@rangojs/router 0.0.0-experimental.ce770c9a → 0.0.0-experimental.d5d3070a

package/skills/cache-guide/SKILL.md

@@ -6,8 +6,140 @@ argument-hint:
 
 # cache() vs "use cache" — When to Use Which
 
-Both mechanisms share the same backing store, cache profiles, and tag-based
-invalidation. They differ in scope, cache key, execution model, and runtime control.
+Both mechanisms share the same backing store and cache profiles, and both accept
+an optional `tags` field (honored by the built-in stores — invalidate with
+`updateTag`/`revalidateTag`; see "Two axes" below). They differ in scope, cache
+key, execution model, and runtime control.
+
+## Two axes — do not conflate
+
+Everything on this page is **axis 1: stored-value freshness** — _is a cached
+value still good?_ There is a second, orthogonal axis it is easy to mistake for
+caching:
+
+1. **Stored-value freshness** — _is a cached value still good?_
+   → `"use cache"` (fn/component), `cache()` (segment), loader `cache()` (loader data).
+   Entries expire by **TTL/SWR** and can be tagged (`cache({ tags })` or runtime
+   `cacheTag(...tags)`). Built-in stores (`MemorySegmentCacheStore`, `CFCacheStore`)
+   index by tag; invalidate on demand with `updateTag(...tags)` (awaitable,
+   read-your-own-writes) or `revalidateTag(...tags)` (background, non-blocking).
+   Both hard-purge; the difference is awaitability, not stale-serving.
+2. **Client-update selection** — _should this segment re-run and stream to the
+   client on this navigation/action?_
+   → `revalidate()`. Covered in `/loader` and `/route`, **not here**.
+
+They are orthogonal and compose: a segment selected by `revalidate()` still
+consults its cache (hit → no recompute); a cache bust does **not** force a client
+update, and `revalidate()` never reads, writes, or expires a cached value. If you
+know React Router, `revalidate()` is `shouldRevalidate`, not `Cache-Control`. See
+`/rango` → "Coming from another framework" for the cross-framework mapping.
+
+## Fast choice
+
+Read this first; use the rest of the page when the choice has edge cases.
+
+1. Do you want to cache an entire route or group of routes?
+   **Yes** -> `cache()`
+2. Do you need runtime conditions, such as skip for authed users or key by
+   locale?
+   **Yes** -> `cache()` with `condition` / `key`
+3. Do you want to cache a data fetch or helper shared across routes?
+   **Yes** -> `"use cache"`
+4. Do you need different cache entries for different function arguments?
+   **Yes** -> `"use cache"` (keyed by args)
+5. Is the expensive part rendering a subtree?
+   **Yes** -> `cache()` (caches rendered segments)
+6. Is the expensive part one query inside a larger live handler?
+   **Yes** -> `"use cache"` on the query function
+
+## Correctness & invalidation
+
+rango's caches are built so a hit can't serve wrong or stale-shaped data. These
+guarantees are mostly automatic — worth knowing so you don't reimplement
+protection the framework already gives you (or assume one it deliberately
+doesn't).
+
+There are two guard models to keep separate. Both block response side effects
+(`ctx.header()`, cookie writes) that would be lost on a hit; they differ in what
+else they allow:
+
+- **`cache()` boundary guard** (route-level) — fires while the handler runs on a
+  miss. `cookies()` and `headers()` throw (request-scoped data would be baked into
+  the shared cached shell), `ctx.get(nonCacheableVar)` throws (a tainted value
+  would be baked in), and response side effects (`ctx.header()`, `ctx.setCookie()`,
+  `ctx.setStatus()`, `ctx.onResponse()`) throw. `ctx.set()` of a cacheable var is
+  **allowed** — children are cached too and can read it. **Loaders are exempt**
+  (they always run fresh) — read request data inside a loader.
+- **`"use cache"` exec-guard** (function-level) — the same request-scoped APIs
+  throw inside the cached function (`cookies()`, `headers()`, `ctx.set()`,
