feat: replace inline nudges with background memory/skill review

[teknium1]

Summary

Replaces the inline memory/skill nudges — which polluted 43% of user messages with backward-looking system instructions — with a background agent that reviews and saves independently after the main response is delivered.

The problem

Memory and skill nudges were appended directly to the user's message content:

User's actual message: "fix this bug"
What the model saw:    "fix this bug

[System: The previous task involved many tool calls. Save the approach as a skill...]
[System: You've had several exchanges. Consider: has the user shared preferences...]"

The model had to choose between the user's forward-looking task and the backward-looking system directives. In 2 confirmed cases, the agent spent 1-3 tool calls on memory/skill work before starting the user's task. The nudges were also permanently stored in conversation history, polluting session transcripts.

The solution

When nudge conditions are met, a background review agent spawns after the main response completes:

# After the agent finishes responding to the user:
threading.Thread(target=_run_review, daemon=True).start()

The review agent:

• Uses the main model (same quality, not auxiliary — skills/memory are high-precision)
• Gets a read-only snapshot of the conversation
• Has only memory + skill_manage tools (5 iteration budget)
• Shares the memory store so writes persist immediately
• Runs with quiet_mode=True, skip_context_files=True
• Never modifies the main conversation or produces user-visible output
• All exceptions caught — can never affect the main session

What doesn't change

• Trigger conditions: still every 10 user turns (memory) and after 10+ tool iterations (skills)
• Token cost: same context processed either way, just on a separate track
• Memory/skill quality: actually better — dedicated prompt for review vs a hint appended to an unrelated message

Changes

run_agent.py:

• Remove nudge injection from run_conversation() (lines 5225-5249 → tracking only, no user_message +=)
• Add _spawn_background_review() method with _BACKGROUND_REVIEW_PROMPT
• Add background fork trigger after response delivery (before return)
• Net: +77 lines, -15 lines

Test plan

5670 passed, 200 skipped, 23 deselected

Closes #2227.