autumn_web/cache/

mod.rs

//! Caching infrastructure for the Autumn framework.
//!
//! This module provides:
//!
//! - [`Cache`] — a trait abstracting over cache backends (moka by default,
//!   swap in Redis, memcached, etc.)
//! - [`MokaCache`] — the default, lock-free, in-process cache powered by
//!   [moka](https://docs.rs/moka) (behind the `cache-moka` feature)
//! - [`CacheResponseLayer`] — a Tower middleware that caches HTTP GET
//!   responses, usable via `#[intercept(CacheResponseLayer::new(...))]`
//! - [`CacheableResult`] — helper trait used by `#[cached(result)]` to
//!   only cache `Ok` values
//!
//! The `#[cached]` proc macro generates a per-function static `MokaCache`
//! for function-level memoization. The `CacheResponseLayer` operates at
//! the HTTP level using a shared `Arc<dyn Cache>`.
//!
//! # Swapping backends
//!
//! Implement the [`Cache`] trait for your backend:
//!
//! ```rust,ignore
//! use autumn_web::cache::Cache;
//!
//! #[derive(Clone)]
//! struct RedisCache { /* ... */ }
//!
//! impl Cache for RedisCache {
//!     fn get_value(&self, key: &str) -> Option<Box<dyn std::any::Any + Send + Sync>> { /* ... */ }
//!     fn insert_value(&self, key: &str, value: Box<dyn std::any::Any + Send + Sync>) { /* ... */ }
//!     fn invalidate(&self, key: &str) { /* ... */ }
//!     fn clear(&self) { /* ... */ }
//! }
//! ```

pub mod coherence;
#[cfg(feature = "maud")]
mod fragment;
mod layer;
#[cfg(feature = "cache-moka")]
mod moka_impl;
mod read_through;

#[cfg(feature = "maud")]
pub use fragment::{
    cache_fragment, cache_fragment_global, cache_fragment_global_in, cache_fragment_in,
};
pub use layer::{CacheResponseLayer, CacheResponseService};
#[cfg(feature = "cache-moka")]
pub use moka_impl::MokaCache;
pub use read_through::{
    CacheFillError, GetOrComputeOptions, ReadThroughMetrics, ReadThroughMetricsSnapshot,
    get_or_compute, get_or_compute_with, jittered_ttl, read_through_metrics,
};

use std::any::Any;
use std::hash::{DefaultHasher, Hash, Hasher};
use std::sync::{Arc, RwLock};
use std::time::Duration;

// ── Global cache registry ────────────────────────────────────────────

/// Process-level shared cache backend.
///
/// Set once at startup by [`set_global_cache`]; read by every
/// `#[cached]`-annotated function to decide which store to use.
static GLOBAL_CACHE: RwLock<Option<Arc<dyn Cache>>> = RwLock::new(None);

/// Register (or replace) the process-level shared cache.
///
/// Called automatically by [`crate::app::AppBuilder`] when
/// `.with_cache_backend(...)` has been used. Also called by
/// [`crate::state::AppState::set_cache`] when a plugin installs a backend
/// during the startup-hook phase.
///
/// With the `reporting` feature the backend is wrapped once, so a failure
/// capsule records its removals and a replay never reaches it. Thus
/// [`global_cache`] returns the wrapper, not the `Arc` given here.
///
/// # Panics
///
/// Panics if the internal `RwLock` is poisoned.
pub fn set_global_cache(cache: Arc<dyn Cache>) {
    *GLOBAL_CACHE.write().expect("global cache lock poisoned") = Some(with_capsule_seam(cache));
}

/// Return a clone of the process-level shared cache, if one is registered.
///
/// `None` means no global backend has been set and `#[cached]` functions
/// fall back to their per-function Moka stores.
///
/// # Panics
///
/// Panics if the internal `RwLock` is poisoned.
#[must_use]
pub fn global_cache() -> Option<Arc<dyn Cache>> {
    GLOBAL_CACHE
        .read()
        .expect("global cache lock poisoned")
        .clone()
}

