ROWS * COLS)
+static uint8_t grid[NCELL];     /* 0 = empty, 1..3 = block colour */
+static uint8_t shadow[NCELL];   /* what's currently on the BG (diff redraw) */
+static uint8_t matched[NCELL];  /* scratch: cells flagged to clear */
 static uint8_t piece[3];
 static int16_t piece_x, piece_y;
 static uint8_t fall_timer;
 static uint16_t score;
-static uint32_t rng = 1;
+static uint16_t rng = 0xACE1;
+
+#define G(r,c) grid[(uint8_t)((r) * COLS + (c))]
+#define M(r,c) matched[(uint8_t)((r) * COLS + (c))]
 
-static uint32_t xorshift(void) {
-    rng ^= rng << 13;
-    rng ^= rng >> 17;
-    rng ^= rng << 5;
-    return rng;
+/* the 4 line directions scanned for matches: horizontal, vertical, and
+ * both diagonals; each line is only walked from its lowest cell. */
+static const int8_t DIRS[4][2] = { {0,1}, {1,0}, {1,1}, {1,-1} };
+
+/* 16-bit xorshift — kept 16-bit on purpose (sm83 has no fast 32-bit
+ * shifts; a wider generator there degenerates toward one value). */
+static uint8_t xorshift(void) {
+    rng ^= rng << 7;
+    rng ^= rng >> 9;
+    rng ^= rng << 8;
+    return (uint8_t)(rng >> 8);
 }
 
 static uint8_t random_colour(void) { return 1 + (xorshift() % 3); }
@@ -88,20 +116,6 @@ static uint8_t tile_for(uint8_t c) {
     }
 }
 
-static void draw_cell(int16_t col, int16_t row, uint8_t cell) {
-    /* Map base $9800, 32 cells wide. Centre the 6-col grid → offset +7. */
-    uint8_t *map = (uint8_t *)0x9800;
-    if (row < 0 || row >= ROWS) return;
-    map[(row + 1) * 32 + (col + 7)] = tile_for(cell);
-}
-
-static void draw_grid(void) {
-    int16_t r, c;
-    for (r = 0; r < ROWS; r++)
-        for (c = 0; c < COLS; c++)
-            draw_cell(c, r, grid[r][c]);
-}
-
 static uint8_t collides(int16_t col, int16_t row) {
     uint8_t i;
     int16_t r;
@@ -109,36 +123,128 @@ static uint8_t collides(int16_t col, int16_t row) {
     for (i = 0; i < 3; i++) {
         r = row + i;
         if (r >= ROWS) return 1;
-        if (r >= 0 && grid[r][col] != 0) return 1;
+        if (r >= 0 && G(r, col) != 0) return 1;
     }
     return 0;
 }
 
