src/n_iso_engine.c

Isometric/axonometric tile engine with height maps, terrain transitions, and A* pathfinding integration. More...

Functions

	Name
int	_tile_covers_height(const ISO_MAP * map, int mx, int my, int z) Check if a tile has geometry covering height level z.
float	_smooth_neighbor_h(const ISO_MAP * map, float h_self, int nx, int ny) Helper: clamp neighbor height for smooth slope rendering.
int	_neighbor_terrain_at(const ISO_MAP * map, int mx, int my) Edge-aware terrain lookup for chunked rendering.
int	_neighbor_should_transition(const ISO_MAP * map, int mx1, int my1, int mx2, int my2)
int	_neighbor_height_at(const ISO_MAP * map, int mx, int my)
float	_iso_preset_angle(int preset) Get the angle in degrees for a projection preset.
int	_iso_object_check_occlusion(const ISO_MAP * map, float obj_fx, float obj_fy, float obj_fz, float obj_sx, float obj_sy, float sprite_h, float sprite_half_w, float cam_px, float cam_py, float zoom, float * out_clip_y) Check if an object is occluded by tiles drawn in front of it.
FILE *	_iso_fopen_write(const char * path, const char * mode)
void	_iso_corner_to_screen(float hw, float hh, float tl, int cx, int cy, float fh, float cam_px, float cam_py, float zoom, float * sx, float * sy) Project a tile corner to screen coordinates (canonical formula).
void	_draw_tile_warped(ALLEGRO_BITMAP * bmp, const float vx[4], const float vy[4], int tile_w, int tile_h, ALLEGRO_COLOR tint) Draw a tile bitmap warped to 4 pre-computed screen vertices.
void	_draw_segment_underside(ALLEGRO_BITMAP * bmp, float base_dx, float base_dy, float phw, float phh, float tile_draw_w, float tile_draw_h, int tile_w, int tile_h) Draw the underside face of a floating segment at 65% brightness.
void	_draw_segment_sides(const ISO_MAP * map, float east_x, float east_y, float south_x, float south_y, float west_x, float west_y, int mx, int my, int seg_bottom, int seg_top, float lift, float height_brightness, int wall_terrain_override) Draw cliff side faces for a single segment {seg_bottom..seg_top}.
void	_draw_height_borders(const ISO_MAP * map, int mx, int my, int seg_idx, int seg_bottom, int seg_top, int has_underside_face, const float vx[4], const float vy[4], float lift) Draw black border outlines on the diamond edges of an elevated tile.
ALLEGRO_COLOR	_compute_tile_tint(int seg_top, int max_height, float intensity, float ambient_r, float ambient_g, float ambient_b, float dyn_r, float dyn_g, float dyn_b)
int	_cmp_draw_entry(const void * a, const void * b)

Attributes

	Name
IsoPrimitiveColorFilter	g_iso_primitive_color_filter
IsoLineJitterCompensationFn	g_iso_line_jitter_fn

Detailed Description

Isometric/axonometric tile engine with height maps, terrain transitions, and A* pathfinding integration.

Author: Castagnier Mickael

Version: 2.0

Date: 01/03/2026

Functions Documentation

function _tile_covers_height

static int _tile_covers_height(
    const ISO_MAP * map,
    int mx,
    int my,
    int z
)

Check if a tile has geometry covering height level z.

Returns 1 if any segment of (mx,my) has bottom <= z AND top > z, meaning the tile's wall/surface extends strictly above z. Used for occlusion: a tile drawn later in painter order covers edges of earlier tiles when its wall passes above the edge height.

Requires the occluding tile to be at least 2 height units taller (top > z + 1) so that single-step stairs (+1 per tile) do not suppress the NW/NE borders, the step only partially covers the edge, and painter order handles the overlap visually.

function _smooth_neighbor_h

static inline float _smooth_neighbor_h(
    const ISO_MAP * map,
    float h_self,
    int nx,
    int ny
)

Helper: clamp neighbor height for smooth slope rendering.

If the height difference exceeds smooth_slope_max, return self height (creating a cliff instead of a slope).

function _neighbor_terrain_at

static int _neighbor_terrain_at(
    const ISO_MAP * map,
    int mx,
    int my
)

Edge-aware terrain lookup for chunked rendering.

Parameters:

• map the map
• mx tile X (may be just outside [0, width))
• my tile Y (may be just outside [0, height))

Return: terrain index at (mx,my), or 0 if both in-chunk and neighbor lookups fail

When (mx,my) falls just outside the chunk on a cardinal edge and the caller has wired up the neighbor_terrains_* arrays, consult them so transition masks can cross chunk seams. Out-of-range without a neighbor array falls back to the standard 0-out-of-bounds semantics. Diagonal neighbors (corners of the 8-neighborhood that fall outside the chunk's own bounds) are not covered, a fully correct diagonal fix would need 4 extra corner-cell fields.

function _neighbor_should_transition

static int _neighbor_should_transition(
    const ISO_MAP * map,
    int mx1,
    int my1,
    int mx2,
    int my2
)

function _neighbor_height_at

static int _neighbor_height_at(
    const ISO_MAP * map,
    int mx,
    int my
)

function _iso_preset_angle

static float _iso_preset_angle(
    int preset
)

Get the angle in degrees for a projection preset.

Parameters:

• preset one of ISO_PROJ_CLASSIC, ISO_PROJ_ISOMETRIC, ISO_PROJ_STAGGERED, ISO_PROJ_MILITARY

Return: angle in degrees

function _iso_object_check_occlusion

static int _iso_object_check_occlusion(
    const ISO_MAP * map,
    float obj_fx,
    float obj_fy,
    float obj_fz,
    float obj_sx,
    float obj_sy,
    float sprite_h,
    float sprite_half_w,
    float cam_px,
    float cam_py,
    float zoom,
    float * out_clip_y
)

Check if an object is occluded by tiles drawn in front of it.

Parameters:

• map the map
• obj_fx object fractional map X
• obj_fy object fractional map Y
• obj_fz object height
• obj_sx object screen X (camera-transformed)
• obj_sy object screen Y (camera-transformed)
• sprite_h sprite height used for occlusion bounding
• sprite_half_w half-width of the sprite footprint (occlusion bbox)
• cam_px camera X offset in pixels
• cam_py camera Y offset in pixels
• zoom current zoom
• out_clip_y optional out-pointer: when non-NULL and a partial overlap is detected, receives the screen-Y above which the sprite should be clipped (for soft-fade rendering). May be NULL.

Return: 1 if occluded, 0 if visible

Tests tiles at higher depth (mx+my) in a forward cone to see if any elevated tile's screen area overlaps the object's screen position.

function _iso_fopen_write

static FILE * _iso_fopen_write(
    const char * path,
    const char * mode
)

function _iso_corner_to_screen

static inline void _iso_corner_to_screen(
    float hw,
    float hh,
    float tl,
    int cx,
    int cy,
    float fh,
    float cam_px,
    float cam_py,
    float zoom,
    float * sx,
    float * sy
)

Project a tile corner to screen coordinates (canonical formula).

Parameters:

• hw map->proj.half_w (unzoomed)
• hh map->proj.half_h (unzoomed)
• tl map->proj.tile_lift (unzoomed)
• cx corner X in tile grid (e.g. mx+1 for the East corner)
• cy corner Y in tile grid
• fh height (float, supports smooth-height interpolation)
• cam_px integer-snapped camera pixel X: floor(cam_x * zoom)
• cam_py integer-snapped camera pixel Y: floor(cam_y * zoom)
• zoom zoom factor
• sx output screen X
• sy output screen Y

Every screen vertex in the renderer must come from this function so that two tiles sharing a corner get bit-identical screen positions regardless of which tile's draw call is computing the vertex.

The camera offset (cam_px, cam_py) must be integer-snapped pixel offsets computed as floor(cam_x * zoom), floor(cam_y * zoom). This ensures the fractional pixel position of each vertex comes only from (wx * zoom), which is constant per vertex, eliminating sub-pixel aliasing shimmer during camera movement.

function _draw_tile_warped

static void _draw_tile_warped(
    ALLEGRO_BITMAP * bmp,
    const float vx[4],
    const float vy[4],
    int tile_w,
    int tile_h,
    ALLEGRO_COLOR tint
)

Draw a tile bitmap warped to 4 pre-computed screen vertices.

Uses two textured triangles mapped to the warped diamond shape. Vertices must come from _iso_corner_to_screen for bit-identical shared edges between adjacent tiles.

function _draw_segment_underside

static void _draw_segment_underside(
    ALLEGRO_BITMAP * bmp,
    float base_dx,
    float base_dy,
    float phw,
    float phh,
    float tile_draw_w,
    float tile_draw_h,
    int tile_w,
    int tile_h
)

Draw the underside face of a floating segment at 65% brightness.

The underside is a diamond drawn at the segment's bottom height.

function _draw_segment_sides

static void _draw_segment_sides(
    const ISO_MAP * map,
    float east_x,
    float east_y,
    float south_x,
    float south_y,
    float west_x,
    float west_y,
    int mx,
    int my,
    int seg_bottom,
    int seg_top,
    float lift,
    float height_brightness,
    int wall_terrain_override
)

Draw cliff side faces for a single segment {seg_bottom..seg_top}.

For ground segments (seg_bottom==0) uses neighbor-aware wall height. For floating segments (seg_bottom>0) draws full segment height.

function _draw_height_borders

static void _draw_height_borders(
    const ISO_MAP * map,
    int mx,
    int my,
    int seg_idx,
    int seg_bottom,
    int seg_top,
    int has_underside_face,
    const float vx[4],
    const float vy[4],
    float lift
)

Draw black border outlines on the diamond edges of an elevated tile.

Rules:

• Only draws edges where height difference > 1 or neighbor missing.
• NEVER draws the top vertex (N) where NW and NE edges meet, it is always hidden behind the tile surface itself.
• At corners E, S, W where two drawn edges meet, they join naturally.
• Occlusion is handled by painter's algorithm: tiles drawn later (closer to camera) paint over borders of earlier tiles.

Diamond vertices: N=vx[0] (top), E=vx[1] (right), S=vx[2] (bottom), W=vx[3] (left)

function _compute_tile_tint

static inline ALLEGRO_COLOR _compute_tile_tint(
    int seg_top,
    int max_height,
    float intensity,
    float ambient_r,
    float ambient_g,
    float ambient_b,
    float dyn_r,
    float dyn_g,
    float dyn_b
)

function _cmp_draw_entry

static int _cmp_draw_entry(
    const void * a,
    const void * b
)

Attributes Documentation

variable g_iso_primitive_color_filter

static IsoPrimitiveColorFilter g_iso_primitive_color_filter = NULL;

variable g_iso_line_jitter_fn

static IsoLineJitterCompensationFn g_iso_line_jitter_fn = NULL;

Source code

/*
 * Nilorea Library
 * Copyright (C) 2005-2026 Castagnier Mickael
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied. See the License for the specific language governing
 * permissions and limitations under the License.
 *
 * SPDX-License-Identifier: Apache-2.0
 */


#include <math.h>
#include <stdio.h>
#include <stdlib.h> /* realloc, free, getenv */

/* Include n_astar.h BEFORE n_iso_engine.h so the iso_map_to_astar_grid
 * function is enabled via the N_ASTAR_H guard */
#include "nilorea/n_astar.h"
#include "nilorea/n_iso_engine.h"

#ifndef __windows__
#include <fcntl.h>
#include <unistd.h>

/* Open a file for writing with explicit owner-only write permission (0644)
 * instead of the world-writable 0666 that fopen() requests before the umask
 * is applied. On Windows the POSIX mode bits do not apply, so fopen() is used
 * directly. */
static FILE* _iso_fopen_write(const char* path, const char* mode) {
    int fd = open(path, O_WRONLY | O_CREAT | O_TRUNC, 0644);
    if (fd < 0)
        return NULL;
    FILE* f = fdopen(fd, mode);
    if (!f)
        close(fd);
    return f;
}
#else
#define _iso_fopen_write(path, mode) fopen((path), (mode))
#endif

/* Per-pass timing accumulators. Host resets these at frame start,
 * iso_map_draw adds elapsed microseconds per pass per invocation. See
 * the doxygen block in n_iso_engine.h for usage. */
double n_iso_pass1_us = 0.0;
double n_iso_pass2_us = 0.0;
double n_iso_pass3_us = 0.0;
double n_iso_setup_us = 0.0;

#ifdef HAVE_ALLEGRO
/* Optional primitive-color filter (sRGB-aware consumers register a
 * design->linear converter here). See header docstring. NULL = identity.
 * Gated on HAVE_ALLEGRO because the prototype mentions ALLEGRO_COLOR
 * which doesn't exist in the no-allegro test build path. */
static IsoPrimitiveColorFilter g_iso_primitive_color_filter = NULL;

void iso_set_primitive_color_filter(IsoPrimitiveColorFilter fn) {
    g_iso_primitive_color_filter = fn;
}

ALLEGRO_COLOR iso_apply_primitive_color_filter(ALLEGRO_COLOR c) {
    if (g_iso_primitive_color_filter) return g_iso_primitive_color_filter(c);
    return c;
}

/* Line-jitter compensation callback. See header doc. */
static IsoLineJitterCompensationFn g_iso_line_jitter_fn = NULL;

void iso_set_line_jitter_compensation(IsoLineJitterCompensationFn fn) {
    g_iso_line_jitter_fn = fn;
}

void iso_apply_line_jitter_compensation(float* out_dx, float* out_dy) {
    if (g_iso_line_jitter_fn) {
        g_iso_line_jitter_fn(out_dx, out_dy);
    } else {
        if (out_dx) *out_dx = 0.0f;
        if (out_dy) *out_dy = 0.0f;
    }
}
#endif /* HAVE_ALLEGRO */

#ifndef NOISOENGINE

/* Height-aware ISO_MAP API (Articles 747/748/934/1269/2026)
 * Core API does not require Allegro and compiles everywhere.
 */

ISO_MAP* iso_map_new(int width, int height, int num_terrains, int max_height) {
    __n_assert(width > 0, return NULL);
    __n_assert(height > 0, return NULL);

    ISO_MAP* map = NULL;
    Malloc(map, ISO_MAP, 1);
    __n_assert(map, return NULL);

    map->width = width;
    map->height = height;
    map->num_terrains = num_terrains;
    map->max_height = max_height;

    size_t total = (size_t)width * (size_t)height;

    Malloc(map->terrain, int, total);
    if (!map->terrain) {
        Free(map);
        return NULL;
    }

    Malloc(map->heightmap, int, total);
    if (!map->heightmap) {
        Free(map->terrain);
        Free(map);
        return NULL;
    }

    Malloc(map->ability, int, total);
    if (!map->ability) {
        Free(map->heightmap);
        Free(map->terrain);
        Free(map);
        return NULL;
    }

    memset(map->terrain, 0, total * sizeof(int));
    memset(map->heightmap, 0, total * sizeof(int));
    for (size_t i = 0; i < total; i++) {
        map->ability[i] = WALK;
    }

    /* default classic 2:1 projection for 64px wide tiles */
    iso_map_set_projection(map, ISO_PROJ_CLASSIC, 64);

    map->segments = NULL; /* allocated on demand by iso_map_set_segments() */

    /* Overlay layer: allocated and zeroed, 0 = no overlay */
    Malloc(map->overlay, int, total);
    if (map->overlay) memset(map->overlay, 0, total * sizeof(int));

    /* default rendering flags */
    map->smooth_height = 0;
    map->smooth_slope_max = 1;
    map->show_grid = 0;
    map->hover_mx = -1;
    map->hover_my = -1;
    map->height_tint_intensity = 0.0f; /* disabled by default; game sets this */

    /* Ambient color defaults: full daylight (no tinting) */
    map->ambient_r = 1.0f;
    map->ambient_g = 1.0f;
    map->ambient_b = 1.0f;
    map->dynamic_light_map = NULL; /* caller-owned, set before draw */

