Replaced concurrent appends to shared CSV files and their file locks with a coordinator-owned, batched execution protocol. Workers now write private batch results atomically and local batches run in isolated callr processes.
Added batch_size to control how many simulations each worker performs before returning to the coordinator.
Summary statistics and model outputs are now buffered and written as bounded, atomic Parquet fragments per batch instead of one temporary file per particle. Fragment metadata is committed with the batch result, and generation finalization uses bounded memory without rescanning previous generations.
Added storage_chunk_rows and storage_chunk_mb to bound worker storage buffers independently of the simulation batch size.
Added consolidate_abc_storage() to export or transactionally replace active fragments with one Parquet file per named object and generation.
Added optional storage of model summary statistics and detailed outputs for ABC rejection and ABC-SMC analyses.
Model functions can now return a structured list containing distances, summaries, and outputs; the historical numeric distance vector remains supported.
Added Parquet storage for vectors, matrices, and data frames, with independent none, retained, accepted, and all retention policies for summaries and outputs.
Added stable attempt identifiers and acceptance/retention metadata to link stored tables to tested particles.
Parquet fragments are finalized atomically so stopping parallel workers cannot expose partially written files.
A tested particle is now committed before it is published as accepted, and Parquet consolidation ignores orphan fragments not present in the committed attempt journal.