-static void lock_piece(void) {
+/* ── match / clear / gravity core (mirrors the GBC reference) ─────── */
+
+/* Flag every cell in a 3+ run (any of the 4 directions) into matched[];
+ * return the count. A run is walked once, from its lowest end only. */
+static uint8_t mark_and_count(void) {
+    uint8_t r, c, d, len, cnt, col, k;
+    int8_t dr, dc;
+    int16_t sr, sc;
+
+    for (r = 0; r < NCELL; r++) matched[r] = 0;
+
+    for (r = 0; r < ROWS; r++) {
+        for (c = 0; c < COLS; c++) {
+            col = G(r, c);
+            if (col == 0) continue;
+            for (d = 0; d < 4; d++) {
+                dr = DIRS[d][0];
+                dc = DIRS[d][1];
+                sr = (int16_t)r - dr;
+                sc = (int16_t)c - dc;
+                if (sr >= 0 && sr < ROWS && sc >= 0 && sc < COLS
+                    && G(sr, sc) == col) continue;   /* not the run's start */
+                len = 1;
+                sr = (int16_t)r + dr;
+                sc = (int16_t)c + dc;
+                while (sr >= 0 && sr < ROWS && sc >= 0 && sc < COLS
+                       && G(sr, sc) == col) {
+                    len++;
+                    sr += dr;
+                    sc += dc;
+                }
+                if (len >= 3) {
+                    sr = (int16_t)r;
+                    sc = (int16_t)c;
+                    for (k = 0; k < len; k++) {
+                        M(sr, sc) = 1;
+                        sr += dr;
+                        sc += dc;
+                    }
+                }
+            }
+        }
+    }
+
+    cnt = 0;
+    for (r = 0; r < NCELL; r++) if (matched[r]) cnt++;
+    return cnt;
+}
+
+static void clear_marked(void) {
     uint8_t i;
+    for (i = 0; i < NCELL; i++) if (matched[i]) grid[i] = 0;
+}
+
+/* collapse each column so survivors rest on the floor — the "rows move
+ * down after a clear" the old template was missing. */
+static void apply_gravity(void) {
+    uint8_t c, r, n, w;
+    uint8_t buf[ROWS];
+    for (c = 0; c < COLS; c++) {
+        n = 0;
+        for (r = 0; r < ROWS; r++)
+            if (G(r, c)) { buf[n] = G(r, c); n++; }
+        for (r = 0; r < (uint8_t)(ROWS - n); r++) G(r, c) = 0;
+        w = 0;
+        for (r = (uint8_t)(ROWS - n); r < ROWS; r++) { G(r, c) = buf[w]; w++; }
+    }
+}
+
+/* clear chime — one short square blip per cascade step, pitch rises with
+ * the chain so combos audibly escalate. */
+static void sfx_clear(uint8_t chain) {
+    uint16_t p = 1797 + (uint16_t)chain * 26;   /* ~C5 rising */
+    if (p > 1980) p = 1980;
+    sound_play_tone(1, p, 6);
+}
+
+/* settle the board after a lock: match → clear → gravity, looping so
+ * cascades chain; score scales with the chain depth. */
+static void resolve_board(void) {
+    uint8_t n, chain = 0;
+    uint16_t amt;
+    while (1) {
+        n = mark_and_count();
+        if (n == 0) break;
+        chain++;
+        sfx_clear(chain);
+        clear_marked();
+        amt = (uint16_t)n * 10;
+        if (chain > 1) amt = amt * chain;
+        if (score < (uint16_t)(65500u - amt)) score += amt;
+        apply_gravity();
+    }
+}
+
+static void lock_piece(void) {
+    uint8_t i, written = 0;
     int16_t r;
-    int16_t c;
-    uint8_t a, b, d;
     for (i = 0; i < 3; i++) {
         r = piece_y + i;
-        if (r >= 0 && r < ROWS) grid[r][piece_x] = piece[i];
+        if (r >= 0 && r < ROWS) { G(r, piece_x) = piece[i]; written++; }
     }
-    for (i = 0; i < 3; i++) {
-        r = piece_y + i;
-        if (r < 0 || r >= ROWS) continue;
-        for (c = 0; c <= COLS - 3; c++) {
-            a = grid[r][c]; b = grid[r][c + 1]; d = grid[r][c + 2];
-            if (a != 0 && a == b && b == d) {
-                grid[r][c]     = 0;
-                grid[r][c + 1] = 0;
-                grid[r][c + 2] = 0;
-                if (score < 65500u) score += 30;
-            }
-        }
+    sound_play_noise(3);
+    resolve_board();
+    if (written == 0) {
+        /* The piece locked entirely ABOVE the well — the stack reached the
+         * top. Without this the game silently softlocks (invisible pieces
+         * locking off-screen forever). Scaffold behavior: low game-over
+         * tone, clear the board, restart the run. */
+        sound_play_tone(1, 1548, 30);
+        for (i = 0; i < NCELL; i++) grid[i] = 0;
+        score = 0;
     }
-    draw_grid();
 }
 
+/* ── rendering (vblank-budgeted; gameplay code never touches VRAM) ── */
+
 static void upload_tile(uint8_t slot, const uint8_t *src) {
     uint8_t *dst = (uint8_t *)(0x8000 + slot * 16);
     /* memcpy_vram (pointer-walk) — NOT an indexed dst[i]=src[i] loop, which
@@ -146,29 +252,106 @@ static void upload_tile(uint8_t slot, const uint8_t *src) {
     memcpy_vram(dst, src, 16);
 }
 
