[abrahamperz]

Follow-up to #10728, and stacked on it: the first commit here is that PR's NUL-termination, which drops out once that PR merges to main.

A C function registered with sqlite3_create_function_v2 that read an argument with sqlite3_value_text got a null pointer when the argument was an INTEGER, REAL or BLOB. SQLite converts the value to text there, so strlen or printf on the result works for any type. The same accessor is used by sqlite3_value_dup copies and sqlite3_column_value, so they had the gap too.

-- a C scalar f that returns strlen(sqlite3_value_text(argv[0])):
SELECT f(42), f(1.5), f(X'414243');
-- sqlite:  2, 3, 3
-- main:    null pointer for all three
-- this PR: 2, 3, 3

Root cause

sqlite3_value_text returned a pointer only when the value already held text; to_text() gives None for a number or a blob. Turso has the bytes but never converted them.

Fix

Coerce the value to text in place, the way SQLite does, so the returned pointer stays valid: a number becomes its decimal text, a blob keeps its bytes verbatim, each NUL-terminated. from_text_bytes builds the text value for a blob without a UTF-8 round-trip. Any previous heap is freed before the text replaces it.

Replacing the value in place is safe because the ExtValue is owned for the duration of the call and freed by type afterwards: a scalar function's arguments live in a Vec that finish_call frees, column_value caches owned values in the statement, and a dup'd value is owned by the caller. So the allocation and its free stay matched.

One visible consequence, which matches SQLite: after sqlite3_value_text reads a value, a later sqlite3_value_type on it reports TEXT.

Testing

test_value_text_coerces_non_text_like_sqlite registers a C function that returns strlen(sqlite3_value_text(argv[0])) and checks that 42, 1.5, X'414243' and 'hello' give 2, 3, 3 and 5. It runs against both Turso and real SQLite. Without the change the numeric and blob cases return a null pointer, so -1 from the test.