What changed
Changes made to the port after design reviews with the original author of the Java program, most recent first.
Seventh round
- An image per plane, placed by hand, replicated by the symmetry. The texture layer becomes decals: a picture is placed on one plane class — the plane the diagram shows — and the stellation symmetry does the rest, the face wearing as many copies as it has symmetries (a cube face four under O, eight under Oh, mirror images among them) and every plane of the class the same picture. The engine is
planeOrbits: each orbit's chart comes with its representative's stabilizer as 2×2 maps of the chart, and a decal's copies are its placement through each of them. The copies reach the fragment shader as uniform tables sized to the GPU's budget, the vertex data carrying only chart coordinates and the orbit's slice, so a drag is one uniform upload. Overlaps are sorted by depth the sympix way — a stack level plus a tilt in the image's own frame, so symmetric copies weave over and under each other without a draw order — and exact ties are mixed as coset colors are, in Oklab under the union of coverages. Place by hand puts SymmHub's transform rig on the copies in the diagram (move, turn-and-scale about a pivot, the pivot, and a tilt arrow); a gesture on any copy writes back through its stabilizer element. The diagram shows the layer through the same shader, read back from an offscreen pass. A symmetry change carries the decals across losslessly — copies become decals under a smaller group and fold back under a larger. Documents carrydisplay.textures, per representative plane with defaults absent; the old one-image form opens as a decal on every plane. lib/core.js · lib/decals.js · lib/decalrig.js · lib/texmix.glsl.mjs · lib/render3d.glsl.mjs · lib/render3d.js · lib/diagram.js · app/js/worker.js · app/js/app.js · app/js/preset.js · test/decals.mjs · test/docformat.mjs - Two layers, each its own: the replace blend, the ink palette, texture opacity and tiling. The texture becomes a full second layer over the faces, independently colored and independently transparent. replace joins the blends: the image IS the face, its alpha the surface's, so a glyph on a transparent ground floats in the air — the renderer routes the whole surface through the back-to-front compositing path, since every pixel now carries its own coverage. opacity is the layer's own dial. tile the image chooses repetition over the wings (patterns) or one copy per face, clear beyond its edge (emblems) — one premultiplied step function in the shader. And the ink palette: while a texture is armed the Colors panel gains an editing switch, and its texture ink rows are a second color and opacity per group — the same override machinery under
tex.-prefixed keys, defaulting to white — multiplied into the sampled image before it meets the face. So a solid red face carries green arrows and a solid green face red ones, each layer coset-colored, neither consulting the other. All of it ridesdisplay.texture(blend/opacity/tile/colors, defaults absent), release 1; the ink tints reach the shader as a second color attribute, so blends and gray mix in Oklab exactly like face colors. lib/render3d.js · lib/palette.js · app/js/colors.js · app/js/app.js · app/js/preset.js · test/colors.mjs · test/docformat.mjs - Textures honor transparency, and learn to stamp. A texture with an alpha channel used to show its transparent background as solid black — the shader multiplied by the texel's rgb and never asked about alpha, and a transparent pixel's rgb is meaningless (usually zero). Now texels ride PREMULTIPLIED — decoded through a canvas, whose backing store is premultiplied by definition, ending every browser's createImageBitmap ambiguity, and filtered and mipmapped fringe-free — and transparency means what it should: the face color shows through untouched, semi-transparent pixels blend by their alpha. And because multiplication can only darken, a white glyph on a transparent ground (the bundled arrow, the SVG figures) was invisible under tinting by construction — so a blend menu joins the texture controls: tint multiplies as before, both formulas one line in premultiplied space, and stamp lays the image over the face in its own colors — the arrow as engraving or as sticker. Saved as
display.texture.blend, written only when stamp, so tint documents write the same bytes as before. lib/render3d.js · app/js/app.js · app/js/preset.js · test/docformat.mjs - An image on every face, replicated by symmetry. The new texture menu beside the merge dial lays one image over the solid: each plane orbit keeps a single chart, drawn on one representative face — origin at the face's pole, one tile spanning the original polygon — and carried onto every other plane by a group element, so every face wears an exact symmetric copy, mirrored where the carrying element is a reflection. The chart belongs to the plane: the picture runs unbroken across a face's facets and repeats outward over the star's wings; scale sets the tile size in face widths. The sampled image is MULTIPLIED under the face colors before lighting, so every coloring on the books tints it — with cosets, each coset colorizes the same picture in its own color — and merge neighbors and painted labels compose for free, because they only ever changed the tint. Charts ship with the mesh as nine numbers per plane and the renderer derives uv per corner, which is why the split mesh needed nothing at all. Textures live in img/textures/ (five light, tileable starter patterns generated by tools/make-textures.mjs, which also writes the menu's index.json from whatever the folder holds); the choice and scale are saved as
display.texture, no release bump. The transported-chart identities are pinned by test/texcharts.mjs — isometry, pole-centering, one scale per orbit, exact symmetric copy, and that Oh's 24 reflections land their copies mirrored. lib/core.js: planeCharts · lib/render3d.js · app/js/worker.js · app/js/app.js · app/js/preset.js · tools/make-textures.mjs · test/texcharts.mjs - The brush reaches the solid. With the brush down, a plain click on the 3D view paints the facet under the pointer — the same stroke a diagram click makes, by the same region identity, expanded over the same orbit with the same permuted cosets. Turning the solid is untouched: a drag is still the trackball, and only a click whose pointer has not travelled counts as a stroke, so painting and orbiting share the left button without stepping on each other. Hovering with the brush down outlines the facet neutrally — no green or red, since the gesture promises neither adding nor carving — and the corner line still names the coset under the pointer. Each surface facet's paint identity now rides along with the mesh, so a click needs no second trip to the worker to learn what it hit. Also fixed: the palette bar's explicit
displaywas beating thehiddenattribute, so the bar never visually left when the brush was put away. lib/render3d.js · app/js/app.js · app/js/worker.js · css/style.css - Painting cosets: the assignment becomes yours, and the symmetry carries the brush. The coset coloring computes which facets can share a color; painting decides which ones do. A brush button on the diagram opens a palette of the current coset colors — the ones the Colors panel edits — and a click paints not one region but its whole orbit under the full group, each copy taking the coset the chosen one becomes under that symmetry. So painting one face solid green turns every face solid in its own coset: the four-color model, made by hand, agreeing exactly with the per-plane reading — a property the tests pin, along with its converse, that painting a region with the label it already wears changes nothing anywhere. The permutation is never computed as group theory: the labeling is already equivariant, so a label's image under g is read off a witness — the computed label of g's image of any facet wearing it. It is a label, not a tint: readouts name it, both exports carry it, the diagram list splits planes a paint has made different, and merge neighbors votes with it — though while the brush is down the merge is set aside, because its majority vote quietly outvoted fresh paint and the brush looked dead. shift-click adds the chosen coset to what a region wears, or takes it out; a region carrying several cosets — by gesture, or forced where a copy's own symmetry straddles them — shows their blend and a small dot, and falls to gray when mixing is off. gray paints no-coset orbit-wide, auto erases orbit-wide, clear takes everything back. Saved with the document (
display.cosetPaint, keyed by the arrangement's own plane.facet order, no release bump); cleared when the coloring subgroup changes or a different arrangement is built. Painting on the solid comes next. lib/core.js: paintOrbit · app/js/app.js · app/js/worker.js · lib/diagram.js · test/paint.mjs - Merge neighbors: the per-facet readings get a legibility dial. Coloring by cosets or orbits per facet is exact and often unreadable — inside one big flat face the classes sit side by side in enumeration order, and the face reads as confetti. The new merge neighbors switch in the coloring subsection melts classes whose facets touch inside one face into solid regions, smallest first, each absorbed into the neighbor it shares the most boundary with; the colors count says where to stop, from every-patch-solid up to the untouched reading, and the readout says what was kept of what there was. Regions wear the color most of their facets already wore, so full merge degrades gracefully toward the per-plane picture, the palette keeps its meaning and edited colors keep applying. The merging runs on the QUOTIENT by the figure’s full symmetry group: its atoms are the finest pieces both symmetric and single-colored, one step of the dial merges every copy of a face at once, and majorities are taken face by face — with an exact tie worn as the blend of the tied labels, the one rule symmetry cannot tell apart on different copies. All of that was learned from a four-coset figure under D3d(O), which is not normal in Oh: the first version merged global label-classes, whose members entangle across faces for a non-normal subgroup, so the whole surface chained into one component and majority-voted itself a single color — and its one-at-a-time greedy merged the copies of a face at different moments, so every intermediate setting was lopsided. Both fixtures are pinned in the tests, beside the compound of five tetrahedra under I/T, whose point-touching spikes must never weld — adjacency is geometric: same plane, same side, a shared boundary of real length. Saved with the document (
display.colorMerge, no release bump); costs nothing when off. lib/core.js · app/js/worker.js · app/js/app.js · app/js/preset.js · test/merge.mjs - A perspective camera that leaves the figure the size it was. Perspective is normally a different camera, and switching to it resizes and re-frames everything at once. This one is a continuous deformation of the parallel camera instead. Put the eye at distance d and let p = R/d, R the figure’s own radius: p = 0 is the eye at infinity, p = 0.1 an eye ten radii out. The unit is the radius of what is actually selected, so the eye stands off in proportion to it and a bare core and a stellation eight times its size look alike at the same setting — at the price of the eye moving as the figure grows, which shifts the interior on screen. R was the polyhedron’s core radius at first, which measured the same absolute distance whatever was built, and put the eye inside a far-reaching stellation’s own spikes: they came through the screen like a fisheye. Choosing the focal length so the plane through the centre keeps its parallel size collapses the whole projection to a depth-dependent magnification about the view axis, m(z) = 1/(1 − p·z/R).
That is what makes it usable here. m(0) = 1 for every p, so turning perspective up never resizes the figure — it leans the near parts toward you and the far parts away, about a plane that stays put. And m ≡ 1 at p = 0, analytically, so there is no jump leaving the parallel camera: the slider starts at orthographic and grows out of it.
