Vladimir described what he called the program's fundamental defect: planes are carried through the pipeline as a single 3-D point — the foot of the perpendicular from the origin — so a plane through the origin cannot be represented at all. That account is accurate for the Java applet, and §3 traces it to the exact line. But in the JavaScript port the story has already half-changed: planes are stored as unit normal + offset {n, d} — four numbers — and the surviving problem is not the representation but three assumptions layered on it: every plane is flipped to d > 0, planes with d ≈ 0 are silently discarded, and the layer/cell semantics presume a core cell that contains the origin.
Measured against the actual goal — how many of Brückner's 146 photographed models can be rebuilt live — the diametral-plane fix turns out to be the smallest of five levers. The census of all 146 figures (§4) gives this ladder:
| Step | Design change | Adds | Running total |
|---|---|---|---|
| — | Today: engine-verified cell selections on catalog solids | — | 16 / 146 |
| D1 | No engine change: derive selections that already exist in shipped arrangements (winding-number automation + per-figure research) | +39 | 55 |
| D2 | Custom plane sets — restore the applet's __PLANES source (or ship kernel geometry files) for compounds and Brückner's “combination” solids | +71 | 126 |
| D3 | Dihedral families: prism/antiprism/stephanoid geometry for the already-shipped Dn symmetry machinery | +13 | 139 |
| D4 | Diametral planes (d = 0): unlocks the octahemioctahedron directly and is a prerequisite inside D1/D2 for Brückner's through-the-center varieties | +1 | 140 |
| D5 | Face-set display mode for the five truly one-sided models (cells can never represent them) | +5 shown | 145 shown |
The 146th, Taf. VIII Fig. 38, remains unidentified — no passage in the book names it. Counts are reachability upper bounds: each derivation in D1 is mechanical but must be run and eyeballed; each kernel in D2 exists mathematically but Brückner's “varieties” sometimes depend on distance ratios he never states numerically (§4.3).
Seven duals that are in today's catalog menu — d23 d26 d44 d55 d61 d68 d78 — currently build silently wrong arrangements. The cause is not diametral planes but a second, previously undiagnosed bug: their faces are antiparallelograms (bow-ties), whose two lobes have equal areas, so the Newell normal the port computes cancels to ≈0 and the face planes come out as noise (§3.4). This is fixable independently of everything else in this report, and should come first.
The clickable-plates page currently rebuilds 16 of 146 figures, every one verified in the engine (vertex/face/volume/outer-radius). The data layer is far larger than the app reveals: geometry.json ships 150 solids — the 75 Kaleido uniforms u06–u80 (file uNN = Maeder U(NN−5)) and their 75 duals — of which the catalog menus expose 121. The 29 orphans are the nine hemipolyhedra, their nine duals, nine duplicate self-dual twins, and u80/d80 (the great dirhombicosidodecahedron and dual). A build census over all 150 files: every file builds under its best symmetry, in 0.03–7 s; the heaviest (d65 d45 d79…) reach ≈23,000 primitive cells at the default depth cap.
Two backdoors matter for what follows. The URL hash loader accepts any geometry key — #u08 loads the octahemioctahedron today, but with a synthesized Ih symmetry that is wrong for an octahedral solid. And parseStel already reads the applet's planes field from .stel files; the preset layer simply discards it. Both are half-built doors into the capabilities this report costs out.
Planes are stored as a point in space — the point of the plane closest to the origin — so a plane passing exactly through the origin cannot be expressed at all. That choice, made at the very beginning, is why a significant number of the polyhedra cannot be reproduced, and probably why many of Brückner’s figures cannot be built either.
“My planes are specified by a point in space… so I cannot specify a plane through the origin. It was a fundamental mistake at the very start of development… a significant share of polyhedra cannot be reproduced… possibly many of Brückner's too.”
— V. Bulatov, feedback session, 6 Aug 2026. His proposed fix: normal (three numbers, not necessarily unit) + distance to the origin — four numbers.
