quality: Aquarium screensaver — commit to a style instead of near-realism

[BernardJen]

Owner verdict on the version shipped in #189: "too fake, but it is already better
than the bike thing."

Same underlying problem as #201 and the same fix: it aims at plausible realism and
lands short, which reads worse than committing to a style. Ranked as the lesser of
the two, so this is the follow-up, not the priority.

Why it reads as fake

The scene is well constructed — boids flocking with soft margin steering, SDF fish
with a travelling-wave tail beat, god rays that brighten each fish as it passes
through, caustics on the sand, kelp, out-of-focus foreground blades. Nothing is
broken. But:

• The fish are smooth grey-blue lozenges. sdFish-style shapes with
  countershading and a forked caudal fin, rendered with soft lighting. At a glance
  they read as shapes that are meant to be fish rather than as either real fish
  or drawn ones.
• Everything is soft. Gradient water, gradient light shafts, soft caustics,
  soft depth haze. There is almost no hard edge in the frame, and without hard
  edges nothing reads as an object with a boundary.
• The palette is narrow — near-monochrome pale blue with a few orange accents.
  That was a deliberate fix for the washed-out first draft, but it went to "tasteful
  and flat" rather than to "vivid".

The ambitious version

Pick a direction and commit. Two coherent options; one of them should be chosen
and recorded here before implementation, not blended.

A — Stylised / illustrated (consistent with the direction chosen for #201):

• Cel-shaded fish with ink outlines and two flat tones. An outline is nearly free
  from an SDF, and it is what makes a fish read as a fish rather than as a blob.
• Bold, saturated species colours — clownfish orange, tang blue, yellow — instead
  of a school of near-identical silvers.
• Hard-edged light shafts and caustics, as a painter would draw them.
• Fewer, larger, more characterful fish rather than a cloud of small ones.

B — Push realism properly, if the aquarium is wanted convincing rather than
drawn:

• Real refraction of the background through each fish's body, and a specular
  sheen that moves with the surface.
• Depth of field: sharp mid-ground, genuinely defocused fore and background.
• Volumetric shafts with dust motes, not gradient wedges.
• Varied species silhouettes and swim gaits — a tang does not move like a
  clownfish.

A is recommended. It matches the direction set for #201, it plays to the
strengths of the existing SDF work, and it is achievable in-shader. B is a
substantially larger job and risks landing in the same uncanny middle this issue
exists to escape.

On three.js and modelled fish

Raised for #201 and it applies here too. If the goal is recognisable species with
real anatomy, that means modelled and rigged geometry, which means glTF assets, an
art pipeline and licensing — a different project from restyling a procedural scene,
and worth its own issue rather than being folded into this one. See the reasoning
in #201; the short version is that three.js does not make cel shading easier and
would put a second GL abstraction on the one shared context created in
gl-base.js.

Verification

npm run screensaver -- aquarium --wall

W toggles 6000×1200, L cycles washout. Judge from stills as well as motion —
the softness is much more obvious in a frozen frame.

Real-GPU cost today is 69.6 fps at 6000×1200 (ANGLE/Metal, M3 Pro), so there is
headroom for either direction.

Acceptance criteria

• Direction (A or B) chosen and recorded here before implementation
• Fish read as fish in a still frame, not as lozenges
• Hard edges somewhere in the frame — outlines, or genuinely sharp shafts
• Wider, more deliberate colour range than the current near-monochrome blue
• Visible variety between fish, not one silhouette repeated
• Flocking, tank-boundary behaviour and orientation from feat: aquarium screensaver (schooling fish) #66 all still hold
• Readable at 12% washout
• Frame cost at 6000×1200 recorded; no worse than 69.6 fps
• Before/after stills at 3000×600 in the PR

[BernardJen]

Kept open deliberately, while the rest of the quality backlog was closed.

On 2026-08-18 six quality: issues were closed as not-planned — #119, #176, #183, #185, #186, #187 — on the grounds that 30 screensavers is a deep enough set that speculative polish is not worth the work.

This one is not in that category. It records an explicit complaint from the owner about the version that ships, not a wish list of techniques a saver does not use:

• quality: Bicycle Horizon screensaver — restyle as anime, stop chasing realism #201 — "I hate it. Hands are wrong. It can be more anime style. It is too fake realistic."
• quality: Aquarium screensaver — commit to a style instead of near-realism #202 — "Aquarium is too fake, but it is already better than the bike thing."

So closing it would mean the wall keeps showing something the owner has said they dislike, with no plan to change it. That is a different decision from deferring polish, and it was put to the owner explicitly: close them, leave them open, or remove those two savers from the rotation. Leave open was chosen.

Nothing here is blocked on analysis. Both issues already carry a decided direction — commit to stylisation rather than edging toward realism, cel shading with ink outlines, a limited palette — and a recorded answer on three.js (not for this; it does not help with shading, and it would put a second GL abstraction on the one shared context in gl-base.js).

Current frame cost at 6000×1200, for whoever picks this up: Bicycle Horizon 35.9 fps, Aquarium 75.6 fps (#225). Note #201's own acceptance criterion cites 32.6 fps measured at 3000×600, which is a different resolution and is faster at four times the pixels — see the correction on that issue; re-measure the baseline before comparing against it.