    /* Draw_order cache. dirty=1 forces a first build. */
    map->cached_draw_order = NULL;
    map->cached_draw_order_count = 0;
    map->cached_draw_order_cap = 0;
    map->draw_order_dirty = 1;

    return map;
} /* iso_map_new() */

void iso_map_free(ISO_MAP** map_ptr) {
    __n_assert(map_ptr, return);
    __n_assert(*map_ptr, return);
    ISO_MAP* map = *map_ptr;
    Free(map->terrain);
    Free(map->heightmap);
    Free(map->ability);
    Free(map->segments);
    Free(map->overlay);
    /* Cached draw_order */
    if (map->cached_draw_order) {
        free(map->cached_draw_order);
        map->cached_draw_order = NULL;
    }
    Free(*map_ptr);
} /* iso_map_free() */

static float _iso_preset_angle(int preset) {
    switch (preset) {
        case ISO_PROJ_ISOMETRIC:
            return 30.0f;
        case ISO_PROJ_STAGGERED:
            return 18.43f;
        case ISO_PROJ_MILITARY:
            return 45.0f;
        case ISO_PROJ_CLASSIC:
        default:
            return 26.565f;
    }
}

void iso_map_set_projection(ISO_MAP* map, int preset, float tile_width) {
    __n_assert(map, return);
    float hw = tile_width / 2.0f;
    float angle = _iso_preset_angle(preset);
    float hh = hw * tanf(angle * (float)M_PI / 180.0f);

    map->proj.half_w = hw;
    map->proj.half_h = hh;
    map->proj.angle_deg = angle;
    map->proj.target_angle = angle;
    map->proj.lerp_speed = 3.0f;
    map->proj.tile_lift = hh;
} /* iso_map_set_projection() */

void iso_map_set_projection_target(ISO_MAP* map, int preset) {
    __n_assert(map, return);
    map->proj.target_angle = _iso_preset_angle(preset);
} /* iso_map_set_projection_target() */

void iso_map_lerp_projection(ISO_MAP* map, float dt) {
    __n_assert(map, return);
    float diff = map->proj.target_angle - map->proj.angle_deg;
    if (fabsf(diff) < 0.01f) {
        map->proj.angle_deg = map->proj.target_angle;
    } else {
        map->proj.angle_deg += diff * map->proj.lerp_speed * dt;
    }
    float rad = map->proj.angle_deg * (float)M_PI / 180.0f;
    map->proj.half_h = map->proj.half_w * sinf(rad) / cosf(rad);
    if (map->proj.half_h < 8.0f) map->proj.half_h = 8.0f;
    if (map->proj.half_h > map->proj.half_w * 2.0f) map->proj.half_h = map->proj.half_w * 2.0f;
} /* iso_map_lerp_projection() */

const char* iso_projection_name(int preset) {
    switch (preset) {
        case ISO_PROJ_CLASSIC:
            return "Classic 2:1";
        case ISO_PROJ_ISOMETRIC:
            return "True Isometric";
        case ISO_PROJ_STAGGERED:
            return "Staggered";
        case ISO_PROJ_MILITARY:
            return "Military";
        default:
            return "Unknown";
    }
} /* iso_projection_name() */

void iso_map_set_projection_custom(ISO_MAP* map, float half_w, float half_h, float tile_lift) {
    __n_assert(map, return);
    map->proj.half_w = half_w;
    map->proj.half_h = half_h;
    map->proj.tile_lift = tile_lift;
    map->proj.angle_deg = atanf(half_h / half_w) * 180.0f / (float)M_PI;
    map->proj.target_angle = map->proj.angle_deg;
} /* iso_map_set_projection_custom() */

int iso_map_get_terrain(const ISO_MAP* map, int mx, int my) {
    __n_assert(map, return 0);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return 0;
    return map->terrain[my * map->width + mx];
} /* iso_map_get_terrain() */

void iso_map_set_terrain(ISO_MAP* map, int mx, int my, int terrain) {
    __n_assert(map, return);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return;
    map->terrain[my * map->width + mx] = terrain;
} /* iso_map_set_terrain() */

int iso_map_get_height(const ISO_MAP* map, int mx, int my) {
    __n_assert(map, return 0);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return 0;
    return map->heightmap[my * map->width + mx];
} /* iso_map_get_height() */

void iso_map_set_height(ISO_MAP* map, int mx, int my, int h) {
    __n_assert(map, return);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return;
    if (h < 0) h = 0;
    if (h > map->max_height) h = map->max_height;
    map->heightmap[my * map->width + mx] = h;
} /* iso_map_set_height() */

int iso_map_get_ability(const ISO_MAP* map, int mx, int my) {
    __n_assert(map, return 0);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return 0;
    return map->ability[my * map->width + mx];
} /* iso_map_get_ability() */

void iso_map_set_ability(ISO_MAP* map, int mx, int my, int ab) {
    __n_assert(map, return);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return;
    map->ability[my * map->width + mx] = ab;
} /* iso_map_set_ability() */

void iso_map_to_screen(const ISO_MAP* map, int mx, int my, int h, float* screen_x, float* screen_y) {
    __n_assert(map, return);
    __n_assert(screen_x, return);
    __n_assert(screen_y, return);
    *screen_x = ((float)mx - (float)my) * map->proj.half_w + map->proj.half_w;
    *screen_y = ((float)mx + (float)my) * map->proj.half_h - (float)h * map->proj.tile_lift;
} /* iso_map_to_screen() */

void iso_screen_to_map(const ISO_MAP* map, float screen_x, float screen_y, int* mx, int* my) {
    __n_assert(map, return);
    __n_assert(mx, return);
    __n_assert(my, return);
    float tw = 2.0f * map->proj.half_w;
    float th = 2.0f * map->proj.half_h;
    *mx = (int)floorf(screen_x / tw + screen_y / th - 1.0f);
    *my = (int)floorf(screen_y / th - screen_x / tw + 1.0f);
} /* iso_screen_to_map() */

void iso_map_to_screen_f(const ISO_MAP* map, float fmx, float fmy, float h, float* screen_x, float* screen_y) {
    __n_assert(map, return);
    __n_assert(screen_x, return);
    __n_assert(screen_y, return);
    *screen_x = (fmx - fmy) * map->proj.half_w + 2.0f * map->proj.half_w;
    *screen_y = (fmx + fmy) * map->proj.half_h - h * map->proj.tile_lift;
} /* iso_map_to_screen_f() */

void iso_corner_to_screen(const ISO_MAP* map, int cx, int cy, float fh, float cam_px, float cam_py, float zoom, float* sx, float* sy) {
    __n_assert(map, return);
    __n_assert(sx, return);
    __n_assert(sy, return);
    float hw = map->proj.half_w;
    float hh = map->proj.half_h;
    float tl = map->proj.tile_lift;
    float wx = (float)(cx - cy) * hw + 2.0f * hw;
    float wy = (float)(cx + cy) * hh - fh * tl;
    *sx = wx * zoom + cam_px;
    *sy = wy * zoom + cam_py;
} /* iso_corner_to_screen() */

int iso_is_in_diamond(int px, int py, int tile_w, int tile_h) {
    float cx = (float)tile_w / 2.0f;
    float cy = (float)tile_h / 2.0f;
    float dx = fabsf((float)px + 0.5f - cx) / ((float)tile_w / 2.0f);
    float dy = fabsf((float)py + 0.5f - cy) / ((float)tile_h / 2.0f);
    return (dx + dy) <= 1.0f;
} /* iso_is_in_diamond() */

float iso_diamond_dist(int px, int py, int tile_w, int tile_h) {
    float cx = (float)tile_w / 2.0f;
    float cy = (float)tile_h / 2.0f;
    float dx = fabsf((float)px + 0.5f - cx) / ((float)tile_w / 2.0f);
    float dy = fabsf((float)py + 0.5f - cy) / ((float)tile_h / 2.0f);
    float d = dx + dy;
    if (d > 1.0f) return 0.0f;
    return 1.0f - d;
} /* iso_diamond_dist() */

void iso_screen_to_map_height(const ISO_MAP* map, float screen_x, float screen_y, int tile_w, int tile_h, int* mx, int* my) {
    __n_assert(map, return);
    __n_assert(mx, return);
    __n_assert(my, return);

    float half_w = map->proj.half_w;
    float half_h = map->proj.half_h;
    float tile_lift = map->proj.tile_lift;
    float tw = half_w * 2.0f;
    float th = half_h * 2.0f;

    for (int h = map->max_height; h >= 0; h--) {
        float adj_sy = screen_y + (float)h * tile_lift;
        float fmx = screen_x / tw + adj_sy / th - 1.0f;
        float fmy = adj_sy / th - screen_x / tw + 1.0f;
        int tmx = (int)floorf(fmx);
        int tmy = (int)floorf(fmy);

        if (tmx < 0 || tmx >= map->width || tmy < 0 || tmy >= map->height)
            continue;

        if (iso_map_get_height(map, tmx, tmy) == h) {
            float tile_sx, tile_sy;
            iso_map_to_screen(map, tmx, tmy, h, &tile_sx, &tile_sy);
            float norm_x = (screen_x - tile_sx) / tw * (float)tile_w;
            float norm_y = (screen_y - tile_sy) / th * (float)tile_h;
            int lpx = (int)norm_x;
            int lpy = (int)norm_y;
            if (lpx >= 0 && lpx < tile_w && lpy >= 0 && lpy < tile_h &&
                iso_is_in_diamond(lpx, lpy, tile_w, tile_h)) {
                *mx = tmx;
                *my = tmy;
                return;
            }
        }
    }

    /* Fallback: flat projection (height 0) */
    *mx = (int)floorf(screen_x / tw + screen_y / th - 1.0f);
    *my = (int)floorf(screen_y / th - screen_x / tw + 1.0f);
} /* iso_screen_to_map_height() */

void iso_screen_to_map_height_f(const ISO_MAP* map, float screen_x, float screen_y, int tile_w, int tile_h, int* mx, int* my, float* out_fx, float* out_fy) {
    __n_assert(map, return);
    __n_assert(mx, return);
    __n_assert(my, return);

    float half_w = map->proj.half_w;
    float half_h = map->proj.half_h;
    float tile_lift = map->proj.tile_lift;
    float tw = half_w * 2.0f;
    float th = half_h * 2.0f;

    for (int h = map->max_height; h >= 0; h--) {
        float adj_sy = screen_y + (float)h * tile_lift;
        float fmx = screen_x / tw + adj_sy / th - 1.0f;
        float fmy = adj_sy / th - screen_x / tw + 1.0f;
        int tmx = (int)floorf(fmx);
        int tmy = (int)floorf(fmy);

        if (tmx < 0 || tmx >= map->width || tmy < 0 || tmy >= map->height)
            continue;

        if (iso_map_get_height(map, tmx, tmy) == h) {
            float tile_sx, tile_sy;
            iso_map_to_screen(map, tmx, tmy, h, &tile_sx, &tile_sy);
            float norm_x = (screen_x - tile_sx) / tw * (float)tile_w;
            float norm_y = (screen_y - tile_sy) / th * (float)tile_h;
            int lpx = (int)norm_x;
            int lpy = (int)norm_y;
            if (lpx >= 0 && lpx < tile_w && lpy >= 0 && lpy < tile_h &&
                iso_is_in_diamond(lpx, lpy, tile_w, tile_h)) {
                *mx = tmx;
                *my = tmy;
                if (out_fx || out_fy) {
                    /* Refine fractional coordinates using interpolated height
                     * so that iso_map_to_screen_f(fx, fy, interp_h) returns
                     * the original screen position.  Without this, sub-tile
                     * positions are off when the bilinear-interpolated height
                     * differs from the tile's integer height (e.g. at edges
                     * between tiles of different heights).
                     *
                     * Only refine in SMOOTH mode.  In CUT mode the rendered
                     * tile sits at the flat integer height, so the initial
                     * fmx/fmy (computed at that height) are already correct.
                     * Refining with bilinear-interpolated heights in CUT mode
                     * pulls the position toward tile corners and causes
                     * click-to-move to land at the wrong spot. */
                    float refined_fx = fmx;
                    float refined_fy = fmy;
                    if (map->smooth_height) {
                        /* Upper bounds: allow the full extent of the last tile.
                         * Tile tmx occupies [tmx, tmx+1) in map coords, and the
                         * last valid position must stay < map->width so that
                         * floor() never yields an out-of-bounds tile index. */
                        float map_max_fx = (float)map->width - 1e-4f;
                        float map_max_fy = (float)map->height - 1e-4f;
                        /* Tile-local bounds to prevent divergence: clamp each
                         * refinement step to the detected tile so that large
                         * height differences with neighbours cannot pull the
                         * position into an adjacent tile. */
                        float tile_min_fx = (float)tmx;
                        float tile_min_fy = (float)tmy;
                        float tile_max_fx = (float)(tmx + 1);
                        float tile_max_fy = (float)(tmy + 1);
                        if (tile_max_fx > map_max_fx) tile_max_fx = map_max_fx;
                        if (tile_max_fy > map_max_fy) tile_max_fy = map_max_fy;
                        for (int iter = 0; iter < 3; iter++) {
                            float interp_h = iso_map_interpolate_height(map, refined_fx, refined_fy);
                            float new_adj_sy = screen_y + interp_h * tile_lift;
                            refined_fx = screen_x / tw + new_adj_sy / th - 1.0f;
                            refined_fy = new_adj_sy / th - screen_x / tw + 1.0f;
                            /* Keep within detected tile */
                            if (refined_fx < tile_min_fx) refined_fx = tile_min_fx;
                            if (refined_fy < tile_min_fy) refined_fy = tile_min_fy;
                            if (refined_fx > tile_max_fx) refined_fx = tile_max_fx;
                            if (refined_fy > tile_max_fy) refined_fy = tile_max_fy;
                        }
                    }
                    /* Clamp to map bounds so border clicks stay valid */
                    float map_max_fx2 = (float)map->width - 1e-4f;
                    float map_max_fy2 = (float)map->height - 1e-4f;
                    if (refined_fx < 0.0f) refined_fx = 0.0f;
                    if (refined_fy < 0.0f) refined_fy = 0.0f;
                    if (refined_fx > map_max_fx2) refined_fx = map_max_fx2;
                    if (refined_fy > map_max_fy2) refined_fy = map_max_fy2;
                    if (out_fx) *out_fx = refined_fx;
                    if (out_fy) *out_fy = refined_fy;
                }
                return;
            }
        }
    }

    /* Fallback: flat projection (height 0) */
    float fmx = screen_x / tw + screen_y / th - 1.0f;
    float fmy = screen_y / th - screen_x / tw + 1.0f;
    *mx = (int)floorf(fmx);
    *my = (int)floorf(fmy);
    if (out_fx) *out_fx = fmx;
    if (out_fy) *out_fy = fmy;
} /* iso_screen_to_map_height_f() */

float iso_map_interpolate_height(const ISO_MAP* map, float fx, float fy) {
    __n_assert(map, return 0.0f);

    /* Clamp to valid map range so border tiles interpolate correctly.
     * Allow the full extent of the last tile (up to width/height minus
     * a tiny epsilon) so that positions in the right half of border
     * tiles are not snapped to the tile corner. */
    if (fx < 0.0f) fx = 0.0f;
    if (fy < 0.0f) fy = 0.0f;
    float max_fx = (float)map->width - 1e-4f;
    float max_fy = (float)map->height - 1e-4f;
    if (fx > max_fx) fx = max_fx;
    if (fy > max_fy) fy = max_fy;

    int ix = (int)floorf(fx);
    int iy = (int)floorf(fy);
    float frac_x = fx - (float)ix;
    float frac_y = fy - (float)iy;

    /* Clamp neighbor indices for edge tiles */
    int ix1 = (ix + 1 < map->width) ? ix + 1 : ix;
    int iy1 = (iy + 1 < map->height) ? iy + 1 : iy;

    float h00 = (float)iso_map_get_height(map, ix, iy);
    float h10 = (float)iso_map_get_height(map, ix1, iy);
    float h01 = (float)iso_map_get_height(map, ix, iy1);
    float h11 = (float)iso_map_get_height(map, ix1, iy1);

    float top = h00 + (h10 - h00) * frac_x;
    float bot = h01 + (h11 - h01) * frac_x;
    return top + (bot - top) * frac_y;
} /* iso_map_interpolate_height() */

void iso_map_calc_transitions(const ISO_MAP* map, int mx, int my, int* edge_bits, int* corner_bits) {
    __n_assert(map, return);
    __n_assert(edge_bits, return);
    __n_assert(corner_bits, return);

    int center = iso_map_get_terrain(map, mx, my);
    *edge_bits = 0;
    *corner_bits = 0;

    /* Edge neighbors: W(-1,0), N(0,-1), E(+1,0), S(0,+1) */
    int w = iso_map_get_terrain(map, mx - 1, my);
    int n = iso_map_get_terrain(map, mx, my - 1);
    int e = iso_map_get_terrain(map, mx + 1, my);
    int s = iso_map_get_terrain(map, mx, my + 1);

    if (w != center) *edge_bits |= ISO_EDGE_W;
    if (n != center) *edge_bits |= ISO_EDGE_N;
    if (e != center) *edge_bits |= ISO_EDGE_E;
    if (s != center) *edge_bits |= ISO_EDGE_S;

    /* Corner neighbors: only set if adjacent edges don't already have that terrain */
    int nw = iso_map_get_terrain(map, mx - 1, my - 1);
    int ne = iso_map_get_terrain(map, mx + 1, my - 1);
    int se = iso_map_get_terrain(map, mx + 1, my + 1);
    int sw = iso_map_get_terrain(map, mx - 1, my + 1);

    if (nw != center && nw != w && nw != n) *corner_bits |= ISO_CORNER_NW;
    if (ne != center && ne != n && ne != e) *corner_bits |= ISO_CORNER_NE;
    if (se != center && se != e && se != s) *corner_bits |= ISO_CORNER_SE;
    if (sw != center && sw != s && sw != w) *corner_bits |= ISO_CORNER_SW;
} /* iso_map_calc_transitions() */

int iso_map_should_transition(const ISO_MAP* map, int mx1, int my1, int mx2, int my2) {
    __n_assert(map, return 0);
    return iso_map_get_terrain(map, mx1, my1) != iso_map_get_terrain(map, mx2, my2);
} /* iso_map_should_transition() */

void iso_map_corner_heights(const ISO_MAP* map, int mx, int my, float* h_n, float* h_e, float* h_s, float* h_w) {
    __n_assert(map, return);
    __n_assert(h_n, return);
    __n_assert(h_e, return);
    __n_assert(h_s, return);
    __n_assert(h_w, return);

    float c = (float)iso_map_get_height(map, mx, my);
    float nw = (float)iso_map_get_height(map, mx - 1, my - 1);
    float nn = (float)iso_map_get_height(map, mx, my - 1);
    float ne = (float)iso_map_get_height(map, mx + 1, my - 1);
    float ee = (float)iso_map_get_height(map, mx + 1, my);
    float se = (float)iso_map_get_height(map, mx + 1, my + 1);
    float ss = (float)iso_map_get_height(map, mx, my + 1);
    float sw = (float)iso_map_get_height(map, mx - 1, my + 1);
    float ww = (float)iso_map_get_height(map, mx - 1, my);

