[jfo84]

Parallel sqlite_query scans of one database crash or return wrong results. SqliteScan reads each column with sqlite3_column_value() and then the sqlite3_value_*() accessors. That value is unprotected, and SQLite documents the sqlite3_value_*() accessors on an unprotected value as safe only when the connection is not used from another thread. Parallel scans share one connection, so a type conversion (an integer read as text allocates) runs outside the connection mutex while another thread prepares a statement, and SQLite's heap is corrupted. This change reads through sqlite3_column_*(), which take the mutex.

• SqliteScan, CheckTypeMatches and CheckTypeIsFloatOrInteger read through sqlite3_column_*(). The error paths also handle a NULL text value.
• test/sql/scanner/parallel_sqlite_query.test runs sixteen sqlite_query scans of one in-memory database at 12 threads, 100 times.

Validation on macOS arm64 (12 cores, load 3–8), DuckDB 0b272c23, duckdb-sqlite b043eb7b with and without this commit, runs alternated:

	without the fix	with the fix
the new test, unittest	10 of 10 failed: 5 SIGBUS, 4 SIGSEGV, 1 wrong sum (3199838845, expected 3199839440)	10 of 10 passed
the reproduction below, CLI	16 of 20 crashed: 13 SIGSEGV, 2 SIGABRT, 1 SIGBUS	20 of 20 completed

The full sqlite_scanner suite passes with the fix: 69 test cases, 3,029 assertions, 18 skipped.

Reproduction:

SET threads = 12;
SET preserve_insertion_order = false;
ATTACH ':memory:' AS meta (TYPE sqlite);
CREATE TABLE meta.t AS SELECT i FROM range(20000) r(i);
SELECT sum(v) FROM (
    SELECT i::BIGINT AS v FROM sqlite_query('meta', 'SELECT i FROM t WHERE i >= 0')
    UNION ALL SELECT i::BIGINT FROM sqlite_query('meta', 'SELECT i FROM t WHERE i >= 1')
    -- ... sixteen scans in all, i >= 0 through i >= 15
);

Run the SELECT repeatedly. The expected sum is 3199839440.

[staticlibs]

Hi, thanks for the PR and for the quick update! I wonder if you will be able to rebase the PR on the latest main after #264 is integrated (in 10-15 minutes, I will add another comment when it is ready) - it should make the CI to run much faster as GH runners are currently overloaded.

[staticlibs]

Ready now, please rebase the PR onto main. Also, where possible, please remove Claude from commit authors.

[jfo84]

Thanks! Rebased onto the latest main and removed the Claude co-author trailers from both commits. The source patch is unchanged. New head: 8ba69f7.

[staticlibs]

PR looks good to me.