+  `ctx.header()`); additionally, tainted `ctx`/`env`/`req` args are excluded from
+  the cache key.
+
+### Cross-deploy safety: version-segmented store keys
+
+`CFCacheStore` prefixes every **physical** store key (the CF Cache API URL and
+the KV key) with the build version — auto-generated from the
+`@rangojs/router:version` virtual module, overridable via the store's `version`
+option. A new deploy reads under a new prefix, so it can **never** read a
+previous build's entries: no cross-deploy shape drift, and no dead client-chunk
+references baked into cached RSC.
+
+The tradeoff to know: **loader/data caches use the same store**, so they're
+version-segmented too. Every deploy is therefore a _cold data cache_ — SWR can't
+soften it, because no stale entry exists under the new key. For high-traffic,
+frequently-deploying, data-bound apps that's a deploy-time origin warm-up. Decide
+deliberately: accept it (correctness over hit-rate), or split the policy — let
+the render/edge cache auto-version while a separate data store gets a stable
+`version` so its entries survive deploys. (Per-process stores like
+`MemorySegmentCacheStore` are cold on every restart anyway; this matters for
+persistent stores.) See `/caching` for store setup.
+
+### Client cache: forward/back is mutation-aware
+
+The browser keeps a history (forward/back) cache of rendered segments. Any
+client-side mutation (a server action) marks those entries **stale** and
+broadcasts it to other tabs. On back/forward (popstate) the router looks up the
+entry, sees it's stale, and revalidates — so your `revalidate()` predicates re-run
+and the segment refreshes (SWR: the stale view paints instantly, fresh data
+streams in). It's the client-side analog of the server-cache correctness problem,
+solved on the partial-render axis.
+
+### Request-scoped data: the `cache: false` taint
+
+`createVar({ cache: false })` (or a `ctx.set(var, v, { cache: false })` write)
+taints a value as request-scoped; reading it **directly** with `ctx.get()` inside
+a `cache()` boundary throws — the guard against the catastrophic "serve user A's
+data to user B" bug. The guarantee is precise and intentionally narrow — see
+"Context Variable Cache Safety" below for exactly what it does and does not catch.
+
+## Stale-while-revalidate
+
+SWR is a first-class cache behavior when the backing store supports it: while an
+entry is within its SWR window the cache serves the **stale value instantly** and
+refreshes it in the **background** (`waitUntil`), so users never wait on a
+recompute for a merely-aging entry.
+
+- **`"use cache"`** resolves to the `default` profile `{ ttl: 900, swr: 1800 }`,
+  so function/component caching gets a 30-minute SWR window **out of the box**.
+  Tune or add profiles via `createRouter({ cacheProfiles: { … } })`
+  (`"use cache: short"` → the `short` profile).
+- **`cache()` DSL and loader caches** take an explicit `swr` in seconds (or
+  inherit `store.defaults.swr`): `cache({ ttl: 60, swr: 300 })` → fresh ≤60s,
+  stale-served 60–360s, miss after 360s in stores that implement SWR for that
+  layer.
+- **Client forward/back** is SWR after a mutation — see "Correctness &
+  invalidation" → Client cache.
+- **Edge / document layer** uses the HTTP `stale-while-revalidate` directive; see
+  `/document-cache`.
+
+SWR softens normal TTL expiry, **not** a cross-deploy cold cache — a new build
+has no stale entry to serve (see version-segmented store keys above).
+
+Store support is layer-specific. `CFCacheStore` supports SWR for segment,
+document/response, and `"use cache"` item entries. `MemorySegmentCacheStore`
+supports SWR for response and `"use cache"` item entries, but its route-segment
+entries expire at TTL and never background-revalidate. Use the memory store for
+local/dev behavior, not as proof that segment SWR is active.
 
