fix(airt): normalize attacks arg in generate_category_attack

[rdheekonda]

Problem

generate_category_attack's attacks parameter was iterated character-by-character when passed as a bare string, producing cryptic errors like Unknown attack: 't' / Unknown attack: '['. This made the entire category-sweep path effectively unusable — no format (bare string, full SDK name, JSON list) worked.

Root cause: the runner read attacks_raw = params.get("attacks", []) and looped for a in attacks_raw directly, so "tap" became ['t', 'a', 'p']. By contrast, generate_attack correctly does attack_type.split(",").

Fix

• scripts/attack_runner.py: add _normalize_attack_names() accepting a list (["tap", "goat"]), comma-separated string ("tap,goat"), single name ("tap"), or stringified-list noise ("['tap']"). Return a clear validation error instead of a single-character failure.
• tools/attacks.py: widen the attacks annotation to list[str] | str.
• tests/test_attack_runner.py: add TestNormalizeAttackNames regression coverage (10 cases + explicit "must not split to chars" assertion).
• skills/error-troubleshooting/SKILL.md: document the single-character failure signature.
• agents/ai-red-teaming-agent.md: add category-tool auto-fallback guidance.

Verification

• py_compile passes on all changed Python files.
• Normalization logic validated against all parametrized cases.
• End-to-end: re-ran the previously-failing generate_category_attack(attacks=["tap"], categories=["violence"], …) — it now parses correctly, loads bundled violence goals, and runs a real 3-goal sweep to completion.