[Bug]: Unsupported expression syntax evaluates to None silently instead of raising #5073

Bug Description

When the workflow expression evaluator meets syntax it does not parse, it does not raise. It treats the text as a variable path, which resolves to None (_resolve_dot_path, reached from src/specify_cli/workflows/expressions.py:969). A condition that uses such syntax then quietly gives the wrong result. The following cases were confirmed:

1. Exponent-form numbers. Numeric literals and numeric strings are parsed with float() only when they contain a ., and otherwise with int() (expressions.py:930-935 for literals, expressions.py:972-976 for _coerce_number). So 1e3 is not read as a number: as a literal it is looked up as a variable and becomes None, and the string "1e3" stays a string in comparisons. Workflow input coercion does accept "1e3" as the number 1000 (src/specify_cli/workflows/engine.py:1449-1456), so the same text means different things in different places.
2. Escaped quotes in string literals. A string literal is recognised only if its first matching quote is its last character (expressions.py:800), so 'it\'s' is looked up as a variable and becomes None. As a result, inputs.name == 'it\'s' is false for name = "it's".
3. Chained indexing. The indexed-segment pattern accepts one index only (expressions.py:304), so matrix[0][1] is looked up as the literal key "matrix[0][1]" (expressions.py:331) and returns None.
4. Word operators next to a quote. and, or and in are only found with a space on both sides (expressions.py:884, expressions.py:893, expressions.py:715). inputs.a=='x'and inputs.b is therefore split only at ==, its right side 'x'and inputs.b is looked up as a variable and becomes None, and the result is false.

Steps to Reproduce

1. From a fresh clone of spec-kit, run uv sync.
2. Save the script below as repro.py in the repository root.
3. Run uv run python -I repro.py.

from specify_cli.workflows.base import StepContext
from specify_cli.workflows.engine import WorkflowEngine
from specify_cli.workflows.expressions import evaluate_condition, evaluate_expression

ctx = StepContext(
    inputs={"n": 5, "s": "1e3", "name": "it's", "a": "x", "b": True},
    steps={"plan": {"output": {"matrix": [["a", "b"], ["c", "d"]]}}},
)
cases = [
    ("1", "{{ 1e3 }}", "value"),
    ("1", "{{ inputs.n < 1e3 }}", "cond"),
    ("1", "{{ inputs.s > 100 }}", "cond"),
    ("2", r"{{ 'it\'s' }}", "value"),
    ("2", r"{{ inputs.name == 'it\'s' }}", "cond"),
    ("3", "{{ steps.plan.output.matrix[0] }}", "value"),
    ("3", "{{ steps.plan.output.matrix[0][1] }}", "value"),
    ("4", "{{ inputs.a == 'x' and inputs.b }}", "cond"),
    ("4", "{{ inputs.a=='x'and inputs.b }}", "cond"),
    ("4", "{{ 'x'in inputs.a }}", "cond"),
]
for num, expr, kind in cases:
    fn = evaluate_expression if kind == "value" else evaluate_condition
    print(f"[{num}] {expr} -> {fn(expr, ctx)!r}")
print("input coercion of '1e3' (type: number):",
      WorkflowEngine._coerce_input("x", "1e3", {"type": "number"}))

Expected Behavior

The module docstring describes the evaluator as "a safe Jinja2 subset" (src/specify_cli/workflows/expressions.py:3), and _evaluate_simple_expression lists "String and numeric literals" as supported (expressions.py:791). For each case, the expression should either evaluate as it would in Jinja2 or raise a ValueError naming the unsupported syntax. Malformed filters already get such a ValueError (docs/reference/workflows.md, "Types are validated, not coerced"). Specifically:

1. {{ 1e3 }} should be 1000.0, and {{ inputs.n < 1e3 }} should be true for n = 5. {{ inputs.s > 100 }} with s = "1e3" should be true, in line with the _safe_compare docstring (expressions.py:1003-1012) and the input-coercion test tests/test_workflows.py:7092 (("1e3", 1000)).
2. {{ 'it\'s' }} should be "it's", and the comparison should be true.
3. {{ steps.plan.output.matrix[0][1] }} should be 'b'. The _resolve_dot_path docstring promises list indexing (expressions.py:312).
4. {{ inputs.a=='x'and inputs.b }} should be true, like the spaced form. {{ 'x'in inputs.a }} should be true.

Actual Behavior

[1] {{ 1e3 }} -> None
[1] {{ inputs.n < 1e3 }} -> False
[1] {{ inputs.s > 100 }} -> False
[2] {{ 'it\'s' }} -> None
[2] {{ inputs.name == 'it\'s' }} -> False
[3] {{ steps.plan.output.matrix[0] }} -> ['a', 'b']
[3] {{ steps.plan.output.matrix[0][1] }} -> None
[4] {{ inputs.a == 'x' and inputs.b }} -> True
[4] {{ inputs.a=='x'and inputs.b }} -> False
[4] {{ 'x'in inputs.a }} -> False
input coercion of '1e3' (type: number): 1000

Specify CLI Version

1.2.4.dev0 (main at 27d14e2)

AI Agent

Not applicable

Operating System

Linux 6.18 (Ubuntu container)

Python Version

Python 3.12.3

Additional Context

When cases 1, 3 and 4 are used as an if condition, validate_workflow returns no errors. Case 2 is flagged in a condition, but as unbalanced delimiters and quotes rather than as an escaped quote.

Related: #5060 proposes replacing the hand-written expression parser with a grammar so that invalid input raises. It does not mention this case.

Also reproduced on macOS (Darwin 25.6, arm64) with Python 3.14.4 and uv run python -I repro.py against 27d14e2; the output matches the above.

AI Disclosure

Found and reproduced with Claude Code (model: Claude Opus 4.8), run autonomously under human supervision. The human reporter re-ran every reproduction above before filing.