/// Set the global cache for `autumn replay`, before the state initializers
/// run (#2351).
///
/// Replay builds no cache backend. When the app builder has one
/// (`with_cache_backend`), production installed it before the initializers,
/// so replay installs the capsule seam over a backend that stores nothing, in
/// the global cache and the app state. A cache call then takes the path it
/// took in production, and the tape answers it. Otherwise the global cache is
/// cleared, as the build clears it in production.
///
/// Returns the cache for the app state.
#[cfg(feature = "reporting")]
pub(crate) fn install_replay_cache(builder_cache: bool) -> Option<Arc<dyn Cache>> {
    if !builder_cache {
        clear_global_cache();
        return None;
    }
    // Wrapped once, so the global and the state hold the same `Arc`.
    let cache = with_capsule_seam(Arc::new(ReplayBackend));
    set_global_cache(Arc::clone(&cache));
    Some(cache)
}

/// After the state initializers: install the replay seam for a cache the
/// capsule recorded and nothing has installed yet. It came from a startup
/// hook (`RedisCachePlugin`), which replay does not run. Production's
/// initializers ran before that hook, so they did not see it either.
#[cfg(feature = "reporting")]
pub(crate) fn install_late_replay_cache(
    state: &crate::state::AppState,
    recorded: &crate::capsule::CapsuleEffects,
) {
    if recorded.state_cache && state.cache().is_none() {
        state.set_cache(Arc::new(ReplayBackend));
    }
}

/// A replay's answer to a coherence namespace removal (#2351).
///
/// The tape answers it, as it does for the installed cache, and no store is
/// reached: a registered store is not on the seam. During `autumn replay`
/// with no tape, nothing is reached either. `None` outside a replay.
#[cfg(feature = "reporting")]
pub(crate) fn replayed_namespace_removal(namespace: &str) -> Option<bool> {
    if let Some(tape) = crate::capsule::effects::current_tape() {
        return Some(tape.cache_invalidate_namespace(namespace).is_ok());
    }
    replay_blocked().then_some(true)
}

/// No capsule support compiled in: never a replay.
#[cfg(not(feature = "reporting"))]
pub(crate) const fn replayed_namespace_removal(_namespace: &str) -> Option<bool> {
    None
}

/// A coherence namespace removal's place on the capsule seam (#2351).
///
/// Coherence clears the registered stores, which are not on the seam, and
/// asks the installed cache. It returns `false` when either part fails. The
/// capsule records that combined answer once, as a replay returns it, so the
/// installed cache's own record of its part is suppressed: call it through
/// [`call_installed`](Self::call_installed).
#[cfg(feature = "reporting")]
pub(crate) struct NamespaceRemoval(Option<(Arc<crate::capsule::CaptureScope>, usize)>);

#[cfg(feature = "reporting")]
impl NamespaceRemoval {
    /// Take the removal's slot before any store is reached.
    pub(crate) fn reserve() -> Self {
        Self(
            crate::capsule::current_scope()
                .and_then(|scope| scope.reserve_cache().map(|index| (scope, index))),
        )
    }

    /// Call the installed cache without it recording the removal. A future
    /// it returns has already read the flag.
    pub(crate) fn call_installed<R>(call: impl FnOnce() -> R) -> R {
        let outer = COHERENCE_REMOVAL.with(|flag| flag.replace(true));
        let restore = RestoreFlag(outer);
        let result = call();
        drop(restore);
        result
    }

    /// Record the combined answer: `failure` names the part that failed.
    pub(crate) fn record(self, namespace: &str, failure: Option<&str>) {
        if let Some((scope, index)) = self.0 {
            scope.fill_cache(
                index,
                crate::capsule::CacheEffect::InvalidateNamespace {
                    namespace: namespace.to_owned(),
                    error: failure.map(|reason| crate::capsule::CacheInvalidationError {
                        attempts: 1,
                        reason: reason.to_owned(),
                    }),
                },
            );
        }
    }
}

// Set while coherence calls the installed cache for a removal it records.
#[cfg(feature = "reporting")]
thread_local! {
    static COHERENCE_REMOVAL: std::cell::Cell<bool> = const { std::cell::Cell::new(false) };
}

