Truthiness/exception-handling bugs allow silent validation bypass and uncaught crashes in validate.py

[jbonofre]

Found while reviewing #373. None of these are regressions introduced by that PR — they're pre-existing bugs in code adjacent to what it touches, and outside its diff hunks — but they're the same bug classes #373 fixes for the "Warning:" text-matching case (silent bypass via a shortcut check, and uncaught exceptions crashing the validator).

1. validate_references silently skips empty-string dataset references (validate.py:221)

if from_ds and from_ds not in datasets:

The truthiness check (and the equivalent to_ds check) skips the unknown-dataset
check whenever from/to is "". Since the schema has no minLength constraint,
this is schema-valid and currently passes validation:

{"from": "orders", "to": "", "from_columns": ["customer_id"], "to_columns": ["id"]}

"" is never a real dataset — empty-named datasets are excluded from the datasets
map — yet no error is raised.

Fix: check is not None (or just always run the not in datasets check)
instead of relying on truthiness.

2. validate_unique_names silently skips empty-string names (validate.py:180)

dataset_names = [d.get("name") for d in datasets if d.get("name")]

Same truthiness-filter pattern (repeated for fields/metrics/relationships below)
drops empty-string names before duplicate detection runs. Two datasets both named
"" are schema-valid and currently pass with no duplicate-name error.

Fix: filter on is not None instead of truthiness.

3. validate_sql_expression only catches ParseError/TokenError (validate.py:295)

Other exceptions sqlglot can raise propagate uncaught and crash main() with a raw
traceback instead of a [SQL] diagnostic.

Repro: a SQL expression with ~5000 nested ( raises RecursionError from
sqlglot.parse_one, uncaught.

Fix: broaden the catch (or catch Exception and re-tag as [SQL]) so malformed
input can't take down the whole validator run.

4. main() YAML load only catches yaml.YAMLError (validate.py:384)

A RecursionError from PyYAML's scanner/composer on deeply nested input propagates
uncaught.

Repro: a file with ~3000 nested [ crashes with a raw multi-frame traceback instead
of "Error: Invalid YAML: ...". Exit code stays non-zero either way, but tooling/humans
expecting clean diagnostic output on stdout get a stack trace instead.

Fix: widen the catch here as well.

[christianeu-db]

@jbonofre i'm happy to take a pass on these if you'd like