SPECTRA 0.7.1: exact search efficiency

SPECTRA 0.7.1: exact search efficiency

Source 9fd6682345b34e992b3cab7e55750e9ef4b04e87; tree 9c93556661e41f6d991310948cd15ed8c42a86e2.

Opt-in indexed CPU search preserves every seeded path and original-clause answer.
Historical search and neural replay implementations are unchanged.

• efficiency-cnf-3.10-34708498065-1: cold/reference ratio 0.300115; formula-bootstrap 95% interval [0.26872913728248904, 0.3668990766160362]; 1,152 answers and 384 unique execution replays checked.
• efficiency-cnf-3.13-34708498065-1: cold/reference ratio 0.321206; formula-bootstrap 95% interval [0.2912564858828442, 0.3846631458346244]; 1,152 answers and 384 unique execution replays checked.

Primary scope: 16 large random/planted formulas, 4,096-flip executions, complete cold call cost.
Large timed cases return UNKNOWN. This is lower cost of identical bounded search, not improved SAT success or external-solver superiority.
Every smaller/shorter cell and memory observation remains in the evidence; regressions are not removed.
Prepared warm calls exclude separately measured reusable indexing. Cold allocation can increase.
Output-only inference uses six untrained artifacts with bitwise-equivalent final outputs. Returned tensor storage is not peak inference RAM or RSS; latency is not universally lower.

Source, wheel and selected raw evidence are attached with SHA256SUMS. See docs/EFFICIENCY_GUIDE.md for reproduction and limitations.
Old slow retraining tests are not included in the fast-suite count. No hiring, novelty, energy or new learned-capability claim.