[TASK] Refactor chat state machine to skip local tool execution in Claude Code mode

[edenreich]

Summary

PR #749 made Claude Code subscription mode (claude_code.enabled: true) a true pass-through for the headless infer agent path: infer no longer re-executes claude's tool calls locally - the adapter captures claude's own streamed tool results (TakeToolCallResults(), internal/infra/adapters/claude_code_client.go) and processSyncResponse (cmd/agent.go) synthesizes the tool messages from them instead of calling executeToolCalls.

The chat TUI path still has the original double-execution bug. Chat runs through the event-driven state machine, and EventDrivenAgent.executeTools (internal/agent/agent_tools.go) unconditionally calls executeToolCallsParallel on the accumulated tool calls - there is no IsClaudeCodeMode gate anywhere in internal/agent/states/, agent_event_driven.go, or agent_streaming.go. In claude_code mode, claude executes a tool internally and streams the call, then infer executes it again locally (or pops an approval prompt for an action claude has already performed).

The system-prompt and reminder gates from #749 already cover chat (shared addSystemPrompt / injectDueReminders in internal/agent/agent_utils.go); only tool execution is missing.

Proposed refactor:

• In claude_code mode, executeTools should skip executeToolCallsParallel and synthesize results from provider.TakeToolCallResults(), matched by tool-call ID with a neutral "executed by Claude Code" fallback, then feed them through handleToolResults unchanged.
• Plumb provider/client access into EventDrivenAgent so it can call TakeToolCallResults().
• Bypass the ApprovingTools state in claude_code mode - the actions already happened inside claude, so a local approval prompt is meaningless.
• Mirror the headless behavior so both paths share the same pre-executed-results handling where practical.

Source PR: #749 (headless fix; this gap was explicitly flagged there as a follow-up).

Acceptance Criteria

• In claude_code mode, chat does not execute tool calls locally; claude's streamed tool results are used as the tool messages
• No approval prompt is shown in chat for tools claude already executed
• Gateway-mode chat behavior is unchanged (local execution + approval flow intact)
• Tests cover the claude_code chat path (skip execution, synthesized results by tool-call ID, fallback for missing IDs)