    *h_n = (c + nw + nn + ww) / 4.0f;
    *h_e = (c + nn + ne + ee) / 4.0f;
    *h_s = (c + ee + se + ss) / 4.0f;
    *h_w = (c + ss + sw + ww) / 4.0f;
} /* iso_map_corner_heights() */

int iso_map_save(const ISO_MAP* map, const char* filename) {
    __n_assert(map, return FALSE);
    __n_assert(filename, return FALSE);

    FILE* f = _iso_fopen_write(filename, "wb");
    if (!f) {
        n_log(LOG_ERR, "iso_map_save: could not open %s for writing", filename);
        return FALSE;
    }

    /* magic */
    fwrite("ISOM", 1, 4, f);

    /* header */
    const int32_t hdr[4] = {map->width, map->height, map->num_terrains, map->max_height};
    fwrite(hdr, sizeof(int32_t), 4, f);

    size_t total = (size_t)map->width * (size_t)map->height;

    /* terrain layer */
    fwrite(map->terrain, sizeof(int), total, f);
    /* heightmap layer */
    fwrite(map->heightmap, sizeof(int), total, f);
    /* ability layer */
    fwrite(map->ability, sizeof(int), total, f);

    fclose(f);
    return TRUE;
} /* iso_map_save() */

ISO_MAP* iso_map_load(const char* filename) {
    __n_assert(filename, return NULL);

    FILE* f = fopen(filename, "rb");
    if (!f) {
        n_log(LOG_ERR, "iso_map_load: could not open %s for reading", filename);
        return NULL;
    }

    /* check magic */
    char magic[4];
    if (fread(magic, 1, 4, f) != 4 ||
        magic[0] != 'I' || magic[1] != 'S' || magic[2] != 'O' || magic[3] != 'M') {
        n_log(LOG_ERR, "iso_map_load: invalid file format");
        fclose(f);
        return NULL;
    }

    int32_t hdr[4];
    if (fread(hdr, sizeof(int32_t), 4, f) != 4) {
        n_log(LOG_ERR, "iso_map_load: truncated header");
        fclose(f);
        return NULL;
    }

    ISO_MAP* map = iso_map_new(hdr[0], hdr[1], hdr[2], hdr[3]);
    if (!map) {
        fclose(f);
        return NULL;
    }

    /* each read only runs after the previous one returned everything it
     * asked for: a short read leaves the stream position indeterminate */
    size_t total = (size_t)map->width * (size_t)map->height;
    if (fread(map->terrain, sizeof(int), total, f) != total ||
        fread(map->heightmap, sizeof(int), total, f) != total ||
        fread(map->ability, sizeof(int), total, f) != total) {
        n_log(LOG_ERR, "iso_map_load: truncated data");
        iso_map_free(&map);
        fclose(f);
        return NULL;
    }

    fclose(f);
    return map;
} /* iso_map_load() */

void iso_map_randomize(ISO_MAP* map) {
    __n_assert(map, return);
    for (int y = 0; y < map->height; y++) {
        for (int x = 0; x < map->width; x++) {
            int idx = y * map->width + x;
            map->terrain[idx] = rand() % map->num_terrains;
            map->heightmap[idx] = rand() % (map->max_height + 1);
            map->ability[idx] = WALK;
        }
    }
} /* iso_map_randomize() */

static inline float _smooth_neighbor_h(const ISO_MAP* map, float h_self, int nx, int ny) {
    float nh = (float)iso_map_get_height(map, nx, ny);
    if (fabsf(h_self - nh) > (float)map->smooth_slope_max) return h_self;
    return nh;
}

void iso_map_smooth_corner_heights(const ISO_MAP* map, int mx, int my, float* h_n, float* h_e, float* h_s, float* h_w) {
    __n_assert(map, return);
    __n_assert(h_n, return);
    __n_assert(h_e, return);
    __n_assert(h_s, return);
    __n_assert(h_w, return);

    float h = (float)iso_map_get_height(map, mx, my);

    *h_n = (h + _smooth_neighbor_h(map, h, mx - 1, my) + _smooth_neighbor_h(map, h, mx, my - 1) + _smooth_neighbor_h(map, h, mx - 1, my - 1)) / 4.0f;

    *h_e = (h + _smooth_neighbor_h(map, h, mx, my - 1) + _smooth_neighbor_h(map, h, mx + 1, my) + _smooth_neighbor_h(map, h, mx + 1, my - 1)) / 4.0f;

    *h_s = (h + _smooth_neighbor_h(map, h, mx + 1, my) + _smooth_neighbor_h(map, h, mx, my + 1) + _smooth_neighbor_h(map, h, mx + 1, my + 1)) / 4.0f;

    *h_w = (h + _smooth_neighbor_h(map, h, mx - 1, my) + _smooth_neighbor_h(map, h, mx, my + 1) + _smooth_neighbor_h(map, h, mx - 1, my + 1)) / 4.0f;
} /* iso_map_smooth_corner_heights() */

/* Per-tile segment API */

int iso_map_set_segments(ISO_MAP* map, int mx, int my, const ISO_TILE_SEGMENT* segs, int count) {
    __n_assert(map, return 0);
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return 0;
    if (count < 0 || count > ISO_MAX_SEGMENTS_PER_TILE) return 0;

    /* Validate invariants */
    for (int i = 0; i < count; i++) {
        if (segs[i].bottom >= segs[i].top) return 0;
        if (i > 0 && segs[i - 1].top > segs[i].bottom) return 0;
    }

    /* Allocate segment array on first use */
    if (!map->segments) {
        size_t total = (size_t)map->width * (size_t)map->height;
        Malloc(map->segments, ISO_TILE_SEGMENTS, total);
        if (!map->segments) return 0;
        memset(map->segments, 0, total * sizeof(ISO_TILE_SEGMENTS));
    }

    int idx = my * map->width + mx;
    map->segments[idx].count = count;
    for (int i = 0; i < count; i++) map->segments[idx].segs[i] = segs[i];
    for (int i = count; i < ISO_MAX_SEGMENTS_PER_TILE; i++) {
        map->segments[idx].segs[i].bottom = 0;
        map->segments[idx].segs[i].top = 0;
        map->segments[idx].segs[i].upper_tile = -1;
        map->segments[idx].segs[i].lower_tile = -1;
    }

    /* Update derived heightmap */
    map->heightmap[idx] = (count > 0) ? segs[count - 1].top : 0;
    /* Draw_order depends on segments + heights; invalidate. */
    map->draw_order_dirty = 1;
    return 1;
}

const ISO_TILE_SEGMENTS* iso_map_get_segments(const ISO_MAP* map, int mx, int my) {
    if (!map || !map->segments) return NULL;
    if (mx < 0 || mx >= map->width || my < 0 || my >= map->height) return NULL;
    return &map->segments[my * map->width + mx];
}

static int _cmp_draw_entry(const void* a, const void* b) {
    const ISO_DRAW_ENTRY* ea = (const ISO_DRAW_ENTRY*)a;
    const ISO_DRAW_ENTRY* eb = (const ISO_DRAW_ENTRY*)b;
    /* Primary: isometric depth (back-to-front) */
    int depth_a = ea->my + ea->mx;
    int depth_b = eb->my + eb->mx;
    if (depth_a != depth_b) return depth_a - depth_b;
    /* Secondary: segment base height (lower drawn first) */
    if (ea->bottom != eb->bottom) return ea->bottom - eb->bottom;
    /* Tertiary: row then column */
    if (ea->my != eb->my) return ea->my - eb->my;
    return ea->mx - eb->mx;
}

int iso_map_build_draw_order(const ISO_MAP* map, ISO_DRAW_ENTRY* out, int max_entries) {
    __n_assert(map, return 0);
    __n_assert(out, return 0);

    int n = 0;
    for (int my = 0; my < map->height; my++) {
        for (int mx = 0; mx < map->width; mx++) {
            int idx = my * map->width + mx;
            if (map->segments && map->segments[idx].count > 0) {
                const ISO_TILE_SEGMENTS* ts = &map->segments[idx];
                for (int si = 0; si < ts->count && n < max_entries; si++) {
                    out[n].mx = mx;
                    out[n].my = my;
                    out[n].seg_idx = si;
                    out[n].bottom = ts->segs[si].bottom;
                    out[n].top = ts->segs[si].top;
                    /* underside: bottom > 0 and no segment directly below */
                    out[n].underside = (ts->segs[si].bottom > 0 &&
                                        (si == 0 || ts->segs[si - 1].top < ts->segs[si].bottom));
                    n++;
                }
            } else {
                /* No segments or count==0: single entry from heightmap */
                if (n >= max_entries) continue;
                int h = map->heightmap[idx];
                out[n].mx = mx;
                out[n].my = my;
                out[n].seg_idx = 0;
                out[n].bottom = 0;
                out[n].top = h;
                out[n].underside = 0;
                n++;
            }
        }
    }

    qsort(out, (size_t)n, sizeof(ISO_DRAW_ENTRY), _cmp_draw_entry);
    return n;
}

int iso_map_should_transition_smooth(const ISO_MAP* map, int mx1, int my1, int mx2, int my2) {
    __n_assert(map, return 0);
    int h1 = iso_map_get_height(map, mx1, my1);
    int h2 = iso_map_get_height(map, mx2, my2);
    if (map->smooth_height)
        return abs(h1 - h2) <= map->smooth_slope_max;
    return h1 == h2;
} /* iso_map_should_transition_smooth() */

static int _neighbor_terrain_at(const ISO_MAP* map, int mx, int my) {
    if (my >= 0 && my < map->height) {
        if (mx == -1 && map->neighbor_terrains_west)
            return map->neighbor_terrains_west[my];
        if (mx == map->width && map->neighbor_terrains_east)
            return map->neighbor_terrains_east[my];
    }
    if (mx >= 0 && mx < map->width) {
        if (my == -1 && map->neighbor_terrains_north)
            return map->neighbor_terrains_north[mx];
        if (my == map->height && map->neighbor_terrains_south)
            return map->neighbor_terrains_south[mx];
    }
    return iso_map_get_terrain(map, mx, my);
}

static int _neighbor_height_at(const ISO_MAP* map, int mx, int my) {
    if (my >= 0 && my < map->height) {
        if (mx == -1 && map->neighbor_heights_west)
            return map->neighbor_heights_west[my];
        if (mx == map->width && map->neighbor_heights_east)
            return map->neighbor_heights_east[my];
    }
    if (mx >= 0 && mx < map->width) {
        if (my == -1 && map->neighbor_heights_north)
            return map->neighbor_heights_north[mx];
        if (my == map->height && map->neighbor_heights_south)
            return map->neighbor_heights_south[mx];
    }
    return iso_map_get_height(map, mx, my);
}

static int _neighbor_should_transition(const ISO_MAP* map, int mx1, int my1, int mx2, int my2) {
    int h1 = _neighbor_height_at(map, mx1, my1);
    int h2 = _neighbor_height_at(map, mx2, my2);
    if (map->smooth_height)
        return abs(h1 - h2) <= map->smooth_slope_max;
    return h1 == h2;
}

void iso_map_calc_transitions_full(const ISO_MAP* map, int mx, int my, int* edge_bits, int* corner_bits) {
    __n_assert(map, return);
    __n_assert(edge_bits, return);
    __n_assert(corner_bits, return);

    int base = iso_map_get_terrain(map, mx, my);
    memset(edge_bits, 0, sizeof(int) * (size_t)map->num_terrains);
    memset(corner_bits, 0, sizeof(int) * (size_t)map->num_terrains);

    /* Cardinal neighbors: only if height-connected. Uses the
     * _neighbor_* helpers so chunk-edge lookups consult the
     * neighbor_terrains_* / neighbor_heights_* arrays instead of
     * falling to the default 0-out-of-bounds and dropping the
     * cross-chunk transition. */
    int t_w = _neighbor_should_transition(map, mx, my, mx - 1, my)
                  ? _neighbor_terrain_at(map, mx - 1, my)
                  : base;
    int t_n = _neighbor_should_transition(map, mx, my, mx, my - 1)
                  ? _neighbor_terrain_at(map, mx, my - 1)
                  : base;
    int t_e = _neighbor_should_transition(map, mx, my, mx + 1, my)
                  ? _neighbor_terrain_at(map, mx + 1, my)
                  : base;
    int t_s = _neighbor_should_transition(map, mx, my, mx, my + 1)
                  ? _neighbor_terrain_at(map, mx, my + 1)
                  : base;

    /* Diagonal neighbors, see _neighbor_terrain_at comment. Diagonals
     * that fall off the chunk still return 0 (water); the visible
     * fallout is single-tile corner artifacts at chunk boundaries
     * rather than whole-edge seams. */
    int t_nw = _neighbor_should_transition(map, mx, my, mx - 1, my - 1)
                   ? _neighbor_terrain_at(map, mx - 1, my - 1)
                   : base;
    int t_ne = _neighbor_should_transition(map, mx, my, mx + 1, my - 1)
                   ? _neighbor_terrain_at(map, mx + 1, my - 1)
                   : base;
    int t_se = _neighbor_should_transition(map, mx, my, mx + 1, my + 1)
                   ? _neighbor_terrain_at(map, mx + 1, my + 1)
                   : base;
    int t_sw = _neighbor_should_transition(map, mx, my, mx - 1, my + 1)
                   ? _neighbor_terrain_at(map, mx - 1, my + 1)
                   : base;

    for (int t = base + 1; t < map->num_terrains; t++) {
        if (t_w == t) edge_bits[t] |= ISO_EDGE_W;
        if (t_n == t) edge_bits[t] |= ISO_EDGE_N;
        if (t_e == t) edge_bits[t] |= ISO_EDGE_E;
        if (t_s == t) edge_bits[t] |= ISO_EDGE_S;

        if (t_nw == t && !(edge_bits[t] & ISO_EDGE_W) && !(edge_bits[t] & ISO_EDGE_N))
            corner_bits[t] |= ISO_CORNER_NW;
        if (t_ne == t && !(edge_bits[t] & ISO_EDGE_N) && !(edge_bits[t] & ISO_EDGE_E))
            corner_bits[t] |= ISO_CORNER_NE;
        if (t_se == t && !(edge_bits[t] & ISO_EDGE_E) && !(edge_bits[t] & ISO_EDGE_S))
            corner_bits[t] |= ISO_CORNER_SE;
        if (t_sw == t && !(edge_bits[t] & ISO_EDGE_S) && !(edge_bits[t] & ISO_EDGE_W))
            corner_bits[t] |= ISO_CORNER_SW;
    }
} /* iso_map_calc_transitions_full() */

/* N_ISO_CAMERA */

N_ISO_CAMERA* n_iso_camera_new(float zoom_min, float zoom_max) {
    N_ISO_CAMERA* cam = NULL;
    Malloc(cam, N_ISO_CAMERA, 1);
    __n_assert(cam, return NULL);
    cam->x = 0.0f;
    cam->y = 0.0f;
    cam->zoom = 1.0f;
    cam->zoom_min = zoom_min;
    cam->zoom_max = zoom_max;
    return cam;
} /* n_iso_camera_new() */

void n_iso_camera_free(N_ISO_CAMERA** cam) {
    __n_assert(cam && *cam, return);
    Free(*cam);
} /* n_iso_camera_free() */

void n_iso_camera_scroll(N_ISO_CAMERA* cam, float dx, float dy) {
    __n_assert(cam, return);
    cam->x += dx;
    cam->y += dy;
} /* n_iso_camera_scroll() */

void n_iso_camera_zoom(N_ISO_CAMERA* cam, float dz, float mouse_x, float mouse_y) {
    __n_assert(cam, return);
    float old_zoom = cam->zoom;
    cam->zoom += dz;
    if (cam->zoom < cam->zoom_min) cam->zoom = cam->zoom_min;
    if (cam->zoom > cam->zoom_max) cam->zoom = cam->zoom_max;
    /* Adjust position so the world point under the mouse stays fixed */
    cam->x += mouse_x / cam->zoom - mouse_x / old_zoom;
    cam->y += mouse_y / cam->zoom - mouse_y / old_zoom;
} /* n_iso_camera_zoom() */

void n_iso_camera_center_on(N_ISO_CAMERA* cam, float world_x, float world_y, int screen_w, int screen_h) {
    __n_assert(cam, return);
    cam->x = -world_x + (float)screen_w / (2.0f * cam->zoom);
    cam->y = -world_y + (float)screen_h / (2.0f * cam->zoom);
} /* n_iso_camera_center_on() */

void n_iso_camera_follow(N_ISO_CAMERA* cam, float target_x, float target_y, int screen_w, int screen_h, float smoothing, float dt) {
    __n_assert(cam, return);
    float target_cx = -target_x + (float)screen_w / (2.0f * cam->zoom);
    float target_cy = -target_y + (float)screen_h / (2.0f * cam->zoom);
    float t = smoothing * dt;
    if (t > 1.0f) t = 1.0f;
    cam->x += (target_cx - cam->x) * t;
    cam->y += (target_cy - cam->y) * t;
} /* n_iso_camera_follow() */

void n_iso_camera_screen_to_world(const N_ISO_CAMERA* cam, float sx, float sy, float* wx, float* wy) {
    __n_assert(cam, return);
    __n_assert(wx, return);
    __n_assert(wy, return);
    *wx = sx / cam->zoom - cam->x;
    *wy = sy / cam->zoom - cam->y;
} /* n_iso_camera_screen_to_world() */

void n_iso_camera_world_to_screen(const N_ISO_CAMERA* cam, float wx, float wy, float* sx, float* sy) {
    __n_assert(cam, return);
    __n_assert(sx, return);
    __n_assert(sy, return);
    *sx = (wx + cam->x) * cam->zoom;
    *sy = (wy + cam->y) * cam->zoom;
} /* n_iso_camera_world_to_screen() */

