Convergence traces from two pre-registered experiments on invariance in metaheuristics

The complete convergence trace of every run of two pre-registered experiments on the invariance properties of metaheuristics: 2,079,000 runs for the main experiment and 191,100 for the repair arm, plus 11,700 runs of a tuning robustness check (5,850 with published defaults and 5,850 with the tuned configuration beside them). Not summaries — the recorded best-so-far value at every improvement, for every run. The archive is self-describing. README.md specifies the container format byte by byte and shows how to read one run's trace with nothing but the Python standard library, so the data can be used without the study's code and after that code stops running. Each arm's summaries.jsonl.gz is the index: one JSON object per run carrying its identity, its seed, its outcome, and the file, offset and length of its trace. Trace containers are packed one per worker process and split into one tar per shard. Filenames are flat and prefixed by arm because a record has a single namespace and the arms number their shards alike. Which registered plan produced which arm, and the manifest digest that pins each, are recorded in README.md and in the related identifiers of this record.