Normalize planner wire aliases new_tasks, name, and node_id

[tanmays369]

What breaks

Live planner calls (especially local/open-weight models) often return a next-frontier object that uses new_tasks instead of add, and task entries that use name / node_id instead of capability / id. After the existing repair pass, _parse still rejected those shapes as unsupported fields or invalid tasks, so the run stopped with a visible planner failure even though the model proposed usable work.

Reproduction

Regression tests feed provider-shaped JSON:

1. {"new_tasks":[{"name":"answer_with_evidence","arguments":{...}}], ...} — previously raised unsupported/new_tasks failure; now normalizes to add with a generated id.
2. {"new_tasks":[{"node_id":"clock","capability":"current_datetime",...}], ...} — previously failed missing id; now keeps the provided node id.

Fix

In GeneralAgentPlanner._parse, before the existing capability shorthand / task-keyed adapters:

• Map new_tasks → add when add is absent.
• Map task node_id → id, and registry-known name → capability.
• Synthesize a stable id from capability when still missing.
• Default empty depends_on for otherwise complete task objects.

Strict capability registry validation and max-new-task limits are unchanged.

Test plan

• uv run pytest tests/test_general_planner.py -q (22 passed)
• uv run ruff check s16code/planner.py tests/test_general_planner.py
• git diff --check

No credentials, environment files, local databases, or message data are included.