feat: FocusOnVitalFewLevers — try full first, fall back to batched+compressed

[82deutschmark]

Problem

Local LLMs with small context windows (8K) can't handle the full enriched lever payload (~40KB for 13 levers). The previous approach (PR #134, now closed) always compressed levers, stripping review, consequences, and options fields — losing context the LLM needs for quality assessment.

Solution

Two-stage approach per Simon's feedback:

1. First attempt: Send all enriched levers with ALL fields in a single LLM call
2. On failure: Fall back to batched processing (groups of 4) with compressed levers (keeping only lever_id, name, description, synergy_text, conflict_text)

This means:

• Cloud models and large-context local models get the full picture → better assessment quality
• Small-context local models still work via the fallback path
• No data loss for capable models

Changes

• 1 file modified: focus_on_vital_few_levers.py
• Refactored execute() into two stages
• Added 3 helper methods: _assess_levers_full(), _compress_lever(), _assess_levers_batched()
• No changes to data models or other files

Replaces

PR #134 (closed) which always compressed levers