/// Restores [`COHERENCE_REMOVAL`], also when the backend panics.
#[cfg(feature = "reporting")]
struct RestoreFlag(bool);

#[cfg(feature = "reporting")]
impl Drop for RestoreFlag {
    fn drop(&mut self) {
        let outer = self.0;
        COHERENCE_REMOVAL.with(|flag| flag.set(outer));
    }
}

/// No capsule support compiled in: nothing to record.
#[cfg(not(feature = "reporting"))]
pub(crate) struct NamespaceRemoval;

#[cfg(not(feature = "reporting"))]
impl NamespaceRemoval {
    pub(crate) const fn reserve() -> Self {
        Self
    }

    pub(crate) fn call_installed<R>(call: impl FnOnce() -> R) -> R {
        call()
    }

    pub(crate) const fn record(self, _namespace: &str, _failure: Option<&str>) {}
}

/// The backend under the replay seam. It stores nothing.
#[cfg(feature = "reporting")]
struct ReplayBackend;

#[cfg(feature = "reporting")]
impl Cache for ReplayBackend {
    fn get_value(&self, _key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
        None
    }
    fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {}
    fn invalidate(&self, _key: &str) {}
    fn clear(&self) {}
    fn invalidate_namespace(&self, _namespace: &str) -> bool {
        true
    }
}

/// Remove the process-level shared cache.
///
/// Primarily useful in tests that need per-test isolation.
///
/// # Panics
///
/// Panics if the internal `RwLock` is poisoned.
pub fn clear_global_cache() {
    *GLOBAL_CACHE.write().expect("global cache lock poisoned") = None;
}

/// A boxed future returned by the async [`Cache`] methods.
///
/// Boxed so that `Cache` stays usable as `dyn Cache`.
pub type CacheFuture<'a, T> = std::pin::Pin<Box<dyn std::future::Future<Output = T> + Send + 'a>>;

/// An invalidation that did not complete. Stale data can still be served.
///
/// Returned by [`Cache::invalidate_async`] and
/// [`Cache::invalidate_namespace_async`].
#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
#[error("cache invalidation failed after {attempts} attempt(s): {reason}")]
pub struct InvalidationError {
    attempts: u32,
    reason: String,
}

impl InvalidationError {
    /// Make an error for an invalidation that failed after `attempts` tries.
    pub fn new(attempts: u32, reason: impl Into<String>) -> Self {
        Self {
            attempts,
            reason: reason.into(),
        }
    }

    /// How many times the backend tried.
    #[must_use]
    pub const fn attempts(&self) -> u32 {
        self.attempts
    }

    /// The last backend error, as text.
    #[must_use]
    pub fn reason(&self) -> &str {
        &self.reason
    }
}

/// A backend read or write that did not complete.
///
/// Returned by [`Cache::try_get_value`], [`Cache::try_insert_raw_bytes`],
/// [`try_get_cached`] and [`try_insert_cached`].
#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
#[error("cache backend failed: {reason}")]
pub struct CacheBackendError {
    reason: String,
}

impl CacheBackendError {
    /// Make an error with the backend's `reason`.
    pub fn new(reason: impl Into<String>) -> Self {
        Self {
            reason: reason.into(),
        }
    }

    /// The backend error, as text.
    #[must_use]
    pub fn reason(&self) -> &str {
        &self.reason
    }
}

/// Invalidations that failed after all retries, in this process.
static INVALIDATION_FAILURES: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);

/// Count one invalidation that failed after all retries.
///
/// Shown as `autumn_cache_invalidation_failures_total`. Call it only where you
/// drop the error. Do not call it where you return the error. Then each
/// failure counts one time. A custom backend calls it from a sync method that
/// cannot return the error.
pub fn record_invalidation_failure() {
    INVALIDATION_FAILURES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
}

/// Total of [`record_invalidation_failure`] calls in this process.
#[must_use]
pub fn invalidation_failures_total() -> u64 {
    INVALIDATION_FAILURES.load(std::sync::atomic::Ordering::Relaxed)
}