 ## Key Differences
 
@@ -18,7 +150,7 @@ invalidation. They differ in scope, cache key, execution model, and runtime cont
 | **Cache key**        | Request type + pathname + params (+ optional custom)  | Function identity + serialized non-tainted args    |
 | **Execution on hit** | All-or-nothing: entire handler skipped                | Partial: function body skipped, calling code runs  |
 | **Runtime control**  | `condition` to disable, custom `key` function         | None — if the directive is present, it caches      |
-| **Side effects**     | No guards needed — handler doesn't run on hit         | `ctx.header()`, `ctx.set()`, etc. throw at runtime |
+| **Side effects**     | Response side effects throw inside the boundary       | `ctx.header()`, `ctx.set()`, etc. throw at runtime |
 | **Handle data**      | Captured and replayed                                 | Captured and replayed                              |
 | **Loaders**          | Always fresh — excluded from cache, opt-in per loader | Can be used inside loaders                         |
 | **Nesting**          | Nest `cache()` boundaries with different TTLs         | Compose by calling cached functions from uncached  |
@@ -144,13 +276,38 @@ On cache hit for the route, the handler doesn't run and `getProductData` is neve
 called. On cache miss, the handler runs and `getProductData` may itself return a
 cached value from a previous call with the same slug.
 
+### Nesting rule: the outer window bounds the inner
+
+A cache's window bounds everything rendered inside it (loaders excepted). An
+inner shorter TTL only takes effect when the **enclosing** cache recomputes — it
+does **not** keep a value fresher than its parent:
+
+- Outer `cache()` **fresh hit** → the subtree is served from stored RSC, so inner
+  `"use cache"` functions are **not consulted** (frozen at the outer's age — no
+  code inside the boundary runs on a hit).
+- Outer **miss / SWR revalidation** → inner caches are consulted, each per its own
+  ttl/swr. With SWR on the outer, a stale subtree serves instantly and refreshes
+  in the background, so under traffic it keeps refreshing rather than rotting to
+  the worst case.
+- **Loaders are the exception** — excluded from the segment cache, re-resolved
+  live even on an outer hit.
+
+So `"use cache: short"` (60s) inside `cache({ ttl: 600 })` yields ~600s freshness
+on hits, **not** 60s. This is not a bug: setting `cache({ ttl: 600 })` declares
+"this subtree may be ~600s stale." **If a value must be fresher than its
+enclosing segment, put it in a loader** (always live). `debugPerformance` prints
+cache hits per layer, so the actual per-request behavior is observable.
+
 ## Headers and Cookies
 
 Neither mechanism caches response headers or cookies.
 
-- **cache()**: Headers set by handlers are naturally absent on hit because no
-  handler runs. If you need headers on every response, set them in middleware
-  (which runs before cache lookup).
+- **cache()**: Response-level side effects throw inside the cache boundary even
+  on a miss: `ctx.header()`, `ctx.setCookie()`, `ctx.deleteCookie()`,
+  `ctx.setStatus()`, `ctx.onResponse()`, and direct `ctx.headers` mutation. On a
+  hit the handler would be skipped, so allowing the write on a miss would produce
+  inconsistent responses. If you need headers or cookies on every response, set
+  them in middleware or a live segment outside the cache boundary.
 - **"use cache"**: cookies() and headers() throw inside the cached function
   (both reads and writes). ctx.header() also throws. Move them outside.
 
