A malformed prediction entry is silently scored against the "financial" database instead of raising

[christian-gheorghe]

In evaluation_utils.package_sqls(), a prediction is split on "\t----- bird -----\t" to recover (sql, db_id). When the split fails the code does not raise — it falls back to db_name = "financial":

try:
    sql, db_name = sql_str.split("\t----- bird -----\t")
except ValueError:
    sql = sql_str.strip()
    db_name = "financial"

So a prediction file that loses its separators (a formatting bug, a bad join, a truncated write) is quietly evaluated entirely against financial, producing a low-but-plausible score. Nothing in the output says the file was invalid.

Reproduce: pass {"0": "SELECT 1 FROM frpm"} (no separator) through package_sqls — it returns normally with db_name == "financial".

Suggested fix: raise on a failed split (or at minimum count and print the number of fallbacks).