feature: attribute tool-pruning savings — tokens and cost saved

[huang195]

Feature Description

The README claims tool-definition pruning saves 4–20% of the prompt on every turn. Nothing
reports the realised figure, so the claim is currently unfalsifiable.

Proposed Solution

• Per request: tokens removed by pruning and the cost that avoided
• Rolled into session and per-model totals
• Distinguish savings from cache effects so the two are not double-counted
• Report per agent, and report zero honestly for agents with no tool inventory — do not
  present "no savings" as "no data" or vice versa

Additional Context

Part of the Cortex v0.9.0 release, stretch scope. Note that toolscan/known.go hardcodes
Claude Code built-in tool names, so pruning currently applies to Claude Code only.