[antonsoo]

Reading list-valued foreign keys from SQLite can change their values and drop references. For example:

from csvw import TableGroup
from csvw.db import Database

tg = TableGroup.fromvalue({'tables': [
    {'url': 'places', 'tableSchema': {
        'columns': [{'name': 'ID'}], 'primaryKey': 'ID'}},
    {'url': 'visits', 'tableSchema': {
        'columns': [{'name': 'ID'}, {'name': 'Places', 'separator': ';'}],
        'primaryKey': 'ID',
        'foreignKeys': [{'columnReference': 'Places', 'reference': {
            'resource': 'places', 'columnReference': 'ID'}}]}}
]})
db = Database(tg)
db.write(places=[{'ID': 'New York'}, {'ID': 'San Francisco'}],
         visits=[{'ID': 'trip', 'Places': ['New York', 'San Francisco']}])
print(db.read()['visits'][0]['Places'])

Before: ['New', 'York']. After: ['New York', 'San Francisco'].

The stored association rows are correct, but group_concat followed by whitespace splitting loses the key boundaries. The same aggregation can split a custom context containing ||. Read the individual rows and group them in Python, preserving each key/context value and repeated association. Add regressions for spaces, tabs, newlines, repeated keys, two foreign-key columns and opaque custom contexts.

Validation: the full suite passes on Python 3.12 (835 passed, 16 expected failures) and 3.9 (838 passed, 13 expected failures), with 100% configured coverage; changed-file Flake8 passes. A separate relational workflow uses published Concepticon labels in 128 multiword configurations across memory/file SQLite, CSV serialization and table/scalar-column translation. All fail before and pass after; all 128 complete single-word controls are unchanged. Raw SQLite association rows and foreign-key checks provide independent references. Source and an installed wheel reproduce the same results.

[xrotwang]

For a bit of context: SQLite was always just meant as a format to allow performant analysis of CSVW data. So round-tripping data back from SQLite to CSVW was more of a proof-of-concept thing. And the "list-valued foreign keys" feature is something that isn't clearly spelt out in the CSVW spec. I implemented it because we explicitly allow this for CLDF datasets - where ID columns are typically very restricted, not allowing whitespace.

Anyway, making this functionality more robust is ceertainly welcome!