@rangojs/router 0.0.0-experimental.f3c21dba → 0.0.0-experimental.f4a8e8de

package/skills/ppr/SKILL.md

@@ -0,0 +1,904 @@
+---
+name: ppr
+description: PPR shell caching — opt a page route in with the `ppr` path option; the router serves the cached HTML shell instantly and resumes the live holes. Use when a page should render instantly from a cached shell while specific parts stay live, or asking about partial prerendering in Rango.
+argument-hint: "[setup]"
+---
+
+# PPR Shell Caching
+
+Caches the rendered HTML **shell** of a page route (React `prerender` prelude
+bytes plus `postponed` state) and, on a later request, flushes those bytes after
+route classification and the complete middleware chain, but before downstream
+tail rendering. It then resumes fizz for just the live holes. The browser sees
+one ordinary streamed document; loaders stay fresh on every request. This is
+the second render axis — the default axis-1 path is untouched, and every
+ineligible request falls open to it.
+
+Compare `/document-cache`, which freezes the WHOLE response including loader
+output. Shell caching is for pages that mix a stable shell with live data: the
+shell is shared per host+URL, the holes are per request.
+
+This is in-function PPR on every deployment. The worker/function serves the
+prelude; it is not a CDN static file. See `/deployment-caching` before combining
+PPR with HTTP shared-cache headers.
+
+## Not this skill if…
+
+- You want the WHOLE response frozen, loader output included — see
+  `/document-cache`.
+- You want build-time Flight segment payloads from `Static()`/`Prerender()` —
+  see `/prerender`. A `Prerender` page may also declare `ppr`; then producer B
+  can bake the HTML shell at build time while loaders stay live.
+- You are unsure which cache layer you need — start at `/cache-guide`.
+
+## Setup: one path option, no PPR middleware to mount
+
+PPR is a DOCUMENT-level property declared on the page route via the `ppr` path
+option. Serving is **integral to the router** — there is nothing to mount. The
+only prerequisite is an app-level `createRouter({ cache })` store that
+implements the shell family (`getShell`/`putShell`): `MemorySegmentCacheStore`
+(dev/tests), `CFCacheStore` (Cache API L1 + KV L2), or `VercelCacheStore`
+(runtime cache). A ppr route on a store without the family stays on axis 1 with
+a once-per-key warning.
+
+```typescript
+import { createRouter, urls } from "@rangojs/router";
+import { CFCacheStore } from "@rangojs/router/cache";
+
+export const urlpatterns = urls(({ path, layout, loader, loading }) => [
+  layout(ProductShell, () => [
+    path(
+      "/products/:id",
+      PricePage,
+      // `ppr` is the whole opt-in AND the policy. `ppr: true` uses the default
+      // ttl (300s); an object sets ttl/swr/tags (PartialPrerenderProps).
+      { name: "product", ppr: { ttl: 600, swr: 120 } },
+      () => [
+        loader(LivePriceLoader),
+        loading(<PriceSkeleton />), // structural hole: the loader subtree stays live
+      ],
+    ),
+  ]),
+]);
+
+const router = createRouter<AppBindings>({
+  document: Document,
+  urls: urlpatterns,
+  cache: (env, ctx) => ({
+    store: new CFCacheStore({ kv: env.CACHE_KV, ctx: ctx! }),
+  }),
+});
+export default router;
+```
+
+That is ONE of two hole mechanisms — the loader one. Do not conclude PPR
+requires a loader:
+
+### The same opt-in with NO loader and NO loading() — promise holes
+
+A route (or its layouts) whose live regions are pending promises under
+`<Suspense>` PPRs with no `loader()` and no `loading()` anywhere. Hand the
+un-awaited promise over as a prop; the consumer suspends under its OWN
+boundary; the boundary postpones at capture and becomes the hole:
+
+```typescript
+// Handler: kick off the fetch, do NOT await it.
+function ProductPage(ctx: HandlerContext) {
+  const reviews = fetchReviews(ctx.params.id); // Promise<Review[]> — pending
+  return (
+    <main>
+      <h1>Product {ctx.params.id}</h1> {/* shell — baked into the prelude */}
+      <ReviewsSection promise={reviews} /> {/* hole — resumes per request */}
+    </main>
+  );
+}
+```
+
+```tsx
+// ReviewsSection.tsx — the consumer owns its Suspense boundary.
+"use client";
+
+import { Suspense, use } from "react";
+
+function Inner({ promise }: { promise: Promise<Review[]> }) {
+  return <ReviewList reviews={use(promise)} />;
+}
+export function ReviewsSection({ promise }: { promise: Promise<Review[]> }) {
+  return (
+    <Suspense fallback={<ReviewsSkeleton />}>
+      <Inner promise={promise} />
+    </Suspense>
+  );
+}
+```
+
+```typescript
+path("/products/:id", ProductPage, { name: "product", ppr: true });
+// No use() list at all: no loader, no loading, still a shell + live hole.
+```
+
+At capture the pending fetch cannot win the task-quantized quiet window, so
+the boundary postpones — fallback in the frozen prelude, value resumed fresh
+on every HIT. This is the PHYSICS class from the hole doctrine below, and it
+is exactly how an existing Suspense-shaped tree (e.g. migrated from Next.js
+PPR) works with zero restructuring. The e2e proof is
+`e2e/test-app/src/components/ShellPhysicsValue.tsx` — a promise hole living in
+a LAYOUT with no loader registration at all.
+
+A route WITHOUT the `ppr` option is pure axis 1: no store read, no capture, no
+logs, zero cost. `ppr` is per page route — declaring it on a layout is not
+supported (subtree inheritance is a possible follow-up).
+
+## Where PPR sits: the cache onion
+
+Rango's caches layer like an onion — each ring stores a progressively more
+"cooked" representation of the same page. From innermost (raw values) to
+outermost (final bytes):
+
+| Ring                    | Primitive                                | What is stored                                     | What stays live on a hit               |
+| ----------------------- | ---------------------------------------- | -------------------------------------------------- | -------------------------------------- |
+| 1. Function values      | `"use cache"`                            | a function's return value                          | everything around the call             |
+| 2. Loader values        | `loader(Fn, () => [cache({...})])`       | one loader's result (opt-in; loaders default live) | all other loaders, handlers, rendering |
+| 3. Segments (Flight)    | `cache()` route / build-time `prerender` | serialized rendered segments + replayed handles    | loaders, HTML render                   |
+| 4. **HTML shell (PPR)** | `ppr` path option                        | rendered prelude bytes + React postponed state     | the holes, hydration payload           |
+| 5. Whole response       | `/document-cache`                        | final response bytes, headers included             | nothing — all-or-nothing               |
+
+PPR is ORTHOGONAL to `cache()` (ring 3): a ppr route may be uncached (its
+handlers run fresh on every serve and during capture), fully `cache()`d (its
+segments replay), or mixed. One useful cache() property to know: the segment
+codec **deep-settles promises at the ring-3 write**, so nothing inside a
+`cache()` boundary can stay live — that is a cache() fact, not a ppr one.