/// Serializes same-process test code that mutates [`GLOBAL_CACHE`] via
/// [`set_global_cache`]/[`clear_global_cache`].
///
/// `cargo test` runs a crate's unit tests on parallel threads in one
/// process, so two tests racing on this process-wide singleton can
/// interleave — e.g. one test's `clear_global_cache()` landing between
/// another's set-then-read sequence. Every same-process caller that
/// mutates the global cache for test purposes — [`crate::test::TestApp::build`]
/// and the `cache::fragment` unit tests — acquires this lock for the
/// duration of its critical section so they mutually exclude each other.
/// Poison-tolerant: acquire with `.lock().unwrap_or_else(PoisonError::into_inner)`
/// so one panicking test doesn't cascade into failing an unrelated one. See
/// issue #2218.
pub(crate) static GLOBAL_CACHE_TEST_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());

// ── Cache trait ──────────────────────────────────────────────────────

/// Raw JSON bytes stored by serializing cache backends (e.g. Redis).
///
/// Backends that cannot store `Arc<dyn Any>` directly (because values must
/// survive across process boundaries) return this from [`Cache::get_value`]
/// instead. [`get_cached`] and [`insert_cached`] transparently deserialize it
/// back into the concrete type `V` using `serde_json`.
#[derive(Clone)]
pub struct RawCacheBytes(pub Vec<u8>);

/// A type-erased, thread-safe cache store.
///
/// Implementations must be `Send + Sync` so they can be shared across
/// handlers and tasks. Values are stored as `Arc<dyn Any>` for type
/// erasure, allowing a single cache instance to store heterogeneous
/// types from different `#[cached]` functions.
///
/// Use the free functions [`get`] / [`insert`] for in-process-only values,
/// or [`get_cached`] / [`insert_cached`] for types that also implement
/// `serde`, which is required for cross-replica backends like Redis.
/// [`CacheResponseLayer`] uses the serde-aware path so HTTP response caching
/// works with both in-process and raw-byte backends.
pub trait Cache: Send + Sync + 'static {
    /// Retrieve a type-erased value by key. Returns `None` on miss.
    ///
    /// Backends that store serialized data (e.g. Redis) may return
    /// <code>Arc<[RawCacheBytes]></code> here; [`get_cached`] handles the
    /// JSON deserialization transparently.
    fn get_value(&self, key: &str) -> Option<Arc<dyn Any + Send + Sync>>;

    /// Store a type-erased value by key.
    fn insert_value(&self, key: &str, value: Arc<dyn Any + Send + Sync>);

    /// Remove a specific key.
    ///
    /// This method cannot report a failure. A backend that can fail must log
    /// the failure and call [`record_invalidation_failure`]. Prefer
    /// [`invalidate_async`](Cache::invalidate_async).
    fn invalidate(&self, key: &str);

    /// Remove all entries.
    fn clear(&self);

    /// Drop every entry belonging to one cached read — i.e. every key under
    /// `"{namespace}:"`, the prefix
    /// [`make_cache_key`] stamps on each of that read's keys.
    ///
    /// Returns whether the backend could actually do it. The default is
    /// `false`: a backend that cannot enumerate or pattern-match its key space
    /// must say so rather than silently leave stale entries behind, because the
    /// build-time coherence gate (issue #1716) reports this answer to the
    /// caller verbatim. Backends that CAN do it — `MokaCache` by iteration,
    /// `RedisCache` by `SCAN MATCH` — override it and return `true`.
    fn invalidate_namespace(&self, _namespace: &str) -> bool {
        false
    }

    /// Store pre-serialized JSON bytes for backends that persist data across
    /// process boundaries (e.g. Redis). The default is a no-op; in-process
    /// backends store values via [`insert_value`] instead.
    ///
    /// `ttl` carries the same time-to-live that was declared on the
    /// `#[cached(ttl = "…")]` attribute so backends can apply native expiry
    /// (e.g. Redis `SET EX`). `None` means no expiry.
    ///
    /// [`insert_value`]: Cache::insert_value
    fn insert_raw_bytes(&self, _key: &str, _bytes: Vec<u8>, _ttl: Option<std::time::Duration>) {}