-/* Draw the well frame around the 6×12 play area. Grid cells live at
- * map[(row+1)*32 + (col+7)] (rows 1..12, cols 7..12), so the frame is the
- * column to each side (6 and 13) and the floor row just below (row 13). */
-static void draw_well(void) {
-    uint8_t *map = (uint8_t *)0x9800;
+#define VRAM_MAP ((volatile uint8_t *)0x9800)
+
+/* Direct cell write — ONLY safe with the LCD off or just after vblank. */
+static void set_cell(uint8_t c, uint8_t r, uint8_t tile) {
+    VRAM_MAP[(uint16_t)(WELL_MY + r) * 32 + WELL_MX + c] = tile;
+}
+
+/* COLLECT/FLUSH split (the reference puzzle's architecture, and the part
+ * that actually fixes "pieces flash / don't render"):
+ *  - collect_well() runs OUTSIDE vblank: scans for grid-vs-shadow diffs
+ *    (bounded), or queues rolling SCRUB cells when nothing changed, into a
+ *    tiny queue of precomputed (map offset, tile) pairs. RAM only.
+ *  - flush_well() runs FIRST thing after wait_vblank: pure pointer writes,
+ *    no scanning, no multiplies — the whole batch lands inside the ~10-line
+ *    vblank window every frame. The scrub means even a write the core drops
+ *    anyway heals itself on the next pass instead of sticking forever. */
+#define WQ_MAX 4
+static uint8_t  wq_n;
+static uint16_t wq_off[WQ_MAX];
+static uint8_t  wq_tile[WQ_MAX];
+static uint8_t diff_cursor, scrub_cursor;
+
+static uint16_t cell_off(uint8_t i) {
+    return (uint16_t)(WELL_MY + i / COLS) * 32 + WELL_MX + (i % COLS);
+}
+
+static void collect_well(void) {
+    uint8_t scanned = 0, i, k;
+    wq_n = 0;
+    i = diff_cursor;
+    while (scanned < NCELL && wq_n < WQ_MAX) {
+        if (grid[i] != shadow[i]) {
+            shadow[i] = grid[i];
+            wq_off[wq_n] = cell_off(i);
+            wq_tile[wq_n] = tile_for(grid[i]);
+            wq_n++;
+        }
+        i++;
+        if (i >= NCELL) i = 0;
+        scanned++;
+    }
+    diff_cursor = i;
+    if (wq_n == 0) {
+        /* idle: queue scrub cells so dropped writes self-heal */
+        for (k = 0; k < 2; k++) {
+            wq_off[wq_n] = cell_off(scrub_cursor);
+            wq_tile[wq_n] = tile_for(grid[scrub_cursor]);
+            wq_n++;
+            scrub_cursor++;
+            if (scrub_cursor >= NCELL) scrub_cursor = 0;
+        }
+    }
+}
+
+static void flush_well(void) {
+    uint8_t k;
+    for (k = 0; k < wq_n; k++) VRAM_MAP[wq_off[k]] = wq_tile[k];
+    wq_n = 0;
+}
+
+/* The falling piece = OAM sprites 0-2 (written to shadow_oam, flushed by
+ * one OAM DMA right after vblank starts). Rows above the well top (r < 0)
+ * park the sprite at Y=0 (offscreen). */
+static void update_piece_sprites(void) {
+    uint8_t i, sy, sx;
+    int16_t r;
+    for (i = 0; i < 3; i++) {
+        r = piece_y + i;
+        if (r >= 0 && r < ROWS) {
+            sy = (uint8_t)((WELL_MY + r) * 8 + 16);
+            sx = (uint8_t)((WELL_MX + piece_x) * 8 + 8);
+            oam_set(i, sy, sx, tile_for(piece[i]), 0);
+        } else {
+            oam_set(i, 0, 0, 0, 0);
+        }
+    }
+}
+
+static void draw_well_frame(void) {
     uint8_t r;
-    for (r = 1; r <= 12; r++) {
-        map[r * 32 + 6]  = T_WALL;   /* left wall  */
-        map[r * 32 + 13] = T_WALL;   /* right wall */
+    for (r = 0; r < ROWS; r++) {
+        VRAM_MAP[(uint16_t)(WELL_MY + r) * 32 + WELL_MX - 1]    = T_WALL;
+        VRAM_MAP[(uint16_t)(WELL_MY + r) * 32 + WELL_MX + COLS] = T_WALL;
     }
-    for (r = 6; r <= 13; r++)
-        map[13 * 32 + r] = T_WALL;   /* floor      */
+    for (r = 0; r < (uint8_t)(COLS + 2); r++)
+        VRAM_MAP[(uint16_t)(WELL_MY + ROWS) * 32 + WELL_MX - 1 + r] = T_WALL;
 }
 
 void main(void) {
-    uint8_t pad, prev = 0, fall_rate, t;
-    int16_t r, c;
-    uint8_t i;
+    uint8_t pad, prev = 0, fall_rate, t, i;
+    int16_t c;
     uint8_t *map;
-    int16_t pr;
 
     lcd_init_default();
+    enable_vblank_irq();
+    sound_init();
+    oam_dma_init_hram();
+    oam_clear();
     LCDC = 0;
+    OBP0 = 0xE4;                 /* DMG sprite palette: 3=black .. 0=white */
 
     upload_tile(T_BLANK, tile_blank);
     upload_tile(T_R,     tile_r);
@@ -185,64 +368,59 @@ void main(void) {
     map = (uint8_t *)0x9800;
     for (i = 0; i < 32; i++) {
         c = 0;
-        while (c < 32) { map[i * 32 + c] = T_BLANK; c++; }
+        while (c < 32) { map[(uint16_t)i * 32 + c] = T_BLANK; c++; }
     }
 