Navigation stays in the image. Pan slides the finished picture, zoom scales it, and fit changes only that zoom — none of them move the eye, so the perspective you chose survives being looked around. Nothing needs capping: the eye sits at one radius over p, which is outside the figure at every setting on the slider. Picking inverts the same formula, so what you click is what you point at, and the camera travels in the saved view as a ninth number that is absent whenever it is zero.
The translucent pass had to learn the same lesson. It orders facets plane by plane — everything on the eye’s side of a plane is nearer than everything beyond it — and it decided which side that was from the normal alone, which is an eye at infinity answering about a direction. With the eye brought in, a plane’s own offset matters, and the planes that pass between the eye and the centre came out exactly backwards. Using the eye as a point rather than a direction costs one term, and it is the same p/R the projection uses, so the two cannot drift apart. On the icosahedron at an eye ten radii out the old rule was wrong for planes within 4.6° of edge-on: better than half of all orientations had at least one.
- Point at a facet and it says which cosets it belongs to. A blend answers “which cosets?” with a color, and a color is the one form of that answer you cannot read back — three colors mixed in Oklab is not three colors any more, it is one muddy one. A label now follows the pointer in the solid and in the diagram, naming the group under it: coset 3, or cosets 1 + 8 for a mix, with the mixed color and the colors that went into it swatched beside the numbers, or no coset fits where nothing does.
A table was considered and does not survive contact with the data. Mixes are not pairs: the icosahedron alone reaches six-way and ten-way mixes, and a single figure can carry hundreds of distinct combinations, so there is nothing to lay out in a grid and no legend that would hold them. Naming the one under the pointer costs a line, works for a mix of any size, and is asked exactly where the question occurs.
The solid’s hover pick deliberately runs only while a modifier is held — it drives the cursor and the outline, which are promises about what a click would do, and a bare hover promises nothing. So the label asks its own quieter question, touching neither, throttled to 40 ms because it now runs on every mouse move; one pick measured 0.36 ms on a three-shell icosahedron.
- Elements belonging to several cosets can be told to look it. A plane can belong to more than one coset of the chosen subgroup at once, and then it has no color of its own. Mixing its cosets is a fair picture of that, but the mix is a color like any other, so the fact was easy to miss: a muddy facet reads as a facet somebody chose a muddy color for, not as one the coloring cannot name. A mix cosets colors switch now sits beside the coloring menus. Turned off, every such element falls to gray — which nothing else in a coset reading wears — so what is genuinely uncolored is visible at a glance.
It lives in the palette itself, which is the one place the solid, the diagram, the diagram’s SVG and the mesh exporters all ask for a color, so all four follow it without knowing it exists. Gray arrives through the same door as any other group, so a gray given its own color in the Colors panel still wins. The switch is offered only for the coset reading: an orbit puts every element in exactly one orbit, so nothing is ever shared there and a switch would have nothing to do.
- The coloring menu asks two questions instead of nine. Nine modes in one list made the reader do the work: six of them were the same two readings — cosets of a subgroup, orbits of a subgroup — repeated once per thing that can wear the color, and the repetition buried the choice that actually matters. The first menu now names the reading — by shell, by face class, by stellation face class, by subgroup coset, by subgroup orbit — and a second, applied, says what wears one color.
The second menu appears only for the two readings that have a choice to offer, and they offer different things: a coset labeling can be split where a mirror cuts a facet, an orbit labeling can paint whole cells. It also remembers each reading’s last choice, so going from coset to orbit and back does not quietly reset it. Nothing changed underneath — the nine values are the same nine, so documents, links and stored preferences all read exactly as before, and an old document opens with both menus set to what it saved.
- An empty selection still shows the symmetry. Deselect every cell and the symmetry elements and the coordinate arrows went with them, which is backwards: with no cells in the way, the axes and mirrors are the whole picture, and looking at the symmetry on its own is a reason to clear the selection rather than an accident. Two things were wrong. The frame returned early on an empty mesh, taking the ink pass — edges, elements, arrows — down with it; and both were sized to the mesh’s own radius, which an empty mesh does not have.
They now fall back to the radius of the core — layer 0, the polyhedron you started from — which is the one length in the arrangement that does not depend on what is selected. So the empty figure has a stable size, the same whatever was on screen a moment before, and the elements sit at the scale the whole figure is read against. A selection still sizes them itself, exactly as before.
- The symmetry elements can be any subgroup’s. Solid › symmetry elements gained a group menu. The axes, mirrors and rotoreflections used to be the stellation group’s with no say in the matter, which answers only one of the questions people ask of them: a subgroup’s axes show what a lower symmetry would preserve, and that is exactly what you are choosing between when you decide to stellate under one. Tetrahedral axes drawn on an icosahedral figure say where its five tetrahedra sit.
Any subgroup of the polyhedron symmetry is offered, independent of the symmetry you are building with, and the diagram’s own element marks follow the same choice. It is the viewer’s preference rather than the figure’s — remembered between sessions like the thickness beside it, never written into a document — and it falls back to the stellation symmetry when a new solid leaves it meaningless. Kinds the chosen group does not have grey out, as before.
- The coloring choice is saved with the document. The entry below left one honest gap: a tie is genuinely undetermined by the selection, so a saved both-hands figure reopened by resolving it from nothing, and could come back wearing the other hand’s colors. The document now records the labeling the coloring wears — one integer per plane, beside the subgroup it belongs to — and a reopened document resolves its tie to exactly the colors it was saved with. The autosave carries it too, so a plain reload keeps the coloring as well as the cells.
It is a preference, not geometry: it only ever decides ties, so a selection that strictly demands a labeling still gets it, and a hand-edited document cannot be forced into wrong colors. The format release does not move — an older build ignores the field, builds the same figure, and merely resolves the tie the way it always did.
- A tied coloring keeps the colors already on screen. Select both compounds of five tetrahedra at once and the two candidate labelings score identically — whichever is chosen, one hand is five solid tetrahedra and the other wears four colors per cell, since no labeling can favor both partitions of the same twenty planes. The tie used to fall to whichever candidate the group enumeration produced first, with a symptom the selection history made visible: adding the second hand to the first could repaint the entire figure, and removing it repainted it back — two full recolors for a round trip that ended where it began, while the same round trip from the other hand changed nothing at all.
Ties are now broken by incumbency: among the equally clean labelings, the one agreeing with the colors on screen wins. Adding cells never recolors what is already there; a recolor happens exactly when the surviving figure strictly demands the other labeling — remove the hand the coloring is committed to and the survivor must switch, which no rule can avoid. One honest consequence, cured in the entry above: a saved both-hands figure used to reopen with no history, so a reload resolved the tie deterministically rather than by the path you took to it.
- The coloring keeps following the figure you are editing. The two compounds of five tetrahedra share the icosahedron’s twenty planes, and the group cannot tell them apart — it maps one onto the other. So the coset coloring is steered by the figure on screen: of the candidate labelings it takes the one leaving the fewest chosen cells wearing more than one color, which is what makes a compound come out as five solid tetrahedra instead of five four-colored ones.
That steer was sent when the subgroup changed, when the color mode changed and when a document opened — but not when a cell was toggled. Turning shell 5’s second cell off and the first on gave the other hand, still labeled for the one before it, and all sixty of its cells came out multicolored. Saving and reopening “fixed” it, which was the tell: the document was right and only the screen was stale, so the same file looked one way before a reload and another way after.
The steer now goes with the selection, from wherever it changed — a click on a cell, a shell shortcut, an undo, a typed cells string. It recomputes only what the current coloring reads: the orbit labelings take no steer at all, the plane labeling costs a few milliseconds, and the per-facet labelings — the only expensive ones — are recomputed just for the modes that show them. A repaint that does not move the selection still costs nothing.
- The build stamps itself, and the worker cannot be stale. The build line in the help dialog was a date somebody typed, so it said what the last person to remember said, and an evening went on a browser that showed the right date and the wrong program. It is now written by
docs/tools/stamp-build.mjson every commit, to the minute, and a hook installed bydocs/tools/install-hooks.mjsruns it.The stamp is also a cache key. A worker is a separate script with a cache entry of its own, and nothing about a fresh page obliges the browser to refetch it: a current app can talk to a worker several commits old, which is a bad failure because the screen looks right — the build line is the page’s, while the planes, the cells and the meshes all come from the worker. The worker is now fetched under the build’s name, so a new build is a new script. What the worker itself imports still carries no stamp, so it announces which build its modules came from and the status bar says so, and keeps saying so, if that is not the page’s.
- Every differently colored diagram can be chosen and exported. The plane menu offers the diagrams that differ, and it decided that by geometry alone: two planes the stellation symmetry carries onto each other carry their cells with them, so they draw the same picture and one of them is enough. That was the whole answer while color came from the shell or the face class, which every symmetry of the figure preserves.
A coloring by cosets or by subgroup orbits is not preserved by them — its subgroup is chosen independently of the editing symmetry — so planes of the same shape carrying the same cells can wear different colors. The compound of five tetrahedra is exactly that: one picture in five colorings, of which the menu offered one, and the export wrote that one. Now each geometric class is split again by what its planes actually carry — the plane’s own label where the coloring labels planes, the multiset of its facets’ labels where it labels facets — so the icosahedron under I with coloring subgroup T offers five diagrams of four planes each, and the export offers five cards. A plain coloring is unaffected: twenty planes, one diagram, as before.
The parts are numbered, in the menu and on the export cards — five entries all reading triangle · 4 planes are five diagrams that look like a menu which has not changed, so they now read coloring 1 of 5 and so on.
- A reload no longer costs you the session. The address bar cannot hold a document. A link names one — a preset, or a file of your own — and stops describing what is on screen the moment you change anything, so the app replaced it with a hash of the geometry alone: the solid, the groups, the depth, the camera, the cells. Everything else the document says — the coloring, the coset subgroup, the opacities, the line weights, an edited palette — had nowhere to go. Reload after an edit and the same solid came back wearing a different face, which is a poor thing to discover while working.
So the working document is written into this browser as you go, in the same JSON a file would hold, and a reload comes back to exactly what you left — edits, look and all. It is filed under the document it belongs to, so opening something else still opens that: the URL is where you say which document you want, and this is only a memory of what you were doing to it. A document that came from a file remembers its file too, so Save still writes where it wrote before rather than offering Save As — and if it had unsaved changes it still has them, and says so.
