Directional coupled peak study: registered synthetic inputs and complete execution evidence

Complete evidence for the standalone negative computational-study report Directional coupled peak penalties with original interval rows. All 210 authored synthetic instances are retained: 48 development, 144 heldout and 18 robustness. There are 7770 raw observations, including warmups and oracle records, and 5670 measured method observations. Full canonical inputs, backend primal/dual journals, CSV, registered source snapshots, independently reviewed figures, checksums and provenance are included. The frozen 144-instance heldout paired reduction score is -212.449875%, with 95% six-seed cluster interval [-225.336618%,-192.873759%] and nine prescribed failed pairs. The 10% criterion is not established. Incompatibilities, common-oracle iteration limits, method backend limits without candidate vectors and supervised timeouts remain recorded. No failed case is retried or removed. These are synthetic numerical problems and repeated thermal forecasts, not operational building measurements or deployed control policies. Protocol SHA-256: 3d8eb822a6b31026c491b66cfb21f474b044a6ca3801074379243b8a3e7ca53f. Archive SHA-256: 45b4eceda69f80a30a4ce349d0e9baefacace0de394fe315f3c392260ea8d418. Registered app source: 246bf87e690f4d010494c556e2021d311d587422. The data record contains one independently reviewed ZIP and its checksum; the manuscript PDF is deposited separately. Data and report are CC BY 4.0; embedded source files retain their included source license.