fix: stop ASCII Doughnut tiling one torus five times across the wall (#182)

[BernardJen]

Closes #182.

The bug

The shader folded x with mod() into a 3.4-unit cell whenever the canvas was wider
than that, so a 5:1 wall got the same doughnut stamped five times at identical phase
— five copies tumbling in lockstep. That reads as a rendering fault rather than a design,
which is why this carried the bug label.

The fold is gone. The width is divided into slots and each slot gets its own solid:
shape, proportions, spin phase, scale, vertical offset and depth, all from a hash of the
slot index. Cost is unchanged — still one distance function per march step — because a
fragment only ever belongs to one slot. What changes is that no two slots look alike or
move together.

Slot width follows the solid, not the reverse

Worth recording because my first attempt got it backwards. I picked slot count first and
fitted solids into them, which produced five small solids adrift in a mostly black
frame — trading this issue's tiling complaint for its emptiness complaint.

A solid's half-extent is ~1.34 world units against a vertical half-span of 1.55, so
height caps its size: the frame is 1200px tall however wide the wall is. Sizing the
solid to the height first and then laying out as many slots as fit gives about nine
across the wall, nearly touching.

The rest of the issue

• Cells 26px → 48px (~53px at 6000×1200): about 113 columns × 22 rows, close to
  donut.c's own 80×22. The characters are readable across a room — the entire point of
  ASCII art, and measured as specks before.
• Perceptually even ramp with dithering. The donut.c ramp is ordered by eye, not by
  area, so a linear index terraced the shading into bands. Each glyph's ink coverage is
  now measured from the atlas the shader samples, and luminance is matched against
  real coverage; a 4×4 ordered dither picks between the two bracketing glyphs per cell.
• Lighting: key + fill + specular, plus ambient occlusion from the distance
  function — which is what finally darkens the inside of the hole, where a plain Lambert
  reports full light.
• Normals by central differences rather than the analytic torus gradient, so shading is
  correct for every shape.
• Colour carries depth: near solids warm amber, far ones cooler green, instead of one
  flat red.
• Brightness raised for ambient light.

A wrong diagnosis, recorded rather than buried

The rightmost solid looked clipped in a downscaled crop and I attributed it to an
off-by-one in the slot index. Measuring ink in the last 40px was identical with and
without the clamp (0.01% of pixels, peak 0.298) — there was no clipping and the
diagnosis was wrong. The clamp stays as a one-instruction guard, and its comment says
plainly that it is defensive rather than a fix.

Acceptance criteria

• No mod()-tiled repetition; the wide canvas is filled by a composition
• Characters individually legible at wall viewing distance
• Bright enough to read at 12% washout — peak luminance 0.910 vs 0.510 before
• Ramp perceptually even, with dithering between steps
• Improved lighting (specular, second light, occlusion in the hole)
• Colour carries depth information
• Variety across activations — two seeds shown below differ in palette, shapes and layout
• Frame cost at 6000×1200 recorded: 120.2 fps, vsync-capped, vs main's 120.0
• Before/after at 3000×600

Verification

• Branched from main (76b8259).
• npm run lint clean.
• npm test — 407 passed.
• npm run shadercheck — ok=150 fail=0; edge density 0.0095 against a 0.0124 baseline,
  a 23% drop inside the 35% tolerance, so no baseline change and the new staleness
  report stays empty.
• Rendered at 3000×600 on seeds 555 and 909, at 6000×1200, and at 12% washout.

Not verified / honest limits

• The dark end of the ramp still washes out. . and , are inherently dim and sparse;
  no amount of gain fixes that without destroying the shading the ramp exists to express.
  The bright end carries legibility, which is what the 0.910 peak measures.
• There is still substantial black between solids. Nine round objects in a 5:1 frame cannot
  fill it without either stretching them or shrinking the glyphs, both of which the issue
  argues against.
• Only two of the shape variants (torus, rounded box) appear in the captures; the
  octahedron branch is reachable by hash but was not observed in these two seeds.