harvester/

Data ingestion layer. Maps on-disk dataset trees into RunDirectory records the engine can consume.

• DumpIngestor — wraps core.discovery.DatasetScanner + RunDiscovery + DumpReader. Accepts datasets laid out as TLS{version}/{scenario}/{library}/{run_dir}/.

• SidecarParser — reads .json (dict root) and .meta (key=value) sidecars. Filenames starting with keylog or timing are skipped (they are parsed by dedicated readers).

• MetadataStore — in-memory record list; materializes an eager polars.DataFrame on demand (no LazyFrame) and flattens scalar sidecar fields into meta_* columns.

Filename conventions recognized:

• Dump: ^(\d{8}_\d{6}_\d+)_(pre|post)_(.+)\.(dump|msl)$ — timestamp, phase prefix, phase name.

• Run dir: ^(.+?)_run_(\d+)_(\d+)$ — library, protocol version, run number.

• Keylog: keylog.csv (configurable via --keylog-filename).

PhaseNormalizer canonicalizes phase names across libraries with different conventions.