Fixes openclaw#50145

Problem

When messages.queue.mode is set to "interrupt", incoming messages that abort an active run cause the previous turn's assistant reply to be re-sent to the user before the new (correct) reply is delivered.

Root Cause

buildEmbeddedRunPayloads falls back to extractAssistantText(lastAssistant) when assistantTexts is empty. After an abort, assistantTexts is empty (model never generated new text), but lastAssistant still references the previous turn's message from messagesSnapshot (full session history). This stale text gets packaged as the current run's output and delivered to the user.

Fix

Two changes (defense in depth):

1. Guard lastAssistant fallback (run.ts)

When the run was aborted and produced no new assistant text, pass undefined instead of the stale lastAssistant:

lastAssistant:
  aborted && attempt.assistantTexts.length === 0 ? undefined : attempt.lastAssistant,

2. Early exit for aborted runs (agent-runner.ts)

• Skip blockReplyPipeline.flush() when aborted (buffer may contain stale content)
• Return early before payload delivery when wasAborted is true

const wasAborted = runResult.meta?.aborted === true;

if (blockReplyPipeline) {
  if (!wasAborted) {
    await blockReplyPipeline.flush({ force: true });
  }
  blockReplyPipeline.stop();
}

// ...

if (wasAborted) {
  return finalizeWithFollowup(undefined, queueKey, runFollowupTurn);
}

Testing

1. Set messages.queue.mode: "interrupt"
2. Send message A, then quickly send message B before A completes
3. Verify: only one reply is delivered (combining context from both A and B), no stale replay of previous conversation