The Java Plane class itself already stores normal + signed distance (v·X − d = 0). The defect is one step downstream: StellationController.makePolyhedronPlanes re-encodes every plane as planes[i] = plane.v.mul(plane.d) — the foot of the perpendicular — and Stellation.makeStellationFaces reconstructs the plane as v = normalize(point), d = |point|. At d = 0 the point is the zero vector and the normal is gone; and since d = |point| ≥ 0, every normal is forced to point away from the origin. Exactly as The original author recalled — one lossy round-trip in the middle of an otherwise four-number pipeline.
// docs/lib/core.js — facePlanes()
if (d < 0) { n = mul(n, -1); d = -d; } // orient away from the origin
if (Math.abs(d) < 1e-9) continue; // plane through the center: skip
The representation is Vladimir's proposed four numbers already; the two lines above are the inherited policy. Note the port converts the Java behavior from loud to silent: the applet's plane-from-point constructor degenerates visibly, while the port just drops the plane — no console message, no count. The vestigial planeFromPoint() has no remaining callers and substitutes an arbitrary normal (0,0,1) at |p| < 1e-12.
Recomputing raw face planes for all 150 files, the solids with genuinely diametral planes are the nine hemipolyhedra plus the great dirhombicosidodecahedron:
| File | Solid | Planes | Through center | Survive the skip |
|---|---|---|---|---|
| u08 | octahemioctahedron orientable | 12 | 4 | 8 |
| u09 | tetrahemihexahedron | 7 | 3 | 4 |
| u20 | cubohemioctahedron | 10 | 4 | 6 |
| u54 | small icosihemidodecahedron | 26 | 6 | 20 |
| u56 | small dodecahemidodecahedron | 18 | 6 | 12 |
| u67 | small dodecahemicosahedron | 22 | 10 | 12 |
| u70 | great dodecahemicosahedron | 22 | 10 | 12 |
| u75 | great dodecahemidodecahedron | 18 | 6 | 12 |
| u76 | great icosihemidodecahedron | 26 | 6 | 20 |
| u80 | great dirhombicosidodecahedron (U75) | 78 | 46 | 32 |
Loading any of these through the hash backdoor today silently produces the arrangement of the surviving planes only — u09 quietly becomes a tetrahedron study. On the Brückner side, the census (§4) shows the fix is a prerequisite for four figures: VIII-1 (the octahemioctahedron — and, being the one orientable hemipolyhedron, it is honestly rebuildable as cells once its diametral hexagons exist), XII-1 (small dodecahemicosahedron), and Brückner's own degenerate-parameter varieties IX-1 and IX-18, where he moves a family of dodecahedron planes until they pass through the center. Vladimir's hunch that “possibly many of Brückner's” fail for this reason is right in kind — the book deliberately visits the degenerate case — but small in count: four figures of 146.
What the real fix requires beyond un-skipping: (a) an explicit orientation convention per diametral plane (both half-spaces touch the center, so “outward” is a choice, not a fact); (b) generalized layer semantics — today layer 0 is the cell containing the origin, and with a diametral plane no cell contains it (for u09 the center sits on three planes at once and the “core” is eight octants); (c) UI: the Cells table, the {0} presets, and the walkthrough all assume a core row. None of this touches the clipping algebra itself — the Java constructions show the arrangement code nowhere needs d > 0.
The census flagged five more “diametral” files — and adversarial verification refuted the diagnosis in the most useful way. d44 d55 d61 d68 (small rhombidodecacron, small dodecicosacron, rhombicosacron, great dodecicosacron) have no diametral faces: they are isohedral, every face at one nonzero distance. But their faces are antiparallelograms — crossed quadrilaterals whose two lobes have equal area with opposite winding — so the Newell area-vector the port uses as the face normal cancels to ≈0, and facePlanes emits garbage normals with scattered distances (spread 80:1 where the true solid has 1:1). One face per file lands at exactly 0 and trips the diametral filter; the other 59 pass through as noise. The same degeneracy exists, undetected by any filter, in d23 d26 d78 — and all seven of these are in the catalog today. (d70, the ninth flag, is the genuine hemi-dual with Maeder's finite fudge for its ideal vertices.)
When |Newell normal| ≈ 0, fit the plane from the (coplanar) vertices instead — any non-collinear triple. One guarded branch in facePlanes repairs seven currently-wrong catalog solids and needs no new semantics at all.
