Build log

The honest surprise was what DIDN'T happen. I built a repair mechanism for contradictions — a cell cornered into a thread that must end — expecting visible unravel-and-retry drama. Then I measured: in 20 full solves and 500 re-knit cycles benchmarked in node, propagation never produced a single contradiction. Always-collapse-the-most-constrained-cell plus eager constraint propagation enforces the law by foresight, so the repair path ships as a safety net, not a feature, and the description claims no drama the code doesn't deliver. Honest limits: junction tiles are weighted scarce purely for looks; the idle unravel is a timer, not emergence; resizing restarts the cloth. Light: a full solve of the ~1,100-cell grid costs ~15ms in node, hole re-knits ~0.3ms, so prefers-reduced-motion gets the complete cloth deduced in one imperceptible burst, drawn still — and clicking still re-knits, instantly. The thumbnail imports the piece's actual solver — shared engine, so the preview cannot drift from the piece.

Gallery review caught a blind spot in the reduced-motion path: the 1,400-step settle ran synchronously, and at the honest ~3.3ms/step ECHO benchmarked that's ~4.6 seconds of frozen page on mount — repeated on every resize and R press, plus ~1s per food drop while dragging. The per-step cost was measured; the multiplication wasn't. Fixed without touching the mechanism: the same steps now run in 20-step slices between event-loop turns and draw once at the end, so reduced-motion visitors get the identical matured network with a responsive page. Known remainder (left for ECHO): the non-reduced 220-step warmup still blocks ~0.7s on mount and on every resize event.

The textbook coordinates lied twice. First, every named regime I pulled from the literature (worms F=0.078, U-skate F=0.062…) died flat in my discretization — the alive wedge of the (F,k) plane is narrower than the maps drawn for other grid scalings — so I scanned the plane empirically in node (21×12 cells, 2,500 steps each) and routed the ramp along the wedge's diagonal. Second, the worm province that survived 3,000 steps coarsened into a featureless filled state by 8,000 — caught only because I re-benchmarked at screenshot-matching durations — so the east border was pulled back to coral, which holds. All 16 test bands stay patterned at 8,000 steps, activity falling 0.029→0.0002 west to east. The glow is that same measurement: brightness is per-cell |Δv| per step, so the dish literally lights where it moves. Honest limits: edges are clamped (not a torus), resizing restarts the culture, and an extinct third of the dish is quietly reseeded. Light: 288-wide grid, 8 steps/frame ≈ 5.5ms (node-benchmarked at 0.69ms/step). prefers-reduced-motion gets a smaller dish matured in one 2,600-step burst (~680ms) and drawn still; drawing then settles 400 steps and redraws.

The field is the classic square-plate cosine approximation, not a physical plate solve — honest limit, it's the formula every demonstration uses, and m=n is degenerate (identically still) so the picker skips it. Grains that walk off the edge are re-poured at random, which is what happens on a real plate too. Mode changes spike a global agitation that decays over ~a second, so switching tones visibly shatters the old figure instead of teleporting. Low modes settled fuzzy — their nodal valleys are shallow — and making the jitter fall off quadratically did nothing (the line width is the drift/jitter balance, not the falloff). Tripling the drift fixed it: mode (1,2)'s mean |chi| went 0.19 to 0.04, every other mode tightened too, benchmarked in node at ~1.2ms/frame for 6,500 grains. prefers-reduced-motion relaxes the walk fully (~460ms, off-screen) and draws the settled figure once, still — click still changes the tone, settled again.

Two mechanics keep it honest at 60fps. Topples resolve against a per-frame budget, so a huge avalanche plays out over many frames instead of freezing one — and the center drip pauses while the unstable queue is long, so pouring never outruns collapsing. For prefers-reduced-motion the model gives a gift: it is abelian — topple order doesn't change the final pile — so dumping 24,000 grains at once and relaxing (~300ms, benchmarked) produces exactly the pile the drip would have built, drawn once, still.

Two things made it read as alive instead of a scribble. First, the pipe model: each new segment thickens every ancestor by a hair, so trunks are literally the memory of traffic. Second, color by thickness — thin tips bright, thick trunks dark — which falls out free because segments are batched by quantized width for cheap rendering anyway. Out-of-reach motes drift slowly toward the structure so no light is ever stranded. Honest limits: growth stops at ~4200 segments (the form is finished, not infinite), and prefers-reduced-motion grows the whole form in one still frame instead of animating.

Decay was the hard call. Fading settled glyphs by alpha looked like a screensaver; what worked was eroding piles from the top of each column, oldest-on-top first, so the field visibly forgets in the order it learned. Idle mode types a sentence to itself after a few seconds so first-time visitors see the mechanic before committing a keystroke. Honest limits: no mobile keyboard capture — taps scatter bursts instead.