bug: async replies that complete after clear-all repopulate history and usage

Problem

clearAll() in LLMExtension.scala clears messageHistory, agentUsage, runUsage and agentProfile under their locks, but nothing invalidates calls that are still in flight. Start llm:chat-async on the observer, call clear-all, then let the reply arrive: its completion callback runs recordUsage and commitExchange, so the previous run's tokens and exchange land in the freshly reset state. The separate history and usage locks also mean a reset can split those two updates.

Found by a Codex review on 2026-09-08. Predates today's changes for history (since #44); token usage (#78) extends the same gap.

Fix

Capture a run generation number when a request is launched; advance it in clearAll(); have recordUsage and commitExchange drop the update when the generation no longer matches. One counter under one lock keeps history and usage consistent.

Test

Start a __TEST_DELAY: async call, clear-all, resolve it with runresult, then assert llm:history is [] and llm:get llm:usage-total "calls" is 0.

Start in LLMExtension.scala, tracing llm:chat-async completion handling through recordUsage and commitExchange, then inspect clearAll() and its locks. Reproduce the __TEST_DELAY: scenario by running clear-all before resolving with runresult; done means llm:history remains [] and llm:get llm:usage-total "calls" remains 0.