fix: skip synthetic zero-usage assistant entries in billing

[jverre]

Details

Second phantom-token mechanism found after the compaction fix (#22), exposed by the Σ lanes vs usage discrepancy that #22 deliberately left visible.

Claude Code writes locally fabricated assistant entries into the transcript when an API call errors or is interrupted: model: "<synthetic>", isApiErrorMessage: true, and a complete usage object where every field is zero. These were never billed by the API. llmCallsInTurn only checked Usage != nil, so each one was treated as a real LLM call whose "request" (the entire prior conversation) got reconciled against a zero-token measured prompt: estimated pieces scaled to zero and the usage-derived pieces overflowed the positional cut into the fresh-input tier.

Two such entries in a real session produced 665,727 phantom input tokens (lane input summed to 781,745 vs 115,337 actually billed - ~$6.7 of phantom spend at fable-5 input rates).

Fix: skip calls whose usage sums to zero. They bill nothing, so nothing is lost - and this covers any future unbilled-entry variant, not just <synthetic>.

Testing

• TestBillingSkipsSyntheticZeroUsageCalls: a zero-usage synthetic entry between two real calls is excluded from llm_calls and Σ lanes still equals usage
• Dry-run on the real session that exposed the bug: all four lane columns now reconcile to API usage token-for-token (input 130,000 / cache_read 328,138,781 / cache_creation 3,004,255 / output 733,462) - including output, whose previous small drift also traced to the synthetic blocks
• Full suite, gofmt, go vet clean; binaries rebuilt

🤖 Generated with Claude Code