perf(ui): reduce Linux workspace switch overhead

[pascalandr]

Refs #330

Summary

• reduce Linux/Tauri workspace-switch overhead by unmounting inactive instance shells instead of keeping every workspace UI tree alive in parallel
• keep sidecar tabs mounted, so the optimization stays focused on workspace shells and does not reset unrelated sidecar state
• add targeted Tauri/Linux perf instrumentation for investigation, but gate the benchmark behind an explicit opt-in flag
• keep the Linux/Tauri light-render experiment opt-in as well, so it can be tested without changing default Linux UI behavior

Measured Improvement So Far

In the Linux VM used for investigation, the measured workspace-switch path improved from roughly ~46ms average to ~33ms average after mounting only the active instance shell.

Perf Experiment Flags

• benchmark: localStorage.setItem(\"perf330-bench\", \"1\")
• light-render experiment: localStorage.setItem(\"perf330-light-render\", \"1\")

Reload after setting either flag.

Still Under Investigation

• prompt/input lag is not fully diagnosed yet
• the VM also shows graphics-stack warnings (libEGL / DRI3), so this PR remains a draft for validation on a real Linux machine

Validation So Far

• npm run build --workspace @codenomad/ui
• cargo build
• built and installed a .deb from this branch for local testing

[pascalandr]

@CodeNomadBot

I want you to perform the gatekeeper checks for latest PR state

check for regressions
check for better possible implementations
Be the PR gatekeeper
Be ruthless about code quality

[codenomadbot]

Gatekeeper review of the latest PR state.

Assumption: you wanted a ruthless PR-quality review of the current diff against dev, focused on regressions and stronger implementation choices, not a request for me to change code in this branch.

Findings:

1. High: the Linux "light-render" experiment is currently unconditional production behavior, not an experiment.
   packages/ui/src/index.css:36-50 removes all transition durations, animation durations, backdrop filters, and several shadows for every Tauri/Linux render. packages/ui/src/main.tsx:20-33 and packages/ui/src/renderer/loading/main.tsx:44-60 stamp data-runtime-os="linux" on every run, so this ships broadly to all Linux users if merged.

Why this blocks:

• The PR description frames this as an experiment to test whether WebKitGTK rendering cost is involved.
• The implementation permanently changes Linux UI behavior instead of making it opt-in.
• That creates obvious UX regressions on Linux and makes it harder to measure whether the workspace-shell change alone was the actual improvement.

Better implementation:

• Gate the CSS behind an explicit experiment flag, dev-only flag, or user setting.
• Keep the runtime OS annotation if needed, but do not tie production styling changes to it by default.

2. High: switching tabs now unmounts all inactive tabs, which resets state for sidecars and instance-local UI, not just the heavy workspace tree.
   packages/ui/src/App.tsx:734-775 changed the tab content from rendering all tabs and hiding inactive ones to rendering only activeAppTab(). That means every switch unmounts the previous tab entirely. For sidecars this is a direct regression: packages/ui/src/components/sidecar-view.tsx:12-14 stores iframe/navigation state locally, and sidecar-view.tsx:80-194 rebuilds the iframe from props.tab.shellUrl on mount, so switching away from a sidecar tab will drop its current page/path/input state and reload it on return.

Why this blocks:

• The PR goal is to reduce workspace-switch overhead.
• The lifecycle change is broader than that goal and changes sidecar behavior too.
• Instance-local UI state also resets on remount (packages/ui/src/components/instance/instance-shell2.tsx:107-125), so drawers/dialog state no longer survives tab switches.

Better implementation:

• Restrict the mount-only-active optimization to instance shells, not all tab kinds.
• If the full-shell unmount is required, preserve sidecar and instance local state in stores before unmounting, otherwise the UX regression is immediate.

3. Medium: the dev benchmark mutates real app state automatically on Linux startup, which is too invasive for routine development runs.
   packages/ui/src/App.tsx:322-435 starts runPerf330Benchmark() on mount. On Linux/Tauri dev it may auto-open the most recent folder (App.tsx:330-336), create another instance (App.tsx:353-364), and create a session if needed (App.tsx:406-417).

Why this is risky:

• Running the app in dev now launches real workspace/session activity without explicit consent.
• It can pollute recents/session state and make normal debugging noisy.
• It also means perf instrumentation is no longer isolated from app behavior.

Better implementation:

• Make the benchmark opt-in via env var, query param, or explicit developer command.
• Keep passive logging on by default if useful, but keep state-mutating benchmark setup off by default.