/* Height border edge visibility (public API, no Allegro dependency) */

ISO_VISIBLE_EDGES iso_map_get_visible_edges(const ISO_MAP* map, int mx, int my) {
    ISO_VISIBLE_EDGES edges = {0};
    if (!map) return edges;

    int h = iso_map_get_height(map, mx, my);
    if (h == 0) return edges;

    /* NW edge: neighbor (mx-1, my) */
    if (mx <= 0) {
        /* At west boundary: use neighbor chunk data if available */
        if (map->neighbor_heights_west) {
            if (abs(h - map->neighbor_heights_west[my]) >= 1)
                edges.draw_nw = 1;
        } else {
            edges.draw_nw = 1; /* world boundary */
        }
    } else if (abs(h - iso_map_get_height(map, mx - 1, my)) >= 1) {
        edges.draw_nw = 1;
    }

    /* NE edge: neighbor (mx, my-1) */
    if (my <= 0) {
        if (map->neighbor_heights_north) {
            if (abs(h - map->neighbor_heights_north[mx]) >= 1)
                edges.draw_ne = 1;
        } else {
            edges.draw_ne = 1;
        }
    } else if (abs(h - iso_map_get_height(map, mx, my - 1)) >= 1) {
        edges.draw_ne = 1;
    }

    /* SE edge: neighbor (mx+1, my) */
    if (mx >= map->width - 1) {
        if (map->neighbor_heights_east) {
            if (abs(h - map->neighbor_heights_east[my]) >= 1)
                edges.draw_se = 1;
        } else {
            edges.draw_se = 1;
        }
    } else if (abs(h - iso_map_get_height(map, mx + 1, my)) >= 1) {
        edges.draw_se = 1;
    }

    /* SW edge: neighbor (mx, my+1) */
    if (my >= map->height - 1) {
        if (map->neighbor_heights_south) {
            if (abs(h - map->neighbor_heights_south[mx]) >= 1)
                edges.draw_sw = 1;
        } else {
            edges.draw_sw = 1;
        }
    } else if (abs(h - iso_map_get_height(map, mx, my + 1)) >= 1) {
        edges.draw_sw = 1;
    }

    return edges;
}

static int _tile_covers_height(const ISO_MAP* map, int mx, int my, int z) {
    if (!map || mx < 0 || mx >= map->width || my < 0 || my >= map->height)
        return 0;
    const ISO_TILE_SEGMENTS* ts = iso_map_get_segments(map, mx, my);
    if (ts && ts->count > 0) {
        for (int i = 0; i < ts->count; i++) {
            if (ts->segs[i].bottom <= z && ts->segs[i].top > z + 1)
                return 1;
        }
        return 0;
    }
    int h = iso_map_get_height(map, mx, my);
    return (h > z + 1) ? 1 : 0;
}

ISO_VISIBLE_EDGES iso_map_get_visible_edges_segment(const ISO_MAP* map, int mx, int my, int seg_idx) {
    ISO_VISIBLE_EDGES edges = {0};
    if (!map) return edges;

    /* Get this tile's segment */
    ISO_TILE_SEGMENT seg;
    const ISO_TILE_SEGMENTS* ts = iso_map_get_segments(map, mx, my);
    if (ts && seg_idx >= 0 && seg_idx < ts->count) {
        seg = ts->segs[seg_idx];
    } else {
        /* Fallback: heightmap as single segment {0, h} */
        int h = iso_map_get_height(map, mx, my);
        if (h == 0) return edges;
        seg.bottom = 0;
        seg.top = h;
    }

    if (seg.top <= seg.bottom) return edges;

    /* Direction offsets: NW->(mx-1,my), NE->(mx,my-1), SE->(mx+1,my), SW->(mx,my+1) */
    static const int ddx[4] = {-1, 0, 1, 0};
    static const int ddy[4] = {0, -1, 0, 1};
    int exposed[4] = {0, 0, 0, 0}; /* edge to draw, per direction */

    int any_side_exposed = 0;   /* for underside check */
    int ntop[4] = {0, 0, 0, 0}; /* neighbor top height per direction */

    for (int d = 0; d < 4; d++) {
        int nx = mx + ddx[d];
        int ny = my + ddy[d];
        int neighbor_matches = 0;
        int neighbor_covers_bottom = 0;

        if (nx < 0 || nx >= map->width || ny < 0 || ny >= map->height) {
            /* Boundary: check neighbor chunk height arrays */
            int nh = -1;
            if (d == 0 && map->neighbor_heights_west) nh = map->neighbor_heights_west[my];
            if (d == 1 && map->neighbor_heights_north) nh = map->neighbor_heights_north[mx];
            if (d == 2 && map->neighbor_heights_east) nh = map->neighbor_heights_east[my];
            if (d == 3 && map->neighbor_heights_south) nh = map->neighbor_heights_south[mx];

            if (nh > 0) {
                /* Boundary treated as single segment {0,nh}: match if same top */
                neighbor_matches = (nh == seg.top);
                neighbor_covers_bottom = (nh >= seg.bottom);
                ntop[d] = nh;
            }
            /* else: world boundary -> no match, not covered, ntop stays 0 */
        } else {
            const ISO_TILE_SEGMENTS* nts = iso_map_get_segments(map, nx, ny);
            if (nts && nts->count > 0) {
                /* Match if any neighbour segment has the same top height.
                 * This applies uniformly to ground and floating segments:
                 * same top = connected surface = no border.
                 * Different top = visible height step = border drawn. */
                for (int si = 0; si < nts->count; si++) {
                    if (nts->segs[si].top == seg.top) {
                        neighbor_matches = 1;
                    }
                    if (nts->segs[si].top >= seg.bottom) {
                        neighbor_covers_bottom = 1;
                    }
                }
                /* Use heightmap value as neighbor top for vertical lines */
                ntop[d] = iso_map_get_height(map, nx, ny);
            } else {
                /* No segment data: heightmap as single segment {0, h} */
                int nh = iso_map_get_height(map, nx, ny);
                if (nh > 0) {
                    neighbor_matches = (nh == seg.top);
                    neighbor_covers_bottom = (nh >= seg.bottom);
                }
                ntop[d] = nh;
            }
        }

        if (!neighbor_matches) exposed[d] = 1;
        if (!neighbor_covers_bottom) any_side_exposed = 1;
    }
    edges.draw_nw = exposed[0];
    edges.draw_ne = exposed[1];
    edges.draw_se = exposed[2];
    edges.draw_sw = exposed[3];

    /* OCCLUSION: NW/NE edges face away from camera.
     * Suppress if the tile drawn later (south in painter order) has
     * geometry covering this segment's top face height.
     * NW edge: covered by SW neighbour (mx, my+1)
     * NE edge: covered by SE neighbour (mx+1, my) */
    if (edges.draw_nw && _tile_covers_height(map, mx, my + 1, seg.top))
        edges.draw_nw = 0;
    if (edges.draw_ne && _tile_covers_height(map, mx + 1, my, seg.top))
        edges.draw_ne = 0;

    /* Underside border for floating segments */
    if (seg.bottom > 0 && any_side_exposed) {
        edges.draw_underside = 1;
    }

    /* Store neighbor top heights for vertical corner line computation */
    edges.neighbor_top_nw = ntop[0];
    edges.neighbor_top_ne = ntop[1];
    edges.neighbor_top_se = ntop[2];
    edges.neighbor_top_sw = ntop[3];

    return edges;
}

/* ISO_MAP rendering (requires Allegro 5) */
#ifdef HAVE_ALLEGRO

#include <allegro5/allegro.h>
#include <allegro5/allegro_primitives.h>

static inline void _iso_corner_to_screen(float hw, float hh, float tl, int cx, int cy, float fh, float cam_px, float cam_py, float zoom, float* sx, float* sy) {
    float wx = (float)(cx - cy) * hw + 2.0f * hw;
    float wy = (float)(cx + cy) * hh - fh * tl;
    *sx = wx * zoom + cam_px;
    *sy = wy * zoom + cam_py;
}

/* Extended tint: combines height tint with ambient color and per-tile dynamic light.
 * ambient_r/g/b: global day/night ambient (1.0 = full daylight).
 * dyn_r/g/b: additive dynamic light contribution at this tile (0.0 = no light). */
static inline ALLEGRO_COLOR _compute_tile_tint(int seg_top, int max_height, float intensity, float ambient_r, float ambient_g, float ambient_b, float dyn_r, float dyn_g, float dyn_b) {
    /* Start with height-based brightness */
    float hb = 1.0f;
    if (intensity > 0.0f && max_height > 0) {
        float ratio = (float)seg_top / (float)max_height;
        hb = 1.0f - intensity * (1.0f - ratio);
        if (hb < 0.0f) hb = 0.0f;
        if (hb > 1.0f) hb = 1.0f;
    }

    /* Multiply by ambient color */
    float r = hb * ambient_r;
    float g = hb * ambient_g;
    float b = hb * ambient_b;

    /* Add dynamic light (additive boost) */
    r += dyn_r;
    g += dyn_g;
    b += dyn_b;

    /* Clamp to avoid overexposure */
    if (r > 1.0f) r = 1.0f;
    if (g > 1.0f) g = 1.0f;
    if (b > 1.0f) b = 1.0f;

    /* The tint multiplier is sRGB-designed (ambient curve + dyn-light
     * literals tuned on a non-sRGB display). Route through the primitive
     * color filter so an sRGB-aware consumer can linearise it before it
     * multiplies the linear-decoded texel inside an sRGB FBO. Default
     * filter = identity, so the non-sRGB path is byte-identical. */
    return iso_apply_primitive_color_filter(al_map_rgba_f(r, g, b, 1.0f));
}

static void _draw_tile_warped(ALLEGRO_BITMAP* bmp,
                              const float vx[4],
                              const float vy[4],
                              int tile_w,
                              int tile_h,
                              ALLEGRO_COLOR tint) {
    /* al_draw_prim samples u/v against the PARENT texture and does NOT
     * auto-translate for a sub-bitmap (unlike al_draw_*_bitmap). When tiles
     * are packed into a terrain atlas (each tile an al_create_sub_bitmap),
     * passing sub-local UVs [0..tile_w] samples the atlas ORIGIN, which is
     * transparent padding, so warped (sloped) tiles render BLACK. Resolve
     * the parent + add the sub-bitmap's atlas offset to every UV and draw
     * against the parent so the warp samples the tile's real atlas region.
     * For a standalone (non-sub) bitmap the offsets are 0 and parent == bmp,
     * so this is byte-identical to the pre-atlas behaviour. */
    ALLEGRO_BITMAP* tex = al_get_parent_bitmap(bmp);
    float ox, oy;
    if (tex) {
        ox = (float)al_get_bitmap_x(bmp);
        oy = (float)al_get_bitmap_y(bmp);
    } else {
        tex = bmp;
        ox = 0.0f;
        oy = 0.0f;
    }
    float un = ox + (float)tile_w / 2.0f, vn = oy + 0.0f;
    float ue = ox + (float)tile_w, ve = oy + (float)tile_h / 2.0f;
    float us = ox + (float)tile_w / 2.0f, vs = oy + (float)tile_h;
    float uw = ox + 0.0f, vw = oy + (float)tile_h / 2.0f;
    ALLEGRO_VERTEX vtx[6];
    vtx[0].x = vx[0];
    vtx[0].y = vy[0];
    vtx[0].z = 0;
    vtx[0].u = un;
    vtx[0].v = vn;
    vtx[0].color = tint;
    vtx[1].x = vx[1];
    vtx[1].y = vy[1];
    vtx[1].z = 0;
    vtx[1].u = ue;
    vtx[1].v = ve;
    vtx[1].color = tint;
    vtx[2].x = vx[2];
    vtx[2].y = vy[2];
    vtx[2].z = 0;
    vtx[2].u = us;
    vtx[2].v = vs;
    vtx[2].color = tint;
    vtx[3].x = vx[0];
    vtx[3].y = vy[0];
    vtx[3].z = 0;
    vtx[3].u = un;
    vtx[3].v = vn;
    vtx[3].color = tint;
    vtx[4].x = vx[2];
    vtx[4].y = vy[2];
    vtx[4].z = 0;
    vtx[4].u = us;
    vtx[4].v = vs;
    vtx[4].color = tint;
    vtx[5].x = vx[3];
    vtx[5].y = vy[3];
    vtx[5].z = 0;
    vtx[5].u = uw;
    vtx[5].v = vw;
    vtx[5].color = tint;

    al_draw_prim(vtx, NULL, tex, 0, 6, ALLEGRO_PRIM_TRIANGLE_LIST);
}

static void _draw_segment_sides(const ISO_MAP* map,
                                float east_x,
                                float east_y,
                                float south_x,
                                float south_y,
                                float west_x,
                                float west_y,
                                int mx,
                                int my,
                                int seg_bottom,
                                int seg_top,
                                float lift,
                                float height_brightness,
                                int wall_terrain_override) {
    int base = (wall_terrain_override >= 0) ? wall_terrain_override
                                            : iso_map_get_terrain(map, mx, my);
    float east_shade = 0.55f;
    float south_shade = 0.35f;

    float cr, cg, cb;
    switch (base) {
        case 0:
            cr = 0.7f;
            cg = 0.25f;
            cb = 0.1f;
            break;
        case 1:
            cr = 0.5f;
            cg = 0.45f;
            cb = 0.35f;
            break;
        case 2:
            cr = 0.6f;
            cg = 0.55f;
            cb = 0.45f;
            break;
        default:
            cr = 0.4f;
            cg = 0.35f;
            cb = 0.3f;
            break;
    }
    ALLEGRO_COLOR east_col = iso_apply_primitive_color_filter(
        al_map_rgba_f(cr * east_shade * height_brightness,
                      cg * east_shade * height_brightness,
                      cb * east_shade * height_brightness, 1.0f));
    ALLEGRO_COLOR south_col = iso_apply_primitive_color_filter(
        al_map_rgba_f(cr * south_shade * height_brightness,
                      cg * south_shade * height_brightness,
                      cb * south_shade * height_brightness, 1.0f));
    /* Keep PASS 1 on ONE blender throughout to avoid held-batch
     * cross-tile contamination. Cliff wall fills use alpha=1.0 ->
     * premul == straight, and edge lines are RGB=0 -> premul ==
     * straight, so this switch is a visual no-op. */
    al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);

    int east_face_h, south_face_h;
    if (seg_bottom == 0) {
        /* Ground segment: use neighbor-aware wall height (existing logic) */
        int h_e = iso_map_get_height(map, mx + 1, my);
        int h_s = iso_map_get_height(map, mx, my + 1);
        int h_se = iso_map_get_height(map, mx + 1, my + 1);

        east_face_h = seg_top - (h_e < h_se ? h_e : h_se);
        // cppcheck-suppress redundantAssignment ; fallback chain: min(h_e,h_se) -> h_e -> h_se
        if (east_face_h < 1) east_face_h = seg_top - h_e;
        if (east_face_h < 1) east_face_h = seg_top - h_se;

        south_face_h = seg_top - (h_s < h_se ? h_s : h_se);
        // cppcheck-suppress redundantAssignment ; fallback chain: min(h_s,h_se) -> h_s -> h_se
        if (south_face_h < 1) south_face_h = seg_top - h_s;
        if (south_face_h < 1) south_face_h = seg_top - h_se;
    } else {
        /* Floating segment: full wall from bottom to top */
        east_face_h = seg_top - seg_bottom;
        south_face_h = seg_top - seg_bottom;
    }

    /* No sub-pixel expansion needed: all vertices come from the canonical
     * _iso_corner_to_screen formula, so adjacent tiles sharing an edge
     * produce bit-identical vertex positions. */

