Both observations are correct, and neither is an accident. Measured over every (solid, G, H) the menus can currently express — six solids, both group choices, all 210 subgroup pairs — the per-plane coset coloring is entirely gray in 76% of cases, rising to 84% when G is a full group (Ih, Oh), which is the default. And per-plane and per-facet produce the identical picture in every case where the plane coloring succeeds at all, because the facet mode inherits plane labels by construction; the two differ only when the planes gray entirely.
The gray is mathematically forced, not a bug (§2): the coset coloring exists on an orbit of pieces precisely when some piece's stabilizer fits inside H, and under a full group every face-plane stabilizer contains mirrors that most subgroups lack. On the icosahedron under Ih, 19 of the 22 offered subgroups gray every plane; the entire icosahedral menu contains exactly one interesting per-plane coloring — I/T, the five tetrahedra.
The way out is not to fix the coset coloring but to widen the family. Five schemes, each computed and verified on the real engine (§4): the current blocks (left cosets — "which copy of the sub-figure"); H-orbits (the honest version of "right cosets" — always defined, never gray); blends (the color-mixing idea, which turns out to be canonical, with a theorem fixing exactly how many colors mix); mirror-cut refinement (split each facet along its own mirror traces, after which the full coloring always exists); and coarsening (the nearest subgroup that does color). Section 5 checks all of them against ten classical pictures — the compounds of five tetrahedra, ten tetrahedra, five octahedra, five cubes, Wenninger's color tables, the antipodal conventions — and §7 distills a recommendation ladder.
Four concrete defects in what ships today (§6): the per-facet fallback labels each orbit independently, so on the gallery's own five-tetrahedra compound under Ih/T none of the 60 spikes comes out monochrome; the facet path cannot be steered to the correct hand of a chiral compound (the plane path can), so switching modes can silently flip the enantiomorph; symmetry.json carries stale subgroups lists naming entries that do not exist (no code reads them, but this survey did — see §6c as corrected); and the "fixed by all of H" rule grays the seed planes of Ih/C3v(I), leaving 18 of an intended 20 colors.
Fix one G-orbit of pieces (planes, facets, or cells) and a representative q₀ with stabilizer S = StabG(q₀). The piece g·q₀ determines g only up to the coset gS, so a labeling by cosets of H is consistent exactly when the coset is constant on gS:
The code implements the left labeling: in both cosetClasses and facetCosetClasses the test is g ∈ rH (the transpose contraction tmul(r, g) is r⁻¹g for orthogonal matrices). Verified directly: with H = T in Ih the code's partition of the 120 group elements equals the left-coset partition and differs from the right one.
When S ≤ H the color classes are the G-translates of the H-orbit of q₀ — a system of imprimitivity with [G:H] blocks of [H:S] pieces each. The five-tetrahedra coloring is this: each block is one T-orbit of 4 planes, a regular tetrahedron. And the gray criterion falls out: an orbit colors iff some conjugate of S lies in H; otherwise no consistent labeling exists and gray is the honest answer.
Now the census. Every icosahedron face-plane stabilizer in Ih is C3v — order 6, and it contains a mirror. A subgroup H can host that stabilizer only if it holds the matching mirrors, which almost none do:
| G on the icosahedron | subgroups offered | per-plane coloring exists for | everything else |
|---|---|---|---|
| Ih (order 120) | 22 | Ih (1 color) · D3d(I) (10) · C3v(I) (18 of 20 — see §6) | 19 of 22 all gray, including every chiral subgroup — T, I, all the Cn/Dn |
| I (order 60) | 9 | I (1) · T (5, the five tetrahedra) · D3(I) (10) | 6 of 9 all gray |
So the coloring the feature was built for exists, under exactly one chiral pair — and the default G, being the full group, can never show it per plane. Across all six solids the proportions are the same in aggregate:
| population | pairs | per-plane all gray | per-facet all gray | facet ≡ plane | facet rescues |
|---|---|---|---|---|---|
| all (G, H) menu pairs, six solids | 210 | 76.2% | 34.8% | 58.6% | 41.4% |
| G = full group (Ih, Oh) | 150 | 84.0% | 40.0% | 56.0% | 44.0% |
| G = rotation group (I, O) | 60 | 56.7% | 21.7% | 65.0% | 35.0% |
| excluding the trivial H = G and H = E | 186 | 79.6% | 32.8% | 53.2% | 46.8% |
Solids: icosahedron, dodecahedron, octahedron, cube, cuboctahedron, icosidodecahedron — full arrangements, every listed subgroup, 8.3 s total. "Facet rescues" = planes gray entirely but the facet fallback produces colors. H-orbit colorings (§4.2) grayed in 0 of 210 rows.