@@ -162,6 +319,88 @@ middleware(async (ctx, next) => {
 });
 ```
 
+## Context Variable Cache Safety
+
+Context variables created with `createVar()` are cacheable by default and can
+be read freely inside cached scopes. A non-cacheable var throws when read
+**directly** with `ctx.get()` inside a `cache()` boundary — where the value would
+otherwise be serialized into the stored segment.
+
+There are two ways to mark a value as non-cacheable:
+
+```typescript
+// Var-level policy — inherently request-specific data
+const Session = createVar<SessionData>({ cache: false });
+
+// Write-level escalation — this specific write is non-cacheable
+ctx.set(Theme, derivedTheme, { cache: false });
+```
+
+"Least cacheable wins": if either the var definition or the `ctx.set()` call
+specifies `cache: false`, the value is non-cacheable.
+
+**Behavior inside a `cache()` boundary:**
+
+| Operation                                 | Inside a `cache()` boundary                            |
+| ----------------------------------------- | ------------------------------------------------------ |
+| `cookies()` / `headers()` (read or write) | Throws (request-scoped, would poison the shared entry) |
+| `ctx.get(cacheableVar)`                   | Allowed                                                |
+| `ctx.get(nonCacheableVar)`                | Throws (would be baked in)                             |
+| `ctx.set(var, value)` (cacheable)         | Allowed                                                |
+| `ctx.header()` / cookie writes            | Throws (response side effect would be lost on hit)     |
+| Any of the above **inside a loader**      | Allowed (loaders always run fresh)                     |
+
+(Both scopes block the same request-scoped APIs — `cookies()`, `headers()`,
+response side effects, and non-cacheable `ctx.get()` — because each would leak
+per-request data into a shared cache entry. The `cache()` boundary tracks the
+scope via `isInsideCacheScope()`; `"use cache"` uses the exec guard and also
+excludes tainted `ctx`/`env`/`req` args from the cache key. Loaders are exempt in
+both — see "Headers and Cookies" and the precise guarantee below.)
+
+Write is dumb — `ctx.set()` stores the cache metadata but does not enforce.
+Enforcement happens at read time (`ctx.get()`), where ALS detects the cache
+scope and rejects non-cacheable reads.
+
+### The guarantee is precise — a direct read inside `cache()`, not propagating
+
+The guard fires on a **direct** `ctx.get(taintedVar)` **inside a `cache()`
+boundary** (the scope `isInsideCacheScope` detects). The taint lives on the
+variable; a value **derived** from it and read **outside** the boundary is not
+tracked:
+
+```typescript
+// CAUGHT — direct read of a tainted var inside a cache() boundary
+cache({ ttl: 60 }, () => [
+  path("/dashboard", (ctx) => {
+    const user = ctx.get(User); // throws: non-cacheable read inside cache()
+    return <Dashboard user={user} />;
+  }, { name: "dashboard" }),
+]);
+
+// NOT CAUGHT — read outside the boundary, derived value cached
+layout((ctx) => {
+  const name = ctx.get(User).name; // allowed — this layout is not cached
+  ctx.set(UserName, name); // now a plain (cacheable) string
+  return <Outlet />;
+}, () => [
+  cache({ ttl: 60 }, () => [
+    // a child reads ctx.get(UserName) and silently caches user-derived data
+  ]),
+]);
+```
+
+So do **not** read this as "you can't cache user data" — that overstates it and
+breeds the false confidence that makes the derived leak _more_ likely. The guard
+is deliberately non-propagating (propagation would cost a wrapper per derivation
+on the hot path), and it is scoped to the `cache()` segment boundary. `"use
+cache"` functions block the same request-scoped reads (`cookies()` / `headers()`
+throw inside them) and additionally exclude tainted `ctx`/`env`/`req` args from
+the cache key. The pattern that stays safe is also the natural one:
+**read tainted context at the point of use, in the path that needs it (a loader or
+live segment) — never extract user data into a plain value and cache that.**
+Loaders are exempt because they run outside the cache scope and resolve fresh
+every request.
+
 ## Loaders Are Always Fresh
 
 Loaders are **never cached** by route-level `cache()`. Even on a full cache hit
@@ -185,7 +424,7 @@ subsequent siblings. Everything below the cache boundary is cached as one unit:
 
 ```typescript
 path("/dashboard", DashboardPage, { name: "dashboard" }, () => [
-  cache("long"),
+  cache({ ttl: 300 }),
   layout(DashboardSidebar, () => [
     parallel("@stats", StatsPanel),
     parallel("@activity", ActivityFeed),
@@ -205,7 +444,7 @@ boundary are not cached and always re-render:
 layout(RootLayout, () => [
   // RootLayout is NOT cached — runs every request
   path("/products/:slug", ProductPage, { name: "product" }, () => [
-    cache("long"),
+    cache({ ttl: 300 }),
     layout(ProductSidebar),
     parallel("@reviews", ReviewsPanel),
     parallel("@related", RelatedProducts),
@@ -240,21 +479,6 @@ data is cached independently from the route's segment cache. Loader caching
 supports custom keys, tags, SWR, conditional bypass, and per-loader store
 overrides — see `/loader` for the full reference.
 
-## Decision Flowchart
-
-1. Do you want to cache an entire route or group of routes?
-   **Yes** -> `cache()`
-2. Do you need runtime conditions (skip for auth users, key by locale)?
-   **Yes** -> `cache()` with `condition` / `key`
-3. Do you want to cache a data fetch shared across routes?
-   **Yes** -> `"use cache"`
-4. Do you need different cache entries for different arguments?
-   **Yes** -> `"use cache"` (keyed by args)
-5. Is the expensive part rendering, not data fetching?
-   **Yes** -> `cache()` (caches rendered segments)
-6. Is the expensive part a single query inside a larger handler?
-   **Yes** -> `"use cache"` on the query function
-
 ## See Also
 
 - `/caching` — cache() DSL setup, stores, nested boundaries

package/skills/caching/SKILL.md

@@ -8,6 +8,46 @@ argument-hint: [setup]
 
 @rangojs/router supports segment-level caching with stale-while-revalidate (SWR) for optimal performance.
 
