fix: reduce macOS popup compositing cost

.project-docs/30-worklog/tasks/20260818-macos-popup-lag-7c4e.md

@@ -0,0 +1,76 @@
+# Task: Diagnose macOS popup lag
+
+## Identity
+
+- Task ID: 20260818-macos-popup-lag-7c4e
+- Mode: Feature
+- Branch: codex/20260818-macos-popup-lag-7c4e-macos-popup-lag
+- Worktree: D:\Datas\OthersProjects\makelore-macos-popup-lag
+- Base commit: 13bdc0c9bab656aa9ff6c990bf6359f7d3210750
+- Owner: developer
+- Status: Completed
+
+## Scope
+
+- Diagnose the reported macOS-wide popup lag in the Electron client, with emphasis on the shared Renderer Dialog/Sheet and transient popup compositing path.
+- Build a repeatable Electron E2E frame-budget/computed-style probe, minimize it to the shared popup seam, implement the smallest platform-safe rendering fix, and add a regression check.
+- Do not change firmware, cloud contracts, authentication, provider/runtime behavior, or the unowned primary worktree.
+
+## Intent And Constraints
+
+- Preserve the single Makelore light visual system and existing popup behavior, focus management, accessibility, and Main-owned boundaries.
+- The available environment is Windows, so macOS behavior must be simulated through the exposed Electron platform marker; real signed macOS physical/package validation remains unverified.
+- Keep the existing global hardware-acceleration policy unchanged unless a reproducible seam proves it is the cause; prefer removing only the expensive transient popup filter path.
+- The shared popup regression owns only its test file and the platform/compositing files assigned during implementation; unrelated peer work remains isolated.
+
+## Outcome
+
+- Diagnosis isolated the shared Renderer popup compositing seam: the real Settings
+  Dialog rendered `backdrop-filter: blur(3px)` on its overlay and
+  `blur(18px) saturate(1.35)` on its glass surface. A repeatable probe observed a
+  62.5 ms maximum frame gap and one frame over 32 ms before the filter-only
+  intervention; disabling those filters reduced the observed gap to about 31.4
+  ms with no frame over 32 ms in this environment.
+- Implemented a platform-scoped fix. Renderer startup writes
+  `html[data-platform]` from the existing Electron platform value. On Darwin,
+  transient popup roots and their `glass-surface` descendants disable both
+  standard and WebKit backdrop filters and use an opaque background. Win32 and
+  other platforms retain the existing glass material.
+- Main/IPC, hardware-acceleration policy, native vibrancy, and unrelated popup
+  behavior were left unchanged. The fix is limited to the assigned Renderer
+  platform/style seam and its regression test.
+
+## Verification
+
+- `corepack pnpm@10.33.4 run build:vite` — passed (existing chunk-size and
+  dynamic-import warnings only).
+- `corepack pnpm@10.33.4 run typecheck` — passed.
+- `corepack pnpm@10.33.4 exec eslint src/main.tsx tests/e2e/popup-performance.spec.ts`
+  — passed.
+- `corepack pnpm@10.33.4 exec vitest run tests/unit/modal-layering.test.tsx` —
+  passed (2 tests).
+- `corepack pnpm@10.33.4 test` — passed (176 files, 2,109 tests).
+- `corepack pnpm@10.33.4 exec node ./node_modules/@playwright/test/cli.js test
+  tests/e2e/popup-performance.spec.ts --reporter=line` — passed (1 test).
+- `git diff --check` — passed.
+- The E2E regression confirms the actual Electron platform marker, Darwin
+  computed filters are `none`, and non-Darwin computed filters remain enabled.
+  It intentionally does not enforce a hard RAF budget because shared CI
+  scheduling is not a stable compositor benchmark.
+- Real signed macOS package/physical GPU-vibrancy A/B validation was not
+  available on this Windows host and remains required before claiming measured
+  macOS frame-time improvement.
+
+## Follow-ups
+
+- Run the popup E2E and a short Performance trace against a signed macOS x64 and
+  arm64 package, including a Sheet and nested Dialog path.
+- If popups remain slow with computed filters disabled, A/B the existing
+  `app.disableHardwareAcceleration()` and Darwin transparent/vibrancy window
+  material on real hardware before changing either global policy.
+
+## Promotion Candidates
+
+- No canonical architecture or product decision is proposed. Keep the
+  platform-scoped rendering rule and the evidence topic task-scoped until real
+  macOS measurements confirm the compositor result.