By construction. facetCosetClasses first computes the plane labeling; if any plane label exists, every facet inherits its plane's label verbatim, and only when the planes gray entirely does it fall back to labeling facet orbits on their own. The inheritance was added for a good reason — the planes of a face-transitive solid form a single orbit, so one anchor makes the whole figure coherent — but it means the second menu entry adds information only in the fallback case. Measured: identical in 58.6% of the 210 pairs, which is 100% of the pairs where any plane colors; the remaining 41.4% are rescues, where planes are all gray and facets still find structure (the chiral hands under Ih/I being the flagship).
The rescue is real but has limits. On the dodecahedron every facet of the full arrangement is mirror-fixed — the arrangement is achiral down to its smallest pieces — so the facet fallback rescues nothing there: 13 of 22 subgroup choices are dead in both modes, and tetrahedral subgroups (the five-cubes partition!) can show nothing at all on the dodecahedron itself.
The compound of five octahedra (cells {0,1,2} of the icosahedron) wants one color per octahedron. Each icosahedron plane carries a hexagram — coplanar triangles of two different octahedra — so no per-plane coloring can be right in principle. Under (I, T) the per-plane labeling nevertheless succeeds numerically (5 colors, zero gray) and paints each whole hexagram one tetrahedron-partition color, splitting every octahedron across colors; the facet mode then inherits that wrong picture. Per-facet blocks with their own anchor would give the right one. Mode choice should be available to the user even when the plane labeling exists (§7, R3).
Everything below is G-equivariant (rotating the figure permutes colors consistently), verified on the engine, and reduces to the current coloring where the current coloring exists.
Color = "which G-translate of the reference sub-figure H·q₀ contains this piece." [G:H] colors when defined; the components-of-a-compound coloring; needs an anchor per orbit and, for chiral figures, the hand choice (which preferCells already solves for planes by scoring candidate anchors against the selected cells). Keep it — it is correct, and correctly gray.
Two pieces share a color iff some h ∈ H carries one to the other. On a free orbit these classes are the right cosets Hg; in general they are the double cosets H\G/S — and unlike blocks they always exist, for every (G, H), with no gray (0 gray rows in the whole sweep). The class count comes from Burnside on the restricted action; class sizes vary, which is exactly why they are not cosets — the removed cell-coloring mode was this scheme, mislabeled. The two colorings answer different questions:
Per plane this scheme already ships under another name — "by stellation face class" is H-orbits of planes with H the editing symmetry. Offering it per facet and per cell, with H the coset subgroup, is new, cheap, and honest. Per cell it is the always-defined replacement for the old cell mode — useful for physical models, where a cell is the unit that gets painted.
The color-mixing idea is not a heuristic; it is the canonical object that remains when blocks fail. For a piece p = g·q₀ the well-defined datum is the set of left cosets C(p) = { g′H : g′·q₀ = p } — the set of translates of the reference sub-figure that contain p. Three verified facts make it well-behaved:
The showcase is the case the complaint started from. G = Ih, H = T, all-gray today. Blended, every face mixes exactly two of the ten coset colors — always one right-handed tetrahedron and one left-handed one, because the ten translates of the reference tetrahedron are precisely the two chiral compounds of five. All 20 duotone pairs are distinct; grouping faces by the rotational member of the pair reproduces one hand's five-coloring, by the reflective member the other hand's. The blend superimposes the two enantiomorphic compounds on one figure — a picture the crisp scheme cannot draw at all. A coarser pastel variant exists one step up: H = Th (the stabilizer of one stella octangula) gives five crisp blocks per facet, or ten distinct two-mixes per plane.
Where blend degenerates it should stand down: under Ih/I every plane mixes both cosets — a uniform 50% pastel, informationally gray — and on the dodecahedron with tetrahedral H every plane touches all five sub-figures (36 of 210 rows degenerate fully). Those cases need refinement (§4.4) or the facet fallback, not blending.
Gray pieces are pieces whose own symmetry straddles cosets. Cutting each such facet along the traces of its stabilizer's mirrors removes the obstruction: the sectors have smaller stabilizers, and in the limit (barycentric flags) the stabilizer is trivial — proven, and spot-checked on the engine — so the full [G:H]-color block coloring exists for every (G, H) with no gray at all. The minimal version only cuts facets that need it: a facet with dihedral stabilizer Cnv splits into 2n sectors wearing crisp alternating coset colors. Under Ih/I this reproduces the classical "one face wearing two colors separated by a mirror line" convention exactly, and it eliminates precisely the residual gray of the facet mode (the 380 mirror-straddling facets of the icosahedron). The cost is real — subdivision in the mesh, the 2-D diagram, and every exporter — which is why it is last on the ladder, not first.