+
+Invalidation crosses rings: `updateTag()`/`revalidateTag()` reach segment,
+shell, loader, and item entries in the same store, and shell entries
+additionally self-invalidate on `React.version` change.
+
+## The serve pipeline: commit after ALL middleware
+
+On a document GET to a ppr route the router runs:
+
+1. **match** — route identified, `ppr` config read from the matched route;
+2. **the WHOLE middleware chain** — the global `router.use()` chain AND route
+   DSL `middleware()`; both are guards, and the COMMIT POINT is after all of
+   them: any rejection/redirect/401 wins before a single shell byte, on MISS
+   and on a warmed HIT alike;
+3. **shell lookup** — `getShell(key)` on the app store (key =
+   host+pathname+sorted search);
+4. **HIT** — the composed response is committed immediately: the stored prelude
+   bytes flush first, while segment resolution, the fresh Flight render (the
+   full hydration payload — there is no Flight-side resume), and the fizz
+   `resume` of just the holes run BEHIND them inside the response stream;
+5. **MISS** — plain axis-1 serve, tagged `x-rango-shell: MISS`, plus a
+   background capture (stampede-guarded, retry-in-place, exponential backoff).
+
+`x-rango-shell: HIT | MISS` is the observability header. Because the commit
+point is after the chain, an unauthorized request NEVER sees shell bytes — put
+auth middleware anywhere (global or route DSL) and it guards PPR for free.
+
+### Soft navigation caches and reuses the handler layer
+
+Ordinary partial RSC navigations to a `ppr` URL use the same handler-layer cache
+contract even when no document request has captured an HTML shell yet. When a
+shell snapshot exists, the server replays its canonical document segment record.
+On a cold partial request, normal matching renders the response and schedules a
+background navigation-only shell capture; later navigations and prefetches
+replay its eligible snapshot. In both cases `matchPartial()`:
+
+- preserves client-owned shared layouts by segment id;
+- returns only new or revalidating destination segments;
+- runs DSL loaders fresh with their normal `loading()` streaming behavior;
+- keeps the existing prefetch key, source scope, and in-flight lock unchanged.
+
+This is deliberately invisible to the browser: the response is the same
+`RscPayload` shape as any other partial navigation. Captured item/response values
+and loader-container pins are NOT replayed on this path, so loader reads stay
+live.
+
+A route's own `cache()` scope — including one inherited from an ancestor, the
+common app-wide storefront shape — COMPOSES with replay instead of disabling
+it. The explicit tier stays authoritative: its lookup runs first with its
+normal store, key, TTL, SWR, tags, and condition, and when it supplies the
+match the response reports `BYPASS; reason=explicit-cache-hit` (never a false
+replay `HIT`). Only when the explicit tier misses does the shell snapshot's
+canonical doc segment record supply the match and report `HIT`. To make that
+possible, a capture of such a route records the doc segment record into the
+shell snapshot IN ADDITION to the scope's normal store write; the record rides
+only inside the shell entry, never the real store. Two opt-outs stay absolute:
+`cache(false)` and a `condition()` returning false mean "do not serve this
+request's segments from any cache" — no doc record is captured for them and
+the seeded fallback never rescues a refused read. `cache(false)` is static, so
+replay reports `BYPASS; reason=cache-disabled` before performing a single
+shell read; a `condition()` refusal is request-time state, decided at the
+lookup itself and reported as the same `cache-disabled` post-match (the gate
+must not pre-decide a predicate that could flap between the two evaluations).
+An errored explicit read — a throwing `key()`, or a built-in store's swallowed
+backend failure (`CACHE_READ_ERROR`) — also never falls back: it renders
+uncached, exactly as without ppr.
+
+`transition({ when })` is evaluated from the
+matched manifest before route handlers on every PPR match, so it can vary by
+URL/params/action or middleware context without disabling replay; handler-set
+context is unavailable by design. Intercepts, handler-live holes, an active
+nonce, and an absent/corrupt segment snapshot fall open to the ordinary partial
+path when encountered by the shell capture. A transition already replayed from
+an explicit cache tier remains frozen by that tier's normal semantics.
+
+Two more decisions are made before any shell-store read, so probes and
+prerendered routes never spend passive `getShell` I/O:
+
+- A partial request carrying neither `X-RSC-Router-Client-Path` nor `Referer`
+  (a curl probe, a synthetic monitor) can never produce a partial match; it
+  reports `BYPASS; reason=no-navigation-context`. Alert on replay hit-rate
+  with this in mind — such probes are not cache misses.
+- A `Prerender()` route's partial is served from the build-time prerender
+  store inside matching (a better-than-HIT outcome); it reports
+  `BYPASS; reason=prerender-store`. Its captures never record a doc segment
+  record (the prerender store short-circuits the cache write), so replay
+  seeding would be impossible anyway.
+
+Fresh and stale-within-SWR runtime shells replay. The stale read is passive: it
+uses `getShell(key, { claimRevalidation: false })`, does not claim SWR ownership,
+and cannot recapture HTML. A later document request owns the background
+recapture; hard-expired entries schedule a navigation-only capture. Production
+may also use a fresh local build manifest. Development does not probe
+`/__rsc_shell`; it schedules the same local background capture instead. Custom `SegmentCacheStore`
+implementations must set `supportsPassiveShellReads: true` and honor the
+non-claiming read option.
+
+The first cold partial request reports `BYPASS; reason=no-entry` because no
+artifact could supply that response; its successful render schedules
+`scheduleShellCapture` with a navigation-only marker. The next request reports
+`HIT` when the capture produced an eligible snapshot. Navigation snapshots use
+a separate shell key, so they cannot overwrite or be served as a document shell.
+The capture runs with the stripped target document URL and loads SSR support in
+the bounded background queue, outside the triggering partial response's latency.
+An `allReady` decision still declines capture.
