proper fix for texture seam issue

[obiot]

follow up to #335, see comments starting here #335 (comment)

workaround implemented here :
6f800a4

[obiot]

Analysis of proper fix options

The current workaround (uvOffset = 1) stretches the destination tile by 1px to paper over gaps. It works but is not mathematically precise. Here are three approaches for a proper fix, ordered by complexity:

Option 1: Half-texel UV inset (simplest)

Apply a 0.5 / textureSize inset to tile UVs so the GPU sampler never reaches the boundary between adjacent tiles in the atlas. This could be done in addUVs() when computing UV coordinates:

const epsilon = 0.5;
atlas[name].uvs = new Float32Array([
    (s.x + epsilon) / w,
    (s.y + epsilon) / h,
    (s.x + sw - epsilon) / w,
    (s.y + sh - epsilon) / h,
]);

Pros: minimal code change, works universally, no shader changes
Cons: applies to sprites too (though the inset is invisible at typical sprite sizes — 0.5/256 = 0.002 UV), doesn't fix geometry-level gaps

Option 2: Precompute tile UVs + per-tile inset

Rearchitect tile rendering so that TMXTileset precomputes UVs with half-texel inset at tileset init time (like sprite atlases already do), and drawTile passes them directly to addQuad instead of going through the lazy drawImage → getUVs("sx,sy,sw,sh") string-key path.

Pros: fixes seams, eliminates per-frame string concatenation and map lookups for tiles, only affects tiles (not sprites)
Cons: significant refactor of the tile rendering pipeline — Tile extends Bounds not Renderable, so the UV storage and rendering path needs rethinking

Option 3: UV clamping in the fragment shader

Pass per-quad UV min/max bounds as additional vertex attributes to the quad shader, and clamp sampling in the fragment shader:

varying vec2 vUvBounds; // (u_min, v_min, u_max, v_max)
// ...
vec2 clampedUV = clamp(vUV, vUvBounds.xy, vUvBounds.zw);

Pros: the most robust GPU-side solution, zero bleeding by design
Cons: requires distinguishing tiles from sprites (both go through addQuad), adds per-vertex data and shader branching, most invasive change

---

Current assessment: the existing uvOffset workaround is good enough for most use cases. Option 1 is the lowest-effort improvement if we decide to revisit this. Option 2 would combine the seam fix with a meaningful performance improvement for tile rendering but requires more architectural work.