fix: remove observable states from agent_state (discover, don't track)

The agent_state field was recording observable state like "running",
"dead", "idle" which violated the "Discover, Don't Track" principle.
This caused stale state bugs where agents were marked "dead" in beads
but actually running in tmux.

Changes:
- Remove daemon's checkStaleAgents() which marked agents "dead"
- Simplify ensureXxxRunning() to use tmux.IsClaudeRunning() directly
- Remove reportAgentState() calls from gt prime and gt handoff
- Add SetHookBead/ClearHookBead helpers that don't update agent_state
- Use ClearHookBead in gt done and gt unsling
- Simplify gt status to derive state from tmux, not bead

Non-observable states (stuck, awaiting-gate, muted, paused) are still
set because they represent intentional agent decisions that can't be
discovered from tmux state.

Fixes: gt-zecmc

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

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

internal/cmd/handoff.go

@@ -182,19 +182,9 @@ func runHandoff(cmd *cobra.Command, args []string) error {
 		}
 	}
 
-	// Report agent state as stopped (ZFC: agents self-report state)
-	cwd, _ := os.Getwd()
-	if townRoot, _ := workspace.FindFromCwd(); townRoot != "" {
-		if roleInfo, err := GetRoleWithContext(cwd, townRoot); err == nil {
-			reportAgentState(RoleContext{
-				Role:     roleInfo.Role,
-				Rig:      roleInfo.Rig,
-				Polecat:  roleInfo.Polecat,
-				TownRoot: townRoot,
-				WorkDir:  cwd,
-			}, "stopped")
-		}
-	}
+	// NOTE: reportAgentState("stopped") removed (gt-zecmc)
+	// Agent liveness is observable from tmux - no need to record it in bead.
+	// "Discover, don't track" principle: reality is truth, state is derived.
 
 	// Clear scrollback history before respawn (resets copy-mode from [0/N] to [0/0])
 	if err := t.ClearHistory(pane); err != nil {