refactor(chat): remove openai_convert, add per-thread sampling params & mode transitions

Remove legacy openai_convert.rs (676 lines) and subchat HTTP endpoints as message
conversion is now handled directly in LLM adapters (OpenAI Chat, Anthropic, Refact).

Add per-thread sampling parameters (temperature, frequency_penalty, max_tokens,
parallel_tool_calls, reasoning_effort) with GUI controls and backend persistence.

Add mode transition system:
- New `mode_transition` subagent extracts context from existing chats
- `/trajectory/mode-transition/apply` endpoint creates linked child chats
- GUI mode selector now supports switching/restarting with context preservation
- New `canonical_mode_id()`/`normalize_mode_id()`/`is_agentic_mode_id()` utilities

Update provider configs with model-specific defaults (default_max_tokens,
default_frequency_penalty) and comprehensive chat model definitions.

Other:
- Extract LLM logging/sanitization utilities
- Add `eof_is_done` model capability
- Preserve citations in chat sanitization
- Mode badges with consistent colors in UI
- Update mode YAML files to schema_version: 4 with enhanced `ask_questions()`
  guidance and subagent delegation patterns

BREAKING CHANGE: Legacy ChatMode enum removed - all modes now use string IDs.
Update any hardcoded enum usage to use canonical_mode_id().

Author: JegernOUTT

refact-agent/engine/src/yaml_configs/defaults/subagents/title_generation.yaml

@@ -15,6 +15,6 @@ subchat:
 
 messages:
   user_template: |
-    Summarize this chat in 2-4 words. Prefer filenames, classes, entities, and avoid generic terms. Write only the title, nothing else.
+    Summarize this chat in 2-5 words. Avoid generic terms. Write only the title, nothing else.
 
 tools: []