Trace Claude Code runs and build a per-matter budget report

Two data paths, one question. claude -p --output-format json returns each run's usage and estimated cost directly, so the report reads those files; the OTel collector receives the same numbers as claude_code.cost.usage and claude_code.token.usage metrics tagged with matter.id, which is how you'd watch a running system. The decision that matters most is what to divide by in the cache-read ratio: cache reads over all input (uncached + cache writes + cache reads). Claude Code reports input_tokens without cache tokens, so dividing by it alone gives ratios far above 100% and hides the misses you're looking for.