    /* Query the client-registered jitter source. Returns (0, 0) when
     * no jitter source is active. When non-zero, we subtract it from
     * every line endpoint so the line lands at the same screen pixel
     * each frame regardless of where in the TAA jitter cycle we are.
     * Filled polygon faces deliberately KEEP the jitter: they are
     * the larger draws TAA's history blend smooths perceptually. */
    float jx = 0.0f, jy = 0.0f;
    iso_apply_line_jitter_compensation(&jx, &jy);

    if (east_face_h > 0) {
        float drop = (float)east_face_h * lift;
        float v[8] = {east_x, east_y, south_x, south_y,
                      south_x, south_y + drop,
                      east_x, east_y + drop};
        al_draw_filled_polygon(v, 4, east_col);
        ALLEGRO_COLOR edge = al_map_rgba(0, 0, 0, 80);
        al_draw_line(east_x - jx, east_y + drop - jy,
                     south_x - jx, south_y + drop - jy,
                     edge, 1.0f);
    }

    if (south_face_h > 0) {
        float drop = (float)south_face_h * lift;
        float v[8] = {south_x, south_y, west_x, west_y,
                      west_x, west_y + drop,
                      south_x, south_y + drop};
        al_draw_filled_polygon(v, 4, south_col);
        ALLEGRO_COLOR edge = al_map_rgba(0, 0, 0, 80);
        al_draw_line(west_x - jx, west_y + drop - jy,
                     south_x - jx, south_y + drop - jy,
                     edge, 1.0f);
    }
}

static void _draw_segment_underside(ALLEGRO_BITMAP* bmp,
                                    float base_dx,
                                    float base_dy,
                                    float phw,
                                    float phh,
                                    float tile_draw_w,
                                    float tile_draw_h,
                                    int tile_w,
                                    int tile_h) {
    al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
    /* Draw the tile bitmap with a darkened tint (65% brightness) */
    ALLEGRO_COLOR tint = al_map_rgba_f(0.65f, 0.65f, 0.65f, 1.0f);
    (void)phw;
    (void)phh;
    al_draw_tinted_scaled_bitmap(bmp, tint,
                                 0, 0, (float)tile_w, (float)tile_h,
                                 base_dx, base_dy, tile_draw_w, tile_draw_h, 0);
}

static int _iso_object_check_occlusion(const ISO_MAP* map,
                                       float obj_fx,
                                       float obj_fy,
                                       float obj_fz,
                                       float obj_sx,
                                       float obj_sy,
                                       float sprite_h,
                                       float sprite_half_w,
                                       float cam_px,
                                       float cam_py,
                                       float zoom,
                                       float* out_clip_y) {
    /* Hoist projection constants out of the inner loop. */
    const float proj_hw = map->proj.half_w;
    const float proj_hh = map->proj.half_h;
    const float proj_lift = map->proj.tile_lift;

    float phw = proj_hw * zoom;
    float phh = proj_hh * zoom;
    float lift = proj_lift * zoom;

    /* Entity head screen Y: sprite_h world units above feet */
    float head_sy = obj_sy - sprite_h * lift;
    /* Entity horizontal extent (screen pixels) */
    float ent_left = obj_sx - sprite_half_w * zoom;
    float ent_right = obj_sx + sprite_half_w * zoom;

    int base_mx = (int)floorf(obj_fx);
    int base_my = (int)floorf(obj_fy);
    int check_range = map->max_height + 2;
    int found = 0;
    float best_clip_y = 1e9f;

    for (int dy = -1; dy <= check_range; dy++) {
        for (int dx = -1; dx <= check_range; dx++) {
            /* Only check tiles drawn after the object in painter's order.
             * Filter by isometric depth (dx+dy) not just grid position. */
            int tdepth = dx + dy;
            if (tdepth < 0) continue;
            if (tdepth == 0 && dy < 0) continue;
            if (tdepth == 0 && dy == 0 && dx <= 0) continue;

            int tmx = base_mx + dx;
            int tmy = base_my + dy;
            if (tmx < 0 || tmx >= map->width || tmy < 0 || tmy >= map->height) continue;

            int th = iso_map_get_height(map, tmx, tmy);
            /* tile must reach above the entity's feet to potentially occlude */
            if ((float)th <= obj_fz) continue;

            /* inlined iso_map_to_screen(map, tmx, tmy, th, &tsx, &tsy) */
            float tsx = ((float)tmx - (float)tmy) * proj_hw + proj_hw;
            float tsy = ((float)tmx + (float)tmy) * proj_hh - (float)th * proj_lift;
            float tile_dx = tsx * zoom + cam_px;
            float tile_dy = tsy * zoom + cam_py;

            /* tile screen bounding box */
            float tile_left = tile_dx;
            float tile_right = tile_dx + 2.0f * phw;
            float tile_top = tile_dy;
            float tile_bot = tile_dy + 2.0f * phh + (float)th * lift;

            /* Check horizontal overlap (entity range vs tile range) */
            if (ent_right < tile_left || ent_left > tile_right) continue;

            /* Check vertical overlap (any part of entity behind tile) */
            if (tile_top <= obj_sy && tile_bot >= head_sy) {
                found = 1;
                /* Track north vertex of occluding tile for clip-Y */
                float nx, ny;
                _iso_corner_to_screen(proj_hw, proj_hh, proj_lift,
                                      tmx, tmy, (float)th,
                                      cam_px, cam_py, zoom, &nx, &ny);
                if (ny < best_clip_y) best_clip_y = ny;
            }
        }
    }
    if (out_clip_y) *out_clip_y = best_clip_y;
    return found;
}

static void _draw_height_borders(const ISO_MAP* map,
                                 int mx,
                                 int my,
                                 int seg_idx,
                                 int seg_bottom,
                                 int seg_top,
                                 int has_underside_face,
                                 const float vx[4],
                                 const float vy[4],
                                 float lift) {
    ISO_VISIBLE_EDGES edges = iso_map_get_visible_edges_segment(map, mx, my, seg_idx);
    if (!edges.draw_nw && !edges.draw_ne && !edges.draw_se && !edges.draw_sw && !(has_underside_face && edges.draw_underside))
        return;

    /* Subtract the client-registered jitter (e.g. the consumer's TAA
     * sub-pixel offset) from every line endpoint so these decorative
     * borders stay pixel-stable across frames while sprites and
     * filled faces continue to inherit the jitter. Returns (0, 0)
     * when no jitter source is active. */
    float jx = 0.0f, jy = 0.0f;
    iso_apply_line_jitter_compensation(&jx, &jy);

    ALLEGRO_COLOR border_col = al_map_rgba(0, 0, 0, 200);
    float thickness = 1.5f;

    /* Top face edges */
    if (edges.draw_se)
        al_draw_line(vx[1] - jx, vy[1] - jy,
                     vx[2] - jx, vy[2] - jy, border_col, thickness);
    if (edges.draw_sw)
        al_draw_line(vx[2] - jx, vy[2] - jy,
                     vx[3] - jx, vy[3] - jy, border_col, thickness);
    if (edges.draw_nw)
        al_draw_line(vx[0] - jx, vy[0] - jy,
                     vx[3] - jx, vy[3] - jy, border_col, thickness);
    if (edges.draw_ne)
        al_draw_line(vx[0] - jx, vy[0] - jy,
                     vx[1] - jx, vy[1] - jy, border_col, thickness);

    /* Vertical corner lines.
     *
     * For each drawn edge where seg_top > neighbor_top (A is higher):
     *   Draw a vertical line at BOTH endpoints of that edge,
     *   from tileToScreen(vertex, seg_top) down to
     *   tileToScreen(vertex, neighbor_top).
     *   Exception: NEVER draw at the North vertex (vx[0]/vy[0]).
     *
     * If seg_top <= neighbor_top: no vertical lines (wall belongs to N).
     *
     * vx[]/vy[] are already at tileToScreen(vertex, seg_top).
     * The bottom endpoint is (seg_top - HN) * lift pixels below.
     *
     * Diamond vertices: N=0, E=1, S=2, W=3.
     * Edge endpoints: NE=(N,E) SE=(E,S) SW=(S,W) NW=(W,N)
     *
     * Multiple edges may want to draw at the same vertex with different
     * drops.  We draw the longest line (smallest HN) at each vertex
     * to avoid duplicate shorter lines.
     */
    {
        /* Per-vertex: track the largest drop (longest line) requested.
         *
         * CONTINUITY CHECK: a vertical corner line at a vertex is only
         * drawn if the wall face ENDS at that vertex, i.e. the adjacent
         * tile in the wall's direction is NOT at the same height.
         * If the adjacent tile IS at the same height, the wall surface
         * continues past the vertex and the line is interior (hidden).
         *
         * For the east-facing wall (SE edge):
         *   E vertex: wall continues northward if NE neighbor same height -> skip
         *   S vertex: wall continues southward if SW neighbor same height -> skip
         * For the south-facing wall (SW edge):
         *   S vertex: wall continues eastward if SE neighbor same height -> skip
         *   W vertex: wall continues westward if NW neighbor same height -> skip
         * For NE edge -> E vertex: wall continues if SE neighbor same height
         * For NW edge -> W vertex: wall continues if SW neighbor same height
         */
        float drop_e = 0, drop_s = 0, drop_w = 0;
        int nt_nw = edges.neighbor_top_nw;
        int nt_ne = edges.neighbor_top_ne;
        int nt_se = edges.neighbor_top_se;
        int nt_sw = edges.neighbor_top_sw;

        /* Clamp neighbor tops to seg_bottom: vertical lines must not
         * extend below the segment's own base (floating segments). */
        int bot = seg_bottom;
        int cl_nw = nt_nw > bot ? nt_nw : bot;
        int cl_ne = nt_ne > bot ? nt_ne : bot;
        int cl_se = nt_se > bot ? nt_se : bot;
        int cl_sw = nt_sw > bot ? nt_sw : bot;

        /* NE edge (N,E): only E vertex (N is never drawn) */
        if (edges.draw_ne && seg_top > cl_ne) {
            /* E vertex: east wall continues if SE neighbor same height */
            if (nt_se != seg_top) {
                float d = (float)(seg_top - cl_ne) * lift;
                if (d > drop_e) drop_e = d;
            }
        }
        /* SE edge (E,S): E and S vertices */
        if (edges.draw_se && seg_top > cl_se) {
            float d = (float)(seg_top - cl_se) * lift;
            /* E vertex: east wall continues northward if NE neighbor same height */
            if (nt_ne != seg_top) {
                if (d > drop_e) drop_e = d;
            }
            /* S vertex: east wall continues southward if SW neighbor same height */
            if (nt_sw != seg_top) {
                if (d > drop_s) drop_s = d;
            }
        }
        /* SW edge (S,W): S and W vertices */
        if (edges.draw_sw && seg_top > cl_sw) {
            float d = (float)(seg_top - cl_sw) * lift;
            /* S vertex: south wall continues eastward if SE neighbor same height */
            if (nt_se != seg_top) {
                if (d > drop_s) drop_s = d;
            }
            /* W vertex: south wall continues westward if NW neighbor same height */
            if (nt_nw != seg_top) {
                if (d > drop_w) drop_w = d;
            }
        }
        /* NW edge (W,N): only W vertex (N is never drawn) */
        if (edges.draw_nw && seg_top > cl_nw) {
            /* W vertex: south wall continues if SW neighbor same height */
            if (nt_sw != seg_top) {
                float d = (float)(seg_top - cl_nw) * lift;
                if (d > drop_w) drop_w = d;
            }
        }

        if (drop_e > 0)
            al_draw_line(vx[1] - jx, vy[1] - jy,
                         vx[1] - jx, vy[1] + drop_e - jy,
                         border_col, thickness);
        if (drop_s > 0)
            al_draw_line(vx[2] - jx, vy[2] - jy,
                         vx[2] - jx, vy[2] + drop_s - jy,
                         border_col, thickness);
        if (drop_w > 0)
            al_draw_line(vx[3] - jx, vy[3] - jy,
                         vx[3] - jx, vy[3] + drop_w - jy,
                         border_col, thickness);
    }

    /* Underside face border (floating segments only).
     * The underside diamond is at seg_bottom, offset below the top face
     * by (seg_top - seg_bottom) * lift pixels.  Only drawn if the
     * underside face itself was rendered (has_underside_face).
     * Only SE and SW edges, NW and NE are always hidden behind the
     * segment's own side walls from the camera angle. */
    if (has_underside_face && edges.draw_underside) {
        float dy = (float)(seg_top - seg_bottom) * lift;
        /* SE edge */
        al_draw_line(vx[1] - jx, vy[1] + dy - jy,
                     vx[2] - jx, vy[2] + dy - jy,
                     border_col, thickness);
        /* SW edge */
        al_draw_line(vx[2] - jx, vy[2] + dy - jy,
                     vx[3] - jx, vy[3] + dy - jy,
                     border_col, thickness);
    }
}

void iso_map_draw(const ISO_MAP* map,
                  ALLEGRO_BITMAP** tile_bitmaps,
                  ALLEGRO_BITMAP*** transition_tiles,
                  int num_masks,
                  ALLEGRO_BITMAP** overlay_bitmaps,
                  int num_overlay_tiles,
                  float cam_px,
                  float cam_py,
                  float zoom,
                  int screen_w,
                  int screen_h,
                  int player_mode,
                  N_ISO_OBJECT* objects,
                  int num_objects) {
    __n_assert(map, return);
    __n_assert(tile_bitmaps, return);

    al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);

    float hw0 = map->proj.half_w;
    float hh0 = map->proj.half_h;
    float tl0 = map->proj.tile_lift;
    float phw = hw0 * zoom;
    float phh = hh0 * zoom;
    float lift = tl0 * zoom;
    float tile_draw_w = 2.0f * phw;
    float tile_draw_h = 2.0f * phh;

    /* cam_px / cam_py are pre-computed pixel offsets provided by the caller.
     * The caller may pass either integer-snapped (`floorf(cam_x * zoom)`) or
     * sub-pixel (`cam_x * zoom`) values. Sub-pixel is preferred when paired
     * with MIN_LINEAR bilinear sampling for smooth camera motion. Integer
     * snap pairs with NEAREST sampling for crisp pixel-art but produces
     * "world jumps by 1 pixel" stutter during motion. Chunk pixel offsets
     * are added on top of whatever value the caller passes: they remain
     * integer at zoom = 1 regardless. */

    int tile_w = tile_bitmaps[0] ? al_get_bitmap_width(tile_bitmaps[0]) : 64;
    int tile_h = tile_bitmaps[0] ? al_get_bitmap_height(tile_bitmaps[0]) : 32;

    int num_edge_masks = ISO_NUM_EDGE_MASKS;

    double _iso_t_setup = al_get_time();

    /* Build segment draw order.
     *
     * The draw_order is cached on the ISO_MAP and only rebuilt when
     * the segment/height layout actually changes (`draw_order_dirty`
     * is set by iso_map_set_segments). For a static-terrain scene
     * this drops to a zero-work hit per frame. obj_order / obj_sx_arr
     * / obj_sy_arr stay file-static because objects' world positions
     * change every frame and the buffer just needs to be big enough;
     * per-map ownership isn't required. */
    int max_draw_entries = map->width * map->height * ISO_MAX_SEGMENTS_PER_TILE;
    ISO_MAP* mut_map = (ISO_MAP*)map; /* the cache fields are deliberately writable */
    if (mut_map->cached_draw_order_cap < max_draw_entries) {
        ISO_DRAW_ENTRY* grown = (ISO_DRAW_ENTRY*)realloc(
            mut_map->cached_draw_order,
            (size_t)max_draw_entries * sizeof(ISO_DRAW_ENTRY));
        if (!grown) return;
        mut_map->cached_draw_order = grown;
        mut_map->cached_draw_order_cap = max_draw_entries;
        mut_map->draw_order_dirty = 1;
    }
    if (mut_map->draw_order_dirty) {
        mut_map->cached_draw_order_count = iso_map_build_draw_order(
            map, mut_map->cached_draw_order, mut_map->cached_draw_order_cap);
        mut_map->draw_order_dirty = 0;
    }
    ISO_DRAW_ENTRY* draw_order = mut_map->cached_draw_order;
    int num_entries = mut_map->cached_draw_order_count;