    /// Whether this backend keeps a shared fill epoch.
    ///
    /// Return `true` if you override [`fill_epoch`](Cache::fill_epoch). Then a
    /// failed [`invalidate_namespace`](Cache::invalidate_namespace) on a store
    /// registered with `coherence::register_namespace_store` makes the
    /// invalidation incomplete. The default is `false`.
    fn shares_fill_epoch(&self) -> bool {
        false
    }

    /// Read the namespace's **shared** fill epoch.
    ///
    /// A cross-replica backend keeps one epoch for each namespace in the
    /// shared store.
    /// [`invalidate_namespace`](Cache::invalidate_namespace) must raise it
    /// **before** it sweeps. A fill reads it after a miss and before it
    /// computes. Then the fill inserts with
    /// [`insert_raw_bytes_if_epoch`](Cache::insert_raw_bytes_if_epoch). Then no
    /// fill can write back a value that an invalidation has dropped. This holds
    /// on all replicas.
    ///
    /// The default is [`FillEpoch::Unsupported`]. The fence stays per process.
    fn fill_epoch(&self, _namespace: &str) -> FillEpoch {
        FillEpoch::Unsupported
    }

    /// Store pre-serialized bytes only if the shared epoch still equals
    /// `sampled`. The check and the store must be one atomic step.
    ///
    /// Returns `true` when stored. A backend that overrides
    /// [`fill_epoch`](Cache::fill_epoch) must override this too. The default
    /// stores without a check.
    ///
    /// When the epoch is [`FillEpoch::Sampled`], `insert_value` is not called.
    /// Fill any local tier here.
    fn insert_raw_bytes_if_epoch(
        &self,
        key: &str,
        bytes: Vec<u8>,
        ttl: Option<std::time::Duration>,
        _namespace: &str,
        _sampled: u64,
    ) -> bool {
        self.insert_raw_bytes(key, bytes, ttl);
        true
    }

    /// Try to acquire a cross-replica fill lock for `key`, used by
    /// [`get_or_compute_with`] to ensure at most one replica refills a hot
    /// key at a time.
    ///
    /// `token` identifies the caller so [`release_fill_lock`] can safely
    /// release only a lock it still owns. `ttl` bounds how long the lock is
    /// held if the caller crashes before releasing it.
    ///
    /// The default implementation reports [`FillLockStatus::Unsupported`],
    /// which degrades callers to in-process-only single-flight protection —
    /// safe for backends (like the in-process Moka cache) that have no
    /// cross-replica visibility.
    ///
    /// [`release_fill_lock`]: Cache::release_fill_lock
    fn try_acquire_fill_lock(&self, _key: &str, _token: &str, _ttl: Duration) -> FillLockStatus {
        FillLockStatus::Unsupported
    }

    /// Release the fill lock previously acquired with `token`, if this
    /// caller still owns it. The default is a no-op, matching the default
    /// [`try_acquire_fill_lock`] returning [`FillLockStatus::Unsupported`].
    ///
    /// [`try_acquire_fill_lock`]: Cache::try_acquire_fill_lock
    fn release_fill_lock(&self, _key: &str, _token: &str) {}

    /// Remove a specific key, and return a failure.
    ///
    /// The default calls [`invalidate`](Cache::invalidate) and returns `Ok`.
    /// A backend that does I/O (for example Redis) overrides it with a real
    /// async call that does not block a runtime worker, and returns the error.
    ///
    /// # Errors
    ///
    /// [`InvalidationError`] when the key could not be removed. The old value
    /// can still be served.
    fn invalidate_async<'a>(
        &'a self,
        key: &'a str,
    ) -> CacheFuture<'a, Result<(), InvalidationError>> {
        Box::pin(async move {
            self.invalidate(key);
            Ok(())
        })
    }

