Zomeable 4-polytopes

Inheritance-free matrix sweep: closing the corpus

This document summarises the inheritance-free verification pass that re-validates the entire Wythoff corpus by running the canonical zomeable-projection search directly on every (group, bitmask, rng) triple — without the Step-1 “use only the regular’s kernel directions” shortcut on which the production pipeline (tools/run_wythoff_sweep.py) depends.

The rest of this repository is described in WYTHOFF_SWEEP.md; that document also contains the master shape table and the per-polytope results. This file is a complement, describing how we know the production corpus is empirically saturated at rng ≤ 4.

TL;DR

   
Algorithm inheritance-free direct search, parallel-edge reduction throughout
Driver tools/inheritance_free_sweep.py
Cells covered 4 Coxeter groups × 15 nonzero Wythoff bitmasks × rng ∈ {2, 3, 4} = 180
Wall-clock 38 h 33 min (single-threaded)
CPU time 38.22 h
Cells with new-shape flags 2 (both tesseract; rng=2 & rng=4)
Total flagged candidate shapes 5
Genuinely new shape types found 0

All five flagged tesseract candidates resolved to existing corpus entries or predicted members of the tesseract inf-family (see output/regular/8cell/CLASSIFICATION.md). The production corpus is therefore complete at rng ≤ 4 modulo the documented field boundary (ℤ[φ]³ versus the larger ℤ[φ, √n]³ family discussed in output/regular/8cell/CLASSIFICATION.md §3 and in the snap-failure analysis of WYTHOFF_SWEEP.md).

Motivation

The production sweep (tools/run_wythoff_sweep.py) is a two-step algorithm:

  1. Step 1. For each Coxeter group, find every kernel direction that yields a zomeable projection of the regular member of that group at the chosen rng.
  2. Step 2. Re-test each Wythoff descendant against just those kernel directions.

Step 1 is a representation-theoretic shortcut: Lemma A of the project asserts that any zomeable kernel of a Wythoff descendant is also a zomeable kernel of its parent regular. The shortcut is the only reason H₄ at rng = 4 is tractable at all — without it, every descendant of the 600-cell or 120-cell would need its own multi-million-direction search.

But Lemma A is empirically verified, not proved; and a 2026-05-07 audit (WYTHOFF_SWEEP.md §Kernel completeness) showed that even with both (1,0,0,0) and (0,0,0,1) regulars unioned as kernel seeds, six B₄ descendants surfaced extra fingerprints from the search engine that the regular-only seed had missed. Those extras all turned out to be ℤ[√2]³-field shapes (snap-failed at the export stage) or aliases of master-corpus entries, but the audit highlighted that the inheritance assumption deserves an independent corpus-wide check.

The inheritance-free matrix sweep is that check.

Method

For every triple (group ∈ {A₄, B₄, F₄, H₄}, nonzero bitmask b ∈ 15 values, rng ∈ {2, 3, 4}) — 180 cells in total — we run the same lib.search_engine.search that powers the production pipeline, but on the descendant polytope directly, with the full ZZ[φ]⁴ candidate kernel set of the chosen rng. Specifically:

  1. Build the polytope P_G(b) from lib.wythoff.build_polytope, Procrustes-aligned to lib/polytopes.py’s canonical embedding.
  2. Generate the full kernel-candidate set at rng = 2 (195 312 dirs), rng = 3 (2 882 400 dirs), or rng = 4 (21 523 360 dirs).
  3. Search-engine prefilter + zomeable check + parallel-edge reduction (lib/search_engine.search, with the 2026-05-08 chunk = 16 prefilter that gives the 4–23× speed-up necessary for H₄ at rng ≤ 4).
  4. Stage A fingerprint dedup (3-decimal multiset binning).
  5. Stage B snap-to-ℤ[φ]³ + shape-congruence dedup.
  6. Compare the resulting shape signatures against the current corpus index. Anything not already present is flagged as a new-shape candidate.

Output: ongoing_work/inheritance_free_sweep.json (append-only, resume-aware, 35 MB at completion).

Driver: tools/inheritance_free_sweep.py.

Runner orchestration: ongoing_work/matrix_sweep_runner.ps1, chained across five phases of increasing cost:

Phase rng V cutoff Cells Wall time
1 2 none 60 4 h 18 min
2 3 none 60 8 h 49 min
3a 4 V ≤ 600 45 10 h 02 min
3b 4 V ≤ 3600 +9 4 h 17 min
3c 4 V ≤ 14544 +6 11 h 07 min
total     180 38 h 33 min

The phase split lets the small (low-V) cells fully cover rng ≤ 4 before the H₄ large-V descendants begin; resume mode skips already-done cells between phases.