    /* Pre-process objects: compute screen positions and sort by depth.
     * Sort key matches segment draw order: (floor(fz), my, mx). */
    int nobj = (objects && num_objects > 0) ? num_objects : 0;
    static int* obj_order = NULL;
    static float* obj_sx_arr = NULL;
    static float* obj_sy_arr = NULL;
    static int obj_buf_cap = 0;
    if (nobj > obj_buf_cap) {
        int new_cap = obj_buf_cap > 0 ? obj_buf_cap : 64;
        while (new_cap < nobj) new_cap *= 2;
        int* g1 = (int*)realloc(obj_order, (size_t)new_cap * sizeof(int));
        float* g2 = (float*)realloc(obj_sx_arr, (size_t)new_cap * sizeof(float));
        float* g3 = (float*)realloc(obj_sy_arr, (size_t)new_cap * sizeof(float));
        if (!g1 || !g2 || !g3) {
            n_log(LOG_ERR, "iso_map_draw: object buffer realloc failed");
            /* Keep whatever grew successfully; on next frame the
             * remaining one(s) will be retried. */
            if (g1) obj_order = g1;
            if (g2) obj_sx_arr = g2;
            if (g3) obj_sy_arr = g3;
            nobj = 0;
        } else {
            obj_order = g1;
            obj_sx_arr = g2;
            obj_sy_arr = g3;
            obj_buf_cap = new_cap;
        }
    }

    if (nobj > 0) {
        for (int i = 0; i < nobj; i++) {
            obj_order[i] = i;
            objects[i].is_occluded = 0;
            objects[i].occlude_clip_y = 1e9f;
            float wsx, wsy;
            iso_map_to_screen_f(map, objects[i].fx, objects[i].fy, objects[i].fz, &wsx, &wsy);
            obj_sx_arr[i] = wsx * zoom + cam_px;
            obj_sy_arr[i] = wsy * zoom + cam_py;
        }

        /* insertion sort by (depth, floor(fz), my, mx) to match segment draw order */
        for (int i = 1; i < nobj; i++) {
            int key = obj_order[i];
            int key_fz = (int)floorf(objects[key].fz);
            int key_my = (int)floorf(objects[key].fy);
            int key_mx = (int)floorf(objects[key].fx);
            int key_depth = key_my + key_mx;
            int j = i - 1;
            while (j >= 0) {
                int cj = obj_order[j];
                int cj_fz = (int)floorf(objects[cj].fz);
                int cj_my = (int)floorf(objects[cj].fy);
                int cj_mx = (int)floorf(objects[cj].fx);
                int cj_depth = cj_my + cj_mx;
                if (cj_depth > key_depth ||
                    (cj_depth == key_depth && cj_fz > key_fz) ||
                    (cj_depth == key_depth && cj_fz == key_fz && cj_my > key_my) ||
                    (cj_depth == key_depth && cj_fz == key_fz && cj_my == key_my && cj_mx > key_mx)) {
                    obj_order[j + 1] = obj_order[j];
                    j--;
                } else {
                    break;
                }
            }
            obj_order[j + 1] = key;
        }
    }
    n_iso_setup_us += (al_get_time() - _iso_t_setup) * 1.0e6;
    double _iso_t_pass1 = al_get_time();
    /* PASS 1: Ground segments only (bottom == 0), no objects
     * Draw all ground-level tiles first so they can never overdraw
     * elevated content (entities on bridges, floating segments, etc.).
     *
     * Wrap the entire pass in al_hold_bitmap_drawing. The previous
     * frame-wide wrap regressed because Pass 2 interleaves tile draws
     * with object-callback sprite draws (each from a different source
     * bitmap), forcing flushes. Pass 1 has no object callbacks, it's
     * pure tile draws from whatever source the caller provided. When
     * the caller groups tiles + transitions + overlays into one atlas,
     * consecutive draws share a source texture and accumulate cleanly
     * between state-change flushes. The internal al_set_blender calls
     * still flush, but runs of same-blender draws within a blender
     * regime now batch.
     *
     * `blender_is_one` tracks whether the current blender is
     * ADD/ONE/INV_ALPHA. Every set to ONE flips it true; every
     * set to ALPHA (walls, overlay, borders, hover, grid) flips it
     * false. This lets us skip the per-tile "reset to ONE" at the
     * head of each iteration when no ALPHA block ran in the
     * previous iteration, the common case for uninstrumented
     * ground tiles. */
    int blender_is_one = 1; /* set by line 2582 above the pass */
    /* al_hold_bitmap_drawing wrap around PASS 1, opt-in via
     * NILOREA_PASS1_HOLD=1. Wrapping the pass batches consecutive
     * bitmap draws (~2 ms speedup) but historically introduced
     * intermittent darker tile-mask diamond edges at the boundaries.
     * After the bitmap/primitive interleave carve-out (see below)
     * the wrap is correct end-to-end, but it remains opt-in by
     * environment variable for A/B perf comparison. */
    static int s_hold_pass1 = -1;
    static int s_hold_per_tile = -1;
    static int s_no_transitions = -1;
    static int s_no_overlay = -1;
    static int s_no_height_borders = -1;
    static int s_no_hover_grid = -1;
    static int s_tile_limit = -1;
    static int s_force_alpha = -1;
    if (s_hold_pass1 < 0) {
        /* Default is ON; set NILOREA_PASS1_HOLD=0 to opt OUT. */
        const char* e = getenv("NILOREA_PASS1_HOLD");
        s_hold_pass1 = (e && e[0] == '0') ? 0 : 1;
    }
    /* PASS 1 instrumentation knobs. Each is env-gated, default-off so
     * production builds are unchanged. Useful when investigating
     * rendering artefacts.
     *
     *   NILOREA_PASS1_HOLD_PER_TILE=1
     *       Replace the pass-wide al_hold_bitmap_drawing(true/false)
     *       wrap with a per-tile cycle. Each tile becomes its own
     *       isolated batch.
     *
     *   NILOREA_PASS1_NO_TRANSITIONS=1
     *       Skip the terrain-transitions loop entirely.
     *
     *   NILOREA_PASS1_NO_OVERLAY=1
     *       Skip the overlay-bitmap draw (ALPHA blender).
     *
     *   NILOREA_PASS1_NO_HEIGHT_BORDERS=1
     *       Skip the height-borders ALPHA draw.
     *
     *   NILOREA_PASS1_NO_HOVER_GRID=1
     *       Skip the hover-highlight and grid-overlay ALPHA draws.
     *
     *   NILOREA_PASS1_TILE_LIMIT=N
     *       Render at most N tiles in PASS 1, then break.
     *
     *   NILOREA_PASS1_FORCE_ALPHA=1
     *       Force ALPHA blender (instead of ONE) for the base-tile
     *       and transitions draws. */
    if (s_hold_per_tile < 0) {
        const char* e = getenv("NILOREA_PASS1_HOLD_PER_TILE");
        s_hold_per_tile = (e && e[0] == '1') ? 1 : 0;
    }
    if (s_no_transitions < 0) {
        const char* e = getenv("NILOREA_PASS1_NO_TRANSITIONS");
        s_no_transitions = (e && e[0] == '1') ? 1 : 0;
    }
    if (s_no_overlay < 0) {
        const char* e = getenv("NILOREA_PASS1_NO_OVERLAY");
        s_no_overlay = (e && e[0] == '1') ? 1 : 0;
    }
    if (s_no_height_borders < 0) {
        const char* e = getenv("NILOREA_PASS1_NO_HEIGHT_BORDERS");
        s_no_height_borders = (e && e[0] == '1') ? 1 : 0;
    }
    if (s_no_hover_grid < 0) {
        const char* e = getenv("NILOREA_PASS1_NO_HOVER_GRID");
        s_no_hover_grid = (e && e[0] == '1') ? 1 : 0;
    }
    if (s_tile_limit < 0) {
        const char* e = getenv("NILOREA_PASS1_TILE_LIMIT");
        s_tile_limit = (e && e[0] >= '0' && e[0] <= '9') ? atoi(e) : 0;
    }
    if (s_force_alpha < 0) {
        const char* e = getenv("NILOREA_PASS1_FORCE_ALPHA");
        s_force_alpha = (e && e[0] == '1') ? 1 : 0;
    }
    /* If per-tile hold is requested, force the pass-wide hold OFF:
     * the per-tile cycle below opens/closes its own holds. */
    int effective_hold_pass = s_hold_pass1 && !s_hold_per_tile;
    if (effective_hold_pass) {
        al_hold_bitmap_drawing(true);
        /* Batch warmup. The first batched draw after
         * al_hold_bitmap_drawing(true) can hit a slightly different
         * fragment-shading path than subsequent draws within the same
         * held context, producing darker fractional-alpha tile-mask
         * edges on the early painter-order tiles.
         *
         * Defence: prepend a fully-transparent 1-pixel draw at an
         * offscreen position so the suspect "first draw in hold
         * cycle" is absorbed into an invisible pixel. The actual
         * tile draws then all run from the second-draw-onwards
         * state, which is the consistent one. Zero visual cost
         * (transparent + offscreen) and negligible perf cost. */
        if (tile_bitmaps[0]) {
            al_draw_tinted_scaled_bitmap(tile_bitmaps[0],
                                         al_map_rgba(0, 0, 0, 0),
                                         0, 0, 1, 1,
                                         -100, -100, 1, 1, 0);
        }
    }
    int _p1_tiles_drawn = 0;
    for (int di = 0; di < num_entries; di++) {
        if (draw_order[di].bottom > 0) continue; /* skip elevated */

        /* Tile limit knob: stop after N ground tiles drawn. */
        if (s_tile_limit > 0 && _p1_tiles_drawn >= s_tile_limit) break;
        _p1_tiles_drawn++;

        /* Per-tile hold cycle: each tile is its own isolated batch.
         * The pass-wide hold is suppressed when this knob is on. */
        if (s_hold_per_tile) al_hold_bitmap_drawing(true);

        int mx = draw_order[di].mx;
        int my = draw_order[di].my;
        int seg_top = draw_order[di].top;
        int h = seg_top;
        int base = iso_map_get_terrain(map, mx, my);

        /* `al_hold_bitmap_drawing` only batches bitmap draws.
         * `al_draw_line` / `al_draw_filled_polygon` (primitives addon)
         * draw IMMEDIATELY regardless of hold. So under the pass-wide
         * hold:
         *   - tile base bitmap is queued in batch
         *   - wall polygon is drawn now at the screen position
         *   - border line is drawn now
         *   - more tiles ...
         *   - end of pass: hold(false) flushes all queued bitmaps
         *
         * The bitmap flush at end draws on TOP of the primitives,
         * hiding height borders / wall edges. Adjacent tiles' diamond
         * edges (where transitions + border lines render at the same
         * sub-pixel position with bilinear vs line-AA coverage) also
         * produce a doubled-outline effect during camera motion.
         *
         * Fix: for tiles that need primitive draws (elevated walls,
         * height borders, hover highlight, grid overlay), step OUT of
         * the held batch: flush the queued bitmaps, draw the tile
         * fully in immediate mode (bitmaps + primitives interleaved
         * correctly), then re-enter the held batch for the next
         * batchable tile. The bulk of tiles (flat ground, no overlay,
         * no hover, no grid) stays inside the held batch unchanged. */
        /* Smooth-mode slope detection, computed HERE, before the hold-batch
         * decision below, because a sloped tile's top is drawn with
         * al_draw_prim (a primitives-addon call) which renders INCORRECTLY
         * inside al_hold_bitmap_drawing(true): the queued-batch state scales
         * + misplaces the warped quad (the "flat<->slope transition tiles black
         * out / shrink toward screen centre" bug). Such tiles must step out
         * of the held batch exactly like wall/hover/grid tiles already do. A
         * FLAT tile takes the batch-safe al_draw_tinted_scaled_bitmap path and
         * stays held. h_n/h_e/h_s/h_w are reused by the vx/vy block below so
         * iso_map_smooth_corner_heights is called only once per tile. */
        float h_n = (float)h, h_e = (float)h, h_s = (float)h, h_w = (float)h;
        int tile_is_flat = 1;
        if (map->smooth_height) {
            iso_map_smooth_corner_heights(map, mx, my, &h_n, &h_e, &h_s, &h_w);
            tile_is_flat = (h_n == h_e && h_e == h_s && h_s == h_w);
        }
        int tile_needs_primitives = (seg_top > 0) ||
                                    (map->smooth_height && !tile_is_flat) ||
                                    (!s_no_hover_grid && mx == map->hover_mx && my == map->hover_my) ||
                                    (map->show_grid && !s_no_hover_grid);
        int unbatched_tile = effective_hold_pass && tile_needs_primitives;
        if (unbatched_tile) {
            al_hold_bitmap_drawing(false);
        }

        /* Per-segment tile overrides: upper_tile for top face, lower_tile for walls */
        int top_terrain = base;
        int wall_terrain = base;
        {
            int si = draw_order[di].seg_idx;
            int tidx = my * map->width + mx;
            if (map->segments && tidx >= 0 && tidx < map->width * map->height &&
                si >= 0 && si < map->segments[tidx].count) {
                int ut = map->segments[tidx].segs[si].upper_tile;
                int lt = map->segments[tidx].segs[si].lower_tile;
                if (ut >= 0 && ut < map->num_terrains) top_terrain = ut;
                if (lt >= 0 && lt < map->num_terrains) wall_terrain = lt;
            }
        }

        float vx[4], vy[4];
        float base_dx, base_dy;

        if (map->smooth_height) {
            /* h_n/h_e/h_s/h_w + tile_is_flat already computed above. */
            _iso_corner_to_screen(hw0, hh0, tl0, mx, my, h_n, cam_px, cam_py, zoom, &vx[0], &vy[0]);
            _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my, h_e, cam_px, cam_py, zoom, &vx[1], &vy[1]);
            _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my + 1, h_s, cam_px, cam_py, zoom, &vx[2], &vy[2]);
            _iso_corner_to_screen(hw0, hh0, tl0, mx, my + 1, h_w, cam_px, cam_py, zoom, &vx[3], &vy[3]);
            base_dx = vx[3];
            base_dy = vy[0];
        } else {
            _iso_corner_to_screen(hw0, hh0, tl0, mx, my, (float)h, cam_px, cam_py, zoom, &vx[0], &vy[0]);
            _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my, (float)h, cam_px, cam_py, zoom, &vx[1], &vy[1]);
            _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my + 1, (float)h, cam_px, cam_py, zoom, &vx[2], &vy[2]);
            _iso_corner_to_screen(hw0, hh0, tl0, mx, my + 1, (float)h, cam_px, cam_py, zoom, &vx[3], &vy[3]);
            base_dx = vx[3];
            base_dy = vy[0];
        }

        /* Culling */
        float min_vy = fminf(fminf(vy[0], vy[1]), fminf(vy[2], vy[3]));
        float max_vy = fmaxf(fmaxf(vy[0], vy[1]), fmaxf(vy[2], vy[3]));
        float max_side = fmaxf((float)h, (float)map->max_height) * lift;
        if (vx[1] < 0 || vx[3] > (float)screen_w) continue;
        if (max_vy + max_side < 0 || min_vy > (float)screen_h) continue;

        /* Height-based tint + ambient + dynamic lighting for this segment */
        float amb_r = map->ambient_r > 0.0f ? map->ambient_r : 1.0f;
        float amb_g = map->ambient_g > 0.0f ? map->ambient_g : 1.0f;
        float amb_b = map->ambient_b > 0.0f ? map->ambient_b : 1.0f;
        float dyn_r = 0.0f, dyn_g = 0.0f, dyn_b = 0.0f;
        if (map->dynamic_light_map) {
            int lidx = (my * map->width + mx) * 3;
            dyn_r = map->dynamic_light_map[lidx];
            dyn_g = map->dynamic_light_map[lidx + 1];
            dyn_b = map->dynamic_light_map[lidx + 2];
        }
        ALLEGRO_COLOR seg_tint = _compute_tile_tint(seg_top, map->max_height,
                                                    map->height_tint_intensity,
                                                    amb_r, amb_g, amb_b,
                                                    dyn_r, dyn_g, dyn_b);
        float seg_brightness = 1.0f;
        if (map->height_tint_intensity > 0.0f && map->max_height > 0) {
            float ratio = (float)seg_top / (float)map->max_height;
            seg_brightness = 1.0f - map->height_tint_intensity * (1.0f - ratio);
        }
        /* Factor in ambient for wall brightness */
        seg_brightness *= (amb_r + amb_g + amb_b) / 3.0f;
        float wall_dyn = (dyn_r + dyn_g + dyn_b) / 3.0f;
        seg_brightness += wall_dyn;
        if (seg_brightness > 1.0f) seg_brightness = 1.0f;

        /* Top surface: ALWAYS set the blender at the start of every
         * tile, dropping any cross-tile cache. Under
         * `al_hold_bitmap_drawing(true)`, `al_set_blender` flushes the
         * current held batch and starts a new one. If the call were
         * skipped when the previous tile ended on ONE, the held
         * batch's flush boundaries would depend on which tiles had
         * overlays/walls/hover/grid, and that set shifts when the
         * camera moves. Inconsistent flush boundaries cause
         * inconsistent GPU-rasterised pixel coverage on shared
         * diamond edges, producing flickering darker borders during
         * camera motion. Forcing the call here makes the flush
         * boundary deterministic at every tile, trading O(1000) extra
         * al_set_blender calls per frame for stable rasterisation.
         * The within-tile transitions loop keeps its conditional `if
         * (!blender_is_one)` skip because within a tile the state is
         * fresh and skipping is safe. */
        if (s_force_alpha) {
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ALPHA, ALLEGRO_INVERSE_ALPHA);
            blender_is_one = 0;
        } else {
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
            blender_is_one = 1;
        }
        if (map->smooth_height && !tile_is_flat) {
            _draw_tile_warped(tile_bitmaps[top_terrain], vx, vy, tile_w, tile_h, seg_tint);
        } else {
            /* Flat tile (or CUT mode): the al_draw_prim warp samples atlas
             * sub-bitmaps differently from al_draw_tinted_scaled_bitmap, which
             * renders flat ground tiles wrong once tiles are packed into the
             * terrain atlas. For a flat tile the warp geometry is identical to
             * this scaled-bitmap draw anyway (verified: corner span == tdw*tdh),
             * so route flat tiles through the proven atlas-correct path and
             * reserve the warp for genuinely sloped tiles. */
            al_draw_tinted_scaled_bitmap(tile_bitmaps[top_terrain], seg_tint,
                                         0, 0, (float)tile_w, (float)tile_h,
                                         base_dx, base_dy, tile_draw_w, tile_draw_h, 0);
        }

