feat: add the Aquarium screensaver (#66)

[BernardJen]

Closes #66

What this is

An aquarium: two or three shoals of fish sweeping a wide tank, a few large slow
independents that the shoals scatter away from, and a scene built around the
light. Asset-free — the fish are an SDF, the plants are analytic curves, the
caustics are a Voronoi web.

Why it looks the way it does

It is boids, but three things had to change before fish read as fish.

1. No wrapping. boids.js lives on a torus. A fish that teleports across
   the tank stops being a fish instantly, so the walls here are soft steering
   forces ramping quadratically through a margin band, with the fish's own turn
   rate boosted while it banks away. The hard clamp at the glass exists only so
   a pathological frame delta cannot put one outside; in normal running it is
   never reached.
2. Exhaustive neighbours, not sampled ones. boids.js draws 32 random
   neighbours out of 4,096 each frame, so the forces flicker (quality: Boids screensaver — spatial hash, oriented birds, and stop the flock tearing #127). For
   abstract dots that reads as energy; for a school it reads as broken. The
   shoal here is deliberately small — a fish is ~40x the screen area of a boid,
   so a few hundred already fill a tank — and at that size every fish tests
   every other fish. 676² = 457k interactions/frame is less work than the bird
   flock does, and there is no sampling noise left to filter.
3. Steering, not force integration. Velocity is a unit heading times a
   cruise speed, and the heading is a rate-limited low-pass of the desired
   heading. That is what a fish physically does, and it means the simulation
   cannot explode: speed is assigned, never accumulated.

The light is the trick. One family of functions — causticWeb() and
shaftField() — lights the surface, the god rays, the sand and the fish, so a
fish visibly brightens crossing a shaft and dims on the far side. That shared
sample is what sells "underwater"; without it the fish look pasted onto a blue
background.

Layers, depth-ordered into one HDR target: water column + surface shimmer +
god rays + caustic-lit sand + kelp silhouettes → the shoals (instanced oriented
quads, sorted far-to-near by construction) → large fish → bubbles, marine snow
and out-of-focus near kelp → bloom / ACES / dither.

Ambient light is answered with a real lightness range (OKLab L runs 0.10 at
the floor to 0.72 under the surface) plus a modest exposure lift on wall-sized
canvases. No flat palette * 0.05 black-level tint.

Things worth knowing, found the hard way

• createGLRuntime() does not size the canvas. It only calls resize()
  inside runtime.start()'s loop, so canvas.width/height are still the DOM
  default 300×150 in start(). Every saver that reads canvasAspect(canvas) or
  particleSide(canvas, …) there — including boids.js, whose comment says
  "createGLRuntime has now sized the canvas" — is computing them for a 2:1
  postage stamp. On the wall that makes the simulation's world space 2.5x too
  narrow. This module calls runtime.resize() explicitly; I have not touched
  the others, but they look wrong.
• orientedQuadOffset must be applied in pixel space. boids.js folds
  uQuadScale into the size before rotating, which scales the rotated
  vector's y component by the x conversion factor — a 5x shear on the wall,
  invisible at 16:9. This module converts after rotating.
• Big bright regions fog, they do not glow. The surface band covers ~15% of
  a 5:1 canvas; with a six-level bloom pyramid it turned the entire frame into a
  milky wash. The threshold is now set so only the ripple crests, the caustic
  filaments and the bubble rims cross it.
• Orientation flip. orientedQuadOffset's perpendicular axis has the sign
  of dir.x, so an asymmetric fish swimming left renders belly-up. The vertex
  shader passes sign(dir.x) through and the fragment shader multiplies local y
  by it.

Acceptance criteria

• Fish school with visible flocking behaviour, oriented along travel
  direction — separation/alignment/cohesion with a long cohesion radius,
  plus a saturating pull toward each shoal's own wandering target. Heading
  comes from the velocity and the sprite is built on it, so orientation is
  correct by construction; the large fish go further and take velocity as
  the exact analytic derivative of their path.
• Fish stay within the tank without visible wrapping or teleporting —
  soft margin forces, no torus anywhere. Checked at t = 15/25/70/110/150/180/240s.
• At least one background layer beyond flat colour — four: OKLab water
  gradient, sheared god rays, caustic-lit sand floor, swaying kelp
  silhouettes. Plus an undulating surface and a near-focus foreground.
• Composition fills 6000×1200 as a wide tank — shoals are assigned
  separate lanes down the tank and their targets sweep it over a couple of
  minutes, so there is a subject and negative space rather than uniform
  texture. Verified at 3000×600 (wall aspect, half res) and 1920×1080.
• Fish readable as fish at 12% washout in wall mode — captured at
  --washout 0.12; the forked tail, dorsal fin and eye all still read.
• Appears in the preview list and the random rotation — one import and
  one array entry in registry.js.
• stop() releases GL resources; no leak across repeated start/stop
  cycles — five programs, two instanced-quad VAOs, the ping-pong pair, the
  attribute texture, the post chain and the runtime are all released and
  nulled. shadercheck drives create/start/stop 5x per saver.

Verification

npm run lint          pass
npm test              285 passed (23 files)
shadercheck           135 runs, 0 fail

structure-baselines.js gains one line ("Aquarium": 0.085), set below the
0.0989 and 0.0926 measured across two full harness runs so seed variance cannot
false-positive. npm run baselines was not run.

Frame cost: the issue does not ask for a number, and the only figure
available here is meaningless for the wall — 2.0-2.8 fps at 3000×600 under
SwiftShader, which is a CPU rasteriser. For scale: the whole scene is five
fullscreen-ish passes plus 400-676 instanced quads, and the O(N²) flocking is
457k interactions/frame at the wall cap, against boids.js' 131k stochastic
ones at 4,096 boids. Needs a look on real hardware before it ships.

Screenshots

Captured with the headless harness at the wall aspect and at 16:9, several
seeds, with and without the ambient-light wash, at times from 12s to 240s.