If no conjugate of S fits in H, the minimal admissible enlargements are the groups ⟨H, Sp⟩ over pieces p; taking the smallest gives "the closest coloring to the one asked for." It genuinely helps for axial subgroups — Oh/C4 on the cube grays, but ⟨C4, S⟩ = C4v colors all six faces; Ih/C2 lands on D3d with ten colors — but it cannot save the marquee cases: ⟨T, C3v⟩ = Ih, so Ih/T coarsens to one color. (I is simple; almost everything in the icosahedral lattice collapses.) Worth offering as a hint ("nearest coloring subgroup: D3d") rather than as a mode.
| scheme | defined | colors on one orbit | G permutes palette | H fixes picture | five tetrahedra | anchor |
|---|---|---|---|---|---|---|
| 1 · blocks (left cosets — current) | iff a conjugate of S ≤ H | [G:H] | yes | no | yes | per orbit + hand |
| 2 · H-orbits ("right cosets", generalized) | always | |H\G/S| | iff H ⊴ G | yes | no (4+4+12) | none |
| 3 · blend over the covering translates | always | mixes of [S:S∩H] | yes | no | degrades gracefully | max |S∩H| |
| 4 · refinement (mirror-cut) + blocks | always | [G:H] | yes | no | yes + split faces | inherited |
| 5 · coarsening to min ⟨H, S⟩ | always | [G:K], may be 1 | yes | no | collapses | minimize over orbit |
| 6 · stabilizer class (diagnostic overlay) | always | # classes | invariant | yes | no | none |
"H fixes picture": applying a motion of H maps every piece to a same-colored piece. Blocks and H-orbits are the two halves of the owner's left-vs-right question; they coincide exactly when H is normal — on the icosahedral menus, only for H = I in Ih.
A coloring system for stellations should be judged by the pictures geometers actually draw. Ten, from Escher's hand-colored Brückner plate to Wenninger's color tables — each with the recipe that produces it and whether today's code can:
| # | picture | recipe (G · H · scheme · pieces) | today |
|---|---|---|---|
| 1 | Five tetrahedra, 5 colors, chosen hand (Escher via Brückner IX-11; Cundy & Rollett; W24) | I · T · blocks · planes, hand from selection | works |
| 2 | Ten tetrahedra, 5 colors by chiral pair / cube (C&R; W25) | Ih · Th · blocks · facets | incoherent §6a |
| 3 | Ten tetrahedra, 10 colors, one per tetrahedron | Ih · T · blocks · facets | incoherent §6a |
| 4 | Ten tetrahedra, 2 colors, right vs left compound | Ih · I · blocks · facets | works |
| 5 | Five octahedra, 5 colors, one per octahedron (Hess 1876; W23) | Ih · Th · blocks · facets, overriding the per-plane labeling | obstructed §3 |
| 6 | Five cubes, 5 colors (C&R "ideally in five colors") | triacontahedron d29 · Ih · Th · blocks · planes | verified: 5 × 6, 0 gray |
| 7 | Chiral halves, 2 colors, mirror-fixed pieces gray | Ih · I · blocks · facets | works |
| 8 | Wenninger's icosahedral 5-color table, any icosahedral stellation (incl. great icosahedron) | I · T · blocks · planes | works |
| 9 | Opposite-faces-alike: 6 colors on dodecahedral, 10 on icosahedral, 15 on triacontahedral stellations | Ih · D5d(I) / D3d(I) / D2h · blocks · planes (H ∋ inversion ⇒ antipodal pairs) | verified, all three |
| 10 | Hess/Brückner isogonal convention; one-tetrahedron highlight (4+4+12) | any · any · H-orbits · planes exists as "face class"; facets/cells new | partial §4.2 |
Two structural notes fall out of building this table. First, Th is the workhorse subgroup nobody can currently reach usefully — it is the stabilizer of a stella octangula, of a cube in the compound of five, and of an octahedron in the compound of five, so rows 2, 5 and 6 all run through it. Second, the blend of §4.3 and the Th blocks are two roads to the same place: blending Ih/T pairs the cosets exactly as ⟨T, σ⟩ = Th merges them (Td is not a subgroup of Ih; Th is the only order-24 option).