Every one of the 146 figures was classified by what would make it clickable, using the book's own per-figure descriptions (145 of 146 are cited in the text), the shipped data files, and computed face inventories. The full table is in the appendix; here is each rung with its mechanism and its evidence.
Three groups need only derivation work against arrangements the engine already builds:
Self-stars (17 figures). Brückner photographed seventeen orientable uniform star polyhedra that are in the catalog right now; each equals the solid union of cells in its own face-plane arrangement. This is provable mechanically: compute the winding number of the star surface around a representative point of each sub-cell orbit. Run tonight as a demonstration, with orbit-consistency checks:
| Figure | Solid | File | Derived selection | Checks |
|---|---|---|---|---|
| IX-13 | dodecadodecahedron | u41 | {0,1,2(1)} | 0 inconsistent orbits; union reaches exactly the vertex radius |
| IX-15 | small ditrigonal icosidodecahedron | u35 | {0,1(1)} | clean |
| IX-2 | truncated great dodecahedron | u42 | {0,1,2(1)} | clean |
| — | small dodecicosahedron (probe) | u55 | — | 4 inconsistent orbits — orbit granularity coarser than the solid; needs per-primitive verification |
The probe failure is the honest caveat: u55 is one of the non-orientable non-hemi uniforms, where winding is ill-defined — the tool must refuse non-orientable inputs and double-check orbit unanimity. The seventeen self-star figures are all orientable.
Derivation-only (9 figures). The §157–158 solids — Brückner's 15th-kind and 45th-kind isogonal-isohedral solids and their polars, plus XI-11 on the deltoidal icositetrahedron — lie in the planes of catalog solids (d29, d15). Same situation the compound of five cubes was in yesterday before it fell to a membership computation ({0,1,2,3,4} on d29); these need Brückner's ring-point constructions read out of pp. 209–212 and turned into selections.
Dual-stars with existing kernels (13 figures). Much of Plate X and XI is the polar (isohedral) mate of a uniform star, and thirteen of those mates are literally files in the catalog — e.g. X-28 is the medial rhombic triacontahedron (d41), X-30 the small stellapentakisdodecahedron (d42), XI-17 the great rhombic triacontahedron (d59). “Rebuild” here means: load the dual file, run the same winding derivation, get the cell string. (Three of the thirteen — those with bow-tie faces — wait on the §3.4 fix first.)
The single largest rung, 71 figures, and the design choice is already half-present. The Java applet has a custom-planes dialog: enter [(nx,ny,nz)…] or [(nx,ny,nz,px,py,pz)…], pick any symmetry group, stellate the arrangement, save as .stel with the __PLANES source. The port's parser already reads that field; the preset layer throws it away and the app rejects any document whose polyhedron is not a catalog name. Wiring that one path through buildStellation — which takes an arbitrary {n,d} list plus matrices today — restores it.
What it unlocks splits in two. First, the famous compounds in “alien” planes: two cubes on a 3-fold axis (VIII-3), three cubes (IX-23, Escher's), three octahedra (VIII-12, Escher's), two dodecahedra (VIII-31), six sphenoids (VIII-10), six boxes (VIII-22), and the polar of the final stellation (XII-10/16). A verified subtlety on VIII-12: the three-octahedra compound is a stellation of a tetrakis hexahedron variety whose 24 planes are the signed permutations of (0,1,√2)/√3 — not the Catalan solid d13 (signed permutations of (0,1,2)/√5, ≈8.7° away; 0/24 planes match). The compound tier needs its own plane sets; no shipped kernel substitutes. Second, the long tail of Brückner's Part-F “combination” solids — cube+octahedron, two-tetrahedra, dodecahedron+icosahedron families in 8–62 planes — where he explores varieties of the same plane arrangement at different distance ratios. That is precisely a custom-planes workflow: same normals, slide the ratios. (It is also where the caveat lives: for several varieties Brückner states the ratio only qualitatively, so the kernels are one parameter short of automatic.)
