Phase 5B commit 4: delete style/shape/voice-matching/model-routing systems

Backend cleanup:
- Delete STYLE_SPEED_MODIFIERS, STYLE_PHONE_QUALITY constants
- Delete _normalize_style_key, _match_voices_to_styles,
  get_style_speed_modifier, get_style_phone_quality functions
- Delete _pick_caller_style, _HEAVY_STYLES, _LIGHT_STYLES,
  _EVASIVE_STYLES (only called by template path)
- Delete _is_informal_style + _INFORMAL_STYLES
- Remove orphaned calls to deleted _assign_call_shape
- Drop shape parameter from _pick_response_budget
- Simplify _assess_call_quality (no more shape/style/pool_name)
- Delete Session fields: caller_styles, caller_shapes,
  caller_model_strategy/pool/map/fallback/models/_cycle_idx
- Remove those fields from Session.reset, checkpoint save/load,
  and session.caller property
- Drop style/shape lookups in chat callsites and promotion gate
- Update config.py caller_dialog model stale comment

backend/config.py

@@ -34,9 +34,8 @@ class Settings(BaseSettings):
     ollama_host: str = "http://localhost:11434"
 
     # Per-category model routing
-    # caller_dialog is overridden by style_matched routing (see Session.caller_model_map)
     category_models: dict = {
-        "caller_dialog": "x-ai/grok-4.1-fast",               # fallback if style_matched disabled ($0.20/$0.50)
+        "caller_dialog": "anthropic/claude-haiku-4.5",       # live caller dialog ($0.80/$4)
         "devon_ask": "x-ai/grok-4.1-fast",                   # Devon matches show energy, cheap ($0.20/$0.50)
         "devon_monitor": "google/gemini-2.5-flash",          # just yes/no decisions, keep cheap ($0.15/$0.60)
         "background_gen": "anthropic/claude-sonnet-4.6",       # backgrounds drive the whole call — worth the quality ($3/$15, ~$0.30/show)