        /* Cliff walls (internally sets ONE blender). */
        if (seg_top > 0) {
            if (map->smooth_height) {
                int h_east = iso_map_get_height(map, mx + 1, my);
                int h_south = iso_map_get_height(map, mx, my + 1);
                int h_se = iso_map_get_height(map, mx + 1, my + 1);
                if ((h - h_east > map->smooth_slope_max) ||
                    (h - h_south > map->smooth_slope_max) ||
                    (h - h_se > map->smooth_slope_max)) {
                    float ce_x, ce_y, cs_x, cs_y, cw_x, cw_y;
                    _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my, (float)h, cam_px, cam_py, zoom, &ce_x, &ce_y);
                    _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my + 1, (float)h, cam_px, cam_py, zoom, &cs_x, &cs_y);
                    _iso_corner_to_screen(hw0, hh0, tl0, mx, my + 1, (float)h, cam_px, cam_py, zoom, &cw_x, &cw_y);
                    _draw_segment_sides(map, ce_x, ce_y, cs_x, cs_y, cw_x, cw_y,
                                        mx, my, 0, seg_top, lift, seg_brightness, wall_terrain);
                    /* blender_is_one stays 1: function leaves state on ONE */
                }
            } else {
                _draw_segment_sides(map, vx[1], vy[1], vx[2], vy[2], vx[3], vy[3],
                                    mx, my, 0, seg_top, lift, seg_brightness, wall_terrain);
                /* blender_is_one stays 1 */
            }
        }

        /* Terrain transitions (same tint as base tile; ONE blender) */
        if (transition_tiles && !s_no_transitions) {
            int edge_bits_arr[16];
            int corner_bits_arr[16];
            int nt = map->num_terrains < 16 ? map->num_terrains : 16;
            memset(edge_bits_arr, 0, sizeof(int) * (size_t)nt);
            memset(corner_bits_arr, 0, sizeof(int) * (size_t)nt);
            iso_map_calc_transitions_full(map, mx, my, edge_bits_arr, corner_bits_arr);

            for (int t = base + 1; t < nt; t++) {
                if (edge_bits_arr[t] != 0 && transition_tiles[t] &&
                    transition_tiles[t][edge_bits_arr[t]] != NULL) {
                    if (!blender_is_one && !s_force_alpha) {
                        al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
                        blender_is_one = 1;
                    }
                    if (map->smooth_height && !tile_is_flat) {
                        _draw_tile_warped(transition_tiles[t][edge_bits_arr[t]],
                                          vx, vy, tile_w, tile_h, seg_tint);
                    } else {
                        al_draw_tinted_scaled_bitmap(transition_tiles[t][edge_bits_arr[t]],
                                                     seg_tint,
                                                     0, 0, (float)tile_w, (float)tile_h,
                                                     base_dx, base_dy, tile_draw_w, tile_draw_h, 0);
                    }
                }

                int cidx = num_edge_masks + corner_bits_arr[t];
                if (corner_bits_arr[t] != 0 && cidx < num_masks &&
                    transition_tiles[t] && transition_tiles[t][cidx] != NULL) {
                    if (!blender_is_one && !s_force_alpha) {
                        al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
                        blender_is_one = 1;
                    }
                    if (map->smooth_height && !tile_is_flat) {
                        _draw_tile_warped(transition_tiles[t][cidx],
                                          vx, vy, tile_w, tile_h, seg_tint);
                    } else {
                        al_draw_tinted_scaled_bitmap(transition_tiles[t][cidx],
                                                     seg_tint,
                                                     0, 0, (float)tile_w, (float)tile_h,
                                                     base_dx, base_dy, tile_draw_w, tile_draw_h, 0);
                    }
                }
            }
        }

        /* Overlay tile (drawn after terrain transitions, before
         * walls/borders). Overlay bitmaps are premultiplied (loaded
         * through the same path as base tiles + transitions), so
         * ALPHA blending would produce quadratic darkening at
         * fractional-alpha edges. Use ONE/INVERSE_ALPHA which matches
         * the premul source format. Also avoids cross-tile
         * contamination under al_hold_bitmap_drawing(true), where
         * al_set_blender does not flush the held batch mid-pass: an
         * ALPHA blender set here would otherwise apply to draws
         * emitted earlier in the batch by other tiles when the batch
         * eventually flushes. */
        if (overlay_bitmaps && map->overlay && num_overlay_tiles > 0 && !s_no_overlay) {
            int ov_idx = my * map->width + mx;
            int ov_tile = map->overlay[ov_idx];
            if (ov_tile > 0 && ov_tile < num_overlay_tiles && overlay_bitmaps[ov_tile]) {
                if (!blender_is_one) {
                    al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
                    blender_is_one = 1;
                }
                if (map->smooth_height && !tile_is_flat) {
                    _draw_tile_warped(overlay_bitmaps[ov_tile], vx, vy, tile_w, tile_h, seg_tint);
                } else {
                    al_draw_tinted_scaled_bitmap(overlay_bitmaps[ov_tile], seg_tint,
                                                 0, 0, (float)tile_w, (float)tile_h,
                                                 base_dx, base_dy, tile_draw_w, tile_draw_h, 0);
                }
            }
        }

        /* Height borders (ONE blender; lines use RGB=0 so premul == straight) */
        if (seg_top > 0 && !s_no_height_borders) {
            if (!blender_is_one) {
                al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
                blender_is_one = 1;
            }
            _draw_height_borders(map, mx, my, draw_order[di].seg_idx,
                                 0, seg_top,
                                 draw_order[di].underside,
                                 vx, vy, lift);
        }

        /* Hover highlight (ONE blender; colors premultiplied below) */
        if (!s_no_hover_grid && mx == map->hover_mx && my == map->hover_my) {
            if (!blender_is_one) {
                al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
                blender_is_one = 1;
            }
            /* Premul: each RGB channel scaled by alpha/255. */
            ALLEGRO_COLOR hc = player_mode
                                   ? al_map_rgba(100 * 120 / 255, 255 * 120 / 255, 100 * 120 / 255, 120)
                                   : al_map_rgba(255 * 120 / 255, 255 * 120 / 255, 100 * 120 / 255, 120);
            float verts[8] = {vx[0], vy[0], vx[1], vy[1],
                              vx[2], vy[2], vx[3], vy[3]};
            al_draw_filled_polygon(verts, 4, hc);
            ALLEGRO_COLOR oc = player_mode
                                   ? al_map_rgba(100 * 200 / 255, 255 * 200 / 255, 100 * 200 / 255, 200)
                                   : al_map_rgba(255 * 200 / 255, 255 * 200 / 255, 100 * 200 / 255, 200);
            al_draw_line(vx[0], vy[0], vx[1], vy[1], oc, 1.5f);
            al_draw_line(vx[1], vy[1], vx[2], vy[2], oc, 1.5f);
            al_draw_line(vx[2], vy[2], vx[3], vy[3], oc, 1.5f);
            al_draw_line(vx[3], vy[3], vx[0], vy[0], oc, 1.5f);
        }

        /* Grid overlay (ONE blender; colors premultiplied below).
         * No blender_is_one = 1 update needed: this is the last block in
         * the iteration that touches the blender, and the next iteration
         * unconditionally re-sets the blender at the top of the loop. */
        if (map->show_grid && !s_no_hover_grid) {
            if (!blender_is_one) {
                al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
            }
            /* Premul: white 60/255 alpha is (60,60,60,60). */
            ALLEGRO_COLOR gc = al_map_rgba(60, 60, 60, 60);
            al_draw_line(vx[0], vy[0], vx[1], vy[1], gc, 1.0f);
            al_draw_line(vx[1], vy[1], vx[2], vy[2], gc, 1.0f);
            al_draw_line(vx[2], vy[2], vx[3], vy[3], gc, 1.0f);
            al_draw_line(vx[3], vy[3], vx[0], vy[0], gc, 1.0f);

            if (h > 0) {
                unsigned char a2 = (unsigned char)(60 + h * 35);
                /* Premul: (R*a/255, G*a/255, B*a/255, a) for (255,200,100,a). */
                ALLEGRO_COLOR hc2 = al_map_rgba(a2,
                                                (unsigned char)(200 * a2 / 255),
                                                (unsigned char)(100 * a2 / 255),
                                                a2);
                float dot_x = (vx[0] + vx[2]) / 2.0f;
                float dot_y = (vy[0] + vy[1] + vy[2] + vy[3]) / 4.0f;
                al_draw_filled_circle(dot_x, dot_y, 2.0f * zoom, hc2);
            }
        }

        /* Per-tile hold cycle: flush this tile's batch before moving
         * on. Each tile becomes its own VBO. */
        if (s_hold_per_tile) al_hold_bitmap_drawing(false);

        /* Re-enter the held batch for the next batchable tile.
         * Resets the blender to ONE for the next tile's base bitmap,
         * since immediate-mode primitive draws inside the unbatched
         * section may have left other state.
         * No blender_is_one = 1 update: the next iteration unconditionally
         * re-sets the blender at the top of the loop. */
        if (unbatched_tile) {
            al_hold_bitmap_drawing(true);
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
        }
    }

    if (effective_hold_pass) al_hold_bitmap_drawing(false);
    n_iso_pass1_us += (al_get_time() - _iso_t_pass1) * 1.0e6;
    double _iso_t_pass2 = al_get_time();
    /* PASS 2: Elevated segments + ALL objects (interleaved)
     * Objects are interleaved at every draw_order entry (ground and
     * elevated) to preserve correct depth ordering.  Only elevated
     * segments (bottom > 0) are actually rendered, ground segments
     * were already drawn in Pass 1. */
    int obj_cursor = 0;
    for (int di = 0; di < num_entries; di++) {
        int mx = draw_order[di].mx;
        int my = draw_order[di].my;
        int seg_bottom = draw_order[di].bottom;

        /* Object interleaving runs at every entry regardless of
         * ground/elevated, this ensures objects at ground-level depth
         * positions are drawn at the correct point in the sort order. */
        if (nobj > 0) {
            int seg_depth = my + mx;
            while (obj_cursor < nobj) {
                int oi = obj_order[obj_cursor];
                int obj_fz = (int)floorf(objects[oi].fz);
                int obj_my = (int)floorf(objects[oi].fy);
                int obj_mx = (int)floorf(objects[oi].fx);
                int obj_depth = obj_my + obj_mx;
                if (obj_depth > seg_depth ||
                    (obj_depth == seg_depth && obj_fz > seg_bottom) ||
                    (obj_depth == seg_depth && obj_fz == seg_bottom && obj_my > my) ||
                    (obj_depth == seg_depth && obj_fz == seg_bottom && obj_my == my && obj_mx >= mx))
                    break;
                if (objects[oi].draw) {
                    objects[oi].draw(obj_sx_arr[oi], obj_sy_arr[oi],
                                     zoom, 1.0f, objects[oi].user_data);
                }
                obj_cursor++;
            }
        }

        /* Ground segments already drawn in Pass 1, skip rendering */
        if (seg_bottom == 0) continue;

        /* Elevated segment rendering */
        int seg_top = draw_order[di].top;
        int h = seg_top;
        int base = iso_map_get_terrain(map, mx, my);

        /* Per-segment tile overrides for elevated segments */
        int etop_terrain = base;
        int ewall_terrain = base;
        {
            int si = draw_order[di].seg_idx;
            int tidx = my * map->width + mx;
            if (map->segments && tidx >= 0 && tidx < map->width * map->height &&
                si >= 0 && si < map->segments[tidx].count) {
                int ut = map->segments[tidx].segs[si].upper_tile;
                int lt = map->segments[tidx].segs[si].lower_tile;
                if (ut >= 0 && ut < map->num_terrains) etop_terrain = ut;
                if (lt >= 0 && lt < map->num_terrains) ewall_terrain = lt;
            }
        }

        float vx[4], vy[4];
        float base_dx, base_dy;
        _iso_corner_to_screen(hw0, hh0, tl0, mx, my, (float)h, cam_px, cam_py, zoom, &vx[0], &vy[0]);
        _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my, (float)h, cam_px, cam_py, zoom, &vx[1], &vy[1]);
        _iso_corner_to_screen(hw0, hh0, tl0, mx + 1, my + 1, (float)h, cam_px, cam_py, zoom, &vx[2], &vy[2]);
        _iso_corner_to_screen(hw0, hh0, tl0, mx, my + 1, (float)h, cam_px, cam_py, zoom, &vx[3], &vy[3]);
        base_dx = vx[3];
        base_dy = vy[0];

        /* Culling, objects already handled above, safe to continue */
        float min_vy = fminf(fminf(vy[0], vy[1]), fminf(vy[2], vy[3]));
        float max_vy = fmaxf(fmaxf(vy[0], vy[1]), fmaxf(vy[2], vy[3]));
        float max_side = fmaxf((float)h, (float)map->max_height) * lift;
        if (vx[1] < 0 || vx[3] > (float)screen_w) continue;
        if (max_vy + max_side < 0 || min_vy > (float)screen_h) continue;

        /* Underside face for floating segments */
        if (draw_order[di].underside && tile_bitmaps[etop_terrain]) {
            float under_sx, under_sy;
            iso_map_to_screen(map, mx, my, seg_bottom, &under_sx, &under_sy);
            float under_dx = under_sx * zoom + cam_px;
            float under_dy = under_sy * zoom + cam_py;
            _draw_segment_underside(tile_bitmaps[etop_terrain], under_dx, under_dy,
                                    phw, phh, tile_draw_w, tile_draw_h,
                                    tile_w, tile_h);
        }

        /* Height-based tint for elevated segment */
        /* Elevated segment: same ambient + dynamic light from base tile */
        float eamb_r = map->ambient_r > 0.0f ? map->ambient_r : 1.0f;
        float eamb_g = map->ambient_g > 0.0f ? map->ambient_g : 1.0f;
        float eamb_b = map->ambient_b > 0.0f ? map->ambient_b : 1.0f;
        float edyn_r = 0.0f, edyn_g = 0.0f, edyn_b = 0.0f;
        if (map->dynamic_light_map) {
            int elidx = (my * map->width + mx) * 3;
            edyn_r = map->dynamic_light_map[elidx];
            edyn_g = map->dynamic_light_map[elidx + 1];
            edyn_b = map->dynamic_light_map[elidx + 2];
        }
        ALLEGRO_COLOR eseg_tint = _compute_tile_tint(seg_top, map->max_height,
                                                     map->height_tint_intensity,
                                                     eamb_r, eamb_g, eamb_b,
                                                     edyn_r, edyn_g, edyn_b);
        float eseg_brightness = 1.0f;
        if (map->height_tint_intensity > 0.0f && map->max_height > 0) {
            float ratio = (float)seg_top / (float)map->max_height;
            eseg_brightness = 1.0f - map->height_tint_intensity * (1.0f - ratio);
        }
        eseg_brightness *= (eamb_r + eamb_g + eamb_b) / 3.0f;
        eseg_brightness += (edyn_r + edyn_g + edyn_b) / 3.0f;
        if (eseg_brightness > 1.0f) eseg_brightness = 1.0f;

        /* Top surface (flat, elevated segments don't use smooth height) */
        al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_INVERSE_ALPHA);
        al_draw_tinted_scaled_bitmap(tile_bitmaps[etop_terrain], eseg_tint,
                                     0, 0, (float)tile_w, (float)tile_h,
                                     base_dx, base_dy, tile_draw_w, tile_draw_h, 0);

        /* Cliff walls + height borders */
        if (seg_top > seg_bottom) {
            _draw_segment_sides(map, vx[1], vy[1], vx[2], vy[2], vx[3], vy[3],
                                mx, my, seg_bottom, seg_top, lift, eseg_brightness, ewall_terrain);
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ALPHA, ALLEGRO_INVERSE_ALPHA);
            _draw_height_borders(map, mx, my, draw_order[di].seg_idx,
                                 seg_bottom, seg_top,
                                 draw_order[di].underside,
                                 vx, vy, lift);
        }

