refactor(models): lazy-load tomli_w (sub-track 2 partial)

Sub-track 2 of startup_speedup_20260606. Removes the top-level
'import tomli_w' from src/models.py and moves it inside save_config().
tomli_w (~30ms cold load) is now loaded only when the user saves
config, not on every src.models import.

This drops the audit violation count from 63 to 62.

Pydantic BaseModel (the other src/models.py violation) is left for
a future sub-track: deferring a class base requires a metaclass or
proxy pattern that's higher risk for the small (~50ms) saving.

3 new tests in tests/test_models_no_top_level_tomli_w.py:
- tomli_w NOT in sys.modules after import src.models
- save_config() still works (because tomli_w loads on-demand)
- save_config() actually triggers the import on first call

17 existing model tests pass (test_persona_models, test_bias_models,
test_context_presets_models, test_per_ticket_model, test_file_item_model).

src/models.py

@@ -43,7 +43,6 @@ import json
 import os
 import sys
 import tomllib
-import tomli_w
 
 from dataclasses import dataclass, field
 from pathlib     import Path
@@ -166,6 +165,11 @@ def load_config() -> dict[str, Any]:
   return tomllib.load(f)
 
 def save_config(config: dict[str, Any]) -> None:
+ # tomli_w is loaded on-demand (sub-track 2 of startup_speedup_20260606).
+ # If it's already in sys.modules (e.g. warmed up or loaded by a prior
+ # call), the import is a fast lookup; otherwise it's a cold load paid
+ # only when the user actually saves config.
+ import tomli_w
  config = _clean_nones(config)
  sys.stderr.write(f"[DEBUG] Saving config. Theme: {config.get('theme')}\n")
  sys.stderr.flush()