From b6dd973a88f36cb40fae925ebe66be56e7ae6e68 Mon Sep 17 00:00:00 2001 From: karl Date: Tue, 13 Oct 2026 10:12:25 +0200 Subject: [PATCH] Record native semantic activation timing retest --- docs/LINUX-DESKTOP-VALIDATION.md | 42 ++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+) diff --git a/docs/LINUX-DESKTOP-VALIDATION.md b/docs/LINUX-DESKTOP-VALIDATION.md index 5ac832f..2abddd5 100644 --- a/docs/LINUX-DESKTOP-VALIDATION.md +++ b/docs/LINUX-DESKTOP-VALIDATION.md @@ -332,3 +332,45 @@ Xvfb exited zero, and process inspection found no surviving app, worker, WebKit, DBus, clipboard, display or proxy helpers. Outer `/dev/dxg` remained real. Physical desktop, installed resource discovery, representative relevance, first-install cold storage, p95, GPU and macOS gates remain open. + +## Semantic-activation clock correction: `19d949f` + +The focused retest used an immutable committed source snapshot and matching +`index-DYGZ2Wy5.js`, separately from subsequent glass-UI work. Native embedded +Release build passed in 3m 23s. Executable SHA-256: +`247746c289cd153e7bd70cadbbdbe3a92aa918caa374ad8557196815bd9182fe`; +frontend SHA-256: +`46b1a070c3f8cbe3b11471f049e071eb8e0e8a30e98be3e045256301f9759aec`. +The normal Linux query-role worker remained the previously verified `b52d50...` +binary. The owned `desktop-qa-activation` session isolated XDG and Rust data, +reused the earlier four-file fixture and complete verified model cache, and made +no model download or proxy changes. Startup established full generation 1 in +the new data directory; the measured rebuild therefore activated 1 to 2. + +Actual native UI checks: + +- Enabled Performance details and typed a new `end the supplier contract` query. + Its copied report measured 60 ms debounce, 111 ms first displayed response, + 845 ms final displayed response and a 632 ms native hybrid observation. + These single samples are not first-launch, cold storage or warm p95 proof. +- Explicitly selected `retry.rs`, then left the completed query idle for about + 87 seconds before clicking the actual Rebuild semantic index button. +- The new generation retained four chunks. The copied refresh report had + `trigger: semantic_activation`, **13 ms first / 135 ms final displayed** and + **2.4 ms native**; it excluded the long idle interval. Its separate + `original_input` preserved the original 60/111/845 ms measurements. +- The report context reflected current `semantic_generation: 2` and + `indexed_chunks: 4`. Automatic lexical and hybrid refresh kept the explicit + `retry.rs` preview/selection, and the warmed CPU query worker stayed PID 387 + through activation. Copied reports excluded query text and source paths. + +Local ignored evidence: `.searchi/qa/linux-activation-{initial-search,search}.json`, +`linux-activation-{initial-trace,selected,refreshed-trace,selection-retained}.png`, +`linux-activation-states.jsonl` and `linux-activation-cpu-audit.json`. The previous +`9aae3f8` 143,928 ms report is retained as historical defect evidence and was not +overwritten. The private software-display audit again found no GPU descriptors +or providers; the worker used Zen4 CPU, one thread and zero swap. Owned shutdown +ended the app group and Xvfb successfully; subsequent process inspection found +no app, worker, WebKit, DBus, clipboard or display helpers remaining, and outer +`/dev/dxg` remained the real device. This closes the tested activation-clock +regression, not the broader native UX, p95, installer, GPU or macOS gates.