    /// Read a value, and report a backend failure.
    ///
    /// [`get_value`](Cache::get_value) answers `None` for both a miss and an
    /// outage. This method keeps them apart. The default calls `get_value` and
    /// cannot fail. A backend that does I/O overrides it.
    ///
    /// # Errors
    ///
    /// [`CacheBackendError`] when the backend could not answer.
    fn try_get_value(
        &self,
        key: &str,
    ) -> Result<Option<Arc<dyn Any + Send + Sync>>, CacheBackendError> {
        Ok(self.get_value(key))
    }

    /// Store pre-serialized bytes, and report a backend failure.
    ///
    /// The default calls [`insert_raw_bytes`](Cache::insert_raw_bytes) and
    /// returns `Ok`. A backend that does I/O overrides it.
    ///
    /// # Errors
    ///
    /// [`CacheBackendError`] when the write did not land.
    fn try_insert_raw_bytes(
        &self,
        key: &str,
        bytes: Vec<u8>,
        ttl: Option<std::time::Duration>,
    ) -> Result<(), CacheBackendError> {
        self.insert_raw_bytes(key, bytes, ttl);
        Ok(())
    }

    /// Remove a key, and report a failure.
    ///
    /// This method blocks, like [`invalidate`](Cache::invalidate). In async
    /// code, use [`invalidate_async`](Cache::invalidate_async). The default
    /// calls `invalidate` and returns `Ok`.
    ///
    /// A decorator that wraps another cache must forward `try_get_value`,
    /// `try_insert_raw_bytes` and `try_invalidate`. If it does not, it uses the
    /// defaults, and the backend failures are lost.
    ///
    /// # Errors
    ///
    /// [`InvalidationError`] when the key could not be removed.
    fn try_invalidate(&self, key: &str) -> Result<(), InvalidationError> {
        self.invalidate(key);
        Ok(())
    }

    /// Whether this backend is the framework's failure-capsule wrapper.
    /// Framework internal. A decorator that wraps an installed backend
    /// forwards it, so the backend is not wrapped twice.
    #[doc(hidden)]
    fn is_capsule_seam(&self) -> bool {
        false
    }

    /// Async form of [`invalidate_namespace`](Cache::invalidate_namespace).
    ///
    /// The default calls the sync method and maps `false` to an error.
    ///
    /// # Errors
    ///
    /// [`InvalidationError`] when the backend cannot drop a namespace, or its
    /// sweep failed. Entries of the namespace can still be served.
    fn invalidate_namespace_async<'a>(
        &'a self,
        namespace: &'a str,
    ) -> CacheFuture<'a, Result<(), InvalidationError>> {
        Box::pin(async move {
            if self.invalidate_namespace(namespace) {
                Ok(())
            } else {
                Err(InvalidationError::new(
                    1,
                    "the backend cannot drop a namespace, or its sweep failed",
                ))
            }
        })
    }
}

/// A namespace's shared fill epoch, as read by [`Cache::fill_epoch`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum FillEpoch {
    /// The backend has no shared epoch. Only the per-process fence applies.
    Unsupported,
    /// The epoch at the time of the read.
    Sampled(u64),
    /// The backend has a shared epoch but could not read it. A fill must not
    /// insert. The result is one extra cache miss.
    Unavailable,
}

/// Outcome of [`Cache::try_acquire_fill_lock`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum FillLockStatus {
    /// The caller now holds the lock and must call
    /// [`Cache::release_fill_lock`] when the fill completes (success or
    /// failure).
    Acquired,
    /// Another replica currently holds the lock.
    Held,
    /// This backend has no cross-replica fill lock; callers fall back to
    /// in-process-only single-flight protection.
    Unsupported,
}

// ── Typed convenience functions ──────────────────────────────────────

/// Typed get: retrieve and downcast a cached value.
///
/// Returns `None` if the key is absent or the stored type doesn't
/// match `V`. Works with any `Cache` implementation.
///
/// For cross-replica backends (Redis) use [`get_cached`] instead, which
/// also handles JSON deserialization of [`RawCacheBytes`].
pub fn get<V: Clone + Send + Sync + 'static>(cache: &dyn Cache, key: &str) -> Option<V> {
    // Failure-capsule seam (#2351 item 2). The value is not serializable, so
    // a replay serves only a recorded miss, and capture can record only a
    // miss.
    if replayed_untyped_get(key) {
        return None;
    }
    let value = helper_read(cache, key).and_then(|arc| arc.downcast_ref::<V>().cloned());
    record_untyped_get(key, value.is_some());
    value
}

