0750 — Benchmarks retain recurring workloads and freeze completed investigations

Context

The root lab now owns benchmark execution. The user approved auditing each workload’s value rather than migrating every historical experiment.

Decision

Keep quiet for steady delivery and max-throughput for sustained capacity. Keep the multi-stream family experimental: its group and instance counts grow with stream count, so it compares deployments rather than one variable.
Merge dev into a one-minute quiet invocation, just reliability-smoke. Require a scenario for just reliability-lab; retire throughput’s short fixed-rate declaration. Report mode reads historical results by name.
Retire independent claim, compaction, fillfactor, scale, trigger-fanout, and native scratch programs. Freeze their tracked source together in .bench/results/retired-source-2026-09-11.tar.gz, preserving original paths.
Keep historical results, design rationale, and local investigation evidence. Consumerdefaults and idempotency already retain results without a tracked runner. Do not recreate their temporary tuning harnesses.
A maintained scenario names a recurring decision, distinct measurements, and the action a changed result would prompt. Completed optimization experiments do not automatically become maintained scenarios.
Retain existing correctness tests. The compaction hot-key deadlock assertion lives in compactiondeadlock e2e; stale commit and cursor-floor invariants have integration tests. Sharded prototype assertions describe unshipped machinery. Scratch identity checks remain frozen evidence, not a claim that aggregate max-throughput now proves identities after retention.

Consequences

No new workload engine or fault mechanism is introduced. Historical report names remain readable; new smoke runs record under quiet. Idle-fleet costs, automatic-batching capacity, and recovery under load remain separate research questions. Historical source is available without being maintained as live code.

Evidence retention superseded by 0751.