With that in place the URL has been relieved of the job it could not do. It no longer describes the figure at all: a document opened from a preset or a file shows that name and keeps it through every edit — the link says where this came from, the backup says what it has become — and anything picked from the catalog leaves it empty. A page can still drive the app with a full state hash, which is what those were for; the app simply stops writing them.
- Coloring folds into Construction; each view gets its own faces. The Coloring section held three things that wanted to be in three different places. The reading the colors carry, and the palette editor, are a coloring box in Construction, beside the symmetry the coloring divides by — they describe the figure, not the view. The opacity went the other way: it is a fill box in Solid, and a second one in Diagram, each a faces fill switch and a number.
Splitting it was the point. One slider used to drive both, so thinning the solid to glass to look inside it faded the diagram beside it for no reason — and a lit solid and a drawing on paper do not want the same weight of paint. The colors are still shared; only the paint is not, and an exported diagram now carries the diagram's opacity rather than the solid's. The switch is not the same as typing zero: zero is a wireframe you asked for and the box keeps it, while the switch puts the faces away and hands the number back when you turn it on again. The document still stores one figure for the solid, as it always did.
The tint on the cells you have not chosen joins them in that box, as unused cells — the one paint on the drawing that is not about the figure itself — and its strength is a control now rather than a pair of constants in the drawing code, 0.17 in the dark theme and 0.13 in the light. How much of the arrangement you want showing behind the figure depends on the figure, on how deep it is built and on what you are looking for, which is not a thing to decide once for everybody.
- The style controls fit on one line each. A switch, its color and its weight were three rows deep — a checkbox, then an indented row with a swatch and a slider, then the reading of the slider — so a boxed group of two edge kinds ran to six rows before the cylinders switches. Each kind is one row now: the switch, the swatch, and the weight as a number you type. The number shows itself, so the little reading beside it is gone; the slider it replaces could not be typed into, and a weight is a thing one usually knows rather than hunts for.
The same throughout: the solid's edges, its symmetry elements (the three toggles now share a line), its coordinate frame, and the diagram's four kinds of line and its element marks. A box being retyped is briefly empty, and an empty box reads as zero — which would draw no line at all — so an unreadable weight falls back to the kind's ordinary one until a real number arrives.
The cylinders switch went with them. It was a checkbox on a line of its own reading "as cylinders", which states one of the two answers and leaves the other unnamed; it is a choice between lines and cylinders now, on the same line as the color and the weight. So each kind of edge is one row that reads across as what it is: whether to draw it, in what color, how heavily, and as what.
- Symmetry & depth becomes Construction, in two boxes. The two halves of making a figure are asked at different moments — which groups it is read under, when you decide what you are looking at; how deep to build and out of which planes, when you decide how much of it there is — so they sit in two thin frames, symmetry and build, the way the solid's edges do. The three menus and the shell box size to their own contents rather than stretching the panel, and Edit planes keeps the width its text asks for.
- The symmetry menus say what they mean. Three menus offer up to ninety point groups between them, named in a notation that tells someone who already knows it everything and everyone else nothing. Each option now explains itself — D5d(I): a 5-fold axis, 5 2-fold axes across it, and diagonal mirrors, set in the icosahedron's frame — 20 symmetries in all — and the menu itself carries the description of whatever is chosen, since not every browser shows an option's own tooltip. The note that used to sit under all three is gone, and what it said has moved onto the rows themselves: each of the three now carries its own explanation of what that choice decides — the polyhedron group fixes which cells count as the same cell, the stellation group splits those orbits again and is the grain a click works at, and the coloring group is the one whose cosets paint. So the label tells you what the choice is for, and the menu beside it tells you what the chosen group is.
The coset subgroup moved up to join them, under the name coloring, because it is the third symmetry choice and belongs with the other two rather than among the colors it happens to drive. The depth is a number to type rather than a slider to aim — 60 still means every shell there is, and the label beside it, now reading shells count, says only that. And the little arrows on the section headers went from nine pixels to fifteen, which is the difference between a decoration and something you can hit.
- The symmetry elements reach the diagram, and the export follows the panel. Marking the group on the drawing plane — a dot where each axis pierces it, a dashed line where each mirror crosses it — was one checkbox living with the solid's elements, and it drew only what the SOLID was already wearing. It is two switches and a weight in the Diagram section now, reading the symmetry itself, so the diagram can be marked with its mirrors whether or not the solid is showing them. The marks keep the per-class colors, which is what makes a dot here and a cylinder there read as one object, and they are written into the exported SVG and PNG as well — over paper, each dot takes a ring of the paper so it stays legible on top of the ink.
The export dialog gains the color of that paper, rather than only white or nothing. And it listens: with the dialog open, changing anything about how the diagram looks repaints its preview cards. The preview promises the picture that will be saved, and now that the picture is the panel's, a weight or a color set there had to arrive here — or the promise would quietly stop being true while the dialog sat open beside the diagram it disagreed with.
- View splits into Solid, Diagram and Coloring — and the diagram gets its lines. One section had come to hold three unrelated jobs: how the solid is drawn, how the diagram is drawn, and the coloring both share. They are three sections now. Solid keeps the edges, the symmetry elements and the coordinate frame; Coloring keeps the face colors, the palette editor, the coset subgroup and the facet opacity, all of which the two views have in common; and Diagram holds what is drawn on the plane — including the diagram face, which had been sitting under Symmetry & depth for want of anywhere better.
The Diagram section is where the new work is. The stellation diagram is made of four kinds of line — the intersections right across the plane, the arrangement chopped into its regions, the figure's inside divisions and the figure's outline — and each now has its own switch, thickness and color, where before only the first two existed at all and none could be styled. The two outline colors are shared with the solid's edges: a facet edge on the solid and a facet outline on the diagram are the same thing seen twice, so setting either sets both. The thicknesses are not shared, because a line measured in pixels on a flat drawing and a cylinder in a lit scene do not want the same number, and cylinders are the solid's business alone.
These are the same settings the diagram export writes. The export dialog used to ask for all of them a second time, which is two descriptions of one picture and exactly the pair that drifts — its preview card promised "the picture that will be saved" while the diagram beside it was free to disagree. Those rows are gone: the file is the picture on screen, and the dialog asks only what an export alone can know — which planes, how big, whether to paint the paper, what to call it and where to put it.
- Escape no longer empties the app. It closed the top-most window that could close, of any kind — so with no dialog open it started on the workspace panels, and a few presses took the solid, the diagram and the cells table away. Worse, because a window remembers whether it was open, the empty screen came back after a reload: not a keystroke that could be undone but a program that appeared to have broken, with only the windows menu to put it right. Escape now closes dialogs and nothing else. A panel is closed deliberately, by its own button or from that menu.
It also picks the right dialog now. It took the most recently stacked one, and a modal sits in a fixed band above everything without ever restacking — so a browser opened after it was dismissed first, from underneath. Escape follows the z-bands: whatever is actually in front.
- A picture at any size, and JPEG. The 2-D export asked for a square or nothing; it now has a width and a height, opening at the size the view is on screen with a view button to put them back. They are not a crop: the camera takes the aspect ratio into account, so a picture written at a different shape is the same solid framed for that shape, and the readout says which of the two you have asked for.
JPEG joins PNG and WebP, and brings the one question the other two do not raise. It has no alpha channel, so a picture written as one has to stand on something — and a canvas asked for a JPEG composites onto black, which would send a figure that looked right on a pale screen back on a black square. It is laid on the background currently behind the view instead, read off the canvas element where the renderer now keeps it: light rgb(246,247,249) or dark rgb(14,15,20), the color you were looking at. PNG and WebP are unaffected and stay transparent.
- Exports have a section, and a picture is not a mesh. Taking the figure out of the app was two buttons at the end of the Files row, among the verbs that save the document — which they are not: a copy for a printer, a picture for a page and the diagram it was chosen on are three different things, wanted for three different reasons. They have their own Exports section now, directly under Files, reading solid 3D, solid 2D and Diagram.
The split is more than a rename. The picture of the view had been one entry in the solid export's format list, sitting under eight mesh formats and above the questions that dialog asks — colors, edge tubes, symmetry elements, orientation — none of which a picture has any use for, because a picture is simply what the view is already showing. So the mesh dialog is mesh formats only, and solid 2D is its own, with the two questions it actually has: PNG or WebP, and whether to write the view at the shape it is on screen or square at a size you name. The background is left out either way, so what lands in the file is the solid alone on transparency.
- The diagram export: no paper unless you ask, intersections of its own, and WebP. The background is a switch now rather than its negation, and it starts off — a diagram comes out as the drawing alone, to be laid on whatever it is going onto, and ticking Background paints the paper back. What used to be "traces right across the plane" is Intersections, moved to the head of the ink list and given the switch and weight every other kind has: it was a style the arrangement could be drawn in, and it is really a different drawing — where the other planes cut this one, run right across the picture — so it can now be had with the arrangement, instead of it, or on its own at its own thickness. Documents that still ask for the old single switch get exactly the picture they always got.
And WebP joins SVG and PNG, in both export dialogs. It keeps the alpha channel exactly and lands at about half the bytes of the PNG; its colors are very slightly lossy on the antialiased edges — a canvas offers no lossless setting, whatever quality it is asked for — so PNG remains the one to reach for when the pixels have to match.
- The diagram too, and it now carries the figure's opacity. The stellation diagram is drawn on nothing, like the solid beside it: its backdrop moved to the canvas element, so what a snapshot holds is the drawing and the transparency of its regions, with no paper baked behind it. The SVG could already be written without a background — the export dialog has had the box for it — and now the file says the rest of the truth as well: the facet opacity slider reaches the diagram, on screen and in the file, so a figure shown half transparent is exported half transparent instead of arriving fully painted. A region given an alpha of its own in the Colors panel multiplies with it rather than being overridden by it, and at full opacity nothing extra is written at all — an ordinary opaque diagram is byte for byte the file it always was.