/// Read through a helper that records the read. The capsule seam takes any
/// other `get_value` call as a direct read that it cannot record.
fn helper_read(cache: &dyn Cache, key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
    #[cfg(feature = "reporting")]
    let _helper = HelperCall::enter();
    cache.get_value(key)
}

// Set while a recording helper reads or writes, so the seam can tell a
// direct call.
#[cfg(feature = "reporting")]
thread_local! {
    static HELPER_CALL: std::cell::Cell<bool> = const { std::cell::Cell::new(false) };
}

/// Marks a helper call for its lifetime, also when the backend panics.
#[cfg(feature = "reporting")]
struct HelperCall(bool);

#[cfg(feature = "reporting")]
impl HelperCall {
    fn enter() -> Self {
        Self(HELPER_CALL.with(|flag| flag.replace(true)))
    }
}

#[cfg(feature = "reporting")]
impl Drop for HelperCall {
    fn drop(&mut self) {
        let outer = self.0;
        HELPER_CALL.with(|flag| flag.set(outer));
    }
}

/// Why a capsule with a direct cache write is not replayable.
#[cfg(feature = "reporting")]
const DIRECT_WRITE_NOTE: &str = "the run wrote the cache with `Cache::insert_value` or \
     `insert_raw_bytes` directly; the write is not recorded, so replay cannot check it";

/// Why a capsule with a direct cache read is not replayable.
#[cfg(feature = "reporting")]
const DIRECT_READ_NOTE: &str = "the run read the cache with `Cache::get_value` directly; the \
     value is not recorded, so replay cannot serve it";

/// Typed insert: wrap the value in an `Arc` and store it.
///
/// Works with any `Cache` implementation.
///
/// For cross-replica backends (Redis) use [`insert_cached`] instead,
/// which also serializes the value for storage across process boundaries.
pub fn insert<V: Clone + Send + Sync + 'static>(cache: &dyn Cache, key: &str, value: V) {
    // Failure-capsule seam (#2351 item 2): a replay never writes, and capture
    // cannot record the value.
    if record_or_replay_untyped_insert(key) {
        return;
    }
    #[cfg(feature = "reporting")]
    let _helper = HelperCall::enter();
    cache.insert_value(key, Arc::new(value));
}

/// Serde-aware get: retrieve a cached value, deserializing from JSON if needed.
///
/// First tries a direct in-memory downcast (fast path for `MokaCache`). If
/// that fails — because the backend stored [`RawCacheBytes`] (e.g. Redis) —
/// the bytes are deserialized with `serde_json`. This is what the `#[cached]`
/// macro uses so that values survive across replicas when a shared backend
/// is configured.
pub fn get_cached<V>(cache: &dyn Cache, key: &str) -> Option<V>
where
    V: Clone + serde::Serialize + serde::de::DeserializeOwned + Send + Sync + 'static,
{
    // Failure-capsule seam (#1634). A replay is served entirely from the
    // capsule: the live backend is never consulted, because the value it holds
    // now is not the value the failing request read.
    match replayed_cache_get::<V>(key) {
        ReplayedRead::NoTape => {}
        ReplayedRead::Miss => return None,
        ReplayedRead::Hit(value) => return Some(value),
    }
    let arc = helper_read(cache, key);
    let value = arc.and_then(|arc| {
        // Fast path: in-memory backend stored the concrete type directly.
        if let Some(value) = arc.downcast_ref::<V>() {
            return Some(value.clone());
        }
        // Slow path: serializing backend (e.g. Redis) stored RawCacheBytes.
        arc.downcast_ref::<RawCacheBytes>()
            .and_then(|raw| serde_json::from_slice::<V>(&raw.0).ok())
    });
    // Recorded from the *typed* value rather than from the stored `Arc<dyn
    // Any>`: the in-process backend (Moka) stores `Arc<V>` and never
    // `RawCacheBytes`, so reading the erased value would record every hit from
    // the default backend as valueless — and replay would then take the miss
    // branch of a handler whose bug is on the hit branch. The `Serialize`
    // bound is what makes the recording possible at all; every caller already
    // pairs this with `insert_cached`, which requires it too.
    record_cache_get(key, value.as_ref());
    value
}

