Independently corroborate the corrected seven-label word exclusion

[petalas]

Summary

research/f7-union-support-audit/
  census.cpp       independent iterative union-support census
  oracle.py        unpruned even-block partition oracle
  verify.py        fresh builds, controls and complete replay
  README.md        pruning, completeness and source-impact proofs
  FULL/QUICK.json  exact reproducible receipts

Corroborates Nicholas Stone's #2973 result 22 <= f(7) <= 24 in the PR943 singleton prefix-walk model with a separately implemented full-domain census. This is not a new bound, supplier admission or selected-frontier change.

The new engine uses complete stack snapshots and a completion bound on the union of same-parity pair supports. It does not import or adapt the corrected C engine. #2984 accelerates the predecessor's representation but explicitly retains its DFS/pruning; this closes the different independent-implementation handoff.

Evidence

• Before, the actual pinned PR2542 engine rejects universal 01021323 with the substituted bit-0 budget. The source-bound positive and false-prune calls are reproduced here.
• After, six uncapped runs from the empty prefix complete with zero hits: all five L20 split representatives and the L18 control. Fresh GCC14 and Clang19 builds give identical complete JSON receipts.
• The independent oracle matches every returned witness in 270 comparisons over eight complete small domains. Positive counts include 1,116 universal five-label length-12 words. Twelve cap/error controls, four compiled incorrect implementations and a complete sanitizer comparison pass.
• Registered Python 3.12 replay passes in 115 seconds locally. Clang's full command passes in 87 seconds. These are local command times, not speedup claims. Ruff and metadata/source-pin checks pass.

python3.12 -B scripts/research_package.py --package research/f7-union-support-audit

Requires an installed C++20 compiler with UBSan/bounds support. The manifest registers separate quick and full obligations. Exact sources, commands, environments and limitations are in the package.

Started and refreshed against canonical main 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11, priority revision 6. Canonical claim. Live overlap includes #2973, #2984 and the corrected historical PR2691 notice; the bounded expansion and compute gate are explicit in TASK.json.

Credit Ross Richardson for the PR943 model, Aaron Wang for PR2542/PR2634 and the retained census/set-cover lineage, Nicholas Stone for PR2691 and PR2973's correction/counterexample/exclusion, Tekalign for #2984's representation work, and Douglas Colkitt for the retained research/source-binding framework. Nick Petalas contributes the new implementation, oracle and replay. Apache-2.0. Substantial automated assistance was used in research, implementation, checking and writing. No independent human or formal-kernel review is claimed.

Merge Danger

Door: two-way

Blast Radius: isolated

Only this new research package is added. Shared code, frozen predecessors, upstream/, selected sources and the frontier are unchanged. Length 22 remains unresolved. This is finite executable evidence plus a written completeness argument, not a kernel proof or full multiplication theorem. Full selected cold replay and hosted CI are not claimed passed.