Background 09 phase 0 verdict

Research note dated 2026-10-10. Read for reasoning, not current fact; later decisions in docs/decisions.md and docs/architecture-decisions.md win.

The spikes it describes were throwaway code (D16) and were never merged. Notes 07 and 08 correct parts of the earlier notes, and note 03 assumes Tauri.

Moved from docs/research/09-phase-0-verdict.md (original at af3de53) on 2026-10-16.

Phase 0 verdict: Syrinx on Windows 11

Written 2026-10-10 from two throwaway spikes built on 2026-10-09 and the owner's own tests. The spikes' code was never merged; their full results are on the local branches spike/gpui (spike-gpui/RESULTS.md, at bfac1c5) and spike/tauri (spike-tauri/RESULTS.md, at 2579d86). This verdict is for Windows 11 only (D2); macOS and Linux were not tested.

Verdict

Go, with GPUI, on two conditions. The engine (WebView2 driven from Rust through wry, D3) works for the owner's hardest needs, and the GPUI chrome beats the Tauri + Svelte control by far more than D4's rule requires. The conditions are:

1. Idle CPU must be fixed. GPUI's Windows backend repaints every window on every display refresh even when nothing changed, so the spike idles at 3.75% of one core against the 0.5% bar (D6). The owner allowed a small, pinned GPUI patch offered upstream (D37). If the patch can't bring idle CPU under 0.5%, GPUI against Tauri goes back to the owner.
2. The untested musts are tested before Syrinx becomes the daily browser. Hulu, the 13 other daily sites, Google sign-in with a password and phone prompt, and staying signed in after a restart were not tried in Phase 0. They move into plan 1's Phase 2 exit criteria.

What worked

Check (D6, D7)	Result	Evidence
Google sign-in, passkey	Pass	Owner, 2026-10-10
Google sign-in, Workspace account	Pass (so no device-check block)	Owner, 2026-10-10
YouTube	Pass	Owner, 2026-10-10
Disney+	Pass	Owner, 2026-10-10
Live palette and sidebar peek over a page and over the editor, Esc returns focus	Pass in both spikes	GPUI 50/50 and Tauri 25/25 automated checks
Ctrl T, W, L, K, Ctrl Tab, Ctrl Shift T from a page and from the editor	Pass in both	Same suites
Tabs, navigation, back/forward, address sync, 10 tabs	Pass in both	Same suites
Fullscreen video, crashed tab with reload, suspended tabs, popups as tabs keeping window.opener	Pass in both	Same suites
Page web views with no IPC handler, no injected script, no custom user agent; crash dumps local; SmartScreen on	Pass in both	Same suites; the Tauri control built page tabs straight from wry

What didn't, or wasn't tested

• Idle CPU: GPUI 3.75% of a core (fail); the Tauri control 0.54% (just over the bar). Condition 1.
• Acrylic blur: every call succeeds but no blur of the page is visible. Blur is optional (D7); the palette stays near-opaque.
• Not tested by the owner: Hulu; Google sign-in with a password and phone prompt; still signed in after a restart; the 13 other daily sites. The owner expects them to work because the engine is the same; D6 counted them, so they are tested in plan 1's Phase 2. Condition 2.
• Start to first page was measured at different moments in the two spikes (page title shown for GPUI, page starting to load for Tauri), so it is re-measured like for like with the idle-CPU patch.
• Smart App Control blocks unsigned debug builds on this machine; release builds run. Development uses a release-style profile (D38).

D4's rule: GPUI against Tauri + Svelte

D4: switch to Tauri if GPUI saves less than about 50 MB and 150 ms, or if the overlays work only in Tauri.

Measure (medians)	GPUI	Tauri + Svelte	GPUI's lead
Chrome on screen	0.52 s	1.44 s	about 0.9 s
Memory, no tabs	27 MB (chrome alone)	175 MB	about 148 MB
Memory, one tab	273 MB	335 MB	about 62 MB
Palette opens	37 ms	84–118 ms	about 50–80 ms
Idle CPU, 8 tabs suspended	3.75% (fail)	0.54%	Tauri leads
Overlays	Work	Work	Neither

GPUI clears both thresholds, and the overlays work in both, so D4 keeps GPUI. Edge, measured by the GPUI spike on a fresh profile, opened its window in 0.95 s and used 980 MB with one tab and 2,049 MB with ten; a fresh Edge profile is a noisy baseline, so these are indications, not the D6 test, which is repeated on the real app in plan 1's Phase 5.

Corrections to earlier notes

• GPUI draws above the WebView2 child window, not below it, on this machine (GPUI spike). docs/dev/trapstoknow.md said the opposite; it is corrected. Overlays still use separate popup windows, which work.
• Note 08's code reading of the popup re-owning (GWLP_HWNDPARENT) and the focus hand-off (WebView::focus()) is confirmed by running it.

Recommendation

Start plan 1 with GPUI. Put the GPUI idle-loop patch and a like-for-like re-measure into its foundations (Phase 1), and the owner's untested sites into Phase 2's exit criteria. Break plan 1 into issues once the owner has created a repository (D16).