Independently audit anti-Fano rank-three completions

[petalas]

Summary

Independently corroborate the anti-Fano minimum raw pairing cost of 12, answering item 1 of the independent-check request.

pinned twelve-line configuration
  -> three universal triangular minors
  -> 786,432 column-basis assignments and 12,288 row-basis assignments
  -> one rank-three completion per coordinate block
  -> credited exchange argument plus an admissible (4,4,4) witness
  -> minimum raw cost 12

The new package is research/antifano-subspace-audit/. It replaces the SAT uniqueness step of DreamingOfClouds' #2111 with two complete standard-library subspace censuses. No predecessor implementation is imported. The configuration, license and source notice are retained unchanged.

Exact predecessor heads:

• Three-total root-unit decoders: reductions, a sharp example and the remaining lift question (no κ claim) #2111: 98da76c2839ba38538b510420dec54546ccabf0b, configuration and credited exchange argument.
• ✈️ Characteristic-2 shadows of rank-3 decoders: the anti-Fano bound by first-order jets, with a Lean-checked core (no κ claim) #2162: 113a718265c62aec301c5695f4bfb9c421bef7ab, configuration source and alternate jet proof.

The original bit-helper-contract draw is explicitly marked as a changed scope after existing H2/K and profile-audit coverage was found. This is an independent decoder-floor audit, not a helper-contract advance.

Evidence

Before, #2111's uniqueness step used SAT factorization and enumeration. After, independent packed-column and packed-row implementations agree entrywise after complete finite censuses, without SAT, jet or floating-point dependencies.

python3.12 scripts/research_package.py --package research/antifano-subspace-audit

PASS_ANTIFANO_SUBSPACE_AUDIT minimum=12 blocks=3 column_trials=786432 row_trials=12288 unique_rank3=1,1,1 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/antifano-subspace-audit outcome=completed

• Ten package test groups pass normally and under -O, including all 4,096 three-by-four binary matrices against a separate dense rank oracle. All 65 research workflow tests pass.
• Source-export replay passes without Git or network. Changed-source and false-receipt CLI controls reject.
• A temporary counting bug that leaves the unique-completion baseline unchanged is rejected by the 64-completion fixture.
• Ruff 0.17.0 passes the package's scoped ALL configuration.
• Canonical main 0fdd7d82, priority revision 4, and both predecessor heads were unchanged at 2026-10-14 23:11 UTC. The refreshed overlap search covered 436 open PRs and 877 discussion records with complete pagination. It does not cover all diffs or unpublished work.

The result is min_B cost(B) = 12, not equality for every admissible B. The full incidence matrix costs 24. No general T=3 theorem, enumeration of all 217 minimizers, complete predecessor replay, selected cold replay, physical supplier or new kappa is claimed. No independent human review or Lean build is claimed. Open next questions and exact source pins are included in the package.

Merge Danger

Door: two-way

Blast Radius: package

Only eleven new package files are added. No selected supplier, frontier record, frozen archive, shared checker or existing replay is changed.