+
+Partial responses expose the actual decision as `x-rango-ppr-replay`:
+`HIT; freshness=fresh|stale` or `BYPASS; reason=<bounded-token>`. With
+performance metrics enabled, the same decision appears as
+`ppr-navigation-replay` in `Server-Timing`. `HIT` means matching consumed the
+seeded segment record, not merely that a snapshot existed. Fragment-capable
+clients decode its stored ReactNode fields after the response arrives; if that
+decode fails, the client retries once through the server decode-and-evict path
+and disables further passthrough for that browser document. The retry replaces
+the fragment-capable response-cache slot, and a corrupt seeded snapshot is
+recaptured at the key that supplied it. An explicit `cache()` scope that supplies
+the match cannot produce a
+false HIT — it reports `explicit-cache-hit`. There is still no Flight resume
+API; this is segment replay followed by normal Flight streaming, not reuse of
+the HTML `prelude`/`postponed` bytes. Replayed segments do ride the partial
+payload as verbatim `__rangoFragment` envelopes (#700 fragment splice — the
+stored per-segment Flight strings are string-copied instead of
+deserialize→re-serialize per request); the client expands them at its
+navigation/prefetch decode chokepoints before anything renders, so they are
+consumer-invisible.
+
+The bounded bypass tokens, grouped by when they are decided:
+
+| Token                                                                             | Decided           | Meaning                                                                                                                                                                                                                                                                   |
+| --------------------------------------------------------------------------------- | ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `method`, `dynamic`, `nonce`, `store-unavailable`, `passive-read-unsupported`     | pre-read          | request/route/store ineligible for replay                                                                                                                                                                                                                                 |
+| `no-navigation-context`                                                           | pre-read          | no `X-RSC-Router-Client-Path`/`Referer`; a partial match is impossible                                                                                                                                                                                                    |
+| `prerender-store`                                                                 | pre-read or match | `Prerender()` route served by its baked artifact (pre-read probe of the normal variant; reclassified post-match when the store actually served, either variant)                                                                                                           |
+| `intercept`                                                                       | match             | the navigation resolved to an intercept — replay is never armed for intercepts (they keep their normal cache path); no heal capture                                                                                                                                       |
+| `cache-disabled`                                                                  | pre-read or match | `cache(false)` (pre-read, static) or `condition()` false (decided at the lookup); consumer opt-out is absolute                                                                                                                                                            |
+| `read-error`, `no-entry`, `invalid-version`, `corrupt-entry`, `stale-build-entry` | shell read        | no usable shell entry (`no-entry`/`invalid-version`/`corrupt-entry`/`stale-build-entry` schedule the navigation-only heal capture)                                                                                                                                        |
+| `handler-live-holes`, `transition-when`, `no-segment-snapshot`                    | eligibility       | entry exists but its snapshot cannot replay (no canonical doc record, or handler/transition liveness); on a route with an enabled `cache()` scope, `no-segment-snapshot` heals when the lookup did not refuse (a `condition()` false-at-capture entry becomes replayable) |
+| `explicit-cache-hit`                                                              | match             | the route's own `cache()` tier supplied the match                                                                                                                                                                                                                         |
+| `snapshot-miss`                                                                   | match             | an eligible snapshot was seeded but matching did not consume it                                                                                                                                                                                                           |
+
+### Capture-generation invalidation
+
+If handler or bake-lane code calls `updateTag()` on one of the shell's own tags
+while capture is running, that generation is rejected. Built-in stores report it;
+Rango warns with the shell key and backs capture off instead of rendering the
+same doomed generation on every request. Move a deterministic self-invalidation
+out of render code if you want the shell to persist.
+
+### Opting out per request with `ctx.dynamic()`
+
+Middleware and handlers can call `ctx.dynamic()` to force this request back to
+axis 1. In middleware it runs before the PPR commit point, so the router skips
+shell lookup, HIT serving, and MISS capture for that request. In handlers it is
+too late to prevent a MISS render from already happening, but it still prevents
+the follow-up shell capture.
+
+During `Prerender` + `ppr` build-shell capture, middleware is replayed with
+`ctx.build === true`, `ctx.waitUntil()` inert, and the same `ctx.dynamic()`
+opt-out. Use that for routes where the shell depends on runtime-only auth,
+cookies, or side-effectful SDK calls. A skipped build shell can still be owned
+later by runtime capture when runtime middleware does not call `ctx.dynamic()`.
+
+`ctx.dynamic()` also **re-permits handler header/cookie writes** (issue #735).
+The header doctrine below forbids handler writes on a ppr route because they
+ride MISSes and vanish on HITs — but a `dynamic()` render never HITs, so its
+write is deterministic. Calling `ctx.dynamic()` clears the header latch for the
+rest of the render, so the SAME handler can write its control-flow header
+directly (no middleware relay). Ordering is a contract: call `dynamic()` BEFORE
+the write — a write before it still throws.
+
+```ts
+function catalogPage(ctx) {
+  ctx.dynamic(); // declare live -> refuses capture AND clears the header latch
+  ctx.headers.set("x-rango-sfra-proxy", "catalog");
+  return <Catalog />;
+}
+```
+
+## Verifying it works
+
+The header exists on DOCUMENT responses only. A bare `curl` gets the HTML
+document (Flight is explicit-opt-in via `Accept: text/x-component`), so it
+sees the header directly; only an explicit Flight request shape lacks it:
+
+```
+curl -s -D - -o /dev/null https://app.example.com/products/1 | grep -i x-rango-shell
+```
+
+- Runtime-captured route: first document GET is `MISS`, plus a background
+  capture; a later request becomes a `HIT`.
+- `Prerender + ppr` route: the shell is produced during `vite build`, so the
+  first production document request can already be a `HIT`. In dev, producer B
+  runs on demand and can return a `HIT`; if capture outlasts the bounded
+  foreground wait, the request falls open to `MISS` and runtime capture.
+- Production (workerd/node): the SECOND request is a `HIT`.
+- Dev: expect a few extra MISSes — cold module transforms abort the capture
+  window (per-attempt breadcrumbs: start the server with
+  `INTERNAL_RANGO_DEBUG=1`). This self-heals; only a route that NEVER flips
+  has a real hole/eligibility problem (the once-per-key warning tells the two
+  apart).
+- A HIT is one ordinary document: the frozen prelude first (view-source shows
+  your baked shell, with hole fallbacks in place), then
+  `<div hidden id="S:0">…` segments as the holes resume, per request.
+- A ppr-declared route that CANNOT be honored (missing shell store family,
+  per-request nonce) serves plain axis 1 with NO header and warns once per
+  key — no header + a declared `ppr` means look for that warning.
