spec-lane adapter's 30s timeout cut off 36.454s cost measurements

The investigation began when a session's cost record simply was not there. Running agent-cost directly on that same Claude Code session succeeded and took 36.454 seconds, while the pre-fix AgentCostTelemetryAdapter failed at 30.019 seconds. The adapter's default timeout was 30_000ms, so it killed a process that needed roughly six more seconds. The error message said only "Command failed", giving no indication that a timeout had occurred or that agent-cost itself might be at fault.

The fix in spec-lane v0.10.1 raises the default to 180_000ms and exposes the setting through a CLI option, with a new message that reports the timeout and SIGTERM. A re-run with a different session completed in 38.01 seconds and produced a CalibrationObservation and a cost_ledger entry, with emit-metrics reporting coverage.status=complete. The author notes this proves the measurement finished and was stored, but not that the estimated cost, the pricing catalog, or the task attribution are correct.

A separate concern is that usage-import can return exit 0 while recording measurement-incomplete and matched: false without writing to the ledger. The author's takeaway is that "measurement failed" and "cost was zero" are different states, and that collapsing them into one value makes it impossible to tell later whether a cost was truly zero or never measured.