- The 3D view renders with an alpha channel, and saves one. The figure is drawn onto nothing now: the drawing buffer is cleared fully transparent, so what it holds is the solid and only the solid. Every pixel the figure does not cover stays empty, a facet at partial opacity leaves a partly transparent pixel, and the antialiased silhouette comes out softly covered rather than blended into a background it no longer has. Saving the view, or a document thumbnail, writes an RGBA PNG that can be laid over anything — the transparency is in the file, not painted out of it.
Nothing on screen changes. The backdrop you see is the canvas element's own CSS color, set from the same per-theme value callers always set, and an export reads the drawing buffer, which never had it. The blending needed one real correction to make this true: coverage has to accumulate down the depth-sorted pass —
a + dst·(1−a)— where before the alpha channel was blended with the same rule as a color, so two half-transparent facets over empty space left a pixel thinner than either. That never showed while the background was opaque, and shows at once when it is not. Video is unaffected: it carries no alpha, so its frames are laid on the backdrop explicitly before encoding. - And their duals, whose vertices are at infinity. The ten solids dual to those — tetrahemihexacron to great dirhombicosidodecacron — were the last ones missing, and they are missing for a reason: a face through the center polarizes to a vertex infinitely far away, so no vertex list can hold them. The catalog's own data for them is a truncated stand-in for drawing, and it shows: 65 of its 270 faces are crossed quadrilaterals with no usable normal, and a stored vertex can miss the plane of its own face by nearly half the figure. Building an arrangement from that would build the wrong solid.
But a stellation never needed vertices — only planes, and those are exact. The face planes of a dual are the polars of the primal's vertices, one apiece: for a solid whose vertices sit on the unit sphere, which every uniform polyhedron's do, the plane is simply the vertex direction at distance one. So these ten are built from their primal, named by a
polarfield in the catalog, and the arrangement they cut is finite and ordinary — not one central plane among them, no truncation anywhere, every shell closing under the full group.What comes out is a pleasant surprise, and a smaller answer than ten: between them these solids cut only four arrangements, each one a solid already here. Duals share a plane set exactly when their primals share a vertex set, and these primals share heavily — all six icosahedral ones carry their thirty vertices on the thirty two-fold axes, so the six cut one arrangement, not six. The tetrahemihexacron cuts the cube's six planes, and therefore has no stellation at all, which is the honest answer; the two octahemioctacrons cut the rhombic dodecahedron; the six cut the rhombic triacontahedron; the great dirhombicosidodecacron cuts the sixty planes of the great hexagonal hexecontahedron. Every plane of a set stands at one distance, so the naming holds up to scale, and scale changes no arrangement. The tests assert the grouping, which is also what would catch a mis-stated primal.
One correction rode along: the previous round had the small and great dodecahemicosahedron the wrong way round. u67 is the small and u70 the great, as the source data says and the numbering confirms.
- The hemipolyhedra join the catalog. Ten uniform polyhedra were missing from the solid picker — the nine whose faces pass through the center, tetrahemihexahedron to great icosihemidodecahedron, and the great dirhombicosidodecahedron, whose sixty squares do too. Their geometry was in the data all along; what they needed was the central planes those faces lie in, so picking one now switches the cuts on by itself and the statusline says so ("7 planes (3 central)"). Making them build correctly took one more engine change: cell growing used to sweep a cell's whole ceiling first and then the floors that touch it, which is sound when "up" means away from the center — but a cut cell can wear ceilings and floors alternating around its boundary, and such a cell could never assemble; one member of a sixty-fold orbit simply failed to exist. The walk is now a boundary crawl by edges: each open edge is closed by the facet reached by the smallest turn around it, ceiling or floor, and a crawl that cannot close gives every facet back for the right cell to collect. A pleasant side effect: the crawl refuses the impossible, and a few three-faceted "cells" that used to appear in the outermost shell of the snub-family solids — rubble from the depth cap, never geometrically real — are gone; a selection saved on those solids by an earlier build may need re-picking in that one outermost shell, and nothing in the shipped presets or examples is affected.
- Planes through the center. A plane with d = 0 was always skipped — with a warning, but skipped — because "oriented away from the origin" says nothing about a plane the origin sits on. The plane editor now has keep planes through the center: on, such planes join the arrangement and cut every cell they cross, so the core splits into pieces, a figure can be halved along a mirror, and the cross-section shows as faces lying in the cut. The engine counts shells by the planes that are not central, which is the same on both sides of every cut — so a piece and its mirror image sit in the same shell, symmetry orbits close, and the cells notation, the depth control and saved selections all read exactly as before. The statusline says when cuts are in ("11 planes (3 central)"), the face chooser lists the cut's own diagram as central cut, and such documents save as format release 4 — older builds refuse them cleanly instead of silently rebuilding a different figure under the same cells string. Behind it all sits one hard-won fix: a central plane running along a polyhedron edge lets two coplanar facets meet at the same count, and the cell-growing walk used to sail across that edge on float noise, stitching neighboring cells into one; a strict margin now says coplanar is never the same cell.
- Presets and Files stay open across a reload. Both windows already remembered where you put them and how big you made them; what they forgot was whether they were on screen at all, so every reload started with the workspace tidied away and both shelves closed. Now a session that left one open opens it again. The document browser comes back with its folder already listed, but only where the browser is still holding the permission you granted — a page load is nobody asking for a folder chooser, so when the grant has lapsed the window returns empty with a line pointing at the folder… button rather than an OS dialog you did not ask for.
- A reload comes back to the document you opened. The address bar carried the geometry — solid, groups, depth, camera, cells — which is what makes the app drivable from a web page, but it meant that opening a preset or one of your own files and then reloading gave you the figure without the document: its coloring, its subgroup, its edge style all reverted, and its name with them. A gallery link already knew better, and now every way in does. Open a preset and the address becomes
#doc=presets/TD-4.json, which is a link anyone can follow. Open a file of your own and it becomes#file=and the path below the folder you granted — not a link anyone else can follow, since it names a place on your machine, and it says so by falling back to the default figure with a word about why when the folder permission has lapsed. Editing drops the link, as it always did: at the first change it stops being that document and the address goes back to describing what is on screen. - Both thumbnail shelves open on a double click, and both own the right button. A single click selects a tile and nothing more; opening takes a double click, the way a folder of files behaves everywhere else — and brushing past a thumbnail is exactly how the wrong document used to get opened. This is one rule for the document browser and the presets sheet together: they are the same furniture with the same tiles, so one opening on a click while the other wanted two was confusing precisely because they look identical. The right button belongs to the panel too, anywhere inside it, tiles and gaps alike, where it used to hand you the page's own menu of reload and view-source. In the document browser that menu offers to delete the document under the pointer, and everywhere in it new folder…, which asks for a name and makes one in the folder you are looking at. Both menus offer copy path, which puts where the thing is on the clipboard: for a document the granted folder's name and the way down to it, which is as much as the browser will ever tell a page about where a file lives, and for a preset the address that fetches it. The file browser also offers rename…, which moves the document, its thumbnail and the name written inside the file together — and, if the document renamed is the one you have open, moves what Save writes to along with them. And the window that browses your folder is called Files now, in its title and in the windows menu, matching the button that opens it — it answered to two names before.
- Declining to open a document no longer aims Save at it. Clicking a thumbnail in the document browser while your own work was unsaved raised the "discard your changes?" prompt, as it should — but declining only stopped the opening. The browser went on to point Save at the file that had merely been clicked, so the next Save wrote the document you had kept over the document you had not opened, destroying the file touched by accident. Opening now reports whether it actually happened, and the save target moves only when it did. Accepting the prompt still moves it, as it always should have.
- Saving no longer makes the document browser blink. Every save re-read the whole folder: the grid was emptied, the folder enumerated again, and every thumbnail in it decoded afresh — fifty documents meant fifty file reads and fifty tiles thrown away to put back forty-nine identical ones, which is what the flicker was. Saving over a document that is already listed now re-reads that one document and repaints its tile in place: one file read, no tiles replaced at all. A document saved under a new name is appended, at the end, and scrolled to — on purpose, rather than filed into the sorted order where it might land off-screen: a file you have just made should be where you are looking. The grid sorts itself afresh on the next refresh, the header button that used to be called reload. Only a save into a different folder redraws the lot, because then the lot really is different.
- New and the plane editor are windows now. They were the last two
<dialog>elements in the app, which put them in the browser's own top layer: they could not be moved aside to look at the figure behind them, they could not be resized, they forgot their size and place between openings, and once dismissed there was nothing anywhere to bring them back. Both are internal windows like every other panel — draggable, resizable, remembered, and listed in the windows menu as New and Plane set. The plane editor in particular is better for it: its preview can now sit beside the solid it is describing rather than on top of it. - Unsaved work is not thrown away without asking. Editing a figure and then opening another document, or starting a new one from a solid, used to discard the first silently — and so did closing the tab or following a link out of the app. Each of those now asks first, naming both the document you would lose and what is about to replace it, and declining leaves everything exactly where it was. The page's own exit is the browser's standard "leave site?" prompt, which is all a page is allowed to raise there.
Only work is counted: the cells, the colors, the symmetries, the depth, and a hand-built plane sheet. The camera is not, and neither are the view switches like facet opacity or edge width — they are saved with the document, but they are a keystroke to redo, and a confirm box that fires because a slider moved is one people learn to click through. Saving, or opening a document, makes it quiet again.
- Opening a document changes nothing until the figure is ready. A document's look — its colors, its opacity, its edge styles — used to be applied the moment the file was read, which meant it landed on the figure still on screen. Open a document built on a 120-plane solid and the OLD solid restyled itself instantly, then sat there wearing the new document's colors for as long as the arrangement took, seconds on a deep one, before finally becoming the right figure. The two halves are separated now: only what the calculation depends on is set beforehand, and everything visible waits until the new figure is on screen to receive it. The picture you are looking at simply stays as it was until it is replaced by the right one.
While an arrangement is being intersected the status bar counts the planes as it always did, and now offers a Stop beside the count. Stopping ends the calculation the only way a long synchronous run in a worker can be ended — the worker is discarded and a fresh one started — so the half-built arrangement goes with it and the app says so rather than pretending the old figure is still there. Choosing a solid again picks up immediately.