+- On Cloudflare, `CFCacheStore` reads PPR shells from the per-colo Cache API,
+  falls through to KV on a miss, and promotes the KV hit back into that colo.
+  WITHOUT a KV namespace its shell family remains inert: every ppr route stays
+  `MISS` forever. The store warns once per isolate — bind KV
+  (`new CFCacheStore({ ctx, kv: env.CACHE_KV })`) or use another store. An
+  inert store also stops captures at the gate: the store declares
+  `shellFamilyInert` and the scheduler skips the background render entirely
+  (`skip-inert-store` on the debug event) instead of burning a full capture per
+  MISS whose write could only no-op.
+- Structured capture diagnostics: `createRouter({ debugShellCapture: true })`
+  logs one line per capture attempt/skip (outcome, durations, prelude and
+  snapshot bytes, backoff state); pass a function to receive each
+  `ShellCaptureDebugEvent` instead. `skip-capacity` means the isolate already
+  has 32 queued/running captures; the dropped best-effort capture can retry on a
+  later request. `skip-queue-timeout` means this capture waited 15s behind the
+  active or same-priority work and was dropped before rendering; document-shell
+  captures outrank queued navigation-only snapshots, but never interrupt the
+  active capture — the skip event and the `rango.background` span carry
+  `queuePriority`/`queueAhead` (class and backlog at enqueue) so a parked
+  capture diagnoses itself. A stored attempt reports `bakeWaitMs`: how long the
+  slowest bake source (a top-level pushed handle promise or a bake-lane loader
+  container) held the capture gate — in dev, past 2s, the source is also named
+  once per key in a console warning with the remedies (nest the promise /
+  `cache()` the work / add `loading()`). In dev, with `debugPerformance` on,
+  the last capture outcome for a key also rides the next document GET's
+  `Server-Timing` as `ppr-capture;desc="…"`.
+- For deployed Cloudflare tier diagnostics, build with
+  `INTERNAL_RANGO_DEBUG=1` and run `wrangler tail`. `[CFCacheStore][shell]`
+  JSON events distinguish L1 hit/miss, KV fallback/promotion, tier writes, and
+  marker rejection, with `cf-ray`/colo and read timings. The flag is baked at
+  build time; setting only a Worker runtime variable is too late.
+
+### Unit / integration testing (public primitives)
+
+Import from `@rangojs/router/testing` (Vitest) or `@rangojs/router/testing/e2e`
+(Playwright):
+
+| Helper                                            | Use for                                                      |
+| ------------------------------------------------- | ------------------------------------------------------------ |
+| `assertShellStatus(res, "HIT" \| "MISS")`         | Document Response from a real RSC serve / e2e `page.request` |
+| `assertPprReplayStatus(res, expected)`            | Partial response fresh/stale replay or bounded bypass        |
+| `parsePprReplayStatus(res)`                       | Read structured replay/bypass status or null                 |
+| `shellCacheKey(url)`                              | Production store key for `store.getShell` / custom stores    |
+| `MemorySegmentCacheStore` + `getShell`/`putShell` | Custom store contract / tag eviction (no faked HIT)          |
+
+```ts
+import { MemorySegmentCacheStore } from "@rangojs/router/cache";
+import { assertShellStatus, shellCacheKey } from "@rangojs/router/testing";
+
+// Unit: store + key (after a real putShell / capture flush in e2e)
+const store = new MemorySegmentCacheStore();
+const key = shellCacheKey("http://localhost/products/1");
+// ... after production putShell ...
+expect(await store.getShell(key)).not.toBeNull();
+
+// E2E: header on a real document GET
+assertShellStatus({ headers: new Headers(res.headers()) }, "HIT");
+```
+
+**Out of unit scope** (stay e2e): live MISS → background capture → HIT,
+browser resume of holes, build-time producer B. `dispatch` never runs PPR.
+`renderHandler` only exposes `ctx.dynamic()` / `build` for the opt-out path.
+Do not invent a HIT Response in unit tests. Full recipe: `/testing` skill →
+`cache-prerender.md` (PPR shell section).
+
+## The hole doctrine (encode this in your head)
+
+Holes are **render-defined**, decided by the shape of the tree, on three rules:
+
+| Class          | What makes the hole                                                                                                                                                                | At capture                                                                             | At serve                           |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- | ---------------------------------- |
+| **STRUCTURAL** | the ENTIRE segment subtree under a `loading()` registration — the loader LIVE lane                                                                                                 | loaders masked; the LoaderBoundary postpones; the fallback bakes in as route structure | loaders run fresh; resume fills it |
+| **PHYSICS**    | any promise NESTED in handed-over data still pending at capture, under the consumer's own `<Suspense>` — handler props, handle containers (`push({ x: promise })`), loader-carried | real I/O cannot win the task-quantized quiet window; the boundary postpones            | the promise settles and streams in |
+| **SHELL**      | awaited handler data, TOP-LEVEL `push(promise)` (awaited before SSR), resolved promises, replayed `cache()` segments, BAKE-lane loader containers                                  | baked into the prelude                                                                 | served from the frozen prelude     |
+
+ONE rule for promises, every lane — handlers, handles, AND loaders: **a promise
+nested inside your data is never baked; the container settles.** A loader
+without `loading()` is the BAKE lane (see below): its settled container is
+shell material, exactly like awaited handler data and top-level handle pushes.
+
+**`loading()` is NOT the gate for holes — it is the LANE SELECTOR for loader
+data.** A pending promise under any plain `<Suspense>` postpones and is a hole
+at whatever level it suspends — the PHYSICS row needs no `loading()` anywhere
+(the e2e fixture's physics and nested-handle holes sit in a layout with none).
+A route without a loader PPRs on pure promise/Suspense holes. For loaders,
+`loading()` picks the lane: present = the GUARANTEED live lane (masked at
+capture, fresh every serve, immune to fast resolution); absent = the bake lane
+(the container is shell material, nested promises stay live).
