fix: stop Wave Tank damping the interference out of existence (#184)

[BernardJen]

Closes #184.

The damping was the bug, and not by a little

The band was 0.972-0.984 per step. At 240 steps/s that leaves 0.5% of a wave after one
second. The scheme's CFL number is 0.5, so a disturbance moves ~0.707 cells/step and
crossing a 2048-cell tank takes ~2900 steps — after which 0.978 leaves 1e-28 of the
wave.

Interference is this saver's entire subject and it was being annihilated twenty-eight
orders of magnitude before two ripples could meet.

damping	half-life	amplitude after one 12s crossing
0.978 (old)	0.03s	1e-28
0.99935	4.4s	15.2%
0.99955	6.4s	27.2%
0.99972	10.3s	44.4%

Settled at 0.99935–0.99955, the low end — a balance, not a maximum. At 0.99972 waves
outlived three crossings and the tank accumulated into exactly the uniform chop this issue
also complains about: fine at 20s, over-energised by 45s. The derivation is in the file.

A latent bug that only existed on the wall

MAX_CELLS was clamped per axis independently, so at 6000×1200 the grid came out
1024×300 — 3.41:1 against a 5:1 canvas, stretching every ripple 1.46× horizontally.

It cannot be seen at 3000×600, which never reaches the cap, so no capture would ever have
shown it. Both axes now scale by one factor.

The rest of the issue

• Resolution: CELL_PX 4 → 2, MAX_CELLS 1024 → 2048. The old comment argued a
  quarter-res grid "looks identical" — true at 1080p; a 4× upsample at 6000×1200 is soft.
• Wall-clock stepping. Two-steps-per-frame made wave speed and decay depend on frame
  rate — the same tank ran twice as fast on a 120Hz panel. Steps now derive from dt.
• Caustics from the Laplacian of the height field: curvature is what converges refracted
  rays, and the four neighbours were already fetched for the gradient. This is the single
  change that makes it read as water.
• Refraction — the floor is sampled displaced by the surface gradient, over a tiled bed,
  because distortion is only visible as distortion of something.
• Fresnel mix into a sky gradient, so a glancing surface mirrors instead of showing floor.
• Choreography: one event at a time, separated by 6–14s of quiet. A single large drop; two
  sustained in-phase sources for fringes (fringes are steady-state — an impulse cannot make
  them); or a line source sweeping in y to build a plane wave.
• A barrier with a slit on 40% of activations, so diffraction is on show.

The barrier was broken, and only forcing it on revealed it

Three seeds in a row happened not to enable it. Rather than report a feature I hadn't seen,
I forced it true — and nothing rendered. Two bugs:

1. the drawn wall was sized from uTexel, making it ~3 pixels wide;
2. the slit term was inverted — it painted the gap dark and left the wall transparent.

The simulation barrier was correct all along, so the physics worked and was invisible. Both
passes now share BARRIER_HALF_W. Verified by measuring the darkest interior column:
0.111 → 0.035, with the slit and downstream diffraction visible.

Measurements

• Dynamic range under 12% washout: 0.020 → 0.812. The old surface had essentially no
  contrast to wash out, which is the "soft oil slick" quantified.
• 120.1 fps at 6000×1200, vsync-capped, against main's 120.0 — despite 6.7× the cells.
  The stencil is bandwidth-bound and was never the cost.
• npm run lint clean, npm test 407 passed, npm run shadercheck ok=150 fail=0.

Acceptance criteria

• Simulation resolution raised; no visible upsample blur at 6000×1200
• Damping reduced so waves cross and reflect; interference clearly visible
• Stable damping/speed band documented in the file, with the derivation
• Caustics rendered on a floor below the surface
• Refraction and Fresnel reflection; real value contrast (0.020 → 0.812 under washout)
• Drop choreography with distinct events, not uniform chop
• At least one structured demonstration — two-source fringes, plus a slit barrier
• Readable at 12% washout
• Frame cost at 6000×1200 recorded: 120.1 fps
• Before/after at 3000×600

Not verified / honest limits

• It is energetic rather than "mostly quiet". With damping that lets a wave cross a 5:1
  tank — which this issue explicitly asks for — waves live 10–25s, so events every 6–14s
  necessarily overlap. Genuine calm would need gaps longer than the wave lifetime, leaving
  the wall flat most of the time. Individual wavefronts are traceable across the full width,
  which is the property that matters; "quiet" is not achievable alongside "crosses the tank".
• Wave Tank still measures 0 at frames 5/12, because the first event lands at 0.15s. Its
  baseline stays 0, so it remains one of the two documented savers the structure check skips.
  This PR does not change that, and deliberately does not force an early event just to make a
  number non-zero — a tiny density would sit under the absolute margin anyway.
• The two-source fringe pattern is visible in motion; a still frame catches it mid-formation,
  so the captures show interference rather than a textbook fringe photograph.