+> SWR support is store-specific. `CFCacheStore` revalidates segment, response,
+> and `"use cache"` entries in the background. `MemorySegmentCacheStore`
+> supports SWR for response and `"use cache"` item entries, but its
+> route-segment entries expire at TTL with no background revalidation — use
+> `CFCacheStore` for real segment SWR. See `/cache-guide`.
+
+## cache() is Partial Prerendering (PPR)
+
+`cache()` caches **everything except loaders**. On a cache hit, the cached
+segments (layouts, route components, parallels — including any resolved
+Suspense) are served from the store, and **loaders re-run fresh on every
+request**, streaming their results into the same response. Loaders are the
+dynamic "holes" of an otherwise-cached tree.
+
+This means a `cache()` boundary at the document root **is** whole-document
+Partial Prerendering: the static shell is cached and served instantly while
+per-request/per-user data stays live — in one streamed response, no extra round
+trip. The browser cannot tell the shell came from cache.
+
+```typescript
+cache({ ttl: 60, swr: 300 }, () => [
+  layout(<RootLayout />), // cached shell
+  path("/dashboard", Dashboard, { name: "dashboard" }, () => [
+    loader(StatsLoader), // DYNAMIC HOLE — re-runs every request
+  ]),
+]);
+```
+
+The consumer rule: **want it cached? render it inline. want it dynamic? put it
+in a loader and read it with `useLoader()` in a client component.** Anything
+read with `cookies()`, `headers()`, or a non-cacheable variable belongs in a
+loader (loaders always run fresh). Reading it directly in a cached handler
+throws; awaiting it with `ctx.use()` and rendering the result in a cached
+handler silently bakes per-request data into the shared shell (see "Cache purity
+& tainted objects" below).
+
+Pre-rendering (`/prerender`) is the build-time twin: it caches the same shell at
+build time instead of on first request. Both feed the segment system
+identically, and loaders always run fresh at request time.
+
 ## Route-Level Caching with cache()
 
