hook: refactor to avoid write status to disk

[imeoer]

The hook currently executes during each layer pull, updating progress info to the status file, which is a disk write operation that may impact model pull concurrency on lower-performance disks.

This PR refactors the hook implementation to avoid writing per-layer progress to the disk-based status file, instead keeping it cached in memory, users can still get the model pull progress via the API as before.

---

This pull request refactors how progress tracking for model pulls is handled, moving responsibility from the service package to a new status.Hook and introducing a centralized HookManager for managing pull progress state. The changes simplify the codebase, improve separation of concerns, and ensure more consistent and thread-safe access to pull progress. The most important changes are grouped below.

Refactoring and Code Organization

• Moved the Hook implementation from pkg/service/puller.go to a new file pkg/status/hook.go, and replaced all references to service.Hook with status.Hook throughout the codebase. This includes updating the Puller interface, constructors, and related usages. [1] [2] [3] [4] [5] [6]
• Removed the progress callback mechanism from the pull logic, and now progress is tracked and accessed via the new HookManager. [1] [2] [3]

Progress Tracking and Status Management

• Introduced HookManager in pkg/status/hook.go and added it to StatusManager in pkg/status/status.go for centralized management of pull progress hooks, including methods to set, get, and delete hooks by key (typically the path to the status file). [1] [2] [3]
• Updated the status retrieval logic so that when StatusManager.Get() is called, it fetches the latest progress from the HookManager and attaches it to the returned status. [1] [2]

Worker and Model Pull Logic

• Refactored the worker logic to use the new status.Hook and HookManager for progress tracking; progress is no longer passed around manually but is managed centrally. Also, progress is no longer set directly during status updates, but is attached automatically when status is retrieved. [1] [2]
• Ensured that when a model is deleted, the corresponding progress hook is also removed from the HookManager, preventing stale progress data.

These changes improve modularity, thread safety, and maintainability of the codebase, especially around model pull progress tracking.

[imeoer]

@gemini-code-assist review

[gemini-code-assist]

Code Review

This pull request effectively refactors the model pull progress tracking to improve performance by avoiding disk writes for each layer. The introduction of HookManager to cache progress in memory is a good approach. The code is cleaner and the concerns are better separated.

I've found a few issues:

1. A critical bug in deleteModel that will cause a memory leak due to incorrect statusPath calculation when deleting a hook from HookManager.
2. A high-severity regression where the functionality to detect and log long-running layer pulls (checkLongPulling) has been removed.
3. A medium-severity issue with some dead code in StatusManager.Get due to an incorrect nil check.

I've provided detailed comments and suggestions for each of these points. Once these are addressed, the PR should be in great shape.

[gaius-qi]

LGTM

[chlins]

lgtm