Gemini agents route shell commands through the request_permission ACP
path instead of emitting standard ToolCall events with populated fields.
When the subsequent ToolCallUpdate(completed) arrives with empty title
and kind, the event translator had no metadata to resolve a proper
command name, falling back to the generic "Tool" display.

This fix captures tool call metadata (title, kind, raw_input) from
permission requests via a new ToolCallMetadata struct on ApprovalRequest,
and seeds the shared pending_tool_calls map in the approval handler.
The event translator can then resolve the actual command name (e.g.,
"echo hello") from accumulated metadata on completion events.

Also adds extract_command_from_permission_title() to strip Gemini
compound title format (command [cwd path] (description)) down to just
the command portion.

The approval overlay showed "Execute: command" for all Gemini shell
commands because format_execute_command and extract_reason_from_tool_call
only looked at raw_input (which Gemini leaves empty for Execute calls).

Add extract_command_from_title() to fall back to parsing the command
from the tool call title, which Gemini populates with the actual command
in compound format (command [cwd path] (description)).