 Use the `cache()` DSL function to cache routes:
@@ -41,10 +81,83 @@ cache(
 );
 ```
 
-## Named Profile Shorthand
+## Tag-Based Invalidation
+
+Tag cached entries, then invalidate them on demand. Tags can be attached three ways:
+
+```typescript
+// 1. Static tags in the cache() DSL
+cache({ ttl: 300, tags: ["products"] }, () => [path("/products", List)]);
+
+// 2. Dynamic tags (function of ctx)
+cache(
+  { ttl: 300, tags: (ctx) => [`product:${ctx.params.id}`, "products"] },
+  () => [path("/products/:id", Detail)],
+);
+
+// 3. Runtime tags inside a "use cache" function
+async function getProduct(id: string) {
+  "use cache";
+  cacheTag(`product:${id}`, "products"); // variadic, additive
+  return db.getProduct(id);
+}
+```
+
+Invalidate with one of two server-only verbs (both variadic, imported from
+`@rangojs/router`):
+
+```typescript
+// Server Action — read-your-own-writes. Await it so the action's own re-render
+// (and the next navigation) sees fresh data.
+async function updateProduct(formData: FormData) {
+  "use server";
+  await db.updateProduct(formData);
+  await updateTag("products");
+}
+
+// Route handler / webhook — background, non-blocking (waitUntil). Hard-purge:
+// the next read re-renders fresh (NOT stale-while-revalidate).
+export async function POST() {
+  "use server";
+  revalidateTag("products");
+  return new Response("ok");
+}
+```
 
-Use a named cache profile string instead of an options object. The profile must be
-defined in `createRouter({ cacheProfiles })`. Unknown names throw at boot time.
+| API                      | Timing                      | Use in                    | Semantics                                             |
+| ------------------------ | --------------------------- | ------------------------- | ----------------------------------------------------- |
+| `updateTag(...tags)`     | awaitable (`Promise<void>`) | server actions            | immediate; next read is fresh                         |
+| `revalidateTag(...tags)` | background (`void`)         | route handlers / webhooks | background (non-blocking); next read re-renders fresh |
+
+Both built-in stores support tags. For `CFCacheStore`, distributed (cross-colo)
+invalidation requires a `kv` namespace — the tag-invalidation markers live in
+that same namespace; there is **no** separate tag-invalidation store to wire.
+If no tag-capable store is configured, `updateTag`/`revalidateTag` warn and no-op.
+
+By default `CFCacheStore` reads the KV marker on every tagged cache read
+(strongest invalidation latency). To cut KV reads on hot tagged routes, set
+`tagCacheTtl` (seconds) to cache each marker in the per-colo edge cache for that
+window — the colo running `updateTag`/`revalidateTag` writes the fresh marker
+into its own edge cache immediately (read-your-own-writes), while other colos
+converge within `tagCacheTtl` (the **maximum extra cross-colo invalidation
+latency** when no purge is wired). Keep it small (e.g. 30–60), or wire a purge
+(below) and set it large. (Contrast `tagInvalidationTtl`, which must be _large_
+— it bounds how long the KV marker itself lives and must exceed your max entry
+TTL+SWR.)
+
+To make other colos prompt without a short `tagCacheTtl`, pass `onRevalidateTag`:
+each cached marker carries a namespaced Cloudflare `Cache-Tag`, and the hook is
+handed exactly those tags (batched, once per `updateTag`/`revalidateTag` call) to
+feed Cloudflare's purge-by-tag API — evicting the cached lookups everywhere.
+Purge-by-tag is available on all plans (since April 2025), subject to per-plan
+rate limits, so the batched single call matters. With a purge wired, `tagCacheTtl`
+becomes a pure read-cost reducer + fallback window.
+
+## Named Cache Profiles
+
+Define named profiles in `createRouter({ cacheProfiles })` so the same TTL/SWR
+values can be shared across the DSL and `"use cache"` functions without repetition.
+Unknown names throw at boot time.
 
 ```typescript
 // Define profiles in router
@@ -55,19 +168,25 @@ createRouter({
     long: { ttl: 3600, swr: 7200 },
   },
 });
+```
+
+In the DSL, pass the profile's options directly to `cache()`:
 
-// Use by name in urls
+```typescript
 export const urlpatterns = urls(({ path, cache }) => [
-  cache("long", () => [path("/blog", BlogIndex, { name: "blog" })]),
+  cache({ ttl: 3600, swr: 7200 }, () => [
+    path("/blog", BlogIndex, { name: "blog" }),
+  ]),
 
-  // Also works without children (orphan cache boundary)
-  cache("short"),
+  // Orphan cache boundary (covers subsequent siblings)
+  cache({ ttl: 60, swr: 120 }),
   path("/feed", FeedPage, { name: "feed" }),
 ]);
 ```
 
-These profile names are shared with the `"use cache: <name>"` directive. See
-`/use-cache` for function-level caching.
+The DSL `cache()` helper does NOT accept a string profile name — strings are only
+valid in the `"use cache: <name>"` directive inside server functions. See
+`/use-cache` for function-level caching with named profiles.
 
 ## Loader-Level Caching
 
@@ -116,7 +235,6 @@ import { MemorySegmentCacheStore } from "@rangojs/router/cache";
 
 const store = new MemorySegmentCacheStore({
   defaults: { ttl: 60, swr: 300 },
-  maxSize: 1000, // Max entries
 });
 ```
 
@@ -173,6 +291,157 @@ const router = createRouter<AppBindings>({
 KV entries require `expirationTtl >= 60s`. Short-lived entries (< 60s total TTL)
 are only cached in L1.
 