+
+The three promise positions, side by side:
+
+```typescript
+async function Handler(ctx: HandlerContext) {
+  const push = ctx.use(MyHandle);
+
+  push(fetchBadge());                    // TOP-LEVEL push: awaited pre-SSR → BAKED
+  push({ label: "x", stat: fetchStat() }); // NESTED in container → container baked,
+                                           //   stat streams → HOLE (consumer Suspends it)
+
+  const data = await fetchHeader(ctx);   // awaited by the handler → BAKED
+
+  return (
+    <section>
+      <Header data={data} />                       {/* shell */}
+      <StatsPanel promise={fetchStats(ctx)} />     {/* un-awaited prop + own
+                                                       <Suspense> + use() → HOLE */}
+    </section>
+  );
+}
+```
+
+### Choosing the hole mechanism
+
+| Your live region is…                            | Use                                                               | Why                                                                                              |
+| ----------------------------------------------- | ----------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
+| loader data, must be fresh EVERY serve          | `loader()` + `loading()` (the live lane)                          | the guaranteed structural hole — masked at capture, fresh every serve, immune to fast resolution |
+| loader data, shell container + live parts       | bake-lane loader (no `loading()`): `{ static, dynamic: promise }` | container bakes (snapshot-pinned per shell); nested promises hole at the consumer's `<Suspense>` |
+| handler-fetched real I/O (db, fetch)            | un-awaited promise prop + consumer `<Suspense>` + `use()`         | no loader needed; real latency postpones by physics                                              |
+| per-segment metadata consumed elsewhere         | handle container with a NESTED promise + consumer `<Suspense>`    | container is shell, nested value streams — "nesting = liveness"                                  |
+| already-resolved / instant / synchronous values | `loader(() => Promise.resolve(x))` + `loading()`                  | a raw promise that settles inside the quiet window BAKES; only the live lane guarantees live     |
+| none of the above                               | nothing                                                           | it bakes — that is what the shell is for                                                         |
+
+The physics caveat in one line: HANDLER-created promise props are holes only
+because the I/O is genuinely pending at capture — if the value can resolve
+near-instantly (memory read, warmed cache), it may bake into the shell; when
+liveness must be guaranteed rather than probable, use a loader. BAKE-LANE
+NESTED promises are exempt from that race: the capture MASKS every thenable
+nested in a bake-lane container regardless of settle timing
+(`maskNestedContainerThenables`, loader-cache.ts), so the consuming boundary
+always postpones as a hole and every HIT streams the FRESH value — the
+promise SHAPE is the liveness declaration, not a bet on latency. (Before the
+mask, a nested promise that settled inside the window pinned its capture-time
+value into the shared shell; found live as a storefront basket — with the
+capturing session's identifiers — served to anonymous visitors.)
+
+### Handles: "nesting = liveness"
+
+- `ctx.use(H)(promise)` — a TOP-LEVEL pushed promise is awaited server-side
+  before SSR (`resolvedHandleStream`) and BAKED into the shell. The capture
+  gate is held open for the same await, so real latency here is safe (bounded
+  by the capture's 5s guard).
+- `ctx.use(H)({ x: promise })` — the container passes through verbatim
+  (resolution is shallow); the nested promise streams to the consumer, who must
+  `<Suspense>` it. Under capture that boundary postpones — a hole — REGARDLESS
+  of settle timing: the capture masks nested thenables in pushed handle
+  containers (the capture store's push wrap, shell-capture.ts), so even an
+  already-resolved nested promise holes instead of baking its value into the
+  shared shell. Same shape-is-the-declaration rule as bake-lane loaders.
+
+### Want a hole for already-resolved data?
+
+Put it in a loader: `loader(() => Promise.resolve(x))` + `loading()`. Loaders
+are always the live lane — masked at capture, fresh on every serve — no matter
+how fast the value settles.
+
+### The bake lane: loaders without loading() on THEIR entry
+
+A loader on an entry with no renderable `loading()` EXECUTES during capture
+(the capture gate holds open for its real latency, bounded by the 5s guard).
+Its settled container bakes into the prelude; every promise nested in it is
+masked at capture (regardless of how fast it settles) and postpones at the
+consumer's own `<Suspense>` — a hole. On every HIT the
+capture snapshot's loader family overlays the recorded container onto the
+fresh run, so the payload matches the frozen prelude byte-for-byte while the
+nested promises run fresh. The return shape is the declaration:
+
+```typescript
+export const StorefrontContextLoader = createLoader(async (ctx) => {
+  const config = await loadSiteConfig(ctx.params.locale); // bakes (pinned per shell)
+  return {
+    config, // shell material
+    basket: fetchBasket(ctx), // hole — consumer <Suspense>s it, fresh per request
+  };
+});
+```
+
+The lane is decided PER TREE NODE, at the entry that REGISTERS the loaders —
+`loading()` on a CHILD route does not change a parent layout's lane, and
+`loading()` IS valid on layout and parallel entries, not just routes.
+
+Four hard edges (each e2e/unit-pinned):
+
+- **Header writes throw (issue #713).** ppr is a document-scoped `cache()`:
+  in any cached scenario ONLY MIDDLEWARE writes response headers. A handler
+  or loader on a ppr route calling `ctx.headers.set()`, `cookies().set()`,
+  `ctx.header()`, `ctx.setTheme()`, or `setStatus()` throws on EVERY render —
+  dev and prod,
+  first render, same guard family as the `cache()` boundary guard. Handlers
+  are replayed on HITs (the write would silently differ between MISS and
+  HIT); loaders are live but settle AFTER the response headers flushed with
+  the shell (dead letters). Move the write into route middleware — it runs
+  on every request, including HITs, and its headers/cookies merge into every
+  response. The one exception: a handler that calls `ctx.dynamic()` FIRST
+  re-permits its own header/cookie write (#735) — a dynamic() render never
+  HITs, so the write is deterministic (see "Opting out per request").
+- **Identity refuses.** `cookies()`/`headers()` inside a bake-lane loader
+  throws during capture and the capture REFUSES (deterministic, once-per-key
+  warned) — identity can never bake into the shared shell. Give that loader's
+  entry `loading()` (the live lane is exempt) or move the identity-dependent
+  part into a nested promise. The guard's scope is EXACTLY those two calls:
+  per-user state read from a middleware-provided object (`ctx.get("session")`)
+  does NOT refuse — it bakes silently as the capturing user's data (see
+  Pitfalls: the session-object bake trap).
+- **A rejecting bake-lane loader refuses.** Error UI never bakes.
+- **Baked containers show CAPTURE-time data** for the shell's lifetime on
+  document GETs. Soft navigations may replay the captured handler segments, but
+  DSL loaders and their item/response reads remain fresh. That IS the bake
+  lane's meaning; if a value must be fresh on every serve, it belongs on the
+  live lane (`loading()`) or in a nested promise.
+
+### The layout-with-loaders playbook (the storefront case)
+
+The most common real-app shape: an app-wide layout registers per-user loaders
+(session context, basket, wishlist) and the page under it declares `ppr`.