Validation notes:

• cargo build did not finish in this environment because the Tauri build expects resources/server to exist.
• npm run build --workspace @codenomad/ui did not run here because the workspace is missing the local vite dependency in this checkout.
• I did not see evidence from this environment that these failures are caused by this PR specifically, so I am not treating them as PR findings.

Gatekeeper verdict: not ready yet. I would want the Linux rendering changes gated, and the tab-lifecycle optimization narrowed so it does not regress sidecars and other non-workspace tab state.

--
Yours,
CodeNomadBot

[pascalandr]

@CodeNomadBot

I want you to perform the gatekeeper checks for latest PR state

check for regressions
check for better possible implementations
Be the PR gatekeeper
Be ruthless about code quality

[codenomadbot]

Gatekeeper review of the latest PR state.

Assumption: you wanted a fresh review of the current branch head after the follow-up fixes, focused on regressions and stronger implementation choices, not a request for code changes from me.

Findings:

1. Medium: opening the folder-selection overlay now unmounts the active instance shell, so instance-local UI state is still reset outside the intended workspace-switch path.
   packages/ui/src/App.tsx:733-776 now keeps sidecars mounted, which fixes the earlier sidecar regression, but instance tabs are wrapped in <Show when={isVisible()} keyed>. isVisible() is defined as activeAppTabId() === tab.id && !showFolderSelection() (App.tsx:735), so the currently active instance shell is unmounted whenever the folder picker opens, then remounted on close.

Why this still matters:

• The optimization goal is to reduce inactive workspace-shell overhead during tab switching.
• Tying mount lifetime to showFolderSelection() broadens the behavior change beyond tab switching.
• Remounting the active shell still drops instance-local ephemeral state such as drawer/dialog state and any other non-store UI state in packages/ui/src/components/instance/instance-shell2.tsx.

Better implementation:

• Keep the instance-shell mount condition scoped to active-tab identity only.
• Let the folder-selection overlay visually cover the shell without participating in shell lifetime.
• If there is a specific reason to unmount during folder selection, that reason should be documented because it is not obvious from the PR goal.

What looks fixed from the previous review:

• Sidecar tabs stay mounted again, so the sidecar iframe/state regression appears addressed.
• The Linux light-render path is now opt-in via data-perf330-light-render.
• The perf benchmark path is now gated through isPerf330BenchmarkEnabled() instead of running on every Linux/Tauri dev startup.

Residual note:

• packages/ui/src/App.tsx is now about 824 lines, which is over the repository’s source-file size guideline. I am not treating that as a blocker by itself, but it is worth keeping in mind if this area changes again.

Gatekeeper verdict: closer, but not fully ready yet. I would want the instance-shell lifetime decoupled from showFolderSelection() before approving this as a tight, regression-safe perf change.

--
Yours,
CodeNomadBot

[pascalandr]

@CodeNomadBot

I want you to perform the gatekeeper checks for latest PR state

check for regressions
check for better possible implementations
Be the PR gatekeeper
Be ruthless about code quality

[codenomadbot]

Gatekeeper review of the latest PR state.

Assumption: you wanted a fresh review of the current branch head after the latest follow-up fix, focused on regressions and stronger implementation choices, not a request for code changes from me.

Findings:

• No blocking regressions found in the latest revision.

What changed relative to the previous review:

• The active instance shell now stays mounted during folder selection. In packages/ui/src/App.tsx:735-750, mount lifetime is keyed off isTabActive() while visual visibility is controlled separately by isVisible(). That fixes the earlier issue where opening the folder-selection overlay unmounted and remounted the active shell.
• Sidecar tabs are still kept mounted, so the earlier sidecar state-reset regression remains addressed.
• The Linux light-render path is still opt-in via data-perf330-light-render.
• The benchmark path is still opt-in via isPerf330BenchmarkEnabled().

Residual risks / testing gaps:

• I was only able to do static review in this environment. I did not re-validate the full UI/Tauri builds here, so I am not independently confirming runtime behavior beyond code inspection.
• packages/ui/src/App.tsx is now about 825 lines, which is over the repository’s source-file size guideline. I am not treating that as a blocker for this PR, but it is a maintenance risk if this area keeps growing.

Gatekeeper verdict: current diff looks acceptable from a regression-scope perspective. I do not have a remaining code-quality blocker on the latest PR state.

--
Yours,
CodeNomadBot