+### Resilience & latency budgets
+
+Every cache read is **fail-safe**: a degraded tier never stalls or fails the
+request — it degrades to the next tier (L1 → L2 → render). Three optional latency
+budgets (milliseconds) bound each tier so a slow colo or KV namespace cannot pin
+a request behind it:
+
+| Option                | Default | Bounds                              |
+| --------------------- | ------- | ----------------------------------- |
+| `edgeLookupTimeoutMs` | `10`    | L1 `cache.match` (the lookup)       |
+| `edgeReadTimeoutMs`   | `20`    | L1 body read (CF streams it lazily) |
+| `kvReadTimeoutMs`     | `170`   | L2 / KV read                        |
+
+Set any to `0` (or a negative value) to disable that budget and always await the
+read. A non-finite value (e.g. `Number(env.UNSET)`) falls back to the default.
+The tag-invalidation marker reads inherit these same budgets and **fail open** on
+a KV timeout — the entry is served rather than wrongly treated as invalidated.
+
+```typescript
+new CFCacheStore({
+  ctx,
+  kv: env.CACHE_KV,
+  defaults: { ttl: 60, swr: 300 },
+  // Raise a budget only if your HEALTHY reads legitimately run slower (large
+  // Flight payloads, far-from-colo regions); measure the p99 first. These are
+  // degradation guard-rails, not tuning levers for "slow is normal here".
+  kvReadTimeoutMs: 250,
+});
+```
+
+Failure handling, by kind — none of these fail the request:
+
+| Failure                         | Behavior                                                                                                                                          |
+| ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Transient read error (5xx/blip) | Degrade to the next tier; entry left intact                                                                                                       |
+| Read budget exceeded (timeout)  | Abandon the read, degrade to the next tier                                                                                                        |
+| Corrupt / unparseable L1 entry  | Reported corrupt; degrade to L2 (served if present). The L1 entry is evicted ONLY when L2 has no copy — so the evict can't race the L2→L1 promote |
+| Corrupt / unparseable KV entry  | Reported corrupt; evicted (self-heal) + render (no tier below it)                                                                                 |
+| Write failure                   | No-op (entry simply not cached); never throws                                                                                                     |
+
+Each is surfaced to the router's `onError` callback (phase `"cache"`, with
+`metadata.category` one of `cache-read`, `cache-corrupt`, `cache-write`,
+`cache-delete`, `cache-invalidate`, `stale-revalidation`) so you can observe
+cache health without affecting users.
+
+### Validating cache behavior with `debug`
+
+Pass `debug` to emit one structured event per L1 read — use it to confirm on a
+real deployment (via `wrangler tail`) that the store behaves as expected before
+relying on it. It is intended for validation, not steady-state production.
+
+```typescript
+new CFCacheStore({
+  ctx,
+  kv: env.CACHE_KV,
+  debug: true, // logs each CFCacheReadDebugEvent to the console
+  // ...or capture programmatically:
+  // debug: (event) => myTelemetry.record(event),
+});
+```
+
+Each event reports which tier answered and why (`outcome`: `l1-fresh`,
+`l1-stale-revalidate`, `l1-revalidating-guarded`, `match-timeout`, `match-error`,
+`body-timeout`, `body-error`, `non-200`, `tag-invalidated`, `l1-miss`, `kv-fresh`,
+`kv-stale`, `kv-stale-suppressed`, `kv-miss`, `kv-timeout`, `error`), the
+staleness / revalidating timestamps, and the measured per-tier durations:
+`matchMs` (the L1 `match`), `markerMs` (the tag-marker resolution tail for a
+tagged entry, between `matchMs` and `bodyReadMs`; absent or 0 for an untagged
+entry or a per-request memo hit), and `bodyReadMs` (the L1 body read). A
+persistently large `markerMs` signals a degraded KV namespace; on a healthy
+deployment KV keeps markers hot in its per-colo edge cache, so it stays a few
+milliseconds. `match-error` (a transient `cache.match` rejection that falls
+through to L2) is kept distinct from a plain `l1-miss`.
+
+## Cache purity & tainted objects
+
+A `cache()` boundary caches everything except loaders, so anything read inside a
+cached handler is **frozen into the shared cache entry** and served to every
+subsequent visitor. To stop one user's request-scoped data from leaking to
