# What was measured

Nothing here was drawn by eye. The reference video was cut at 10 frames per
second, then each state measured off the frames: silhouettes by sub-pixel ray
casting, eyes by capsule fitting (PCA), colours and stroke widths by direct
sampling.

**The numeric constants of the bot are measurements, not settings.** Gaze angles,
eye sizes, radii, timings, colours: all of it comes from frame-by-frame analysis.
Rounding them off, simplifying them, or replacing them with values that look
tidier breaks the resemblance, which is the only success criterion here.

## Verified traps, not to be "corrected"

- The eyes lean like `\\`, not `//`.
- The body is a **perfect circle**, not a squircle, radial deviation under 0.7%.
- Transitions are **exponential ease-outs**; the body **never** overshoots. The
  one spring is local and written into the state concerned (the notification
  pastille's +14% pop). There is deliberately no spring engine.
- The vertical `!` has a **tapered** bar (top/bottom = 1.76); the leaning `!` has a
  **capsule** bar. They are not the same shape.
- The leaning `!`'s dot is a **teardrop**, round end towards the bar, point away,
  not a disc.
- In the comet state the dot **does not move**: the trail orbits it.
- At rest the avatar **doesn't float**: the video measures the centre stable to
  ±0.003. The code keeps a deliberate trace of it (a drift of a few thousandths of
  the radius, and a 0.5% breath) purely so the image isn't completely frozen. The
  visible life is gaze drift and blinking. Don't add a float on top.

## The eyes live on a sphere

The eye nearer the edge is about 0.69 times the width of the other and 0.663 times
its area, exactly the depth factor of a point on a sphere at that distance from
the centre. So each eye takes the sphere's tangent frame, projected
orthographically: the compression, the tilt and the passage behind the limb all
follow on their own.

The gaze poses (`REST_GAZE` and the per-state `gaze`) come from fitting that model
to the measured positions, with a residual error of about 1 px on a 190 px ball.

Two notes for anyone comparing numbers. The far/near width ratio appears as 0.69 in
`face.ts`, as 0.674 in `face.test.ts` (the raw video figure) and computes to about
0.708 from the model: source versus fit, and the test's tolerance covers the
spread. Likewise the rest tilt is not a constant anywhere: it emerges from
`REST_GAZE` through the tangent frame, at about 26° off vertical.

## Every shape change is hidden by a blink

That's the morph-damping mechanism in the original, reproduced by `blinkIn` on the
states concerned. The forced blink lasts 0.2 s; the scheduled idle blink is
`BLINK_DUR = 0.18`.

## Regenerating the profiles

`src/bot/profiles.ts` is generated from the video's frames. Don't edit it by hand.

```bash
mkdir -p frames
ffmpeg -i reference.mp4 -vf fps=10 frames/h_%04d.png
pip install numpy pillow
python tools/extract-profiles.py frames/ > src/bot/profiles.ts
```

The script composes exact filenames (`h_0164.png` and friends, zero-padded to
four), so the `h_` prefix and the `fps=10` cut both matter. It extracts the three
profiles that can't be built analytically (egg, hexagon, triangle), not all 14
states; the rest are either the measured circle or constructed in `skins.ts`.

`reference.mp4` and `frames/` are local inputs and are not in the repository.

## The favicon is not an approximation

`public/favicon.svg` is not a lookalike drawing: the circle and **both eye
matrices** are what `engine.sample(1)` returns for `idle`, byte for byte. Hence the
right eye being narrower than the left (0.64 against 0.87). That's depth
compression, not a typo.

`favicon.ico` and `apple-touch-icon.png` are rasterised from it. The `.ico` is
still needed: Safari only reads the SVG from version 26, iOS not before 18.7.