SFS bottleneck audit: executed audit under the frozen protocol (Release B, unrefereed candidate)

Executed identifiability audit of the disputed severe human bottleneck 930–813 kya, run byte-for-byte under the frozen pre-registered protocol (doi:10.5281/zenodo.21893436, GitHub tag v0.1.1-candidate). Nine published spectra — seven 1000 Genomes populations from the bottleneck claimants' own deposit, plus the Deng et al. and Cousins–Durvasula YRI processings — were verified against binding hashes and evaluated across the declared overdispersion and complexity ladders. Finding: universal class rejection. Every one of the 216 declared cells rejects its spectrum: no single-population Kingman history on the declared grid reproduces any of the nine spectra within the declared error model, at any declared budget. The rejections are certified infeasibilities, independently corroborated by an LP-free least-squares analysis; their failure signatures decompose into ancestral-misidentification tails (Hu and Cousins processing) and singleton excess (CEU, CHB, Deng). Clearly-labelled post-hoc diagnostics quantify the rejection at 13–83× inflation of the declared confidence boxes, an effective information content of roughly 5×102–3×104 independent sites — about ten orders of magnitude below the ~5×1013 required for severe-versus-continuous discrimination in the disputed window. The pre-registered conclusion: the severity claim is not evaluable from these spectra under any error model that treats their millions of segregating sites as independent evidence. Status: anonymous-workflow, unrefereed, producer-run. Deviations from the frozen protocol are recorded in DEVIATIONS.md (execution harness, recorded before results; post-hoc diagnostics, labelled). Archives the exact bytes of the GitHub release asset (sha256 b4667574143a0f863c98eeb3bf9ab20b142a9d2738bc78c7ab91e90532e886b9).