diff --git a/.planning/phases/04-harden-clean-up-optional/.plan-phase-preferences.md b/.planning/phases/04-harden-clean-up-optional/.plan-phase-preferences.md
new file mode 100644
index 0000000..e281b50
--- /dev/null
+++ b/.planning/phases/04-harden-clean-up-optional/.plan-phase-preferences.md
@@ -0,0 +1,75 @@
+---
+context: phase-4-plan-preferences
+captured: 2026-05-20
+expires_when: /gsd-plan-phase 4 completes
+---
+
+# Phase 4 plan-phase preferences (captured pre-UI-SPEC exit)
+
+User invoked `/gsd-plan-phase 4` 2026-05-20 and answered both initial gate
+questions before workflow exited at UI-SPEC gate. These preferences should be
+re-honored on next `/gsd-plan-phase 4` invocation (after `/gsd-ui-phase 4`
+completes):
+
+## UI Design Contract gate
+
+**Decision:** Generate UI-SPEC.md first → run `/gsd-ui-phase 4` before
+re-running plan-phase.
+
+**Caveat to revisit:** Unlike Phase 3 (false-positive null-spec), Phase 4 has
+genuine UI work — specifically the dark-surface logo contrast fix (Plan 01-10
+obs 2026-05-20). The cursor visibility constraint is purely behavioral
+(getDisplayMedia `video: {cursor: 'always'}` — 1 line in src/offscreen/recorder.ts)
+and does not need designer iteration. /gsd-ui-phase 4 should produce a
+thin-but-real UI-SPEC covering ONLY:
+- dark-surface logo contrast strategy (media query + alt-color SVG vs wrapper
+  background vs neutral palette swap — designer picks)
+- (informational only) cursor visibility constraint — list as inherited
+  behavioral change, not a design surface
+
+If /gsd-ui-phase 4 surfaces the "is this a UI phase?" question, answer
+"yes — but narrow scope (dark-logo design only; cursor is behavioral)".
+
+## Research gate
+
+**Decision:** Research first (light) → spawn gsd-phase-researcher before
+planning.
+
+**Scope for researcher (~10-20 min):**
+1. **setimmediate polyfill replacement strategy** — current source is
+   `vite-plugin-node-polyfills` (per Plan 01-12 disclosure). Options:
+   (a) inline manual polyfill `globalThis.setImmediate ||= (fn, ...args) =>
+       queueMicrotask(() => fn(...args))` — bypasses plugin entirely
+   (b) configure vite-plugin-node-polyfills to disable setimmediate inclusion
+   (c) document acceptance with explicit CSP-allow rationale (already-done
+       Phase 1 acceptance log path)
+   Recommend best path for v1 close.
+
+2. **SW state persistence 5-min idle test patterns under MV3** —
+   ROADMAP SC #1 requires "extension idle for >5 minutes, then exporting,
+   the archive still contains a non-empty video buffer". Best practices for:
+   - Forcing SW unload in Puppeteer (`chrome.runtime.reload()`? or just wait
+     for natural idle eviction at 30s?)
+   - Verifying buffer survives across reload (IndexedDB? `chrome.storage.local`?
+     existing keepalivePort architecture from Plan 01-05?)
+   - 5min Puppeteer timeout considerations (CI lane vs default)
+
+3. **chrome.scripting.executeScript world:'ISOLATED' best practices** for
+   the A29 cs-injection-world fix (per Plan 03-02 introduced pattern).
+   Confirm: re-targeting the rrweb probe to a fresh probe tab (instead of
+   reading iana.org leftover) is the canonical fix; identify any pitfalls.
+
+Researcher should NOT investigate:
+- rrweb v2 upgrade implementation path (deferred to v1.1/v2 per D-P3-03 +
+  D-P4-01)
+- Full rrweb maskInputFn or data-sensitive masking (out of scope v1 per
+  charter shift)
+- Programmatic SW-RAM measurement via chrome.devtools Memory API
+  (deferred to v1.1/v2 per D-P3-04 + D-P4-01)
+- Any audit P1 #11/#14/#15 deep-dive (already-understood fixes; planner
+  works directly from CONTEXT.md `<specifics>`)
+
+## Auto-cleanup
+
+Delete this file when /gsd-plan-phase 4 successfully completes planner
+spawn (after honoring these preferences).