- The Files row keeps its buttons the size of their words. Every button in that panel stretched to fill the width, so a row that wrapped left one or two buttons sprawling across the whole panel, looking like a different kind of control from the line above; the two export buttons did the same on their own row. They now take the width their text asks for. The export pair also sat flush against the buttons above it, with no gap at all: the rule that spaces one row from the next matches only adjacent siblings, and the hidden file picker sat between the two rows, so it never applied. All seven verbs are one row now and wrap as one group, which is what they always looked like they were. Wrapped lines also get a little more air between them than the buttons have between each other, so a second line can never read as sitting on the first. Save As and Save New gained the ellipsis their neighbors carry, both opening a dialog. Presets… has left the row — it is one click away in the windows menu, where every panel lives — and so has the note about the preset shape.
- No green outline where there is nothing to add. Holding shift over the outermost facets outlined them green, and green means "this click adds a cell" — but those facets have nothing further out, so the click could only report failure. The outline was dimmed there rather than dropped, on the reasoning that you still want to see which face you are pointing at; a faint green outline is still a green outline, and it still makes the promise. Now a gesture with nothing to act on draws no outline at all, and the status line beside it says why. A bare hover with no modifier still outlines neutrally, because pointing at a face is not a promise about a click.
- Any group of facets can take a color of its own — and an opacity. Every coloring the app offers is a function of a group: which shell, which face class, which coset, which orbit. Edit colors…, beside the face-color menu, opens a panel listing those groups one row each, and any of them can be given a color the palette would not have chosen — and an opacity of its own. Turn one coset to glass and the cells behind it show through; turn it all the way down and its fill is gone while its edges stay, so the group becomes a wireframe of itself and the figure keeps its shape as you look past it. The colors follow the figure everywhere it goes: the 3-D view, the stellation diagram, the diagram's SVG export and every mesh format that carries color. The facet opacity slider is unchanged and now reads as what it always was — a modifier over the lot. A document saves its palette as a flat array of hex numbers, one per group with its alpha, so it reopens wearing its own colors and can be lifted whole into another figure through the panel's copy box; positions are group numbers, so a palette copied from a figure with ten cosets still lands correctly on one with five. Nothing is saved unless it differs from the palette, so documents that never touched color read exactly as before.
- Spin is the Animation panel's job now. The spin the solid checkbox in View has gone: the Animation panel below it does the same thing and more — an axis in the figure's own frame, a rate in turns per second either way round, and a recording of exactly one turn. Two controls for one motion was one too many. Documents no longer carry the flag; older ones that do are simply read past.
- A coloring tutorial, with every picture one click away. A new Coloring by symmetry page walks the four colorings — cosets, subgroup orbits, blends, mirror splits — through twelve classical pictures on the icosahedron, the dodecahedron, the rhombic triacontahedron and the cube, from the five tetrahedra to Wenninger's five-color table to what gray means. Every example is a saved document, so each "open in the app" link arrives already built, colored and editable; all twelve are engine-verified. Linked from the front page.
- The figure outranks the plane frame, and the mirror cut finishes the job. The last two steps of the coloring report. Per-facet coset colors are now computed in their own right rather than copied from the planes, with the plane labeling demoted to the final tie-break: where nothing contradicts it, per-facet still equals per-plane exactly, but where the figure knows better the figure wins — the classical test is the compound of five octahedra, whose five components no per-plane coloring can separate (every icosahedron plane carries a hexagram of two octahedra), and which now comes out one color per octahedron under I/T with the compound selected. And a new mode, by cosets — per facet, mirror-split, removes the last gray: a facet that grays because its own symmetry straddles cosets is cut along that symmetry's mirror lines (or fanned about its center when the symmetry is a pure rotation), and the pieces — whose stabilizers are trivial, hence inside every subgroup — take crisp coset colors. The icosahedron under Ih by cosets of I, whose per-facet mode grays 380 mirror-straddling facets, now splits every one into halves or sextants wearing the two hands in equal measure — the classical "one face, two colors, parted by a mirror". The cut follows the figure into the 3-D view, the diagram, the diagram export and the mesh exports.
- Blends where the cosets cannot decide, and orbits where they cannot exist. The coset coloring grayed far more often than it colored — under a full symmetry like Ih almost every subgroup grays every plane, the mirrors being nowhere to put — and a research report traced exactly why and what to do. Three of its steps are now in. Where no single coset can name a piece but a small set can, the piece wears the mix of that set, averaged in Oklab: the icosahedron under Ih by cosets of T — flat gray before — now shows twenty distinct two-color pastels, one right-handed and one left-handed tetrahedron each, both compounds of five legible on one figure; a mix that would need every color stays honestly gray. A new family by subgroup orbits — per plane, per facet, per cell colors pieces by whether some motion of the subgroup carries one onto the other; it exists for every subgroup with no gray at all, and its small classes are exactly the sub-figures the subgroup holds — T on the icosahedron's twenty planes gives 4+4+12, the two invariant tetrahedra picked out against the rest. Per cell is the piece a model painter actually holds. And two defects went: the per-facet fallback now labels all its orbits in one shared frame, stitched together through the cells they cap — on the gallery's compound of five tetrahedra under Ih/T, all sixty spikes are monochrome where before not one was — and it takes the same hand-steering from the selection as the per-plane path, so switching between the two modes can no longer flip a chiral compound into its mirror image. The stored subgroup lists in the symmetry data, long-stale generator residue that had misled the report's own survey, are regenerated from actual containment and tested so they cannot rot again.
- A named view now looks down the axis of the group with the same name. The views ±i5, ±i3 and ±o2 exist to lay a symmetry axis along the line of sight, but the groups named for those axes were stored on different ones — the icosahedron has six five-fold axes, and D5(I) sat on a different one from the +i5 camera, while the icosahedral-frame 3-folds sat on the cube diagonal, which is the o3 axis. Choosing D5 and pressing +i5 therefore showed the figure from an angle unrelated to its own symmetry. Nine groups have been turned onto the axes their views use — each by conjugation with an element of its own parent, so it is the same subgroup type and still literally inside the parent — and the rule is now simply that a group tagged (O) sits on an octahedral axis and one tagged (I) on an icosahedral one, in both cases the one the identically-named view looks down. A test checks every case, and that nothing frame-tagged sits on an axis with no name.
- The fifty-nine icosahedra, and the rules that make them fifty-nine. A page for all 59, generated from Du Val's symbols rather than built by hand — each symbol is parsed into shells, the shells into cells, and the result checked against the counts the literature records before anything is written. Seven more pages take the same twenty planes under different rules: demand that a face plane not be broken and 59 becomes 34, that an edge not be segmented and it becomes 25, or drop Miller's ban on shells separated from the core and it becomes 50, admitting figures with a shell hanging free of the middle. 84 distinct stellations in all, each one a document the app opens.
- Stellation diagrams, on the page and out of it. Every figure in those galleries now carries its diagram beside its picture, and the gallery switches between the two. The drawing became a pure function — geometry in, SVG out, no canvas and no renderer — so the same code draws the card on the page and the file you export, and the generators can draw all 84 on the command line. A new export dialog asks what you want: which of the distinct diagrams, SVG or PNG, at a size and a weight of line you choose. Each SVG carries the document that produced it, so the picture can be rebuilt. The folder is asked for once and remembered, and everything goes into it — a single diagram as much as twenty. Where the browser cannot be given a folder, one file downloads and several come back as one zip.
- Four kinds of ink, and a preview of the file. A stellation diagram is really four drawings laid over each other, and which you want depends entirely on what the picture is for. Each now has its own switch and weight: the diagram lines, the arrangement the figure was chosen out of; the faces fill; the facets outline, the divisions inside the figure where two chosen facets meet; and the faces outline, the edge with the figure on one side and nothing on the other — the line you would cut. A plate for reading is the first alone; a net is the last two with the arrangement gone. A page that would come out blank — a face the figure never reaches, or settings that draw nothing at all — is skipped, and the dialog says so before the button is pressed rather than writing an empty file.
- How big, how full, and how heavy are three different questions. The export now asks them separately. Width and height are the picture's resolution in pixels — the PNG's dimensions, and the SVG's viewBox, which comes to the same thing because the viewBox is one unit per pixel. Scale is how much of that frame the figure fills: it multiplies the drawing's coordinates and nothing else, so 1 fits the figure with a small margin and anything above about 1.05 crops it. And a line weight is now an absolute width in pixels of the finished picture — a tenth of a pixel to ten pixels, the same range whatever the resolution — so a 0.7 px hairline is 0.7 px whether the plate is 500 pixels or 4000, where before it was a fraction of the frame and thickened as the frame grew. A picture that is not square fits the figure to the shorter side and centers it in both, and the previews take the picture's own proportions.
- The top-left corner says which document you are in. Its name, and under it the three things that decide what it can be: the solid, the symmetry its planes were laid out with, and the symmetry its cells are chosen under — icosahedron · Ih · I. A document that has never been saved is named after exactly those, so it reads
u28-Ih-Iuntil you save it. The corner it took belonged to the Stellation link, which has moved to the far right past the ?; the walkthrough, the history and Brückner's plates are all one click further, on that page, so their three buttons have gone. So has the vertex-and-face count, which was a readout nothing depended on. - Coloring by cosets of a subgroup. A fourth face coloring: by coset. A menu that appears beside the color choice picks a subgroup of the polyhedron symmetry, and every facet is painted by which coset its plane belongs to — the classical pictures are plane-partitions: the compound of five tetrahedra IS the icosahedron's twenty face planes falling into five sets of four, and a spike is one color precisely because its whole surface lies in one set's planes. Pick coset subgroup T on the icosahedron and the compound wears its five colors, each tetrahedron one of them. The coloring belongs to the figure, not to the editing: it ignores the stellation symmetry entirely, so regrouping to carve out a single tetrahedron does not repaint what you made. A mirror-free subgroup counts its cosets among the polyhedron group's rotations, which is what the classical pictures do — cosets of T in I, on a solid whose full symmetry is Ih. The twenty planes carry the partition twice, once per hand of the compound, and the selection on screen decides which hand the colors follow. Pieces the cosets cannot name are gray rather than wrong. The coloring is offered two ways — one color per plane, where a whole face plane matches and the five tetrahedra appear, or one per facet, a single flat piece of surface at a time, which can separate things a whole plane cannot — color the icosahedron under Ih by cosets of I and the two hands of its chiral facets come out as the two colors. Where the per-plane labeling exists the facets inherit it, so the smaller pieces never contradict the larger. Both work in the 3-D view, the diagram, the diagram export and every mesh format that carries color.
