feat(hooks): add centralized HookCenter for named hook points and plugin discovery

[aiguozhi123456]

Summary

Add a centralized HookCenter infrastructure that lets internal developers declare named hook points and external plugins attach handlers via Python entry_points. The system supports full interceptor semantics: handlers can observe, mutate data, short-circuit, or cancel the flow. Existing AgentHook/CompositeHook remain untouched.

Context

nanobot already uses entry_points for channel plugin discovery (nanobot.channels group) and has AgentHook/CompositeHook for agent-lifecycle callbacks. Neither mechanism provides a generic, named hook point system that external plugins can tap into. Contributors who want to inject behavior at arbitrary points must modify core code.

Why

• No generic hook registry: Each feature area (channels, agent loop) rolls its own callback pattern. A shared registry eliminates duplication.
• External plugins have no extension surface: Third-party packages cannot intercept or augment behavior without forking.
• Inconsistent error handling: Some callback paths crash on handler errors; others silently swallow them. A centralized system standardizes error isolation.

Solve

Introduce nanobot/hooks/ — a self-contained package with four components:

1. HookContext / HookResult — A mutable data bag passed between handlers, and a Literal-typed control-flow signal (continue / short_circuit / cancel).
2. HookCenter — Global singleton registry with register_point(), register_handler(), and async emit(). Handlers run in registration order with per-handler error isolation (failing handler is logged and skipped).
3. Plugin discovery — Scans the nanobot.hooks entry_points group. Callable plugins must declare a hook_points attribute; dict plugins map {point_name: handler}. Invalid plugins are logged and skipped.
4. AgentLoop integration — _init_hook_center() runs on startup, discovers external plugins, and logs registration count.

Changes

Component	File	Description
Types	nanobot/hooks/context.py	HookContext dataclass (mutable dict), HookResult with Literal["continue","short_circuit","cancel"] action
Types	nanobot/hooks/types.py	HookHandler protocol definition (`async (HookContext) -> HookResult
Core	nanobot/hooks/center.py	HookCenter class: register_point, register_handler, emit, reset; global get_center()/reset_center()
Discovery	nanobot/hooks/discovery.py	discover_hook_plugins() and register_discovered() via importlib.metadata.entry_points
Init	nanobot/hooks/__init__.py	Public re-exports: HookCenter, HookContext, HookResult, HookHandler
Integration	nanobot/agent/loop.py	_init_hook_center() called in AgentLoop.__init__
Docs	docs/hook-center-guide.md	Developer guide for internal and external plugin authors

Test

Suite	File	Count	Coverage
Core	tests/hooks/test_center.py	22	HookContext, HookResult, register/emit, error isolation, global singleton
Discovery	tests/hooks/test_discovery.py	9	entry_points load, dict/callable plugins, missing hook_points rejection
Integration	tests/hooks/test_integration.py	4	AgentLoop init, plugin discovery on startup, AgentHook backward compat, plugin failure resilience
Regression	tests/agent/test_hook_composite.py	18	Existing AgentHook/CompositeHook unchanged

Total: 53 tests, all passing. No regressions in existing test suites.

---