fix(tools): harden UpdateTodoList for local LLM task JSON quirks

[JessicaMulein]

Summary

Hardens UpdateTodoList so local models can pass tasks as a JSON string, char-split arrays, or brace-prefixed strings without breaking format_output or writing garbage one-character “tasks” to todo.txt.

Problem

On Ollama and similar local models we routinely see:

• tasks passed as a JSON string instead of a parsed array
• Character-split “arrays” from the model
• Strings that start with { but are not valid JSON

main already has normalize_json_array() in cecli.tools.utils.helpers (used by Grep and others), but UpdateTodoList still used naive json.loads + a loop that treated each character as a task in format_output.

Solution

• normalize_task_items() / format_task_lines() — shared coercion for execute and format_output
• coerce_task_item() — dict passthrough; invalid {…} strings fall back to plain task text
• format_output — json.loads guard; surface ToolError from normalization instead of silent corruption
• LIST_PARAMS = ["tasks"] preserved for base tool list handling

Tests

New file: tests/tools/test_update_todo_list.py (7 cases)

Test	Covers
test_normalize_task_items_does_not_split_characters	Char-split regression
test_format_task_lines_accepts_json_string	Tasks as embedded JSON string
test_format_output_accepts_tasks_as_json_string	End-to-end format_output
test_format_output_reports_invalid_tasks_json	Clear error on bad tasks
test_coerce_task_item_plain_string_starting_with_brace	Invalid { prefix fallback

pytest tests/tools/test_update_todo_list.py -v
# 7 passed (local run)

pytest tests/helpers/monorepo -q
# 24 passed — no monorepo regression on this branch

Scope

In scope: cecli/tools/update_todo_list.py, tests/tools/test_update_todo_list.py, minor import-order touch in tests/tools/test_get_lines.py.

Out of scope: General glued-tool JSON (parse_tool_arguments), Grep, session encryption — separate PRs if still needed on main.

Checklist

• Uses existing normalize_json_array (no duplicate helper stack)
• execute and format_output share the same normalization path
• Tests added and passing locally
• No change to clone/multi-repo workspace behavior