feat(analytics): add cc-economics --audit for cache-lifecycle savings

[thebigjc]

Summary

• Adds an opt-in --audit mode to rtk cc-economics that attributes token savings by reading Claude Code / pi session logs (~/.claude/projects/**/*.jsonl and ~/.pi/agent/sessions/**/*.jsonl) instead of using a flat weighted-CPT estimate.
• Credits each compressed command's saved tokens across subsequent turns based on the turn's actual Anthropic prompt caching tier:
  • Cache write (1.25x): tool-result appearance + cache-eviction rebuilds
  • Cache read (0.10x): steady cached turns
• Scopes matching to the same project (cwd) with a 60-second window, capped by the authoritative count of actual rtk commands invoked.
• Removes the vestigial $0 input savings row from the audited output since tool results structurally always enter the cached prefix via cache creation (writes), leaving plain input savings at $0.
• Fixes a parallel-test environment/database race condition in src/core/tracking.rs tests by synchronizing database-touching tests with a shared Mutex.

Example Output

Running rtk cc-economics --audit displays the session-log audited prompt caching savings breakdown:

[cost] Claude Code Economics
════════════════════════════════════════════════════

  Spent (ccusage):              $8728.40
  Token breakdown:
    Input:                      42.7M
    Output:                     51.0M
    Cache writes:               168.6M
    Cache reads:                9255.0M

  RTK commands:                 7893
  Tokens saved:                 1421.1M

  Audited Savings (Prompt Caching):
  ┌─────────────────────────────────────────────────┐
  │ Cache write savings:   $242.01  (2.8%)          │
  │ Cache read savings:    $925.66  (10.6%)         │
  │ ─────────────────────────────────────────────── │
  │ Total savings:         $1167.67  (13.4%)        │
  │ Claude rtk calls:      2236/2260 matched        │
  │ Derived input CPT:     $6.09/MTok               │
  └─────────────────────────────────────────────────┘

  How it works:
  RTK compresses CLI outputs before they enter Claude's context.
  Savings derived using API price ratios (out=5x, cache_w=1.25x, cache_r=0.1x).

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test (all 2437+ tests pass successfully)
• Manual testing:
  • Verified that rtk cc-economics --audit parses session logs correctly, maps rtk commands to Claude turns, and displays precise caching savings.

[CLAassistant]

All committers have signed the CLA.