weight confidence updates by reciprocal rank of result position

[prathkan03]

Summary

• Confidence boosts/penalties now scale by 0.1 / rank instead of a flat 10%
• Top results receive the full signal; lower-ranked results receive proportionally less
• Rank is passed from the call site in mcp_tools.py via a rank map built from ordered results

Test plan

• Give positive feedback on a query — confirm rank 1 result gets a larger confidence boost than
  rank 5
• Give negative feedback — confirm rank 1 result gets a larger penalty than rank 10
• Omit node_ids in feedback — confirm all results are updated with correct ranks
• Pass specific node_ids — confirm only targeted nodes are updated, with correct ranks
  preserved