+Those loaders are on the BAKE lane (no `loading()` on the layout), so the page
+captures and HITs — the question is which parts of their data should bake vs
+stay live. Your levers, in order of preference:
+
+1. **Shape the return value.** Shared/config data returns as plain values
+   (bakes, pinned per shell); per-request data returns as NESTED promises
+   consumed under the widget's own `<Suspense>` (live holes). No `loading()`,
+   no restructuring. One wall: a bake-lane loader that reads
+   `cookies()`/`headers()` refuses the capture — identity belongs in a nested
+   promise or on the live lane.
+2. **Do NOT put `loading()` on the layout itself** — that flips the WHOLE
+   layout to the live lane and the LoaderBoundary fallback wraps the layout's
+   ENTIRE subtree: chrome (header, nav, footer) falls out of the shell into
+   the skeleton. Technically PPR, practically pointless.
+3. **Guaranteed-fresh widgets: a parallel slot with its OWN `loading()`.**
+   Parallel-owned loaders get their own per-slot boundary (`fresh.ts` tags
+   them with the slot's loading; `segment-system.tsx` builds a per-slot
+   LoaderBoundary), so the chrome bakes into the shell and each widget is an
+   independent, widget-sized hole:
+
+   ```typescript
+   layout(StoreChrome, () => [
+     // chrome renders NO loader data itself — it bakes into the shell
+     parallel({ "@basket": BasketBadge }, () => [
+       loader(BasketLoader),
+       loading(<BadgeSkeleton />), // hole the size of a badge, not a page
+     ]),
+     // Descriptor form when the slot handler needs ctx (annotate it —
+     // StaticHandlerDefinition in the union blocks inference there):
+     parallel({
+       "@wishlist": {
+         handler: (ctx: HandlerContext) => (
+           <WishlistBadge listUrl={ctx.reverse("wishlist")} />
+         ),
+         use: () => [loader(WishlistLoader), loading(<BadgeSkeleton />)],
+       },
+     }),
+     path("/", HomePage, { name: "home", ppr: true }),
+   ]),
+   ```
+
+   Slot-owned loaders are masked at capture and GUARANTEED fresh per serve —
+   use this where the bake lane's physics (a fast resolve bakes) or pinning
+   (capture-time data for the shell's lifetime) is not acceptable, at the cost
+   of a widget-sized fallback in the shell. The slot handler must hand the
+   loader to a CLIENT component (`useLoader` in a `"use client"` component)
+   for the freshness guarantee to reach the rendered value: server-side
+   `await ctx.use(...)` in the handler is the BAKED lane (the consumption-lane
+   rule, `/rango` → Invariants) — it executes at capture with identity reads
+   permitted, but the value it renders is a capture-time copy wherever it is
+   not shielded by the slot's masked LoaderBoundary.
+
+4. **Shared layout data can also leave the loader lane entirely**: an
+   un-awaited handler promise under the consumer's `<Suspense>` (a physics
+   hole) or `cache()`/`"use cache"` to bake it with tag-invalidation.
+
+The identity rule, stated once: per-user data on a PPR page lives in a NESTED
+promise (a hole, fresh per request) or behind `loading()` with CLIENT-side
+consumption (the live lane). Reading `cookies()`/`headers()` where the value
+would bake as SEGMENT material — handler/render code or a bake-lane loader
+container — refuses the capture by construction. The one exemption is
+handler-INVOKED loader bodies (`await ctx.use(loader)`): they execute at
+capture with identity reads permitted, and the value bakes as a shared
+capture-time copy — mirroring `cache()` semantics (the consumption-lane
+rule; semantic-matrix row PPR3).
+
+### Designing routes for cheap captures (the cost model)
+
+The hole doctrine has a cost corollary that bites silently: **everything that
+bakes is awaited at capture, and that wait recurs on EVERY capture of the
+route**. A top-level pushed handle promise that takes 5s and a Meta promise
+chained 2s further make every capture occupy the per-isolate serialized queue
+for ~7s — while the served response still reads 13ms TTFB, because captures are
+detached. You will not see this cost in any request timing; you see it as slow
+MISS→HIT flips, `skip-queue-timeout` under prefetch pressure, and (in dev, past
+2s) the bake-cost warning naming the source.
+
+Three levers, in preference order:
+
+1. **Nest it** (`{ data: promise }` instead of the promise): the value becomes
+   a hole — per-request, streamed under the consumer's `<Suspense>`, zero
+   capture cost. Choose this whenever the value may be dynamic anyway.
+2. **`cache()` the work inside it**: the value stays baked (shell material,
+   tag-invalidatable), but the capture replays the cached value instead of
+   re-executing the expensive body (mixed-chain). Choose this for shared,
+   expensive-to-compute shell content.
+3. **`loading()` on the entry** (loaders only): moves the loader to the live
+   lane — masked at capture, fresh per serve — at the cost of a fallback in
+   the shell.
+
+Head material (Meta) generally cannot be a hole — the head is shell — so its
+promises bake by design; make them cheap with lever 2. `bakeWaitMs` on the
+capture debug event tells you what each capture actually paid.
+
+## Execution matrix
+
+| Phase            | MISS (foreground)      | Background capture                                                 | HIT (foreground)                                            |
+| ---------------- | ---------------------- | ------------------------------------------------------------------ | ----------------------------------------------------------- |
+| Middleware chain | runs (full)            | **NOT re-run** — inherits the request's post-middleware context    | runs (full) — commit point is after it                      |
+| `router.match`   | runs                   | re-runs under a derived context                                    | runs (behind the flushed prelude)                           |
+| Handlers         | run                    | run on UNCACHED segments; `cache()`d segments replay (mixed-chain) | run (same mixed-chain rules as any render)                  |
+| Loaders          | run **fresh**          | LIVE lane (`loading()`): MASKED; BAKE lane: execute + snapshot-pin | run **fresh** (bake containers overlaid from the snapshot)  |
+| Flight render    | full                   | full                                                               | full (hydration needs the whole payload — no Flight resume) |
+| HTML production  | full fizz              | `prerender` + abort → prelude + postponed                          | `resume` only the holes — O(paths to holes)                 |
+| Shell store      | schedules a bg capture | `putShell(key, …)`                                                 | `getShell(key)`; a stale/SWR hit also schedules a recapture |
+| Prelude bytes    | —                      | —                                                                  | flushed FIRST, before segment resolution starts             |
+
+Middleware is not re-run during capture because it already ran for the
+triggering request — the capture's derived context inherits the
+post-middleware state (`ctx` variables included, which is what makes
+middleware-derived shell content photograph correctly). Guarding is
+serve-time: the commit point runs the full chain on EVERY serve.