Verified as a theorem, not an analogy: any finite union of closed orientable solids equals a union of cells in the arrangement of its own face planes (each open cell is disjoint from every boundary, hence entirely in or out). That is why five cubes fell to {0,1,2,3,4} on d29, and why every compound on the plates is reachable once its plane set can be entered. The converse boundary is just as firm: a one-sided surface bounds nothing, so no cell union will ever be XII-5 (§4.5).
Thirteen figures are prisms, antiprisms, crown polyhedra and stephanoids — Brückner's axial chapter (VIII-4, 13, 15–20, 24, 40, 41; X-17's four-tetrahedra compound; X-21). The remarkable fact from the parity audit: the port ships the entire dihedral symmetry apparatus already — symmetry.json carries C2…C12 and D2…D12 in h/d/v variants — and it is dead data, because no catalog solid has an axial point group and there is no custom-planes path. Kaleido's prism families (#1–5) were never shipped as files, in the applet either. The unlock is parametric: generate {p/q}-prism and antiprism plane sets (or geometry) on demand. One repair rides along: twenty of the 85 symmetry entries (C8–C12, C8h–C12h, C8v–C12v, D8h–D12h) ship with empty matrix arrays in both the applet and the port — selecting them silently yields no group. The stephanoids need exactly these Dn entries to be real.
Covered in §3.3: one direct figure (VIII-1, the orientable octahemioctahedron), plus prerequisite status inside earlier rungs for IX-1, IX-18, XII-1, and honest arrangements for ten orphan files. Ranked here not because it is deep — the arrangement algebra never uses d>0 — but because its blast radius is semantics and UI (layer 0, presets, the Cells table) rather than one function.
Five figures can never be cell unions, by the theorem above: VIII-29 (extended triakis octahedron), IX-10, X-16 (Möbius solids from truncated solids), XII-1, XII-5. The book is explicit that these are einseitig — Brückner's §143/§148/§163 one-sided chapter is where his subject leaves Coxeter's future cell-world entirely. Showing them needs a face-set renderer — draw the self-intersecting faces as given in geometry.json, no arrangement involved — which is a small, self-contained viewer mode, and incidentally the honest way to show any star polyhedron “as Brückner built it”: his paper models are surfaces, not solids.
The session's other capability thread was symmetry. Two findings sharpen it. First, the port's subgroup menu is computed (by matrix containment over top-level groups), where the applet used a hand-written pairing table — and the computation loses all ten icosahedral-frame subgroups (D5d(I) D5(I) S10(I) C5v(I) C5(I) D3d(I) D3(I) S6(I) C3v(I) C3(I)): they exist only as axis-aligned variants inside the applet's table, not as top-level entries. Since 109 of the 121 catalog solids are icosahedral, the practical loss is every 5-fold and 3-fold partial-symmetry stellation — the “pillar” and chiral-axis selections. Second, the granularity story: the engine's two-level design (full group for cell types, one chiral subgroup for hand-splitting) is what made the five-tetrahedra compound selectable; each further subgroup exposed multiplies selectable subsets and multiplies UI rows in equal measure — the transcript's C3 complaint (a group list long enough that it no longer fits without scrolling) is the cost curve of the same choice that Miller's rules were invented to cap (§6).
The design space this report walks was mapped a century before the software. Three traditions correspond to three representation choices:
Brückner thinks in faces. His method is the Grenzfläche: read a boundary polygon off the complete figure (the stellation diagram), then build the paper surface. His §153 description of the icosahedron's complete figure handles the twentieth, parallel plane projectively — “Die Ebene 20 schneidet die Zeichenebene in der unendlichweiten Geraden” — the plane at infinity is just another line of the figure. The irony writes itself: the 1900 book is more projectively complete than the affine d > 0 shortcut in the 2000s software; and Brückner's one-sided §148 models with faces through the center are exactly the cases the point encoding excludes. His face-first stance is also why one-sided solids belong in his book at all: a surface needs no interior.
Coxeter–Du Val think in cells. The Fifty-Nine Icosahedra (1938) replaced the boundary-polygon reading with the cell-set reading this engine implements — Du Val's a…h layer notation is literally the app's {0,…,7}. Miller's rules are then a selection-criteria design choice over the same arrangement: symmetry constraints (his rules fix the rotation group I, admitting the two chiral hands) plus connectivity filters to cut 2N subsets to a publishable 59. Every later fight — Pawley's fully-supported criterion, Inchbald's critique that cells can't distinguish the medial from the great triambic icosahedron — is an argument about that same lever, and the app's shift-click “add with support” is Pawley's criterion as a gesture.