-    for (r = 0; r < ROWS; r++)
-        for (c = 0; c < COLS; c++)
-            grid[r][c] = 0;
+    for (i = 0; i < NCELL; i++) { grid[i] = 0; shadow[i] = 0; }
 
     score = 0;
     fall_timer = 0;
+    rng ^= DIV;                  /* a dash of boot-time entropy */
     new_piece();
-    draw_well();
-    draw_grid();
+    draw_well_frame();
 
-    LCDC = LCDC_LCD_ON | LCDC_BG_ON | LCDC_TILE_DATA_LO;
+    LCDC = LCDC_LCD_ON | LCDC_BG_ON | LCDC_OBJ_ON | LCDC_TILE_DATA_LO;
 
     while (1) {
-        wait_vblank();
-
-        /* Erase current piece visual (overwrite with what's underneath). */
-        for (i = 0; i < 3; i++) {
-            pr = piece_y + i;
-            if (pr >= 0 && pr < ROWS)
-                draw_cell(piece_x, pr, grid[pr][piece_x]);
-        }
-
         pad = joypad_read();
 
         if ((pad & PAD_LEFT)  && !(prev & PAD_LEFT)
-            && !collides(piece_x - 1, piece_y)) piece_x--;
+            && !collides(piece_x - 1, piece_y)) { piece_x--; sound_play_tone(1, 1899, 2); }
         if ((pad & PAD_RIGHT) && !(prev & PAD_RIGHT)
-            && !collides(piece_x + 1, piece_y)) piece_x++;
+            && !collides(piece_x + 1, piece_y)) { piece_x++; sound_play_tone(1, 1899, 2); }
         if ((pad & PAD_A) && !(prev & PAD_A)) {
             t = piece[0]; piece[0] = piece[1]; piece[1] = piece[2]; piece[2] = t;
+            sound_play_tone(1, 1923, 2);
+        }
+        if ((pad & PAD_B) && !(prev & PAD_B)) {
+            t = piece[2]; piece[2] = piece[1]; piece[1] = piece[0]; piece[0] = t;
+            sound_play_tone(1, 1923, 2);
         }
         if ((pad & PAD_START) && !(prev & PAD_START)) {
             while (!collides(piece_x, piece_y + 1)) piece_y++;
             lock_piece();
             new_piece();
-            prev = pad;
-            continue;
-        }
-        prev = pad;
-
-        fall_rate = (pad & PAD_DOWN) ? 4 : 30;
-        if (++fall_timer >= fall_rate) {
-            fall_timer = 0;
-            if (collides(piece_x, piece_y + 1)) {
-                lock_piece();
-                new_piece();
-            } else {
-                piece_y++;
+        } else {
+            fall_rate = (pad & PAD_DOWN) ? 4 : 30;
+            if (++fall_timer >= fall_rate) {
+                fall_timer = 0;
+                if (collides(piece_x, piece_y + 1)) {
+                    lock_piece();
+                    new_piece();
+                } else {
+                    piece_y++;
+                }
             }
         }
+        prev = pad;
 
-        /* Re-draw piece in new position. */
-        for (i = 0; i < 3; i++) {
-            pr = piece_y + i;
-            if (pr >= 0 && pr < ROWS) draw_cell(piece_x, pr, piece[i]);
-        }
+        /* COLLECT (RAM only, runs in active display) … */
+        update_piece_sprites();
+        collect_well();
+        /* … then FLUSH right after vblank starts: OAM DMA first (sprites
+         * tear if it slips out of vblank), then the queued BG writes. */
+        wait_vblank();
+        oam_dma_flush();
+        flush_well();
     }
 }

package/examples/gb/templates/racing.c

@@ -95,11 +95,23 @@ static void reset_run(void) {
   game_over_timer = 0;
 }
 
+/* Galois LFSR (taps $B8), period 255 -- real per-spawn randomness.
+ * The old code derived the spawn column from spawn_timer, but the caller
+ * resets spawn_timer just before calling here, so it was CONSTANT and
+ * every enemy spawned in the same left column/lane. */
+static uint8_t rng_state = 0xA5;
+static uint8_t rand8(void) {
+  uint8_t lsb = (uint8_t)(rng_state & 1);
+  rng_state >>= 1;
+  if (lsb) rng_state ^= 0xB8;
+  return rng_state;
+}
+
 static void spawn_obstacle(void) {
   uint8_t i;
   for (i = 0; i < MAX_OBSTACLES; i++) {
     if (!obstacles[i].alive) {
-      obstacles[i].x = lane_x[(spawn_timer * 13) % 3];
+      obstacles[i].x = lane_x[rand8() % 3];
       obstacles[i].y = 0;
       obstacles[i].alive = 1;
       return;