+
+Because handlers on uncached segments EXECUTE during capture — and BAKE-lane
+loaders now do too — the `cookies()`/`headers()` capture guard is load-bearing:
+those reads THROW during a capture render (`assertNotInsideShellCapture`), so
+identity can never leak into a shared shell through them. Live-lane loaders
+(behind `loading()`) are exempt: masked at capture, they never run there.
+
+## allReady: the SEO/bot story
+
+`ssr: { resolveStreaming: ... }` returning `"allReady"` (e.g. for bot user
+agents) bypasses PPR entirely — the request gets one complete, fully-buffered
+axis-1 document. Crawlers that dislike streamed shells get a finished page;
+regular users get the streamed shell. No configuration interaction: allReady
+wins.
+
+## Security
+
+Shell caching shares one shell per host+URL across all users:
+
+**(a) Access control is sound by construction.** The commit point is after ALL
+middleware on every serve. A 401/redirect short-circuit returns before any
+shell byte.
+
+**(b) Identity can't leak via cookies/headers.** `cookies()` and `headers()`
+THROW during the background capture render — in handlers AND in bake-lane
+loaders (whose containers would bake). A shell that reads them is
+PPR-ineligible by construction; the live lane (`loading()`) stays exempt.
+
+**(c) Residual hazard — middleware-derived per-user state.** A `ctx` variable
+set by an upstream auth middleware and rendered by shell material is
+photographed into the SHARED shell (the capture inherits post-middleware
+state). That is scope fidelity working as designed — for shared values. The
+same hazard reaches BAKE-LANE LOADERS: a loader reading a middleware-provided
+session object (`ctx.get("session")`) never calls `cookies()` itself, so the
+guard cannot see it — whatever it returns as settled container data is
+photographed as the CAPTURING user's state. If the
+value is per-user: shell-cache only public/shared pages, keep per-user content
+in nested pending promises or live-lane (`loading()`) loaders — NOT in a
+bake-lane container — or key per variant at the CDN tier.
+
+## What always stays on axis 1
+
+Non-GET, non-partial RSC/action/loader fetches, partial requests without an
+eligible captured segment snapshot, per-request CSP nonce, `streamMode:
+"allReady"`, redirects, 404s, error renders, routes without `ppr`, and any store
+without the shell family. A stored shell is invalidated when
+`React.version` changes (postponed state is build-coupled), so deploys
+self-heal via recapture.
+
+The per-request CSP nonce guarantee covers BOTH ways a nonce arrives — the
+`createRouter({ nonce })` provider AND a direct `ctx.set(nonce, value)` token
+write in middleware (the `nonce` token from `@rangojs/router`). Either way the
+nonce ends up rendered into the document (`useNonce()` puts the provider nonce
+on every nonced script/style/meta; a token nonce is rendered by whatever app
+code reads `ctx.get(nonce)`), so a shell shared per host+URL cannot bake it
+without freezing one request's nonce for every visitor (the browser's CSP
+would then reject the frozen nonce for all but the capture request). The serve
+gate reads the token off the post-middleware request variables at the commit
+point (which runs after the whole middleware chain), so a middleware-set nonce
+blocks capture the same as a provider one. Because the route DECLARED `ppr`
+but cannot be honored, it logs a once-per-key worker warning (same
+declared-intent-cannot-be-honored doctrine as the missing-store warning) and
+serves pure axis 1 with no `x-rango-shell` header. An undeclared route stays
+silent.
+
+### The proper way to supply a nonce
+
+`createRouter({ nonce })` is the canonical path — supply the nonce THERE, not
+via a token write. The provider value is threaded into the router's own SSR
+machinery: `NonceContext`/`useNonce()`, automatic nonce attributes on
+`<Scripts />` and `<MetaTags />` output, and the inlined Flight payload
+scripts. It ALSO sets the `nonce` token, so `ctx.get(nonce)` works in
+middleware and handlers for the CSP response header. A direct
+`ctx.set(nonce, value)` write in middleware is app-managed only: the router
+resolves its SSR nonce BEFORE middleware runs, so a token-set value is
+readable via `ctx.get(nonce)` and gates PPR (this section), but the router
+will NOT apply it to its own scripts — `useNonce()` stays undefined and the
+Flight payload scripts carry no nonce, which a nonce-only `script-src` policy
+would then block. If you need a per-request nonce, use the provider; reserve
+the token for READING the value.
+
+## Options: PartialPrerenderProps
+
+```typescript
+path("/products/:id", Page, { name: "product", ppr: true }, use);
+path(
+  "/products/:id",
+  Page,
+  { name: "product", ppr: { ttl: 600, swr: 120, tags: ["catalog"] } },
+  use,
+);
+```
+
+| Field              | Default | Notes                                                                                                                                                     |
+| ------------------ | ------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `ttl`              | `300`   | shell freshness window in seconds (`ppr: true` uses the default)                                                                                          |
+| `swr`              | —       | stale window: serve the stale shell + background recapture                                                                                                |
+| `tags`             | —       | operational tags UNIONED with the tags the capture render auto-collects — see "Invalidation" below                                                        |
+| `maxSnapshotBytes` | 8 MiB   | cap on the entry's capture data snapshot; over it the snapshot is skipped (shell still stored, warned once per key) so the entry stays under store limits |
+| `captureTimeout`   | 15000ms | capture settle budget; a timed-out capture is refused rather than storing a partial shell                                                                 |
+
+The shell store is always the app-level `createRouter({ cache })` store; the
+default key is `${host}${pathname}${sortedSearch}:shell` (host-scoped so
+multi-tenant shells never collide).
+
+These options control the in-function shell entry only. They do not emit HTTP
+`Cache-Control`. Adding `s-maxage` separately allows a platform CDN to cache the
+completed response, including the live-hole output, and CDN hits bypass Rango
+middleware entirely. Only do that for a fully public response whose complete
+output is shared; see `/deployment-caching`.
+
+## Invalidation: tags vs revalidate()
+
+`updateTag()`/`revalidateTag()` is the ONLY lever that changes the frozen shell
+HTML; `revalidate()` is a DATA lever that never touches it.
+
+A captured shell auto-carries the UNION of the non-loader tags recorded during
+the capture render — every `cacheTag(...)` that ran as shell material, whether
+from a `"use cache"` function, a `cache()` segment, or a render-callable
+`cacheTag()` in a plain server component (no `"use cache"`/`cache()` in its
+tree). Loader tags never attach (the holes are already live). `ppr.tags` adds
+operational tags the render cannot know (a tenant id, a deploy marker).