What no scheme reproduces — declared out of scope rather than papered over: the map-style chromatic colorings (the 4-coloring of the dodecahedron's 12 planes, three faces per color, adjacent faces distinct, used in many published models). Ih has no index-4 subgroup and the partition is not G-invariant; it is graph coloring, not symmetry, and would need a separate manual-palette facility if ever wanted.
a · The facet fallback is incoherent across orbits. When planes gray, each facet orbit picks its own anchor independently, so labels do not agree between orbits. Measured on the gallery's own compound-of-five-tetrahedra document under Ih/T: 10 colors appear, but 0 of the 60 spike cells wear a single color — every spike mixes. (Under Ih/I the fallback happens to cohere — index 2 leaves no room to shuffle — which is why the flagship chirality picture works and hid the defect.) The fix is the same anchoring idea that already works for planes: one global reference, with per-orbit anchors chosen compatibly (equivalently: label a facet by which translate of the H-invariant component contains it).
b · The facet path cannot be steered. cosetClasses accepts preferCells and picks the hand of a chiral compound from the on-screen selection; facetCosetClasses has no such parameter and the worker passes none (worker.js:339 vs :341). Consequence: on the five-tetrahedra document, per-plane picks the correct hand and per-facet may pick the wrong one — switching modes can silently flip the enantiomorph.
c · The stored subgroup lists are stale generator residue. Corrected after implementation began. The first version of this paragraph claimed the coset menu offered two broken names; in fact the app never reads the subgroups fields at all — its menus compute literal matrix containment — and the C3/S6 subgroup types were reachable inside Ih all along as the aligned cubic-frame copies C3(O) and S6(O) (the 18-of-20 coloring works today as I/C3(O)). The real defect is dead, wrong data: across the whole file the lists named 60-odd phantom entries (C3(I), Cs(C2v), C2(D), …) and wrong-frame names not actually contained — harmless to the app, but poison to anything that trusted them, including this report's own survey, which enumerated its subgroup pairs from them.
d · The seed-plane wart. The rule "a piece fixed by all of H is gray" is right for the core cell (the owner's original spec) but fires on well-defined singleton blocks too: under Ih/C3v(I) the two planes on H's own axis gray, leaving 18 of the intended 20 colors. A singleton block has an unambiguous label; coloring it and marking "sits on the symmetry" in the legend loses nothing. Judgment call — but it should be a decision, not a side effect.
| step | change | unlocks | size |
|---|---|---|---|
| R0 | Correctness: regenerate the stale subgroups lists from containment (see §6c as corrected); give facetCosetClasses the preferCells steering; anchor the facet fallback globally instead of per orbit | rows 2, 3 of §5 become coherent; mode switching stops flipping hands; the data stops misleading its readers | S–M |
| R1 | Blend as the graceful degradation of the existing modes: crisp where blocks exist → duotone/mix where [S:S∩H] is small (anchor at max |S∩H|, average in Oklab or draw two-tone) → gray only where the mix is total | Ih/T shows the superimposed chiral compounds instead of gray; most of the 84% headline becomes informative; the legend can say why something is pastel | M |
| R2 | H-orbit coloring as a named scheme, per plane / facet / cell ("by subgroup orbits") | never-gray option for every (G, H); the one-tetrahedron highlight (4+4+12); the honest form of "right cosets"; per-cell coloring for physical models | S–M |
| R3 | Let per-facet blocks be chosen even when per-plane succeeds (compute facet blocks directly with the global anchor; agreement with per-plane then emerges where it should instead of being imposed) | five octahedra (§5 row 5); removes the last redundancy between the two menu entries | M |
| R4 | Mirror-cut refinement (facet sectors along stabilizer mirror traces; barycentric as the universal fallback) | every (G, H) colorable with zero gray; the classical split-face pictures; eliminates the 380 residual gray facets of Ih/I | L |
UI shape that falls out: one coset entry with three small controls — subgroup (as now), pieces (planes / facets / cells), scheme (cosets / subgroup orbits), with blend as the built-in degradation of "cosets" rather than a separate mode. The legend earns its keep here: it should show the palette with [G:H] or the orbit count, mark blended swatches as pairs, and name gray precisely ("fixed by the subgroup" vs "no coset can label this orbit"). A one-line hint for dead cases — "nearest subgroup that colors: D3d" (§4.5) — turns the remaining grays from a wall into a signpost.