+another, request-scoped APIs are guarded inside a cache scope:
+
+| Inside a `cache()` boundary                                     | Behavior                                            |
+| --------------------------------------------------------------- | --------------------------------------------------- |
+| `cookies()` / `headers()` (read or write)                       | **throws** — request-scoped, would poison the entry |
+| `ctx.header()` / `setCookie()` / `setStatus()` / `onResponse()` | **throws** — response side effects lost on a hit    |
+| `ctx.get(var)` where the var is `{ cache: false }`              | **throws** on read                                  |
+| `ctx.set(var, value)` for a cacheable var                       | allowed (children are cached too)                   |
+| Any of the above **inside a loader**                            | **allowed** — loaders always run fresh              |
+
+**Tainted objects.** Request-scoped objects (`ctx`, `env`, `request`) carry an
+internal taint symbol so they are excluded from `"use cache"` cache keys, and
+the cache scope is tracked via async-local state. Two flags back the guards:
+`INSIDE_CACHE_EXEC` (set while a `"use cache"` function runs) and the `cache()`
+DSL scope (`isInsideCacheScope()`). `isInsideCacheScope()` deliberately returns
+`false` inside loaders — which is exactly why loaders are the dynamic holes:
+they may read `cookies()`/`headers()` and re-run on every request.
+
+The fix for "I need request data in a cached route": register a `loader()` and
+**consume it with `useLoader()` in a client component**. The loader is the
+dynamic hole — its data rides the fresh (never-cached) loader segment and is
+rendered in the client component, so it never lands in the cached shell.
+
+This is NOT the same as awaiting the loader in the handler. A cached handler
+that does `await ctx.use(Loader)` and renders the result bakes that per-request
+data straight into the shared cached segment — the loader running "fresh" does
+not help, because its output was inlined into the cached parent, and `ctx.use()`
+is **not** guarded. `ctx.use()` is a server-side escape hatch for non-rendered
+uses (set a ctx var, make a routing decision); never render its result inside a
+cached handler.
+
+```typescript
+// WRONG — throws: cookies() read directly in a cached handler
+cache({ ttl: 60 }, () => [
+  path("/me", () => <Profile id={cookies().get("uid")?.value} />),
+]);
+
+// ALSO WRONG (unguarded, but leaks) — the awaited loader data is rendered into
+// the cached handler, so the user's data is frozen into the shared shell.
+cache({ ttl: 60 }, () => [
+  path(
+    "/me",
+    async (ctx) => {
+      const { user } = await ctx.use(MeLoader); // runs fresh…
+      return <Profile user={user} />; // …but inlined into the CACHED segment → leak
+    },
+    { name: "me" },
+    () => [loader(MeLoader)],
+  ),
+]);
+
+// RIGHT — consume the loader in a CLIENT component via useLoader(). The cached
+// route segment holds only the <Profile/> reference; the user data rides the
+// fresh loader segment and renders client-side.
+
+// profile.tsx (client component)
+"use client";
+import { useLoader } from "@rangojs/router/client";
+
+export function Profile() {
+  const { user } = useLoader(MeLoader); // fresh per request; never cached
+  return <span>{user.name}</span>;
+}
+
+// urls — register the loader; MeLoader reads cookies() inside the loader (allowed)
+cache({ ttl: 60 }, () => [
+  path("/me", () => <Profile />, { name: "me" }, () => [loader(MeLoader)]),
+]);
+```
+
+See `/cache-guide` for the full decision guide and the `cache()` vs `"use cache"` comparison.
+
 ## Nested Cache Boundaries
 
 Override cache settings for specific sections:
@@ -209,6 +478,7 @@ cache({ store: checkoutCache }, () => [
 ```typescript
 import { urls } from "@rangojs/router";
 import { MemorySegmentCacheStore } from "@rangojs/router/cache";
+import * as CartActions from "./actions/cart";
 
 // Custom store for checkout (short TTL)
 const checkoutCache = new MemorySegmentCacheStore({
@@ -237,7 +507,7 @@ export const urlpatterns = urls(({ path, layout, cache, loader, revalidate }) =>
     path("/shop/product/:slug", ProductPage, { name: "product" }, () => [
       loader(ProductLoader, () => [cache({ ttl: 120 })]),
       loader(CartLoader, () => [
-        revalidate(({ actionId }) => actionId?.includes("Cart") ?? false),
+        revalidate((ctx) => ctx.isAction(CartActions) || undefined),
       ]),
     ]),
   ]),