Performance

This directory is the durable record of rozi performance investigations. It separates:

• the audit playbook, which explains how to reproduce measurements;
• benchmark documentation, which describes the permanent Criterion and memory harnesses;
• dated audit reports in audits/, which record evidence and conclusions for one revision and environment.

Current assessment

The latest audit found no release-blocking performance problems and classified rozi as ready with minor improvements. Broad-scope scrollback search now runs in bounded cooperative slices, epoch-gated workers resolve the workbar command lifecycle concern, and an isolated framework comparison confirms a roughly 38-39% snapshot-rebuild reduction. Saturation and durable resurrection now have permanent measured harnesses. Durable snapshots write off the server loop and reuse unchanged panes' replay files, cutting maximum 16-pane 5,000-row server-loop blocking from 207.7 ms to 12.0 ms. Remaining work is idle server backoff/readiness design, gated image-budget work, bounding persisted history for sessions where every pane is busy, slow-storage evidence, and a full memory soak.

Audit ledger

Date	Revision	Verdict	Report
2026-08-04	2b85924 plus recorded audit worktree changes, with resurrection follow-ups at 9f66a6a and 4d45cf4	ready with minor improvements	Performance follow-up evidence
2026-08-03	c07b6be plus recorded worktree changes	ready with minor improvements	Performance and resource-efficiency audit

Recording future audits

Create a new audits/YYYY-MM-DD.md rather than overwriting an older result. Every report should record:

1. exact Git revision and whether the worktree was dirty;
2. OS, CPU, Rust version, release profile, and unavailable tools;
3. exact commands, dimensions, durations, sample counts, and statistics;
4. application memory separately from child-process memory;
5. confirmed findings, deliberate trade-offs, harmless details, and unverified risks;
6. prioritized follow-up work and the measurement that would prove each improvement;
7. one of the verdicts ready, ready with minor improvements, needs optimization before release, or insufficient evidence.

Small factual corrections may update an existing report. New measurements, changed workloads, or a different revision require a new dated report so historical comparisons remain meaningful.

Raw Criterion data, Samply profiles, memory-matrix output, logs, sockets, and terminal captures stay under ignored target/ paths and are never committed. Only the summarized report belongs here.