agent_end: auto-complete in_progress tasks when LLM finishes turn

[ZGltYQ]

Adds two related fixes:

1. Auto-complete on agent_end: When the LLM stops producing tool calls (agent_end event), mark any in_progress tasks as completed. This matches the natural semantic of 'the LLM is done for this turn, so tasks it was working on are done'.

   Previously, only subagent-executed tasks were auto-completed via subagents:completed. Regular tasks required an explicit TaskUpdate call, which the LLM often forgot, leaving tasks stuck in_progress.

   The very first agent_end of the session (fired on startup before the user sends any message) is skipped via the isFirstAgentCycle guard.

2. Fix task loss across store upgrade: In session scope, the store starts in-memory before the session ID is known, then upgrades to file-backed once upgradeStoreIfNeeded() fires. Previously, the upgrade created a brand new empty TaskStore and the old in-memory tasks were GC'd \u2014 so any task created before the first turn_start was silently lost.

   Now we persist the old store's tasks to the new session file (atomic temp-file + rename, same pattern as TaskStore.save()) before swapping the store reference.

Without fix #2, fix #1 is only partially effective: agent_end fires but the in-memory task created in the same turn is in a different store instance than the file-backed one used later.

[tintinweb]

Thanks for the PR. going to close this one.

The premise that "LLM stops producing tool calls ⇒ tasks are done" doesn't hold. agent_end also fires when the LLM pauses for a clarifying question, reports partial progress, hits an error, or gets interrupted — in all those cases this would silently flip unfinished work to completed, which defeats the point of the task widget surfacing a mismatch between claimed and actual status.