Era media: one frame, ten picture technologies

The scene renders one realistic eye. The medium is a final full-screen pass that reads that finished frame as a texture and repaints it. One era's branch runs per frame (M.cur is a uniform; the same for every pixel, so the branch is free).

The input the medium gets

• The finished frame, display-encoded (gamma 1/2.2), 8 bits a channel, mipmapped, as a texture. Why: every medium works in display values (that is what screens quantised); 8 bits halves bandwidth against half-float and needs no decode per tap; the mips are a free blur for frost and for reading a cell's mean.
• Its alpha is depth: 0.5 at the focal plane, ±20 mm mapped to 0..1, near = 1. Use it for elevation, shadows and parallax.
• Every sample in a branch names its level: tex.sample(uv).level(n). Why: implicit derivatives are undefined inside non-uniform branches.
• Output is linear: wrap the result in lin().

Order of operations inside a branch (always)

1. Scale. Everything is sized in reference pixels at 1440×900 and multiplied by kc = irisPx / 245. Grids use cellPx(n) (whole device pixels); strokes px(n) (≥ 2 px); hairlines hair(n) (≥ 1 px). Never size anything in raw device pixels. Why: the GPU watchdog changes the pixel ratio, and a fractional or device-px pitch beats into moiré.
2. Anchor grids on the eye, floor((p − e) / cell), where e is the globe's resting centre. Why: a saccade never drags the grid; a screen-anchored grid crawls over the face.
3. Develop the tone before quantising: develop(l, E, Ec, cornea) pushes the dark face up by E and holds the page in the cornea in the middle greys (0.12..0.9). Why: a dark face on a dark ground dithers to nothing, and a bright reflected page blows out to a white disc.
4. Restore local contrast with lc(L, mean, g) against a few taps at ±2 cells, stronger over the cornea. Why: quantising flattens the lid line, lashes and limbus first.
5. Quantise in the era's palette, with an ordered dither fixed in screen space. Add +1/128 (bayer8) or +1/32 (bayer4) to the threshold. Why: a threshold of 0 leaves one lit cell in 64 in the blackest ink, a screen-door lattice.
6. Draw the sign (eye-interaction), clipped by almond() so the lids close over it, and faded in through the era's own dither (dissolve), never alpha. Why: a grey fade leaves a 1-bit or 4-grey palette.
7. Departure / arrival / enter (eye-era-cut) in the era's own physics.

The kit (post/medium-kit.ts)

Values used by more than one branch (kc, e, pc, R) are built once before the If chain and pinned with .toVar() (eye-tsl-craft).

Quantisation decision tree

Is the era a 1-bit or few-level display?
├── Yes → cells of cellPx(3–4), develop, lc, ordered dither, ink edges as strokes (1981, 1991)
└── No
    ├── A fixed palette (web-safe, 16-bit)? → per-channel levels with bayer4 (1995) or bit depth (1997)
    ├── A panel or camera? → a grade: black floor, gamut, curves, falloff; no cells (2000, 2003, 2009)
    ├── A component system? → a quadtree of parts sized by detail (2013, references/2013-system.md)
    └── A measured plate? → B&W plate plus instruments (2026, references/2026-measured.md)
Does a pattern sit on skin mid-tones near 50%?
├── Yes → rotate it 45°, use round spots, a finer pitch, ink spread; never a square grid at full contrast
└── No → fine

Quality gates (check every era, every change)

• Detail: at rest, brow hairs, lashes, lid fold and the page in the cornea are readable at 1:1.
• No moiré at DPR 1, 1.5, 2 and on a 390×844 phone. 1970's scanlines are never finer than 3.46 device px (phones aliased).
• No black holes: shadows never crush to pure black on dark skin; lift darks the way 1970's tone curve does.
• Eye reads as an eye: the iris stays darker than the skin round it; the reflected page never fills the iris.
• The surround is the device's physics, never a shape: no frame, ring, card, inset or vignette that reads as a border. (2013's lens vignette was set to 0 because it read as a frame.)
• Compile cost: a new branch adds less than ~150 KB of WGSL; components and helpers are Fn().setLayout functions (eye-tsl-craft).

Realism pass: what the media read, and what to re-check

Reference: references/realism-through-media.md (per-phase re-check list, per-medium floor, kit changes).

1. When a realism phase lands, re-check the branches that file lists for it, at rest and mid-cut.

2. The kit reads the eye you see:

   • the apparent iris (F.R, pupilR);
   • the elliptic limbus (F.Rx, F.Ry, limbus());
   • the twin's margin samples;
   • the real crease and brow points.

   Never re-derive a feature's position from a formula inside a medium. Why: the guessed almond() and 2026's fixed landmark offsets drift as the eye gets more real.

3. Every era's medium declares floorPx (about two cells). The scene fades micro-detail to it, so nothing speckles the ink or crawls in the rows.

4. Never retune a medium to hide a realism regression. Fix the eye, or log the conflict in the plan's drift log.

Per-era recipes

See references/era-recipes.md for every era's medium, sign and cut in one table, and the two system-scale eras:

• references/2013-system.md: the face built from UI components.
• references/2026-measured.md: the plate under live instruments.