fix(workflows): stop list-literal-then-index expressions silently corrupting

[Noor-ul-ain001]

Summary

_evaluate_simple_expression's list-literal detection (workflows/expressions.py) is expr.startswith("[") and expr.endswith("]"). This is true not only for a genuine literal like [1, 2, 3] but also for a list literal immediately followed by an index suffix, e.g. [1,2,3][1] — which reads as "index 1 of [1,2,3]" (i.e. 2).

Naively stripping the outer brackets from "[1,2,3][1]" produces "1,2,3][1", which the comma-splitter then breaks into ["1", "2", "3][1"]. The malformed last segment resolves to None via the existing dot-path fallback, so the whole expression silently evaluates to [1, 2, None] — no error, no warning, just wrong data.

>>> evaluate_expression("{{ [1,2,3][1] }}", ctx)
[1, 2, None]          # should read as index 1 of [1,2,3], i.e. 2

This is the exact "grabs the wrong span" failure mode the adjacent string-literal check already guards against a few lines above (verifying the matching quote is the last character, not just present anywhere) — the list-literal branch never got the same rigor.

Fix

Added _is_single_list_literal: a quote/bracket-depth scan (mirroring _split_top_level_commas's existing tracking style in the same file) that confirms the opening [ closes exactly at the final character before treating the expression as one literal. A misclassified expression now falls through to the existing dot-path resolution and evaluates to None — consistent with how every other unresolvable expression in this module already behaves, not a new failure mode. This does not add support for indexing directly into a literal list (an author would still write steps.x.output.list[0], which the dot-path resolver already supports); it only stops the misclassification from silently corrupting the result.

Test plan

• ruff check . clean
• New test fails against the pre-fix code (assert [1, 2, None] is None → confirmed via test-the-test) and passes with the fix
• Full test_workflows.py suite (985 tests) passes