The compounds tradition thinks in plane sets. Five cubes as a stellation of the triacontahedron (verified here as shells 0–4 of d29) goes back through Hess to Cauchy's 1812 memoir; Luke (1957) did the rhombic dodecahedron, Ede (1958) and later Messer the triacontahedron — each of these is a catalog choice: which plane arrangements are worth naming as a “solid to stellate.” The applet's custom-planes dialog was Bulatov's generalization of exactly that decision to the user, and D2 is its restoration.
| # | Change | Kind | Payoff |
|---|---|---|---|
| 1 | Bow-tie normal fix in facePlanes (fit plane from vertices when |Newell| ≈ 0) | correctness, ~10 lines | 7 catalog solids stop being silently wrong; unblocks 3 dual-star figures |
| 2 | Surface the skip: log/status “N of M planes used” whenever planes are dropped or deduped | honesty, trivial | no more silent wrongness anywhere; makes #6 testable |
| 3 | Winding-number derivation tool (refusing non-orientable input) + run over the 17 self-stars and 13 dual-stars; add results to the grid page | tooling, no engine change | 16 → ~55 clickable figures; three selections already derived and verified |
| 4 | §157–158 research derivations on d29/d15 | research | the last gray “stellation of the triacontahedron” entries go live |
| 5 | Custom plane sets: honor the .stel planes source end-to-end; textarea UI later | feature, port-side plumbing exists | +71 figures reachable (compounds incl. both remaining Escher solids; combination varieties); old applet files load |
| 6 | Diametral planes: un-skip, define orientation + layer-0 semantics, expose hemis with correct symmetries | engine semantics + UI | VIII-1 buildable; IX-1/IX-18/XII-1 varieties honest; 10 orphan files real |
| 7 | Dihedral families: parametric prism/antiprism planes; fill the 20 empty symmetry stubs; restore icosahedral-frame subgroups | feature | +13 axial figures; partial-symmetry stellations return |
| 8 | Face-set display mode | viewer feature | the five one-sided models shown; “as built” view for every star polyhedron |
Perf guardrails from the census: everything convex is instant; the big star duals run 2–7 s and ≈5,000–23,000 primitive cells at the default depth caps; 120-plane custom arrangements will need the existing suggestDepth caps surfaced in the UI (“showing N of M layers”) rather than raised.
(1) Grid page VIII-6 is labeled “truncated great dodecahedron”; by Brückner's face inventories the uniform truncated great dodecahedron is IX-2, and VIII-6 is the non-uniform truncation of the small stellated dodecahedron — one of the two labels is wrong, likely VIII-6's. (2) VIII-12's tooltip implies the Catalan tetrakis hexahedron; it should say “a tetrakis hexahedron variety” (§4.2). (3) bruckner.js PLATES notes for Tafel VI/VII were already corrected in-session; no further action.
Tier assignments from the classification pass over the book's own figure descriptions, cross-checked against computed face inventories of the shipped geometry; kernels named as “n planes, group” are described, not yet shipped. Figures are Taf. plate-fig.
Sources: the ported source (docs/lib/core.js, docs/app/js/app.js, preset.js) and data files; the reverse-engineering specs with their Java line references; the per-figure citation sweep of Brückner's text (145/146 figures). Computations run for this report: a 150-solid build census (planes, diametral counts, cells, timings); exact plane-fitting to detect the bow-tie degeneracy; plane-set comparison of the three-octahedra compound vs d13 (0/24 matches); winding-number derivations for u41, u35, u42 (clean) and u55 (refused, correctly). Historical framing follows the project's sourced research notes on the 59 icosahedra, Miller's rules, and the post-1938 literature. Load-bearing claims were checked by independent adversarial passes; two of this report's own working hypotheses were refuted that way (the Catalan-d13 kernel for VIII-12, and “diametral planes” as the cause of the d44-family breakage) and are reported in their corrected form.