python scripts/benchmark_throughput_realpath.py owns the Kafka → Flink →
serving bridge → ClickHouse throughput measurement family. The default command
writes both mutable outputs under the ignored runtime artifact directory:
.artifacts/throughput/realpath-current.md— human-readable report;.artifacts/throughput/realpath-current.json— machine-readable results.
Re-running the command replaces those runtime files. The script refuses to write either this lifecycle page or the archived baseline, so a benchmark run cannot silently mutate tracked documentation.
The former pre-Q1.2 report stored at this path is preserved as the immutable 2026-07-09 S10 burst baseline: 400 events, 699 events/s produce, 7.97 events/s bridge apply, zero duplicates and apply failures, with lag returning to zero on the measured Mac/Colima stand.
Later date-stamped records remain separate evidence identities. The Q1.2 re-measure begins the narrow Q1.2 → Q1.3 → Q1.4 optimization chain, while the four-hour paced r4 PASS owns the closed pre-materializer serving-path endurance result. These modes do not form one direct supersession chain and do not establish a production SLA or production acceptance.
Real-path throughput measurements require Kafka, Flink, the serving bridge,
ClickHouse, and host-specific runtime state. Run that infrastructure on
deproject-mac, not the Windows development host. Measurements have no
byte-for-byte regeneration check; promote a meaningful run only under a
date-stamped filename with its JSON companion and exact run, host, source, and
configuration provenance.