feat: support plan mode in headless investigation executor

[Anthony-Bible]

Background

Plan mode (activated via enter_plan_mode) improves AI accuracy by separating the planning and execution phases. Currently, plan mode is explicitly denied for investigations via a hard-blocking callback added in #58 to prevent the AI from silently stalling a headless investigation.

Goal

Enable the investigation executor to use plan mode in a way that is safe for headless operation:

• The AI should be able to enter plan mode to produce a structured investigation plan
• The plan phase should complete (AI writes all planned steps as plan files)
• The executor should then automatically execute the plan without requiring terminal input to approve each step
• If execution of a plan step fails, fall back to logging the error and continuing (non-interactive)

Proposed Approach

1. Replace the hard-deny callback with a headless-aware plan mode handler
2. After plan mode produces plan files (.agent/plans/), automatically replay them as real tool executions
3. Add a config flag (e.g. AGENT_INVESTIGATION_PLAN_MODE) to opt-in, defaulting to disabled until the replay logic is stable

Why This Matters

Planning before acting is a known technique to increase LLM reasoning accuracy on multi-step tasks (chain-of-thought, ReAct, etc.). Investigations are exactly the kind of multi-step diagnostic task that benefits most from an explicit planning phase.

Acceptance Criteria

• Investigation executor can enter plan mode without blocking
• Plans are auto-executed after the planning phase completes
• No terminal interaction required at any point
• Unit tests cover the headless plan→execute transition
• Existing SetPlanModeConfirmCallback(false) guard is replaced, not just removed