fix: raise clear parse error for invalid derivation expressions

[nielspardon]

Problem

evaluate() builds the ANTLR parser in _parse() but installs no error-raising listener. On invalid syntax, ANTLR's default error recovery prints a warning to stderr and returns a partial parse tree containing None children, which _evaluate then walks into — surfacing as the confusing Exception: Unknown token type <class 'NoneType'>.

This is easy to hit with any precision-parameterized type used without its required argument, e.g. evaluate("interval_day"), evaluate("precision_timestamp"), evaluate("decimal"). Per the type grammar all of decimal, interval_day, interval_compound, precision_time, precision_timestamp, and precision_timestamp_tz require explicit parameters, so these are genuine syntax errors.

Fix

Attach a custom ANTLR error listener to both the lexer and parser in _parse() (after removing the default listeners, which also silences the stray stderr warnings). On any syntax error it raises a dedicated DerivationExpressionParseError naming the offending input, chaining the underlying ANTLR exception:

DerivationExpressionParseError: Could not parse derivation expression 'interval_day' at line 1:12: mismatched input '<EOF>' expecting {'<', '?'}

Fallout

Enabling strict parsing surfaced that test_nstruct_simple / test_nstruct_nested were only passing because of the silent recovery this change removes: they used nStruct<a i32, b i32>, but the grammar (substrait-antlr 0.96.0) requires colons — nStruct<a: i32, b: i32>. ANTLR had been inserting the missing : and continuing. Updated those tests to the valid colon syntax; the asserted results are unchanged.

Testing

Added a parametrized test covering every precision-parameterized type used without arguments, plus a general malformed-input case. pytest — 519 passed, 30 skipped; ruff check and ruff format --check clean.

Fixes #216.

🤖 Generated with AI