dsh-rerun-metrics

Folds a harvested session log into the efficiency and process halves of an attempt's measurement.

Every figure is derived from the durable log and nothing else. That is deliberate: the harness under test runs in another process, and the only thing rerun is allowed to depend on is what DSH already guarantees to write down. It also means a metric added later can be folded over stored logs from months ago.

Time figures follow the session-stats semantics so a rerun number and the product's own stats strip cannot disagree.

Packed chunk rows are expanded HERE, by shape, rather than through the session package's decoder — the log was written by a different build, and the packing vocabulary exists in DSH 0.1.1 and is gone by 0.1.5.

A library, not a Cordis plugin: in this design the orchestrator measures logs another process already wrote, so there is no live session to project from.

Model Experience

None. This package never contributes to a model request: it runs in the orchestrator process, and the model calls it cares about happen inside a separate attempt subprocess running a different build of DSH entirely.

KV Cache effect

None; it neither assembles nor sends a provider request.

Known Limitations and Deferred Work

• stepsToFirstEdit is name-based, so a task solved entirely through shell heredocs reports none.
• Unreadable rows are counted, not recovered. A log written by a much newer harness may carry rows this build cannot expand; the count says the numbers are a floor.
• Subagent child logs are not pooled into the parent's counts. Delegation is measured through the parent's own events.