fix: clear live messages.jsonl during async commit to prevent duplicate commits

[voidborne-d]

Summary

Fixes #580 — async session commit can re-commit old messages before live session switch completes.

Root cause

commit_async() uses a Copy-on-Write pattern:

1. Phase 1 (Copy): Copy live session to temp directory
2. Phase 2 (Write): Archive messages + extract memories in temp, clear self._messages in memory
3. Phase 3 (Switch): Enqueue to SemanticProcessor for atomic temp→live directory switch

The problem: after Phase 2 clears the in-memory message list, the live session directory still contains the original messages.jsonl. The atomic switch only happens later when SemanticProcessor processes the queue. If the session object is evicted from memory and reloaded from disk before that switch (or if the process restarts), the old messages.jsonl is read back — causing a subsequent commit to re-archive the same messages.

Fix

Write an empty messages.jsonl to the live session directory immediately after Phase 2 completes (step 2.6), before enqueuing to the semantic queue. This closes the race window:

• In-memory: self._messages is already cleared (step 2.1)
• On-disk: messages.jsonl is now also cleared (step 2.6)
• Even if the session is reloaded before the semantic switch, it finds no messages to re-commit

The fix is intentionally minimal — one write_file call + logging. It uses the existing _viking_fs.write_file() API that is already used during session initialization (line 151).

What about the task completion semantics?

The issue also mentions that task.status=completed should imply the live session has been switched. This PR addresses the data safety concern (no duplicate commits) but does not change the task tracking semantics. A follow-up could defer tracker.complete() until the semantic switch finishes, if stricter task completion guarantees are desired.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[yangxinxin-7]

Hi, thanks for the detailed writeup and the fix!

After digging into the current codebase, I think this PR may no longer be needed. PR #584 reverted the COW pattern that was introduced in #535, which is the root cause of the race condition described in #580. The # 2.5 Update active_count context this diff targets was removed as part of that revert and no longer exists in main.

In the current commit_async(), the race window no longer exists — after self._messages.clear() (session.py:328), _write_to_agfs_async(self._messages) is called immediately (session.py:349), which writes empty content directly to the live messages.jsonl via a synchronous write_file call (viking_fs.py:1268), no async queue involved.