diff --git a/.planning/phases/04-harden-clean-up-optional/04-VALIDATION.md b/.planning/phases/04-harden-clean-up-optional/04-VALIDATION.md
new file mode 100644
index 0000000..e79746f
--- /dev/null
+++ b/.planning/phases/04-harden-clean-up-optional/04-VALIDATION.md
@@ -0,0 +1,96 @@
+---
+phase: 04
+slug: harden-clean-up-optional
+status: draft
+nyquist_compliant: false
+wave_0_complete: false
+created: 2026-05-21
+---
+
+# Phase 04 — Validation Strategy
+
+> Per-phase validation contract for feedback sampling during execution.
+
+**Phase 4 character:** Final v1 hardening phase. Mix of bug fixes + flake stabilization + audit P1 polish + visual polish + build hygiene + closure aggregator. Wave 0 RED test scaffolds for new audit-P1 fixes; existing harness extended for new A33+ assertions; spike-first for SW state persistence (Plan 04-03 per RESEARCH finding 2).
+
+---
+
+## Test Infrastructure
+
+| Property | Value |
+|----------|-------|
+| **Framework** | vitest 4.x (unit) + custom Puppeteer harness (UAT — `npm run test:uat`) |
+| **Config file** | `vitest.config.ts` + `tests/uat/harness.test.ts` (orchestrator) |
+| **Quick run command** | `npm test -- --run tests/<focused-file>.test.ts` |
+| **Full suite command** | `npm test -- --run` (vitest) + `HEADLESS=1 SKIP_PROD_REBUILD=1 npm run test:uat` (UAT harness) |
+| **Estimated runtime** | ~50s (vitest 171→~190 tests post Wave 0) + ~95-300s (UAT harness 33→~37 assertions; 5-min idle test adds ~300s to single run if not gated) ≈ 2.5-6 min full sweep |
+
+---
+
+## Sampling Rate
+
+- **After every task commit:** Run focused test command (vitest single-file OR `npm run test:uat -- --grep A<NN>` for harness)
+- **After every plan wave:** Run full vitest + full UAT harness — both MUST be GREEN
+- **Before `/gsd-verify-work 4`:** Full suite GREEN + pre-checkpoint bundle gates 6/6 PASS (per saved memory `feedback-pre-checkpoint-bundle-gates.md`)
+- **5-min idle test (Plan 04-03):** dedicated `npm run test:uat:long` lane with 6-min timeout; NOT in default sample
+- **Max feedback latency:** ~2.5 min (default full sweep); ~10s (focused vitest); ~25s (focused UAT assertion)
+
+---
+
+## Per-Task Verification Map
+
+| Task ID | Plan | Wave | Requirement | Threat Ref | Secure Behavior | Test Type | Automated Command | File Exists | Status |
+|---------|------|------|-------------|------------|-----------------|-----------|-------------------|-------------|--------|
+| _Filled by planner during Plans 04-01..08 creation_ | _via planner_ | _via planner_ | _via planner_ | _via planner_ | _via planner_ | _via planner_ | _via planner_ | _via planner_ | ⬜ pending |
+
+*Status: ⬜ pending · ✅ green · ❌ red · ⚠️ flaky*
+
+**Planner instructions:** Populate one row per task. Per RESEARCH finding 5 (6 new Wave-0 test files anticipated), expect ~6 unit-test rows + ~4 harness-A33+ rows + ~4 bundle-gate rows + ~3 docs rows. Format per Phase 3 03-VALIDATION precedent.
+
+---
+
+## Wave 0 Requirements
+
+Per RESEARCH finding 5, 6 new test files anticipated as Wave 0 RED scaffolds for audit-P1 / ROADMAP-SC items:
+
+- ⬜ `tests/build/no-new-function-in-sw-chunk.test.ts` — grep `dist/assets/index.ts-*.js` for `new Function(` count = 0 after setimmediate polyfill replacement
+- ⬜ `tests/build/dead-code-grep.test.ts` — ROADMAP SC #4: rg `permissions\.request` + duplicate offscreen inline string in src/ = 0 hits
+- ⬜ `tests/content/fetch-interception.test.ts` — audit P1 #11: Request-arg vs string-arg case for `args[0]`
+- ⬜ `tests/content/navigation-tracking.test.ts` — audit P1 #14: module-level previousUrl tracking
+- ⬜ `tests/content/rrweb-timestamps.test.ts` — audit P1 #15: epoch vs page-load-relative
+- ⬜ `tests/welcome/inline-svg.test.ts` — UI-SPEC: `?raw` import + DOMParser inline-SVG + currentColor resolves via CSS
+
+Existing infrastructure already in place (inherited from Phases 1-3):
+- ✅ `tests/uat/extension-page-harness.ts` — page-side assertA* host (extend with assertA33+)
+- ✅ `tests/uat/lib/harness-page-driver.ts` — host-side driveA* host (extend with driveA33+ + improved driveA29)
+- ✅ `tests/uat/harness.test.ts` — orchestrator (extend)
+- ✅ `tests/uat/lib/assertions.ts` — shared helpers
+- ✅ `tests/uat/lib/zip.ts` — jszip-based archive parsing
+- ✅ `tests/uat/lib/launch.ts` — Puppeteer Chrome launch + extension load
+- ✅ `tests/background/no-test-hooks-in-prod-bundle.test.ts` — FORBIDDEN_HOOK_STRINGS lockstep
+
+`wave_0_complete: false` until 6 new test files committed per planner Wave 0.
+
+---
+
+## Manual-Only Verifications
+
+| Behavior | Requirement | Why Manual | Test Instructions |
+|----------|-------------|------------|-------------------|
+| Dark-surface logo visual contrast judgment (WCAG ratio not specified in UI-SPEC) | UI-SPEC acceptance criterion #2 + operator empirical | Aesthetic judgment of deep-indigo stroke on madder-orange wrapper in dark-OS context — UI-SPEC defers to operator empirical checkpoint | Load extension in OS dark mode; open welcome page; verify mark wrapper still madder, stroke is deep-indigo, contrast is acceptable. Plan 04-06 operator empirical UAT cycle covers this. |
+| Cursor visibility verification | ROADMAP cursor visibility item per D-P4-03 | Already shipped at src/offscreen/recorder.ts:285 per Plan 01-09 (RESEARCH finding 4) — Plan 04-06 downgrades to verification + stale-note correction; minor manual check of one captured frame to confirm cursor present | Load extension, start recording, click anywhere, SAVE archive; open `video/last_30sec.webm` in Chrome; verify cursor visible in playback. |
+
+*All other Phase 4 behaviors have automated verification via vitest (Wave 0 unit tests) + UAT harness (A33+) + bundle gates.*
+
+---
+
+## Validation Sign-Off
+
+- [ ] All tasks have `<automated>` verify or Wave 0 dependencies — pending planner fill-in
+- [ ] Sampling continuity: no 3 consecutive tasks without automated verify — verify after planner fills table
+- [ ] Wave 0 covers all MISSING references — 6 new test files anticipated; planner confirms count
+- [ ] No watch-mode flags — verify in planner output (focused commands use `--run`)
+- [ ] Feedback latency < ~2.5 min default (5-min idle test on dedicated lane) — confirmed by RESEARCH
+- [ ] `nyquist_compliant: true` set in frontmatter — pending sign-off after planner completes
+
+**Approval:** pending (planner fills per-task map; checker validates)