EvolMoves: smoke tests and failure-mechanism probes

tiny smoke runs for execution → decode → burn-in → pooling. raw-OK can still leave zero usable samples... useful failure to catch before paying for the bigger run.

• simulation panel: six original conditions, twenty datasets each, 27 taxa, with low/medium/high indel settings that vary by mutation scale. The extra seventh condition lacks original raw data. WAG and JTT inference are planned comparisons, not interchangeable condition labels.

• smoke-test results: all four first-panel chains ran/decoded individually, but one or two raw samples per chain did not survive required burn-in/pooling into usable analysis cells. In the six-condition panel, 23/24 chains were raw-OK; one stopped before its first thinning boundary. These test failure handling, not alignment accuracy.

• exploratory mechanisms: two families × two seeds × three starts show finite-time initialization sensitivity in short, dirty-snapshot probes. A separate MTM probe reports 132 finite proposals among 200, all with positive target differences but none accepted; normalizer asymmetry is a hypothesis to investigate, not proof of a kernel defect. Do not add the target term twice when reading the MTM-weight decomposition.

• files: simulation/PROTOCOL.md, SOURCE-VALIDATION.md, RUNNER-ACCEPTANCE.md, FINALIZATION.md under benchmarks/nute2019/; historical failure reports under benchmarks/prospective/historical/output/t0-failure-investigations-2026-08-28/.

• review question: keep engineering completion, usable posterior output and scientific convergence as separate gates.

Smoke-test execution and usable posterior-output counts.