Performance archive
This directory records performance audits for specific revisions and machines. Dated reports are historical evidence. They are not current benchmark guidance and their source, dependency, or architecture descriptions may no longer match the repository.
The latest report is Compact scrollback in Alacritty, 2026-09-15.
Use the benchmark guide for permanent harness commands and definitions. Use the audit playbook to run and record a new audit.
Audit ledger
| Date | Measured revision | Recorded verdict | Report |
|---|---|---|---|
| 2026-09-15 | Alacritty d692748d plus archived port e8e5aee6 | opt-in port halves Alacritty PSS with short-line history and is faster on short lines; other vtebench within 3%; upstream closed the proposal; not adopted | Compact scrollback in Alacritty |
| 2026-09-14 | 52cccdc plus archived experiment | application ingest cost within noise for short lines, +7% for full-width styled logs; 55-122 MiB application PSS saved; technical go, adoption waits on an engine-fork decision; found two tui-lipan hang-up exit bugs | Compact history ingest decision |
| 2026-09-14 | cedd4a6 plus sibling tui-lipan worktree | attr-keyed runs cut realistic snapshots ~46-49%; live compact gap gone at the new baseline; ready to ship in tui-lipan, compact still blocked on ingest | Attr-keyed terminal spans |
| 2026-09-14 | cedd4a6 plus archived experiment | Style-per-cell mapping and PartialEq dominate released snapshots; attr-keyed runs cut assembly ~43%; not adopted | Dense span-assembly profiling |
| 2026-09-14 | cedd4a6 plus archived experiment | bulk reads recover scrollback and cut the live snapshot gap in half; live path still about 15% slower; not adopted | Render-oriented history reads |
| 2026-09-14 | cedd4a6 plus archived experiment | batching is slower and retains more memory; active snapshot regression confirmed; not adopted | Bounded batch history compaction |
| 2026-09-13 | 824d925 plus archived experiment | 86% wide short-line allocation saving and zero steady allocation calls; short-line ingestion still slower; not adopted | Dense viewport and compact history |
| 2026-09-10 | 824d925 | compact-history prototype saves 86% on wide short-line allocations, but fails the ingest gate; not adopted | Terminal history storage investigation |
| 2026-09-05 | 1e8f4a0 plus recorded worktree changes | fixed: default-history 253×64 pane with one client uses 14.1% less application PSS | Pane retained-memory audit |
| 2026-09-04 | ca5f34b, fix measured at 151f175 plus the framework change | confirmed defect, fixed: one changed character repainted the whole window; damaged-row repaint cuts the reported case from 8.3% to 2.7% of a core | Idle CPU under a non-fullscreen agent spinner |
| 2026-09-04 | 4b8f076 plus recorded framework changes | five changes land a 32% cut | Element-tree allocation in view and layout |
| 2026-09-03 | c9e9881, clean worktree | ready with minor improvements | Render and ingest CPU attribution |
| 2026-08-29 | 4dedfc8 before and 0638f41 after, plus recorded benchmark and wait tuning | ready | Idle server wakeup evidence |
| 2026-08-04 | 2b85924 plus recorded audit worktree changes, with 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 |
Create a new audits/YYYY-MM-DD.md for measurements from another revision, machine, or workload. Do not overwrite old results to make them describe the current tree. Small corrections may clarify the original record if they do not change its measured meaning.
Keep raw Criterion data, profiles, memory-matrix output, logs, sockets, and captures under ignored target/ paths. Commit only the summarized report.