diff --git a/docs/deferred-improvements.md b/docs/deferred-improvements.md
new file mode 100644
index 0000000..b73b1d9
--- /dev/null
+++ b/docs/deferred-improvements.md
@@ -0,0 +1,96 @@
+# Deferred Improvements
+
+This document tracks review follow-ups that were investigated but intentionally
+deferred, along with the reasoning and the minimum work required to pick them up
+later. Complements the merged PRs from the 2026-04-17 full-project review
+(PRs #38–#49).
+
+## Next.js 16 Cache Components adoption
+
+**Status**: attempted, reverted. Tracked for a dedicated future PR.
+
+**Goal**: replace `export const revalidate = 3600` on public pages with
+`'use cache' + cacheLife + cacheTag`, enabling Partial Prerendering (PPR) and
+targeted tag-based invalidation from admin mutations.
+
+**Attempt**: branch `feat/next16-cache-components` set `cacheComponents: true`
+in `next.config.ts` and migrated:
+
+- `app/sitemap.ts` — fully cached (no user state).
+- `app/(public)/lists/page.tsx` — extracted data fetch into a cached helper.
+- `app/(public)/lists/[slug]/page.tsx` — `getListBySlug` + `getListProblems`
+  as cached helpers; per-user progress stays dynamic in the page body.
+- `app/(public)/problems/page.tsx` — catalog query cached; solved-IDs fetch
+  stays dynamic.
+- `app/(public)/problems/[slug]/page.tsx` — `getProblemBySlug` cached with
+  `cacheTag('problems', 'problem:<slug>')`.
+- `app/page.tsx` — made dynamic (auth-aware CTA).
+- `app/api/{cron/push,health}/route.ts` — removed `dynamic = 'force-dynamic'`
+  (incompatible with cacheComponents; defaults to dynamic).
+- `lib/actions/admin.ts` — `deleteProblem` wired to
+  `revalidateTag('problems', 'max')` + `revalidateTag('lists', 'max')`.
+
+**Why reverted**: `next build` with `cacheComponents: true` enforces that
+**every uncached data access must sit inside a `<Suspense>` boundary during
+static generation**. Admin pages (each one fetches user-specific or live data)
+and several public pages (per-user solved checkmarks, progress bar) fail the
+build with errors like:
+
+```
+Route "/admin/channels": Uncached data was accessed outside of <Suspense>.
+Route "/lists/[slug]": Uncached data was accessed outside of <Suspense>.
+```
+
+Making this work cleanly requires Suspense-wrapping every dynamic region on
+every page that currently blends static and user-specific data — roughly
+20+ files in `app/(admin)/**` plus the public routes that show solved status.
+That exceeds what the original review item scoped and is a meaningful UI
+refactor (streaming fallbacks, layout shifts, etc.) rather than a drop-in
+migration.
+
+**Resume checklist** when someone picks this up:
+
+1. Re-apply the changes from the `feat/next16-cache-components` attempt as a
+   starting point (see git history for the branch or this document's summary).
+2. Audit each page under `app/(admin)/**` and `app/(public)/**`. For each
+   Server Component that mixes cached and dynamic data, wrap the dynamic
+   section in `<Suspense fallback={...}>` and move the dynamic fetch into the
+   child component.
+3. Design skeleton fallbacks consistent with the current layout to avoid
+   perceived layout shift.
+4. Verify `next build` with `cacheComponents: true` succeeds for every route.
+5. Extend admin mutations to call `revalidateTag('problems'|'lists', 'max')`
+   alongside the existing `revalidatePath` so updates propagate instantly.
+6. Remove every remaining `export const revalidate` from app routes.
+
+**Why this is worth revisiting**: when complete, content edits in the admin UI
+show up on public pages within seconds instead of up to an hour, and static
+shells render immediately while dynamic regions stream in — real UX win for
+SEO-critical problem and list pages.
+
+## Branded channel-identifier types
+
+**Status**: deferred in PR #42. Low-value-per-effort.
+
+**Goal**: prevent swapping a Telegram chat-id for a LINE user-id or an email
+address at compile time by branding the `string` identifier type per channel.
+
+**Why deferred**: would require threading branded types through every channel
+call site across shared + web. With only three channel types and the bundled
+push pipeline (PR #40) already failing loudly on mismatched registry entries,
+the compile-time gain is narrow.
+
+## Node 22 -> 24 LTS, pnpm 9 -> 10
+
+**Status**: deferred in PR #43. Needs manual install + regression verification.
+
+**Goal**: stay on current LTS (Vercel's default is Node 24 as of 2026-03).
+
+**Resume checklist**:
+
+1. Update `.nvmrc` and `.github/workflows/ci.yml` to `24`.
+2. Bump `packageManager` in root `package.json` to `pnpm@10.x` and re-run
+   `pnpm install` to regenerate the lockfile.
+3. Full regression: `pnpm build`, `pnpm test`,
+   `pnpm --filter @caffecode/web exec playwright test`.
+4. Watch for pnpm 10 breaking changes (strict peer deps, different hoisting).