immundata

• agg_strata() now uses the argument names schema and prefix instead of by and strata_name_prefix.
• Removed the ImmunData$metadata accessor. Use idata$repertoires for the repertoire definitions and summaries, and use manifests for annotations associated with input repertoire files.
• For extension developers, imd_schema("metadata_filename") is now imd_schema("manifest_filename"), and the unused imd_files()$receptors entry has been removed.

New features and improvements

• read_repertoires() now works approximately 60 times faster by combining CSV, TSV, and compressed text inputs into one temporary Parquet file before processing by default. This avoids repeated text scans in downstream duckplyr queries while retaining original input paths in provenance. Use prematerialize = FALSE to disable it or prematerialize_folder to select the temporary storage directory. I recommend you to use it pretty much always.
• Added read_manifest() for CSV, TSV, TXT, and in-memory manifests. It infers common delimiters, resolves file-relative paths, validates file availability, and adds normalized source paths for joining to repertoire data. The special repertoire_schema = "<manifest>" value defines repertoires from all manifest columns.
• Promoted strata to first-class ImmunData state. Objects now expose schema_strata and a $strata table; agg_strata() and rename_strata() update this state, and snapshots persist and restore it.
• Added grouped mutation through .by in mutate_immundata() and dplyr::mutate() methods for ImmunData, including a duckplyr-compatible fallback for grouped summary expressions.
• Added conflicts = c("error", "replace") to the annotation functions. Existing annotation columns are protected by default, while intentional replacement is allowed for columns that do not define core ImmunData state.
• mutate(), compute(), and annotation operations now preserve repertoire, strata, and provenance state when the biological grouping has not changed. Filtering and downsampling rebuild affected repertoire and strata summaries and retain existing stratum labels when possible.
• Added consistent progress control to manifest reading, repertoire ingestion and aggregation, and snapshot reading and writing. Use verbose = FALSE for individual calls or options(immundata.verbose = FALSE) globally.
• Snapshot metadata now stores repertoire and strata definitions and validates them against the Parquet annotation columns when loading. Older metadata formats remain readable and are upgraded in memory when necessary.
• Provenance now includes derived artifact locations (artifacts_root and artifacts_path) associated with the project home and current snapshot.

Bug fixes

• Corrected repertoire-level cell and receptor counts for paired-chain data and prevented chain rows from inflating n_barcodes, n_receptors, receptor proportions, and repertoire-occurrence counts.
• Made imd_repertoire_id assignment deterministic by ordering repertoire schema values before assigning identifiers.
• Scoped single-cell barcodes by source filename during chain selection and pairing, preventing identical barcode strings from different input files from being treated as the same cell.
• Sequence filters now retain every chain belonging to a matched receptor, including exact, regular-expression, Hamming, and Levenshtein matching. Distance calculations no longer use the k-mer prefilter, and temporary DuckDB table names are unique across repeated operations.
• Fixed downsampling with DuckDB 1.5 and later, preserved annotation columns in bulk count mode, retained provenance and strata state, and made n = 1 mean an absolute sampling depth of one.
• Hardened ingestion against duplicate manifest paths, negative bulk counts, missing argument columns, and collisions between custom and canonical locus columns.
• Prevented mutate() and annotation replacement from overwriting system, receptor-schema, repertoire-schema, or strata-defining columns.
• Fixed Windows path handling in the test and example infrastructure.

Documentation and maintenance

• Reworked the package documentation around biological units, lazy duckplyr workflows, ingestion, aggregation, filtering, annotation, snapshots, and provenance, and moved website generation to altdoc.
• Removed the unused ggplot2 dependency and raised the minimum supported dplyr version to 1.2.1.