fix: prevent filtering of assistant messages with tool completions

The main session LLM was unable to see task tool output because
we were aggressively filtering assistant messages that followed
subtask-only user messages. The tool output ONLY exists in the
assistant message's ToolPart (part.state.output), so filtering
those messages deleted the only copy of subtask results.

Removed:
- isSubtaskOnlyMessage() function
- isToolCompletionMessage() function
- skipNextAssistant variable and related logic

Now only filters summary leak messages while preserving all tool
completion messages. This restores:
- Return prompt functionality
- [name] captures
- Multi-step return chains

Fixes critical bug introduced in v0.3.4