package/examples/gb/templates/shmup.c

@@ -86,11 +86,23 @@ static void fire(void) {
     }
 }
 
+/* Galois LFSR (taps $B8), period 255 -- real per-spawn randomness.
+ * The old code derived the spawn column from spawn_timer, but the caller
+ * resets spawn_timer just before calling here, so it was CONSTANT and
+ * every enemy spawned in the same left column/lane. */
+static uint8_t rng_state = 0xA5;
+static uint8_t rand8(void) {
+  uint8_t lsb = (uint8_t)(rng_state & 1);
+  rng_state >>= 1;
+  if (lsb) rng_state ^= 0xB8;
+  return rng_state;
+}
+
 static void spawn(void) {
     uint8_t i;
     for (i = 0; i < MAX_ENEMIES; i++) {
         if (!enemies[i].alive) {
-            enemies[i].x = ((spawn_timer * 37) & 0x7F) % (160 - 16) + 8;
+            enemies[i].x = rand8() % (160 - 16) + 8;
             enemies[i].y = 0;
             enemies[i].alive = 1;
             return;

package/examples/gb/templates/sports.c

@@ -130,15 +130,21 @@ void main(void) {
 
   while (1) {
     wait_vblank();
+    /* OAM DMA FIRST — at the leading edge of vblank. The old order staged
+     * 45 oam_set CALLS before the DMA; the SDCC call overhead pushed the
+     * DMA ~a third of the frame into ACTIVE display, so the sprites tore
+     * on one fixed scanline ("horizontal line a 3rd of the way down").
+     * Sprites now display the state staged LAST frame (1 frame of latency,
+     * imperceptible in Pong). */
+    oam_dma_flush();
 
-    /* Stage OAM: left paddle (2 stacked), right paddle (2 stacked), ball. */
-    for (i = 0; i < 40; i++) oam_set(i, 0, 0, 0, 0);
+    /* Stage next frame's OAM (RAM only — safe any time). Slots 5-39 were
+     * zeroed once by oam_clear() at boot and never change. */
     oam_set(0, (uint8_t)(p1y + 16),         (uint8_t)(PADDLE_X1 + 8), 1, 0);
     oam_set(1, (uint8_t)(p1y + 16 + 8),     (uint8_t)(PADDLE_X1 + 8), 1, 0);
     oam_set(2, (uint8_t)(p2y + 16),         (uint8_t)(PADDLE_X2 + 8), 1, 0);
     oam_set(3, (uint8_t)(p2y + 16 + 8),     (uint8_t)(PADDLE_X2 + 8), 1, 0);
     oam_set(4, (uint8_t)(by + 16),          (uint8_t)(bx + 8),        1, 0);
-    oam_dma_flush();
 
     pad = joypad_read();
     if (pad & PAD_UP    && p1y > COURT_TOP)            p1y -= 2;

package/examples/gba/templates/platformer.c

@@ -65,18 +65,6 @@ static int on_platform(s16 px, s16 py) {
     return 0;
 }
 
-static int blocked_below(s16 px, s16 py) {
-    for (int i = 0; i < N_PLATFORMS; i++) {
-        const Rect *p = &platforms[i];
-        if (py + 8 <= p->y && py + 9 > p->y
-            && px + 8 > p->x
-            && px     < p->x + p->w) {
-            return 1;
-        }
-    }
-    return 0;
-}
-
 int main(void) {
     /* ── BG palette (for the platform tile) ──────────────────────────
      * pal_bg_mem[i] is the BG palette. */
@@ -182,7 +170,13 @@ int main(void) {
         /* Vertical with platform-stop. */
         s32 np = py + vy;
         s16 npy = np >> 4;
-        if (vy > 0 && blocked_below(ipx, ipy)) {
+        /* THE fall-through-the-floor fix: this used to be additionally
+         * gated on blocked_below(), which only matches when a platform
+         * top is within ONE pixel of the feet — but falls reach 20 px/
+         * frame, so the (correct) crossing test below almost never got
+         * to run and the player tunnelled through every platform. The
+         * crossing test alone is the right check. */
+        if (vy > 0) {
             for (int i = 0; i < N_PLATFORMS; i++) {
                 const Rect *p = &platforms[i];
                 if (ipy + 8 <= p->y && npy + 8 >= p->y