        /* Hover highlight */
        if (mx == map->hover_mx && my == map->hover_my) {
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ALPHA, ALLEGRO_INVERSE_ALPHA);
            ALLEGRO_COLOR hc = player_mode
                                   ? al_map_rgba(100, 255, 100, 120)
                                   : al_map_rgba(255, 255, 100, 120);
            float verts[8] = {vx[0], vy[0], vx[1], vy[1],
                              vx[2], vy[2], vx[3], vy[3]};
            al_draw_filled_polygon(verts, 4, hc);
            ALLEGRO_COLOR oc = player_mode
                                   ? al_map_rgba(100, 255, 100, 200)
                                   : al_map_rgba(255, 255, 100, 200);
            al_draw_line(vx[0], vy[0], vx[1], vy[1], oc, 1.5f);
            al_draw_line(vx[1], vy[1], vx[2], vy[2], oc, 1.5f);
            al_draw_line(vx[2], vy[2], vx[3], vy[3], oc, 1.5f);
            al_draw_line(vx[3], vy[3], vx[0], vy[0], oc, 1.5f);
        }

        /* Grid overlay */
        if (map->show_grid) {
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ALPHA, ALLEGRO_INVERSE_ALPHA);
            ALLEGRO_COLOR gc = al_map_rgba(255, 255, 255, 60);
            al_draw_line(vx[0], vy[0], vx[1], vy[1], gc, 1.0f);
            al_draw_line(vx[1], vy[1], vx[2], vy[2], gc, 1.0f);
            al_draw_line(vx[2], vy[2], vx[3], vy[3], gc, 1.0f);
            al_draw_line(vx[3], vy[3], vx[0], vy[0], gc, 1.0f);

            if (h > 0) {
                ALLEGRO_COLOR hc2 = al_map_rgba(255, 200, 100,
                                                (unsigned char)(60 + h * 35));
                float dot_x = (vx[0] + vx[2]) / 2.0f;
                float dot_y = (vy[0] + vy[1] + vy[2] + vy[3]) / 4.0f;
                al_draw_filled_circle(dot_x, dot_y, 2.0f * zoom, hc2);
            }
        }
    }

    /* Draw any remaining objects past the last entry */
    while (obj_cursor < nobj) {
        int oi = obj_order[obj_cursor];
        if (objects[oi].draw) {
            objects[oi].draw(obj_sx_arr[oi], obj_sy_arr[oi],
                             zoom, 1.0f, objects[oi].user_data);
        }
        obj_cursor++;
    }

    n_iso_pass2_us += (al_get_time() - _iso_t_pass2) * 1.0e6;
    double _iso_t_pass3 = al_get_time();
    /* Step 7: Occlusion detection and clipped overlay pass.
     * For each occluded entity, restrict the ghost draw to the region
     * below the occluding tile's north vertex using a clipping rectangle. */
    for (int i = 0; i < nobj; i++) {
        int oi = obj_order[i];
        if (objects[oi].occluded_alpha <= 0.0f) continue;
        if (!objects[oi].draw) continue;

        float clip_y = 1e9f;
        if (!_iso_object_check_occlusion(map,
                                         objects[oi].fx, objects[oi].fy, objects[oi].fz,
                                         obj_sx_arr[oi], obj_sy_arr[oi],
                                         objects[oi].sprite_h,
                                         objects[oi].sprite_half_w,
                                         cam_px, cam_py, zoom,
                                         &clip_y))
            continue;

        objects[oi].is_occluded = 1;
        objects[oi].occlude_clip_y = clip_y;

        /* Clip ghost overlay to below the occluder's top edge. The
         * ghost is drawn at alpha < 1 to make the occluded sprite
         * look semi-transparent. Use straight-alpha blending here:
         * with the premultiplied (ADD/ONE/INV_ALPHA) blender the
         * 0.4-alpha overlay is ADDED on top of the already-drawn
         * full-bright sprite, brightening it instead of revealing
         * what's behind. ALPHA/INV_ALPHA does the standard
         * src*alpha + dst*(1-alpha) transparent compositing the
         * caller meant. */
        int prev_cx, prev_cy, prev_cw, prev_ch;
        al_get_clipping_rectangle(&prev_cx, &prev_cy, &prev_cw, &prev_ch);
        int clip_top = (int)floorf(clip_y);
        if (clip_top < prev_cy) clip_top = prev_cy;
        al_set_clipping_rectangle(prev_cx, clip_top,
                                  prev_cw, prev_cy + prev_ch - clip_top);
        al_set_blender(ALLEGRO_ADD, ALLEGRO_ALPHA, ALLEGRO_INVERSE_ALPHA);
        objects[oi].draw(obj_sx_arr[oi], obj_sy_arr[oi],
                         zoom, objects[oi].occluded_alpha, objects[oi].user_data);
        al_set_clipping_rectangle(prev_cx, prev_cy, prev_cw, prev_ch);
    }

    n_iso_pass3_us += (al_get_time() - _iso_t_pass3) * 1.0e6;

    /* The draw_order / obj_* buffers are file-static and reused
     * across calls; no per-call free. They survive for the process
     * lifetime which is fine: max draw entries is bounded by
     * map->width * map->height * ISO_MAX_SEGMENTS_PER_TILE (~2 KB
     * for typical 32x32x2 chunks), and obj_buf_cap grows once to the
     * scene's peak entity count. */
} /* iso_map_draw() */

/* Transition mask & tile generation (Article 934) */

void iso_mask_tile_to_diamond(ALLEGRO_BITMAP* bmp, int tile_w, int tile_h) {
    __n_assert(bmp, return);

    ALLEGRO_LOCKED_REGION* lr = al_lock_bitmap(bmp,
                                               ALLEGRO_PIXEL_FORMAT_ABGR_8888_LE, ALLEGRO_LOCK_READWRITE);
    if (!lr) return;

    for (int py = 0; py < tile_h; py++) {
        unsigned char* row = (unsigned char*)lr->data + py * lr->pitch;
        for (int px = 0; px < tile_w; px++) {
            if (!iso_is_in_diamond(px, py, tile_w, tile_h)) {
                row[px * 4 + 0] = 0;
                row[px * 4 + 1] = 0;
                row[px * 4 + 2] = 0;
                row[px * 4 + 3] = 0;
            }
        }
    }
    al_unlock_bitmap(bmp);
}

void iso_generate_transition_masks(ALLEGRO_BITMAP** masks, int tile_w, int tile_h) {
    __n_assert(masks, return);

    float cx = (float)tile_w / 2.0f;
    float cy = (float)tile_h / 2.0f;

    ALLEGRO_STATE state;
    al_store_state(&state, ALLEGRO_STATE_TARGET_BITMAP | ALLEGRO_STATE_BLENDER);

    /* Edge masks (0..15) */
    for (int mask = 0; mask < ISO_NUM_EDGE_MASKS; mask++) {
        masks[mask] = al_create_bitmap(tile_w, tile_h);
        al_set_target_bitmap(masks[mask]);
        al_clear_to_color(al_map_rgba(0, 0, 0, 0));

        ALLEGRO_LOCKED_REGION* lr = al_lock_bitmap(masks[mask],
                                                   ALLEGRO_PIXEL_FORMAT_ABGR_8888_LE, ALLEGRO_LOCK_WRITEONLY);
        if (!lr) continue;

        for (int py = 0; py < tile_h; py++) {
            unsigned char* row = (unsigned char*)lr->data + py * lr->pitch;
            for (int px = 0; px < tile_w; px++) {
                if (!iso_is_in_diamond(px, py, tile_w, tile_h)) {
                    row[px * 4 + 0] = 0;
                    row[px * 4 + 1] = 0;
                    row[px * 4 + 2] = 0;
                    row[px * 4 + 3] = 0;
                    continue;
                }

                float alpha = 0.0f;
                float nx = ((float)px + 0.5f - cx) / ((float)tile_w / 2.0f);
                float ny = ((float)py + 0.5f - cy) / ((float)tile_h / 2.0f);
                float gw = 0.65f;

                if (mask & ISO_EDGE_W) alpha = fmaxf(alpha, fminf(1.0f, fmaxf(0.0f, (-nx - ny) / gw)));
                if (mask & ISO_EDGE_N) alpha = fmaxf(alpha, fminf(1.0f, fmaxf(0.0f, (nx - ny) / gw)));
                if (mask & ISO_EDGE_E) alpha = fmaxf(alpha, fminf(1.0f, fmaxf(0.0f, (nx + ny) / gw)));
                if (mask & ISO_EDGE_S) alpha = fmaxf(alpha, fminf(1.0f, fmaxf(0.0f, (-nx + ny) / gw)));

                if (alpha > 1.0f) alpha = 1.0f;
                alpha = alpha * alpha * (3.0f - 2.0f * alpha);

                unsigned char a = (unsigned char)(alpha * 255.0f);
                row[px * 4 + 0] = a;
                row[px * 4 + 1] = a;
                row[px * 4 + 2] = a;
                row[px * 4 + 3] = a;
            }
        }
        al_unlock_bitmap(masks[mask]);
    }

    /* Corner masks (16..31)
     * In isometric diamond projection, diagonal neighbors map to diamond tips:
     *   NW (mx-1,my-1) = North tip  (0, -1) in normalized coords
     *   NE (mx+1,my-1) = East tip   (1,  0)
     *   SE (mx+1,my+1) = South tip  (0,  1)
     *   SW (mx-1,my+1) = West tip   (-1, 0) */
    for (int mask = 0; mask < ISO_NUM_CORNER_MASKS; mask++) {
        int idx = ISO_NUM_EDGE_MASKS + mask;
        masks[idx] = al_create_bitmap(tile_w, tile_h);
        al_set_target_bitmap(masks[idx]);
        al_clear_to_color(al_map_rgba(0, 0, 0, 0));

        ALLEGRO_LOCKED_REGION* lr = al_lock_bitmap(masks[idx],
                                                   ALLEGRO_PIXEL_FORMAT_ABGR_8888_LE, ALLEGRO_LOCK_WRITEONLY);
        if (!lr) continue;

        for (int py = 0; py < tile_h; py++) {
            unsigned char* row = (unsigned char*)lr->data + py * lr->pitch;
            for (int px = 0; px < tile_w; px++) {
                if (!iso_is_in_diamond(px, py, tile_w, tile_h)) {
                    row[px * 4 + 0] = 0;
                    row[px * 4 + 1] = 0;
                    row[px * 4 + 2] = 0;
                    row[px * 4 + 3] = 0;
                    continue;
                }

                float alpha = 0.0f;
                float nx = ((float)px + 0.5f - cx) / ((float)tile_w / 2.0f);
                float ny = ((float)py + 0.5f - cy) / ((float)tile_h / 2.0f);
                float corner_radius = 0.55f;

                if (mask & ISO_CORNER_NW) {
                    float dy = ny + 1.0f;
                    float dist = sqrtf(nx * nx + dy * dy);
                    float c_alpha = fmaxf(0.0f, 1.0f - dist / corner_radius);
                    alpha = fmaxf(alpha, c_alpha);
                }
                if (mask & ISO_CORNER_NE) {
                    float dx = nx - 1.0f;
                    float dist = sqrtf(dx * dx + ny * ny);
                    float c_alpha = fmaxf(0.0f, 1.0f - dist / corner_radius);
                    alpha = fmaxf(alpha, c_alpha);
                }
                if (mask & ISO_CORNER_SE) {
                    float dy = ny - 1.0f;
                    float dist = sqrtf(nx * nx + dy * dy);
                    float c_alpha = fmaxf(0.0f, 1.0f - dist / corner_radius);
                    alpha = fmaxf(alpha, c_alpha);
                }
                if (mask & ISO_CORNER_SW) {
                    float dx = nx + 1.0f;
                    float dist = sqrtf(dx * dx + ny * ny);
                    float c_alpha = fmaxf(0.0f, 1.0f - dist / corner_radius);
                    alpha = fmaxf(alpha, c_alpha);
                }

                if (alpha > 1.0f) alpha = 1.0f;
                alpha = alpha * alpha * (3.0f - 2.0f * alpha);

                unsigned char a = (unsigned char)(alpha * 255.0f);
                row[px * 4 + 0] = a;
                row[px * 4 + 1] = a;
                row[px * 4 + 2] = a;
                row[px * 4 + 3] = a;
            }
        }
        al_unlock_bitmap(masks[idx]);
    }

    al_restore_state(&state);
    n_log(LOG_INFO, "Generated %d transition masks (%dx%d)", ISO_NUM_MASKS, tile_w, tile_h);
}

void iso_generate_transition_tiles(ALLEGRO_BITMAP*** tiles,
                                   ALLEGRO_BITMAP** masks,
                                   ALLEGRO_BITMAP** tile_bitmaps,
                                   int num_terrains,
                                   int tile_w,
                                   int tile_h) {
    __n_assert(tiles, return);
    __n_assert(masks, return);
    __n_assert(tile_bitmaps, return);

    ALLEGRO_STATE state;
    al_store_state(&state, ALLEGRO_STATE_TARGET_BITMAP | ALLEGRO_STATE_BLENDER);

    for (int t = 0; t < num_terrains; t++) {
        for (int m = 0; m < ISO_NUM_MASKS; m++) {
            /* Skip mask 0 (no edges active) and mask 16 (no corners active) */
            if ((m < ISO_NUM_EDGE_MASKS && m == 0) ||
                (m >= ISO_NUM_EDGE_MASKS && m == ISO_NUM_EDGE_MASKS)) {
                tiles[t][m] = NULL;
                continue;
            }

            tiles[t][m] = al_create_bitmap(tile_w, tile_h);
            al_set_target_bitmap(tiles[t][m]);
            al_clear_to_color(al_map_rgba(0, 0, 0, 0));

            /* Draw the terrain tile first */
            al_set_blender(ALLEGRO_ADD, ALLEGRO_ONE, ALLEGRO_ZERO);
            al_draw_bitmap(tile_bitmaps[t], 0, 0, 0);

            /* Multiply by the alpha mask: keeps terrain color
             * but applies the mask's alpha for the transition gradient */
            al_set_blender(ALLEGRO_ADD, ALLEGRO_DEST_COLOR, ALLEGRO_ZERO);
            al_draw_bitmap(masks[m], 0, 0, 0);
        }
    }

    al_restore_state(&state);
    n_log(LOG_INFO, "Pre-composited transition tiles for %d terrains x %d masks", num_terrains, ISO_NUM_MASKS);
}

#endif /* HAVE_ALLEGRO - iso_map_draw */

#ifdef N_ASTAR_H
ASTAR_GRID* iso_map_to_astar_grid(const ISO_MAP* map, int max_height_diff, int start_x, int start_y) {
    __n_assert(map, return NULL);

    ASTAR_GRID* grid = n_astar_grid_new(map->width, map->height, 1);
    if (!grid) return NULL;

    int total = map->width * map->height;

    /* First pass: all cells unwalkable, set costs */
    for (int y = 0; y < map->height; y++) {
        for (int x = 0; x < map->width; x++) {
            n_astar_grid_set_walkable(grid, x, y, 0, 0);
            int h = map->heightmap[y * map->width + x];
            int cost = ASTAR_COST_CARDINAL + h * 100;
            n_astar_grid_set_cost(grid, x, y, 0, cost);
        }
    }

    /* Flood-fill from (start_x, start_y) to mark reachable cells as walkable.
     * This ensures height constraints are checked per-edge (between consecutive
     * cells) rather than per-cell, so tiles on a plateau edge remain walkable
     * when approached from the same elevation. */
    if (start_x < 0 || start_x >= map->width || start_y < 0 || start_y >= map->height)
        return grid;

    int start_ab = map->ability[start_y * map->width + start_x];
    if (start_ab != WALK && start_ab != SWIM)
        return grid;

    int* queue = (int*)malloc((size_t)total * 2 * sizeof(int));
    if (!queue) {
        n_astar_grid_free(grid);
        return NULL;
    }

    uint8_t* visited = (uint8_t*)calloc((size_t)total, sizeof(uint8_t));
    if (!visited) {
        free(queue);
        n_astar_grid_free(grid);
        return NULL;
    }

    int qfront = 0, qback = 0;

    /* seed the start cell */
    visited[start_y * map->width + start_x] = 1;
    n_astar_grid_set_walkable(grid, start_x, start_y, 0, 1);
    queue[qback++] = start_x;
    queue[qback++] = start_y;

    /* 8-directional expansion (matches ASTAR_ALLOW_DIAGONAL) */
    const int dirs[][2] = {{-1, 0}, {1, 0}, {0, -1}, {0, 1}, {-1, -1}, {-1, 1}, {1, -1}, {1, 1}};

    while (qfront < qback) {
        int cx = queue[qfront++];
        int cy = queue[qfront++];
        int ch = map->heightmap[cy * map->width + cx];

        for (int d = 0; d < 8; d++) {
            int nx = cx + dirs[d][0];
            int ny = cy + dirs[d][1];
            if (nx < 0 || nx >= map->width || ny < 0 || ny >= map->height) continue;
            if (visited[ny * map->width + nx]) continue;

            int nab = map->ability[ny * map->width + nx];
            if (nab != WALK && nab != SWIM) continue;

            int nh = map->heightmap[ny * map->width + nx];
            if (max_height_diff >= 0 && abs(ch - nh) > max_height_diff) continue;

            visited[ny * map->width + nx] = 1;
            n_astar_grid_set_walkable(grid, nx, ny, 0, 1);
            queue[qback++] = nx;
            queue[qback++] = ny;
        }
    }

    free(queue);
    free(visited);

    return grid;
}
#endif /* N_ASTAR_H */

#endif /* #ifndef NOISOENGINE */