- A turntable, and a video of one turn. A new Animation panel under View spins the solid about an axis given in its own coordinates —
0 1 0for y, or type in one of the figure's own symmetry axes — at a chosen number of turns per second, negative for the other way. The axis holds still on screen while the solid goes round it. Save to video records exactly one full turn from the current angle, as WebM or MP4 — an MP4 only where the browser can encode one, and the menu says so where it cannot. The turn is counted in frames, not clocked: sixty per second of turn, each exactly its share of the circle, so a slow machine takes longer to record but the video is perfectly even — every frame is in the file. The size is yours — SD, HD, Full HD, 4K, or a custom width and height snapped to the nearest even numbers — and the turn is rendered at that size rather than scaled up from the window, so a 4K file is genuinely 4K sharp. While it records, the view letterboxes itself to the video's shape, so what you watch is the framing the file is getting. The file is named in the panel, or after the document when the field is left empty; it goes into a folder you choose once and it remembers, like the exports — and a name already in that folder is a question, never a silent replacement, in the video, the solid export and the diagram export alike. And because a turn is a loop, the last frame meets the first and it repeats without a seam. The axis and speed are remembered; the spin itself always starts off, because a page that begins moving the moment it loads is a surprise, not a preference. The Files panel moved to the top of the column, where the way in and the way out belong. - The solid goes out through a dialog, and in five more formats. STL, OBJ, OFF,
.steland view PNG were five buttons that each wrote immediately, under a name nobody chose, into the downloads folder however capable the browser was. They are now one Export Solid…, which asks the three questions the diagram export asks: which format, what the file is called, and — where the browser allows it — which folder, once, remembered. Three formats join them: 3MF, which is what a modern slicer would rather have than an STL; and glTF, GLB, VRML 2 and X3D, for anything that displays models. - And it takes the colors and the edges with it. Every format that has somewhere to put a color now carries the one you are looking at — by shell, or by which face of the original solid a facet lies in, whichever the view is using. OFF puts it after each face, 3MF as base materials a slicer reads, glTF and GLB on the vertices, VRML and X3D in a
Colornode indexed per face. STL has nowhere to put it and OBJ would need a second file beside it, so those two are the shape alone, and the dialog says so rather than graying a box without a reason. And when the view is drawing the edges as cylinders rather than lines, those cylinders go into the file too, at the radius and in the color they are drawn in: if you are looking at tubes then they are part of the figure, not an annotation on it. The same for the symmetry elements and the coordinate frame — the rotation axes as their cylinders, the mirror planes as their rims, the frame as its three arrows, whichever of them the view is showing. And a save orientation switch writes the model facing the way the screen does, x right, y up, z toward you; left off, it is written in its own frame, which is what two figures have to share to be compared. - One row of file buttons. New… is the solid picker, where every new document starts. Files… is what Browse… was called, and it now answers "open a document" on any browser: the folder browser with previews where the File System Access API exists, and the plain file picker where it does not — before, the folder button was simply hidden there and opening lived under a different label. Then Save, Save As, and Save New, which is the Save As dialog opened on a freshly generated name instead of this document's: it writes a copy and leaves the original where it was, and from then on Save writes the copy.
- Which diagrams to export is now asked with the diagrams. A solid has one diagram per kind of face — one for the icosahedron, three for the truncated cuboctahedron — and which of them you want is not a thing anyone can answer from the words all and the one shown, least of all under a low symmetry, where faces that look alike give genuinely different diagrams. So each is drawn, on its own card, with the checkbox that decides whether it is written. Every card is the file: the same drawing code on the same geometry with the same options, at the export's own resolution and shrunk to fit, so a hairline looks like a hairline instead of the several times heavier it looked when the preview was redrawn small. A face the figure never reaches shows its blank card and says so, rather than the button quietly counting one lower than the ticks suggest.
- A picture no longer depends on the window it was made in. Thumbnails were cropped out of the live canvas, and the camera accounts for the aspect ratio — it must, or a narrow canvas would cut the sides off — so the same solid at the same zoom filled a different fraction of a wide window than a tall one. They are now drawn into a fixed square buffer, which makes the picture a property of the document rather than of the reader's screen. The same fault, and the same fix, for exported diagram PNGs.
Sixth round
- The panels can float as windows. A toggle in the header switches the app between the docked three-pane layout and a desktop of movable, resizable windows — the solid, the diagram, the cells table and the settings each in one. Docked stays the default; a phone is always docked, floating panels there being a worse kind of scrolling. The trick underneath is that the windows adopt the existing panels rather than rebuild them: the live DOM is moved into the windows and back, so every control, every wired handler and the live WebGL context come along untouched. Window positions and which windows are open persist between visits.
- The solid can be made translucent. A facet opacity slider in the View panel takes the facets from solid down to nothing. Below 100% the solid turns to glass and its hidden edges come through with it, so the cells inside — the one part of a stellation you can never otherwise look at — are visible through the ones in front. At 0% only the edges are left, which is a wireframe of the whole arrangement. The glass composites in true depth order, using the sorting algorithm from the original Java applet: a stellation is cut from a fixed set of planes, and those planes order the facets exactly, where a general depth sort would get these meshes wrong. Everything opaque — the edges, the symmetry axes, the mirror rims — draws first and writes real depth, and the glass then composites over it: an edge or axis deep inside is dimmed by every layer of glass in front of it, anything it hides stays hidden, and the front surface stays crisp. So the solid reads as a translucent object rather than an x-ray. The setting is saved with the document, so a figure can be kept the way it reads best.
- The missing symmetries. The icosahedron offered thirteen stellation symmetries where it has twenty-two. The absentees were real subgroups stored about the wrong axis — D3d, the symmetry of a triangular antiprism, sits inside an icosahedron, but the stored D3d has its 3-fold axis where a textbook puts it and the icosahedron's lie along body diagonals, so the program could not see that one was inside the other. Twelve groups have been added in the frames that fit, and every symmetry the catalog uses now offers its complete subgroup lattice: Ih 13 → 23, Oh 25 types all reachable, and the same for I, O and Td. A test enumerates each lattice and fails if anything is unreachable again.
- A third coloring: by stellation face class. Coloring by face class asks which faces are the same kind of face of the solid you started from — a property of that solid, so an icosahedron is one color however you stellate it. The new setting asks the other question: which faces the symmetry you are building under can still carry onto one another. The two agree at full symmetry and part company below it, so dropping the icosahedron to a subgroup splits its twenty faces into two colors under Th, and into twenty under E. Switching symmetry recolors immediately, without rebuilding.
- A gallery of examples. A new Examples page: a card per stellation, and clicking one opens the app with that arrangement built and those cells chosen. It reads a catalog that keeps the curation — tags, notes, where a figure comes from — outside the documents themselves, so a document stays a thing to open rather than an opinion about which gallery it belongs to. Links name the document rather than describing it, which is what lets an example built from a hand-made plane set be linked at all.
- One rule for a document's name. Save As offers the document's own name — the file it came from, or the preset it came from — and otherwise a generated one, now beginning
stel-. Picking a solid from the catalog starts a new document, so it no longer arrives wearing the last one's name, and saving onto a name that already exists asks first. - Coordinate axes, and a home button that tours. The View panel can draw the x, y and z axes through the origin — red, green and blue arrows spanning the solid, each tipped with a 60° cone, at whatever thickness reads best. They show the frame the symmetry groups are written in: in the +z view x points right, y up, z toward you, and they are sized to the selection you have rather than to everything the planes could build. Beside the home button there is now a menu of the directions a solid can be seen from square-on: the coordinate axes ±x, ±y, ±z, and the symmetry axes themselves — ±i2, ±i3 and ±i5 for an icosahedron's two-, three- and five-fold axes, ±o2, ±o3 and ±o4 for a cube's. It names the one you are in, blanks itself when you turn the solid off it, and home puts you back on it: the button restores your orientation rather than moving you to a new one, so it can always undo a drag. Orientation only — your zoom stays where you left it. The isometric views are the true ones, the body diagonal landing exactly on the line of sight so the three axes leave the origin 120° apart and equally foreshortened; and the symmetry views are where a solid looks like what it is — seen down its five-fold axis, an icosahedral stellation is exactly five-fold on the screen. Each view is the shortest turn from the +z frame to its direction and nothing more, so none of them spins the model on the way: reaching the (0, 1, 1) view is a 45° tip about x, which leaves x lying across the screen where it began. The isometrics keep their own roll, the one that stands y upright — that pose is what an isometric view is for.
- Plane sets are stored as data, not as text. A document's custom plane set used to be saved as the lines of the editor — nx, ny, nz, d and a group name, re-parsed on every read. It is now a list of proper JSON objects with a normal, a distance, a symmetry and a factor, each checked when it is read and when it is written, naming the row at fault. A mistyped plane set now refuses to open instead of quietly building a different solid, and numbers keep their full precision because they are never turned back into text. Documents saved in the old form still open: their lines are read once and re-saved as data.
- One New dialog, and planes from anywhere. The polyhedron picker is now the single front door: New starts from a polyhedron, a preset, a file on your computer, or your own plane set, and hands off to whichever of those you choose. The plane editor imports from the same three sources — a catalog solid, a preset, or a document from disk, each reduced to one row per symmetry class — so a plane set can be assembled out of anything the program can already open. Each row also gains a factor that scales its distance from the origin, which slides the plane along its own normal: the same solid's planes can be imported twice and set at two depths without retyping a single number.
- Four more of Brückner's models, rebuilt. The Brückner page gains the compound of five cubes (Taf. XII, 24) — computed to be exactly the first five shells of the rhombic triacontahedron — the other three regular star polyhedra, which are the dodecahedron's complete set of stellations (drawings Taf. VII, 21–23; models Taf. IX, 7 and 16 and Taf. X, 5), and Kepler's stella octangula (Taf. VII, 20), the octahedron's one and only stellation. Every entry quotes the passage of the book that places the model on its plate.
