fix: use clean user message for all memory provider operations

[teknium1]

Problem

When a skill is active, user_message contains the full SKILL.md content injected by the skill system. This bloated string (often 10K+ chars) was being passed to memory provider operations, causing providers with query size limits (e.g. Honcho's 10K char limit) to fail.

Relates to #4889 and #4902.

Fix

Both memory provider call sites now use original_user_message (the clean user input, already defined earlier in run_conversation()) instead of the skill-inflated user_message:

• Pre-turn prefetch (line ~6695): prefetch_all() query used for context recall before the LLM call
• Post-turn sync (line ~8672): sync_all() for turn persistence + queue_prefetch_all() for next-turn recall

original_user_message is derived from persist_user_message (the clean original) with fallback to user_message when no override exists — this pattern was already established at line 6516 for plugin hooks.

Why not just the post-turn fix?

PR #4902 only fixed the post-turn sync/prefetch. The pre-turn prefetch_all() had the same issue — it ran the skill-inflated message as the recall query, returning irrelevant context and potentially exceeding provider limits.

Test plan

• 269 memory/agent tests pass
• Single file change, +9/-5 lines