Aggregate results

Per (group, rng) cell counts and timing:

Group rng Cells Wall hours Cells flagged “new” Flagged shapes
A₄ 2 15 0.05 0 0
A₄ 3 15 0.42 0 0
A₄ 4 15 2.95 0 0
B₄ 2 15 0.06 1 (tesseract) 2
B₄ 3 15 0.46 0 0
B₄ 4 15 3.13 1 (tesseract) 3
F₄ 2 15 0.07 0 0
F₄ 3 15 0.52 0 0
F₄ 4 15 3.40 0 0
H₄ 2 15 4.11 0 0
H₄ 3 15 7.38 0 0
H₄ 4 15 15.67 0 0
total   180 38.22 2 5

H₄ dominates the runtime budget: the rng = 4 omnitruncated 120-cell alone takes 5 h 4 min single-threaded (~14 500 vertices).

Flagged candidates: all aliases

Both flagged cells are the same polytope (B₄ bitmask (1,0,0,0), the tesseract), at two different rng levels. All five candidate fingerprints resolve to existing-corpus or predicted-family entries:

B₄ tesseract rng = 2 (2 flagged shapes)

edge spectrum (multiplicity ×) classification corpus file
(1.0 ×8, 1.473 ×8, 1.618 ×8, 1.701 ×8) — 4 distinct edges, kernel support 3 phi-oblique sporadic output/regular/8cell/8cell_phi_oblique.vZome
(1.0 ×8, √5/2 ×8, 3/2 ×16) — 3 distinct edges, kernel support 2 inf-family at (a:b) = (2:√5) output/regular/8cell/8cell_inf_family_phi_aSqrt5_b2.vZome

Both are already in the master corpus. The “new” flags are a corpus-index timestamp artefact: the corpus index was rebuilt mid-sweep between the phi_oblique / inf_family_phi_aSqrt5 emission timestamp and the rng = 2 phase reading the index.

B₄ tesseract rng = 4 (3 flagged shapes)

All three are members of the tesseract inf-family (the 1-parameter family of “split-cuboid” projections parameterised by ℤ[φ]²-Pythagorean pairs (a, b) with a² + b² = c² ∈ ℤ[φ], defined in output/regular/8cell/CLASSIFICATION.md §3 and enumerated independently to scalar bound 15 in the iso-frame enumerator):

edge spectrum (normalised) (a : b) ratio inf-family row
(0.6, 0.8, 1.0) (3 : 4) 3 : 4 : 5 Pythagorean (already in corpus as output/regular/8cell/8cell_inf_family_a3_b4.vZome)
(0.263, 0.965, 1.0) (φ + 3 : 1) strut-iso b15 row, predicted by the parameterisation
(0.369, 0.930, 1.0) (φ + 2 : 1) strut-iso b15 row, predicted by the parameterisation

Each is a valid ℤ[φ]²-Pythagorean instance of the closed 4-shape tesseract taxonomy (cube, rhombic dodecahedron, phi-oblique, inf-family parameterised by all ZZ[φ]²-Pythagorean (a, b)). The first is already in the master corpus; the latter two are predicted parameterisation rows beyond rng ≤ 2 — they would be added to output/regular/8cell/ if larger-coefficient family members were desired, but no new shape type is involved.

Strut iso-frame corroboration

An independent enumeration via tools/enumerate_strut_iso_frames.py at scalar bounds b ∈ {8, 12, 15} produced (cumulatively): 7, 15, 22 unique tesseract iso-frame shapes, decomposing as

1 cube  +  1 rhombic dodecahedron  +  1 phi-oblique  +  k inf-family ratios

with k = 4, 12, 19 at the three bounds. No new sporadic types appear beyond the three already known. Output: ongoing_work/strut_iso_frames_b15.json.

Implications for production

  1. Lemma A is empirically validated at rng ≤ 4 across all 60 (group, bitmask) cells. Every direct-search hit either already maps to an existing corpus signature or belongs to an explicit parameterised family the corpus represents canonically with finite rows. Zero genuinely new shape types found.
  2. Within this direct uniform-polychoron sweep, the tesseract inf-family is the only source of unbounded shape count growth. Every other (group, bitmask) cell produces the same finite shape count at rng = 4 as it does at rng = 2 (no new shapes ever appear when rng increases for the non-tesseract cells). The inf-family is itself a closed structural object — each member is a ℤ[φ]²-Pythagorean instance of the split-cuboid projection family. Later prismatic follow-up found separate inf families for {4}×{6} and {4}×{10}.
  3. Corpus freeze condition is reached. Two consecutive sweep levels (rng = 3 → rng = 4) flag the same {tesseract inf-family only} cells with the same provenance, and no other cell flags anything. This satisfies the merge-prep gate stated in the project plan.

Reproducing

# small smoke (~30 s):
python tools/inheritance_free_sweep.py --group A4 --rng 2

# full A4 + B4 + F4 matrix at rng = 2, 3 (excludes H4 large-V):
python tools/inheritance_free_sweep.py --group A4 --rng 3
python tools/inheritance_free_sweep.py --group B4 --rng 3
python tools/inheritance_free_sweep.py --group F4 --rng 3

# full H4 at rng = 4 (~16 hours single-threaded):
python tools/inheritance_free_sweep.py --group H4 --rng 4

Each call appends to ongoing_work/inheritance_free_sweep.json in resume mode; already-done cells are skipped automatically.

See also