- Make planes, made friendly. The plane editor is rows instead of a raw sheet: each row holds one plane as four copy-and-pasteable numbers (normal and distance, full double precision), a symmetry group of its own, and the computed orbit size. Groups mix freely — icosahedron planes beside cube planes in one arrangement. A live 3D preview shows every plane as a spoke from the origin with a disc lying in it, colored by row, and an import menu appends any catalog solid's planes reduced to one row per symmetry class — the icosahedron imports as a single Ih row of orbit twenty. Documents are unchanged: the rows are the same plane sheet they always saved.
- The stellation symmetry no longer clears your cells. Changing it used to reset the selection to the core and wipe the undo history. Now the selection, the history and the camera all stay put, and the symmetry means what it should have meant all along: how big a bite the next click takes. Every editing operation toggles the whole orbit of the clicked cell under the current group — so you can rough out a stellation under the full symmetry, drop to a subgroup (E takes single cells) for the asymmetric touches, and come back, without losing a thing. A box in the Cells table holding only some of its pieces shows a faint fill, says "k of n pieces" on hover, and fills whole on a click. Documents whose selection is whole orbits save exactly as before; one that is not saves in a new notation an old build declines to guess at, and exports to .stel under E, which the original Java program reads exactly.
- First visits no longer open on a wireframe. A first visit — any browser with nothing remembered yet — read the absent opacity setting as 0% and booted into an invisible wireframe: the blank 3D view reported on Android. And on a phone the two views tiled the whole screen while swallowing every touch gesture, so the page could not be scrolled at all. The views now leave the statusbar and the top of the settings panel showing, which gives the page a place to scroll from.
- Symmetry elements, smoother and adjustable. The mirror-plane rims and fills are drawn at four times the segments, so they read as circles rather than polygons, and a thickness slider under the symmetry checkboxes scales the axis tubes and rims from hairline to emphatic — the reach of each element stays put, only the girth changes.
- Edges can be thin cylinders. Each kind of edge — face and facet, independently — can be drawn as slender capped cylinders instead of flat lines: real geometry with thickness in the model, shaded by the same lights as the solid, so an edge catches a highlight and shades round. The choice is per kind because the two read differently: the solid's own edges carry a cylinder well, while the hairline facet subdivision can shimmer as sub-pixel geometry and often stays cleaner as lines. The width sliders now move in steps of 0.1, down to 0.1; being real geometry, cylinders grow as you zoom in where the flat lines keep a constant screen width. Saved with the document like the rest of the edge styling.
- One menu for every window. The windows menu in the header lists them all — the four panels, the presets and the document browser — with a tick beside the ones that are open, and it opens and closes any of them. It is there in both layouts, because the two browsers float over the docked layout just the same. Closing a window with its ✕ therefore always has a way back.
- Presets with pictures, instead of four buttons. The Presets… button opens a window of thumbnails for ready-made stellations the site ships; picking one loads it. The four numbered sample buttons are gone, converted into the first four presets. They are ordinary
.jsonfiles with a.json.pngthumbnail beside each, served as plain images, so the browser does all the work and this runs everywhere, phones included. - Documents can live in a folder, not just downloads. In browsers that allow it (Chrome, Edge — the File System Access API), Browse… opens a thumbnail view of a folder you choose, Save overwrites the file a document came from and refreshes its preview, and Save As… writes a json + thumbnail pair into a folder you pick. The folder is remembered, so later visits need only a one-click permission, not the picker again. Elsewhere — Firefox, Safari, mobile — those buttons simply do not appear, and Save stays the
.jsondownload it always was. Nothing breaks on any browser. - The narrow-screen layout works now. The responsive rules had been dead for a while — a stylesheet ordering bug meant the two views never stacked on a small screen. They do again: below a certain width the views stack, then the side panel drops below them, then on a phone the whole thing is one scrolling column. The window chrome is built on pointer events, so it drags with a finger; the file and preset windows become near-fullscreen sheets where a floating window would be pointless.
Fifth round
- Facets can be colored by which face of the original solid they lie in. The 3D view had one coloring, a hue per shell outward from the core. A second is now offered in the View panel: a hue per symmetry class of original face, so every facet lying in the same kind of face matches however far apart they are, with undersides in the same hue darkened. An icosahedron has one kind of face and so takes two colors; an icosidodecahedron has two and takes four. The classes are orbits of the face planes under the polyhedron's group, not the stellation group chosen below it — the kind of face is a property of the solid you started from. The diagram follows the same choice: it is drawn on one plane, so it takes that plane's color and the only variation left is the above/below split it already draws.
- Face edges and facet edges are drawn separately. Every edge of every facet used to be drawn in one color and one weight, which buries the shape of the solid in the arrangement that produced it. Where two facets of different planes meet there is a real crease, an edge of the solid; where two facets of the same plane meet, the surface runs flat across the join and the line only records how the arrangement was cut up. Each kind now has its own toggle, color and width. How much this matters depends on the selection: whole consecutive shells have no facet edges at all, their per-plane regions being disjoint, while the shipped samples are close to half and half — 240 of each for sample_02.
- The solid follows a virtual pointer on a spring. A mouse reports in uneven jumps and applying each one directly makes the model stutter, so the trackball now drives a unit mass pulled by a spring whose far end is the real pointer, damped by fluid friction — the AnimatedPointer from the author's SymmHub, with its physics unchanged. Releasing the button switches the spring off and the mass keeps its speed, so a throw spins as fast as it was thrown, with no cap. This replaced three separate mechanisms whose release behavior split at a threshold: a throw twice as hard used to produce exactly the same spin as one just over the line. A drag that decelerates before release still stops dead, which is the fault the Java drag fix diagnosed.
- A parallel projection, so shape no longer depends on distance. Selecting a cell in the outermost layer turned the view inside out: the fitted camera sat inside the selected cell and 20 of its 152 vertices fell at or behind the eye point. A parallel projection has no eye to fall behind, so the failure cannot recur at any depth — and it settles the three framing problems logged on 6 August, which were one problem, since a camera that moves changes the perspective of everything that did not change.
- The tutorial figures re-fit only when the frame no longer suits. Following the content on every change meant that switching one outer cell resized the whole model, and a jump that large reads as the model changing rather than the selection. The frame is now left alone while the content sits comfortably inside it, and eases to a new fit only when the model would be cut off or has shrunk to a speck.
- A capped arrangement no longer builds shells it cannot support. The snub dodecahedron came apart past layer 11 — cells scattered through the outer shells instead of tiling them. Not precision: the facet cap was dropping facets whose split would carry them past it, so at layer 12 only 150 cells of 1650 survived. Comparing capped against uncapped builds showed the first differing layer is always the cap itself and everything below identical, so the build now stops there: fewer layers, all of them true. Raising the depth control returns the rest.
- The site has a front door. Arriving at the project put a three-pane workspace in front of you, with the documentation reachable only from a header link inside it. The app is now stellation_app.html and the address you land on is a short description with a card for each page. Links shared before the move still work — the app keeps its whole state in the URL hash, and the landing page forwards a hash that looks like app state.
- The JavaScript is in three trees rather than one. docs/app/js is the application, docs/lib the layer it shares with the tutorial and sample pages, docs/js those pages' own scripts. Imports go through barrels that list every name explicitly, so a future move is one file to edit rather than every consumer, and a collision between two shared modules is an error at load instead of a binding that quietly becomes undefined.
Fourth round
- The camera never moves within an arrangement. The third-round fix pinned foreshortening to the selection's bounding sphere, but the camera still traveled as that sphere changed — so toggling an outer cell subtly changed the perspective of everything else, on both machines this time. Distance is now set once per build, from the radius of the whole buildable arrangement; selecting and deselecting cells changes it by exactly nothing, and only zoom — a flat projection scale — does the framing.
- Each view now works at its own scale. The solid adds or removes one cell: shift the cell sitting on the clicked face, ctrl the one behind it, green and red as before. The diagram toggles the cell on either side of its plane: shift the one beneath, ctrl the one on top — in gold and blue, because those are toggles and green/red would lie. The group operation — a box together with its whole supporting set — lives only in the Cells table, on either modifier, and it too is a toggle. The previous "same gestures everywhere" was this port's invention, not the design.
- The Cells table counts again. Every plain box drew the digit 0 — it was showing its sub-cell index, which is always zero for a box with one sub-cell, instead of its cell index. Visible the moment a low symmetry put several cells in one row.
- Du Val's letters, where they exist. For the icosahedron under Ih the boxes now read a, b, c, d, e₁, e₂, f₂, f₁, g₁, g₂, h rather than positions in a sorted list. The identification of which orbit is which letter is the derivation in the research notes, the same one the walkthrough's Ef₁ preset was verified against. Every other solid keeps indices — no other solid has an accepted lettering.
- Scrollbars in the Cells table, horizontal and vertical, drawn when the table overflows and draggable. And switching to a smaller table resets the scroll, so it no longer comes back cropped to where a bigger table left it.
- Symmetry elements in class colors, with a legend. Inequivalent elements — axes no operation of the group carries onto each other — get different colors; for Ih the legend reads C₅ ×6, C₃ ×10, C₂ ×15, m ×15, S₁₀ ×6, S₆ ×10, which is the textbook census. Mirror planes are rimmed with thin tori of the same thickness as the axis cylinders, so an edge-on plane reads as a bar instead of vanishing; the faint translucent fill stays, having been looked at and kept. Checkboxes gray out when the group has none of that kind.
- The elements are marked on the diagram too. A dot where each rotation axis pierces the drawing plane, a dashed line where each mirror plane crosses it, in the same class colors — what the Java applet's diagram checkboxes drew. One toggle: mark on the diagram.
- A real flick keeps the solid spinning. Release while moving fast enough and it spins on at that rate until you catch it — the Java behavior. A gentle release still just coasts to a stop.