+
+| Lever                                         | Reaches the frozen shell?                                     | Reaches the holes?                                  |
+| --------------------------------------------- | ------------------------------------------------------------- | --------------------------------------------------- |
+| `updateTag` / `revalidateTag` on a SHELL tag  | YES — drops the shell → MISS → recapture                      | n/a (holes are already live)                        |
+| `updateTag` / `revalidateTag` on a LOADER tag | no — loader tags never attach to a shell                      | drops that loader's cached value (if it `cache()`s) |
+| `revalidate()` (named revalidation contract)  | **no** — re-runs segments/loaders for the PAYLOAD, never HTML | yes — the hole re-renders with fresh data           |
+
+A server action's automatic invalidation refreshes the CLIENT only — it re-runs
+the holes and streams a fresh payload, but does NOT evict the server shell.
+Shell-baked data stays stale until the shell's TTL unless you tag-invalidate it
+(`updateTag` on a shell tag). Data baked into the shell WITHOUT a tag cannot be
+evicted by tag at all — move always-fresh data under a `loading()` hole.
+
+## Pitfalls
+
+- **A bake-lane loader that reads `cookies()`/`headers()`**: the capture is
+  REFUSED (deterministic, once-per-key warned) — the route stays on axis 1.
+  Move identity into a nested promise or behind `loading()`.
+- **A bake-lane container that must be fresh per document GET**: it is
+  snapshot-pinned for the shell's lifetime by design. Use the live lane
+  (`loading()`) or a nested promise instead.
+- **A bake-lane loader slower than `ppr.captureTimeout` (15s by default)**: the
+  capture is refused rather than storing a partial shell. Increase the route's
+  budget only when the deployment can keep the background/build work alive.
+- **Per-user value in shell material**: baked into the shared shell —
+  deterministically, not by race (handler promises deep-settle at the ring-3
+  write on cached chains; awaited/resolved values bake everywhere). Put
+  per-user data in a nested pending promise or a live-lane (`loading()`)
+  loader — a BAKE-lane loader container bakes just like handler material.
+- **The session-object bake trap (the guard cannot save you here)**: the
+  capture guard sees `cookies()`/`headers()` calls ONLY. A bake-lane loader
+  reading a middleware-provided session object (`ctx.get("session")`) refuses
+  nothing. Per-user data survives ONLY behind a nested promise — the shape is
+  the declaration, and it holds for BOTH branches regardless of settle timing
+  (nested thenables are masked at capture):
+
+  ```typescript
+  const CartLoader = createLoader(async (ctx) => {
+    const basketId = ctx.get("session")!.get("basketId");
+    if (!basketId) return { cart: Promise.resolve(null) }; // nested thenable → masked → hole, fresh per HIT
+    return { cart: fetchBasket(basketId) }; // nested thenable → masked → hole, fresh per HIT
+  });
+  ```
+
+  The remaining trap is returning per-user data as PLAIN container material:
+  `return { user: session.user }` bakes it into the shared shell like any
+  other settled value — deterministically, not by race. Wrap it in a promise
+  (even an already-resolved one) or put the loader on the live lane.
+
+- **Theme on a HIT is capture-then-corrected**: the resume tree replays the
+  CAPTURE's `initialTheme` (resume requires it to match the frozen prelude);
+  the visitor's cookie theme is applied pre-paint by the FOUC script and
+  re-synced post-mount by ThemeProvider. Nothing to configure — but a themed
+  component in the shell may briefly render the captured theme's markup before
+  the post-mount re-sync.
+- **Shell shows CAPTURE-time data for the shell's lifetime**: a `cache()`/`"use
+cache"` value baked into the shell is PINNED at capture (the capture data
+  snapshot) and replayed on every HIT, so the shell stays byte-identical to the
+  frozen prelude even after that cache entry expires, gets recomputed, or is
+  tag-invalidated. This is deliberate — parity beats freshness inside the shell.
+  If a shell region needs to be fresh, put it under a hole — `loading()` for
+  loader data, or an un-awaited promise under the consumer's `<Suspense>`
+  (holes are never pinned) — or make the SHELL itself invalidatable by tagging
+  it: call `cacheTag(...)` from the shell-material render code (the render-time
+  lever), or add the tag to `ppr.tags` (operational tags the render cannot know —
+  a tenant id, a deploy marker). Ring-1/ring-3 tag invalidation does NOT drop the
+  shell. Tags are optional: if TTL/SWR is the complete freshness policy, leave
+  the shell untagged. Rango does not warn for that choice; with
+  `debugShellCapture` enabled, a stored event reports `untaggedBake: true` when
+  bake-lane loader material uses TTL/SWR-only invalidation.
+- **Uncached nondeterminism in the shell is a hydration hazard**: a raw
+  `Date.now()` / `Math.random()` / uncached `fetch` rendered directly in shell
+  material (outside any cache ring) drifts between capture and hit and the
+  snapshot CANNOT pin it — it was never a cache read. It will mismatch the frozen
+  prelude and detonate hydration. Wrap it in `cache()`/`"use cache"` (then it is
+  pinned) or move it under a hole (`loading()`, or a pending-promise
+  `<Suspense>` region).
+- **Stacking with `/document-cache` or HTTP CDN caching**: both cache the
+  completed composite, including live-hole output, so PPR becomes redundant on
+  a hit. A platform CDN also bypasses every Rango middleware. Restrict this to
+  fully public, shared responses; see `/deployment-caching`.
+- **Dev + HMR**: works, but edits produce stale shells until TTL/recapture.
+- **Dev cold-start cadence**: expect `MISS -> (in-place retry) -> HIT`. A
+  refused capture is negatively cached with an exponential window (1s doubling
+  to a 60s cap), so declaring `ppr` on an ineligible route never re-renders it
+  on every request.
+- **HIT status is committed at the flush**: a failing hole cannot become a
+  500/redirect after the first shell byte — error UI renders inline via
+  Suspense/error boundaries (the same property any streamed SSR page has after
+  its shell flushes).
+
+## Related
+
+- `/defer-hydration` — keep the full body HTML in the shell while moving a
+  heavy subtree's hydration off the initial main-thread task (gated boundary,
+  content-as-fallback)
+- `/document-cache` — whole-response edge caching (no live holes)
+- `/caching` and `/cache-guide` — segment/function caching (axis 1 data)
+- `/shell-manifest` — replayed handles as cache metadata read by live loaders
+- Design doc: `docs/design/ppr-shell-resume.md` in the package