R0–R2 landed the day this report was published. The subgroup lists are regenerated from containment and pinned by a test; the facet fallback labels all its orbits in one frame stitched through the cells they cap (all sixty spikes of the gallery compound now monochrome under Ih/T, where before none was) and takes the same hand-steering as the plane path; blends ship exactly as §4.3 specifies, mixed in Oklab, with the twenty Ih/T duotones verified live; and by subgroup orbits — per plane, per facet, per cell — is a menu family of its own. Implementation corrected two of this report's numbers, marked in place: §6c's mechanism, and 44 cell orbits under T (Burnside: 528/12), not the survey's 40. R3–R4 remain open.
An adversarial review of the implementation (four independent lenses, each finding verified by refutation) then surfaced and fixed six more defects: the blend path skipped the fixed-by-the-subgroup gray rule (the cube under Oh/C4 blended its two axis planes); the fallback's tie-break settled orbits with a choice before the forced orbits that would have informed it (112 → 52 multicolored caps on the showcase, the reviewers' computed optimum); the fallback cost |orbit| group sweeps per orbit — a 16-second stall on the great dodecahemicosahedron under E, now 0.6 s via an integer multiplication table; per-cell orbit colors disagreed between the diagram and the solid on exposed undersides; the diagram-export cache ignored the coloring subgroup; and a custom-planes document reopened with its hand steered by the wrong selection.
R3 and R4 followed. Per-facet blocks are now computed directly with the plane labeling as the last tie-break — with nothing selected the two modes still agree exactly, and with the compound of five octahedra selected each octahedron wears its own color, the picture §5 row 5 said the inherit rule obstructed. The mirror-cut refinement of §4.4 ships as a menu mode ("per facet, mirror-split"): the 380 mirror-straddling facets of Ih/I split into halves and sextants wearing the two hands in equal measure, pieces tiling their facets to 10⁻⁹, no gray left, in the 3-D view, the diagram and the exports. Every rung of §7's ladder is now built.
Everything numerical ran on the shipped engine (docs/lib/core.js, buildStellation / cosetClasses / facetCosetClasses) against data/symmetry.json matrices, tolerance 10⁻⁶. The sweep: six solids at full arrangement depth, G ∈ {full, rotation}, H = every listed subgroup — 210 pairs, five schemes each, 8.3 s. Verification was adversarial and independent: the mathematical framework was reduced to 15 numbered claims, each confirmed by a separate pass that re-derived the group theory and re-ran the computations from scratch (15/15 confirmed, two cosmetic corrections); the sweep's golden rows were recomputed independently, which caught one real interpretive error — a blend measured at an arbitrary anchor rather than the optimal one — now §4.3's warning. The figures were asserted in code before being drawn (block sizes, tetrahedron angles, orbit sizes, blend cardinalities). The repo's own suite (docs/test/cosets.mjs, 25 checks) passes throughout.
| solid · G · H | per plane | per facet | H-orbits (pl / fc / cells) | blend (planes) |
|---|---|---|---|---|
| icosa · I · T | 5 colors, 0 gray | identical | 3 / — / — | crisp (= blocks) |
| icosa · Ih · T | all gray | 10 colors, 28% gray, incoherent | 3 / 113 / 44 | 20 duotones (2 of 10 each) |
| icosa · Ih · Th | all gray | 5 colors, 1.5% gray | — | 10 duotones |
| icosa · Ih · I | all gray | 2 colors + 380 gray (480/480/380) | — | degenerate (total mix) |
| icosa · Ih · C5v(I) | all gray | 12 colors, 1.5% gray | — | 20 distinct triples |
| icosa · Ih · D3d(I) | 10 colors, 0 gray | identical | — | crisp |
| dodeca · I · T | all gray | 5 colors, 6% gray | — | degenerate (each plane meets all 5) |
| dodeca · Ih · T / I | all gray | all gray (every facet achiral) | colors fine | degenerate |
| dodeca · Ih · D5d(I) | 6 colors (antipodal pairs), 0 gray | identical | 2+10 | crisp |
| triaconta · Ih · Th | 5 colors × 6 planes, 0 gray (five cubes) | identical | — | crisp |
| cube · Oh · C4 | all gray | — | — | coarsens to C4v: 6 colors |
Full data: 210 rows in results.json (sweep scripts alongside), plus the verification scripts — kept with the analysis work, not shipped in the repo.
Prepared for the stellation app — a JavaScript port of Vladimir Bulatov's 1998–2001 Stellation applet. Palette note: coset i wears … hsl(i·137.508°, 60%, 55%) — the golden angle, so any index count stays distinguishable — and gray marks a piece no coset can honestly claim. All figures in this report use that palette, so they match the app.