[petalas]

Summary

conorholds' pinned cycle_search.cpp
  -> literal Q3 paths and relation oracle
  -> explicit path-tuple feasibility oracle
  -> actual Engine::propagate preserves every feasible choice

Add research/connector-path-propagation-audit/, taking PR3224's released propagation handoff. The README proves preservation under the declared simple-word contract, including coordinate totals and conditional coverage across repeated low blocks. No defect was found in this scope.

Base and refreshed canonical main: 463113d675af7410ac91d53a4fed51ec6494eb9a, priority revision 6. Contribution: 5964d7f22ad9cd9c27a1d2eb0b897f52b9425a13. Canonical investigation.

Evidence

• Before: PR3224 checks capacity and explicitly leaves endpoint, same-low disjointness and conditional-coverage propagation for independent audit.
  After: all 4,314,624 native binary-relation entries agree with literal vertex predicates; all 82 emitted templates' incidence is checked. Exhausting 5,114,880 bounded Cartesian tuples retains all 65,934 feasible candidate occurrences. Conditional-coverage deletion is exercised; zero feasible choices are lost.
• Registered ordinary/optimized replays, four compiled unsound-source mutants, source-pin and receipt-corruption controls pass. ASan/UBSan report no diagnostics. GCC 14.2 and Clang 19.1 produce identical counters.
• All nine committed package/input files pass publication-byte checks. A read-only, Git-free committed export passes direct replay under Python 3.13, separately from the Python 3.12 registered replay.

python3.12 -I -B scripts/research_package.py \
  --package research/connector-path-propagation-audit

Exact source hashes, environment, domain bounds, proof, commands and unrun obligations are in the package. This does not establish search/template completeness, the full B census, the literature/classification premises, the complete f(7)=24 claim, supplier admission or a new kappa. A separately budgeted independent B search remains a next question.

Credit conorholds for the source and fixture, PR3224 for the handoff, Nicholas Stone, Aaron Wang, James Reed, Nick Petalas, Ross Richardson and Douglas Colkitt for the retained model and correction chain. Substantial automated assistance is disclosed in TASK.json; no independent human review is claimed.

Merge Danger

Door: two-way

Blast Radius: research

Only one new scoped package. Frozen inputs, shared code, selected result and others' work remain unchanged.