docs: record startup-time research; plan re-check after Fyne 2.7.x

Add docs/PERFORMANCE.md documenting the post-Phase-4 startup slowdown:
the ~290ms increase is entirely Fyne 2.5.3->2.6.3's w.Show() cost, not
the Service/domain/UI refactor. Includes the GOSENTRY_TIMING method and
warm-run results table so it can be reproduced.

Wire the startup re-measurement into the existing Fyne 2.6.3->2.7.x
upgrade task in ROADMAP.md so 2.7's effect on w.Show() gets checked.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/PERFORMANCE.md

@@ -0,0 +1,57 @@
+# Performance Notes
+
+Measured performance findings for GoSentry. Each entry records the method so the
+numbers can be reproduced and re-checked after relevant changes.
+
+## Startup Time
+
+### Finding (2026-06-22)
+
+After the Phase 4 refactor, cold startup time (the History "Window shown in …"
+metric) increased by **~290 ms**. The increase is caused entirely by the
+**Fyne v2.5.3 → v2.6.3 upgrade** (task T4.1), **not** by the Service / domain /
+UI restructuring.
+
+### Method
+
+Env-gated phase timers (`GOSENTRY_TIMING`) were added across the startup path
+(`Run` in `src/ui/run.go` and `newMainView` in `src/ui/mainwindow.go`) and the
+equivalent points in the pre-refactor entry point (`src/gui/app.go` at commit
+`c5e0ef9`, the last commit before T4.1). Both were built with the CGO / MSYS2
+UCRT64 toolchain and run 5× each; the first run of each is a cold-disk outlier
+and is excluded. The timed span (`started` → `w.Show()`) is identical in both
+builds, so the comparison is fair.
+
+### Results (warm-run averages)
+
+| Phase (cumulative from start) | Old (Fyne 2.5.3) | New (Fyne 2.6.3) | Δ |
+|-------------------------------|------------------|------------------|--------|
+| after single-instance check   | ~0.5 ms          | ~0.6 ms          | —      |
+| after Fyne app + window + tray | ~277 ms         | ~285 ms          | +8 ms  |
+| `app.Open()` done             | +3 ms            | +3 ms            | 0      |
+| views built + `svc.Start()`   | +42 ms           | +43 ms           | ~0     |
+| after `SetContent`            | ~348 ms          | ~353 ms          | +5 ms  |
+| **after `w.Show()` (TOTAL)**  | **~348 ms**      | **~644 ms**      | **+~290 ms** |
+
+### Interpretation
+
+- Everything up to and including `SetContent` costs the same in both versions
+  (~350 ms). The refactor-specific code — `app.Open()` (~3 ms) and the new
+  `app.Service` plus view construction (~42 ms) — is unchanged, so the
+  restructuring added no measurable startup cost.
+- The entire regression lands in **`w.Show()`**: ~0 ms under Fyne 2.5.3,
+  ~290 ms under 2.6.3. Fyne 2.6 reworked main-thread marshaling (the change that
+  introduced `fyne.Do`) and front-loads first-window GL/driver realization into
+  the `Show()` call.
+- The cost is a fixed, one-time Fyne expense, not a leak in GoSentry code, and
+  the upgrade cannot be reverted because `fyne.Do` requires Fyne ≥ 2.6.
+- The tray / autostart path (`--start-in-tray`) skips `w.Show()` until the user
+  opens the window, so it is unaffected.
+
+### Next check
+
+Re-measure with the same method after the planned **Fyne 2.6.3 → 2.7.x upgrade**
+(see [ROADMAP.md](ROADMAP.md) → Tray Interaction). The goal is to learn whether
+2.7's driver/threading changes recover any of the ~290 ms `w.Show()` cost or
+hold it steady. Reuse the `GOSENTRY_TIMING` instrumentation pattern above; do not
+commit the temporary timers.