Six core runtime files exceed the 600-line hard cap (config.py is 2151 lines)

Problem

CONTRIBUTING.md rule 3 states a 600-line hard cap — split into submodules. The core runtime/model layer violates it six times over:

Why it matters

• config.py is a grab-bag of ~80 getters/setters spanning XDG paths, model settings, banner colors, autosave/TTY session tracking, API keys, and command history — at least six unrelated concerns in one module.
• callbacks.py duplicates its PhaseType Literal (~55 entries) in the hand-maintained _callbacks dict (lines 6–111); these can drift. It could be _callbacks = {p: [] for p in typing.get_args(PhaseType)}.
• cli_runner.py:interactive_mode is a ~700-line god function (see separate issue).
• model_factory.py:ModelFactory.get_model is a ~450-line if/elif chain over model types; a dispatch table ({"openai": _build_openai, "anthropic": _build_anthropic, ...}) would make each builder independently testable and let the plugin handler path share the same mechanism it already uses for custom types.

Suggested fix

Split along existing seams, e.g.:

• config/ package: paths.py, core.py (get_value/set_value), model_settings.py, sessions.py, colors.py, toggles.py — re-export from config/__init__.py to preserve the public API.
• model_factory.py: extract one _build_<type>() function per model type into model_builders.py and dispatch via a dict.
• callbacks.py: derive _callbacks from get_args(PhaseType) and consider grouping the thin on_* wrappers.

Filed by Zen Reviewer A (code-puppy-60635a)

Start with rule 3 in CONTRIBUTING.md, then map imports and public APIs for config.py, callbacks.py, cli_runner.py, model_factory.py, claude_cache_client.py, and gemini_model.py. Review the proposed config/ package and model builder seams before changing structure. Done means the six runtime files are split below 600 lines without breaking their public APIs or behavior; run the repository’s existing checks afterward.