rp-source-csv

Fingerprint the header row against vendor anchor columns: required anchors must match exactly, strong anchors raise the score, and negative anchors penalize it. A match needs a score of at least 0.70, two strong anchors, and a clear margin over the runner-up; otherwise the file falls back to custom.

Run the csv-discovery script on the export to capture its layout, vendor, and column statistics into fileset.json and raw-capture.json. The discovery stage synthesizes a source-schema.json from that capture, and codegen produces a reader that groups rows by Handle and emits per-entity NDJSON streams for import.

Yes. Files sharing a role and the same normalized column-name set are concatenated into one logical stream, with part order determined by numeric-suffix-aware sorting. A subset or superset header halts with the exact diff unless you pass --allow-header-superset to union the parts.

Yes. The parser implements RFC-4180 quoting with embedded delimiters, newlines, and escaped quotes, and auto-detects comma, semicolon, tab, and pipe delimiters scored by field-count consistency. It also strips BOMs, handles all line-ending styles, and detects whether empty cells mean absent values.

The capture halts with halt: true and records the ranked layout candidates instead of guessing. Discovery stops and asks the user the question named in the warning, because synthesizing a schema from an ambiguous grouping would corrupt the entity model.

Cross-file foreign-key join detection is deferred to v1.1, and only Shopify has a canonical record field map so far. Media columns holding local paths or private URLs are not reachable by Wix and must be rewritten to public URLs or deferred.