feat(registry): filmed-screen curves, orbits and follows typing by default - #5499

Post-merge review of 433fe03e, which landed as 3886aaec: same parent, and the three files' patches are identical. No blocker. I'd have approved it. Two small clamps are worth a fix-forward, and nothing needs reverting.

1. Curve normalisation is correct.

• s = s0·(1 + k|s0|²) / (1 + k|corner|²) maps the frame's corners to themselves.

• For every pixel in frame, |s0|² ≤ |corner|², so the factor is ≤ 1. Each sampled ray therefore lies on the segment from the centre to the flat ray, and the curve can't show any of the page that the flat shot didn't. For k ≥ 0 the map is monotonic in radius, so nothing folds.

• What it costs is a zoom. At 16:9 with the default 0.15:

  • the centre is magnified 1.156×;
  • the top and bottom edge midpoints are pulled in about 10% (0.897);
  • the side midpoints are pulled in about 3%.

  (At 9:16 it's 1.049× in the centre.) So a framing that fit flat can lose content near the top or bottom centre edge, and this stacks with the shot's zoom. It's worth one line in the README.

• Should-fix: clamp curve at 0. A negative value makes the factor > 1, so it samples outside the flat frame, and at 16:9 the map folds below about −0.32. Math.max(0, …) where the uniform is set, or saying ≥ 0 in the docs.

2. Computing the vignette on s0 is the right choice. Vignetting is the camera's falloff in image space. On the curved s it would weaken toward the centre and change shape with curve. With curve: 0, s0 == s, so vignette: 0.42 reproduces the old look exactly.

• The new default is much stronger, though. At 16:9 the corners go from 62% to 10% brightness and the side midpoints from 76% to 43%. On a white page that reads as a lens; on a dark page the corners go black.
• Should-fix: above vignette ≈ 1.11 (16:9) the factor goes negative at the corners. col then goes below zero before the scanline and grain terms are applied. col *= max(0.0, 1.0 - uVig * dot(v, v)); fixes it.

3. Changing the defaults is acceptable without an opt-in.

• The registry copies the snippet into the project (compositions/components/filmed-screen.html), so existing projects keep their copy and render exactly as before. Only a fresh add gets the new look.
• #5487 landed earlier today, no release tag contains it, and the item is marked experimental, so almost nobody has it yet.
• One caveat for the README: passing the old values doesn't fully restore the old look. The old values are blur 0.9, rack 0.7, vignette 0.42, aberration 1, led 0.9, curve 0. But the LED fade-in threshold (cells spanning 2.2–4.5 px, now 4.0–6.5) and the quarter of fringing on the focused element are hard-coded. aberration also scales the cut tear's fringing, so tears are 2.2× too.
• A short "Previous look" line with those values would help anyone who liked #5487's version.

Also checked

• tiltTo unset gives lerp(a, a, x) == a exactly, so the rotation is unchanged.
• A glide evaluates the previous shot past its end, and smoother clamps, so it starts from that shot's tiltTo. Orbits chain correctly.
• The LED grid's new fade-in (only once a cell spans ~4 output px) only fades the grid in later, so on light pages it can only alias less.

Tests and CI. filmed-screen.test.mjs passes 6/6 locally at 433fe03e. All 39 checks are green, and the required set is green with the Windows jobs skipped by path.

No render from me, so I haven't looked at the moiré or the lumen demo on screen.

— Rames