- Make planes, ported. The original program's plane editor: one plane per line —
nx ny nz dand optionally a symmetry group to multiply that plane by — seeded from the current solid's own planes, built into an arrangement like any catalog solid, saved into the JSON document. One line,0 0 1 1 Ih, builds the rhombic triacontahedron's thirty-plane arrangement. - Winding repair moved into the engine. The face-orientation repair the derivation tool needed lives in
core.jsnow (orientFaces), because it is a property of the catalog data, not of one tool: most solids in the file are not consistently wound, and anything that ever culls back faces or integrates a volume will need it. - Crowding relief for the zoomed-out diagram. Far below 1× the facet strokes thin and fade with the zoom, so the bundle of near-parallel traces shimmers less. This is mitigation, not elimination — at that scale there are more lines than pixels and no stroke width is honest.
- Documented, not changed: the ten hemi-dual solids whose faces are approximations of faces with vertices at infinity (clamped to ~100 in the data, hence genuinely non-planar) are now documented, along with a short how-to for serving this site from GitHub Pages.
Third round
- Resizing the 3-D pane no longer changes the perspective. The camera distance was derived from the narrower of the two field-of-view angles, so it depended on the shape of the canvas — drag the splitter and you could watch vertical lines bend. Distance now depends only on the solid's radius; a narrow canvas is framed by scaling the projection, which cannot bend a straight line. Measured: the ratio distance ÷ radius is now bit-identical from 1600×600 to 300×1200.
- Toggling a cell no longer changes the perspective either. Same root cause seen from the other side: the radius was floored at 1, so any selection smaller than the first one sat too far from the camera and flattened. The Windows-only symptom was the Cells panel gaining a scrollbar and resizing the canvas. Perspective is now invariant under every selection.
- Symmetry elements are solid geometry, so the solid hides what is behind them. Axes are drawn as cylinders into the same depth buffer as the stellation instead of being painted over the top. Three separate toggles: rotation axes, mirror planes, and rotoreflection (Sn) axes — the three-dimensional cousin of a glide reflection. Mirror planes are drawn as a solid rim with only the faintest wash inside, because fifteen translucent discs compound into an opaque ball that hides the very thing they explain.
- Turning the solid is smoothed, and a flick coasts. Pointer motion is banked and spent a fixed fraction per frame, which evens out the uneven steps a mouse reports; on release the recent speed is added to the buffer so the motion carries on and settles. A first-order filter rather than a spring, so it cannot overshoot and there is no ringing to tune out.
- Undo and redo, as buttons beside core / + layer / all / clear and on ctrl+Z and ctrl+shift+Z (cmd works too). Every change to the cell set is undoable, a hundred deep. Carving from a high cell takes its whole supporting set, which can reach the core, and there is no way to reconstruct what was there from what is left.
- Green adds, red removes, in all three views. Holding shift outlines the target green, the carve modifier outlines it red, and the same two colors are used in the 3-D view, the diagram and the Cells table. Releasing the modifier clears the outline immediately. When the gesture would do nothing — nothing further out to add, nothing behind the face to remove — the outline is dimmed rather than dropped, so you still see where you are pointing but know the click is spent.
- Diagram face is a dropdown of the faces that actually differ. It was a number with no upper limit, so most values did nothing. The list is now the orbits of the face planes under the chosen symmetry, each named by the polygon at the center of its diagram. The cuboctahedron under Oh offers exactly two — triangle · 8 planes and square · 6 planes — and lowering the symmetry splits them, as it should.
- The collapse arrow points back the way it opened — ▸ to open a group of sub-cells, ◂ to close it, rather than a downward arrow that read as unrelated. The arrow and the count of hidden parts are larger, heavier and in a color you can find.
- The view is kept in the URL and in saved documents. The hash now carries the orientation and zoom, so a reload comes back to the same angle and a link shows the recipient the same picture. Saved JSON carries them too.
- A stale build cannot be served silently any more. A good part of the session went on a bug that had been fixed an hour earlier — the browser had an old
app.jsand neither ctrl+shift+R nor ctrl+F5 shifted it. The site now sendsCache-Control: no-cachefor code and markup, so every visit revalidates, and the help panel prints which build you are looking at. - Crossed faces get the right plane. Newell's method sums a signed area, and a bow-tie face has two lobes that cancel — so its normal came out of rounding error. 413 faces across 15 solids were affected. Where the sum collapses, the plane is now fitted to the vertices instead. Seven solids whose faces are otherwise planar go from wrong to correct to machine precision; the rest are the hemi-duals, whose face data is genuinely non-planar.
- Dropped planes are declared. When a solid's faces do not all yield a usable plane, the status line reads ⚠ 8 of 12 planes rather than simply 8 planes, and the tooltip says why. Building an arrangement out of fewer planes than the solid has is the one failure that looks perfectly healthy: you get a stellation, of a different solid.
Second round
- The hover outline sat off the face it marked. A regression from making the model scale sticky: the highlight still computed its own scale, so the yellow triangle floated over the wrong place.
- Fit and home are on the view itself. Fit existed but was inside a collapsed panel section, so it could not be found. Both are now always visible in the corner of the 3-D view. home is the canonical orientation — x right, y up, z toward you.
- Zoom scales the projection instead of moving the camera, so magnifying no longer swings the viewpoint round the solid, and the range is wide enough to frame the largest arrangements. Picking was corrected to match the new frustum.
- Fit eases into place rather than snapping, over about four tenths of a second, and home turns the solid to the canonical orientation along the shortest arc. Any drag or wheel cancels the movement, so it never fights the pointer.
- The camera now retreats as the selection grows. Holding it at a fixed distance was wrong once a stellation reached past radius 6 while the camera sat at 3.3 — the viewpoint ended up inside the solid. It now backs off by the same factor the projection is scaled by, which leaves the on-screen size and the perspective untouched.
- ctrl is the exact inverse of shift. Both now walk inward: shift adds a cell with everything supporting it, ctrl removes that same set. Before, shift reached inward and ctrl reached outward, so the pair never canceled.
- Sub-cell groups start collapsed, with an arrow beside each showing how many parts are hidden. Sub-cells are drawn smaller than their parent so the grouping reads at a glance. Under a low symmetry one orbit can hold 120 sub-cells, which used to run off the panel.
- Symmetry axes can be drawn in the 3-D view — the rotation axes of the current stellation group, through the origin. Checked against group theory: 31 for I and Ih, 13 for O and Oh, 7 for T, 3 for D₂.
- Inward-facing regions in the diagram are less faint. They were pale enough to be missed.
Controls
The largest group, and the one that changes how the app feels.
- Shift always adds; ctrl always removes. Neither toggles any more. Shift-clicking an already-selected cell used to undo the previous shift-click, so the same gesture sometimes added and sometimes removed depending on what you had just done. The two are now plain inverses.
- A bare click no longer edits in the 3-D view or the diagram. It turns the solid and pans the diagram respectively; only the Cells table spends a plain click on toggling. Previously a click meant one thing in one pane and something else in another.
- Right-click does nothing. It used to carve. On macOS ctrl-click is the system right-click, so the two are told apart by whether the ctrl key was actually held — a real right-click is ignored, a ctrl-click removes.
- One vocabulary on both platforms. The remove key is now called ctrl everywhere, because ctrl-click now works on macOS too. Option-click still works there as an alternative.
The Cells table
- Turned the right way up. The core is now the bottom row and each shell sits on the one below it, so the table grows the way the solid does — from the inside out. Row numbers are unchanged.
The diagram
- Inward-facing surfaces are drawn pale and dashed. The diagram shows the surface of the solid, not its volume, and a filled region invited the reading that there is material behind it — when an inward-facing face means the opposite, that there is a cavity. Solid fill now means outward, pale means inward.
Symmetry
- Only genuine subgroups are offered. The dropdowns listed all 85 groups regardless of the solid; for the icosahedron they now list the 13 that are actually subgroups of Ih, and the stellation symmetry is restricted to subgroups of whatever the polyhedron symmetry is set to. Membership is decided by testing the matrices, not from a hand-written table.
- That also fixes the "everything is broken" case. The all-zeros table came from choosing T (order 12) as the polyhedron symmetry with I (order 60) as the stellation symmetry — not a lowering at all, so every group had exactly one sub-cell. That pairing can no longer be selected. The sub-cell splitting itself was correct: under Ih → T an icosahedron's layers split 3/5/10, and under Ih → E into all 120.
Layout and view
- The panes can be dragged. Both dividers resize, and the positions are remembered — a deep arrangement can run a row of the Cells table far wider than the default column.
- The solid no longer resizes itself as you build. The scale was recomputed on every change, so adding a shell shrank everything already on screen. It is now fixed when the arrangement is built, with a fit button to rescale on demand.
- Edge weight now defaults to 1.
Documentation
- A short Controls & views page — the two gestures, the three views and what each pane is for, on one screen. The Cells panel now links there instead of to the walkthrough, under a shorter label.
- This page.
Investigated, not changed
Deliberately left alone, with the findings written down.
- Planes through the origin. Described in the session as a fundamental defect: a plane stored as a single point cannot represent one that passes through the center, which is why twenty hemipolyhedra and their duals are commented out of the catalog — the reason is written in the source, in the author's hand.
Investigated rather than fixed, on the instruction not to start it yet. The short version: the JavaScript port already stores four numbers, so the representation is not the blocker — one guard line is. Removing it does let all ten hemipolyhedra build, u80 included, in under a third of a second. But the guard turns out to be doing a second, unrelated job, and a prototype that removed it naively destroyed seven working solids. The discriminator that separates the two cases cleanly is face planarity, which differs between them by sixteen orders of magnitude. Written up with both prototypes, the measurements and an acceptance test in the design notes. - Clicking a region that has cells on both sides. On the diagram a plane has a cell above and a cell below, and there was a proposal to reach them separately — shift for one, ctrl for the other. Left as it is for now: shift and ctrl are spent on add and remove, which is the more valuable pair, and the session concluded "counterintuitive but ok, can leave as is".
- 3-D rotation in the original Java applet. Noted as a separate to-do against the original program, not this port.
Questions still open — what the "polyhedron" symmetry is for, what "compensate completely" should mean, the cube's orientation, and the cells-versus-surface confusion — are written up in the design notes. Source and notes on GitHub. See Controls & views for how the app works now.