Refactor startup paths to use RuntimeConfig (gt-j0546)

Replaced all hardcoded 'claude --dangerously-skip-permissions' invocations
with configurable helpers from internal/config:

- GetRuntimeCommand(rigPath) - simple command string
- GetRuntimeCommandWithPrompt(rigPath, prompt) - with initial prompt
- BuildAgentStartupCommand(role, bdActor, rigPath, prompt) - generic agent
- BuildPolecatStartupCommand(rigName, polecatName, rigPath, prompt) - polecat
- BuildCrewStartupCommand(rigName, crewName, rigPath, prompt) - crew
- BuildStartupCommand(envVars, rigPath, prompt) - custom env vars

Files updated:
- internal/cmd/start.go (4 locations)
- internal/cmd/crew_lifecycle.go (2 locations)
- internal/cmd/crew_at.go (2 locations)
- internal/cmd/deacon.go
- internal/cmd/witness.go
- internal/cmd/up.go (2 locations)
- internal/cmd/handoff.go (2 locations)
- internal/daemon/daemon.go (3 locations)
- internal/daemon/lifecycle.go
- internal/session/manager.go
- internal/refinery/manager.go
- internal/boot/boot.go

This enables future support for alternative LLM runtimes (aider, etc.)
via rig/town settings configuration.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

internal/refinery/manager.go

@@ -192,7 +192,7 @@ func (m *Manager) Start(foreground bool) error {
 	// Start Claude agent with full permissions (like polecats)
 	// NOTE: No gt prime injection needed - SessionStart hook handles it automatically
 	// Restarts are handled by daemon via LIFECYCLE mail, not shell loops
-	command := "claude --dangerously-skip-permissions"
+	command := config.GetRuntimeCommand("")
 	if err := t.SendKeys(sessionID, command); err != nil {
 		// Clean up the session on failure (best-effort cleanup)
 		_ = t.KillSession(sessionID)