fix: use cached raw user message as auto-recall query

[raw34]

Summary

• Auto-recall's recallQuery uses event.prompt directly, which on Slack includes Conversation info JSON metadata (message_id, sender_id, conversation_label, etc.) prepended by the platform adapter
• This metadata pollutes the embedding vector, causing irrelevant memories to score higher than actually relevant ones
• The fix uses lastRawUserMessage.get(cacheKey) || event.prompt — the same pattern already applied to gatingText (line 2245) — so the retrieval query uses clean user text instead of the noisy assembled prompt

Context

The gatingText variable (used for short-message skip and greeting detection) was already correctly patched to prefer lastRawUserMessage. However, the actual recallQuery passed to retrieveWithRetry() was missed, so retrieval itself was still affected.

The lastRawUserMessage Map is populated in the message_received hook (line 2211–2219), which strips bot mentions and caches clean user text keyed by channelId. The before_prompt_build hook shares the same cacheKey derivation, so the cached value is always available when needed.

Reproduction

1. Deploy with Slack channel adapter
2. Send a short CJK message (e.g. "我们团队谁负责项目管理")
3. Observe that event.prompt in before_prompt_build contains Conversation info: {"message_id":"...","sender_id":"...","conversation_label":"..."} prefix
4. Auto-recall returns irrelevant memories because the metadata dominates the embedding

Test plan

• Verify Slack messages trigger auto-recall with clean query text (no Conversation info JSON in embedding input)
• Verify non-Slack channels (Telegram, API) still work correctly (falls back to event.prompt)
• Verify gatingText and recallQuery now use the same source for consistency

🤖 Generated with Claude Code