/// Serde-aware insert: store the value both in-memory and as JSON bytes.
///
/// Calls [`Cache::insert_value`] (for in-process backends like Moka) **and**
/// [`Cache::insert_raw_bytes`] (for cross-replica backends like Redis). This
/// is what the `#[cached]` macro uses so that the stored value is accessible
/// both within the same process and on other replicas.
///
/// `ttl` is forwarded verbatim to [`Cache::insert_raw_bytes`] so backends
/// like Redis can apply a native entry expiry (e.g. `SET EX`). In-process
/// backends (Moka) manage TTL via the per-function static cache instance
/// and ignore this parameter.
pub fn insert_cached<V>(cache: &dyn Cache, key: &str, value: V, ttl: Option<std::time::Duration>)
where
    V: Clone + serde::Serialize + Send + Sync + 'static,
{
    let bytes = serde_json::to_vec(&value).ok();
    // Failure-capsule seam (#1634): the write is recorded on capture, and on
    // replay it lands in the tape (so a read-back in the same run finds it)
    // instead of in a live backend.
    if record_or_replay_cache_insert(key, bytes.as_deref(), ttl) {
        return;
    }
    #[cfg(feature = "reporting")]
    let _helper = HelperCall::enter();
    // In-memory path (MokaCache, CountingCache in tests, …)
    cache.insert_value(key, Arc::new(value));
    // Serialized path (RedisCache, any cross-replica backend)
    if let Some(bytes) = bytes {
        cache.insert_raw_bytes(key, bytes, ttl);
    }
}

/// [`get_cached`], with a backend failure kept apart from a miss.
///
/// Unlike [`get_cached`], stored bytes that cannot be decoded are an error,
/// not a miss. The key stays unreadable until a writer replaces it. A failed
/// read marks a failure capsule incomplete, so replay does not run it.
///
/// # Errors
///
/// [`CacheBackendError`] when the backend could not answer, or the stored
/// bytes cannot be decoded.
pub fn try_get_cached<V>(cache: &dyn Cache, key: &str) -> Result<Option<V>, CacheBackendError>
where
    V: Clone + serde::Serialize + serde::de::DeserializeOwned + Send + Sync + 'static,
{
    match replayed_cache_get::<V>(key) {
        ReplayedRead::NoTape => {}
        ReplayedRead::Miss => return Ok(None),
        ReplayedRead::Hit(value) => return Ok(Some(value)),
    }
    let read = {
        #[cfg(feature = "reporting")]
        let _helper = HelperCall::enter();
        cache.try_get_value(key)
    };
    let found = match read {
        Ok(found) => found,
        Err(error) => {
            mark_failed_cache_op(FAILED_READ_NOTE);
            return Err(error);
        }
    };
    let Some(arc) = found else {
        record_cache_get::<V>(key, None);
        return Ok(None);
    };
    if let Some(value) = arc.downcast_ref::<V>() {
        let value = Some(value.clone());
        record_cache_get(key, value.as_ref());
        return Ok(value);
    }
    // Bytes the backend holds but this type cannot read are a fault, not an
    // absent key. The text is fixed: a guest can read it.
    let value = if let Some(raw) = arc.downcast_ref::<RawCacheBytes>() {
        let Ok(value) = serde_json::from_slice::<V>(&raw.0) else {
            mark_failed_cache_op(FAILED_READ_NOTE);
            return Err(CacheBackendError::new("the stored value cannot be decoded"));
        };
        Some(value)
    } else {
        None
    };
    record_cache_get(key, value.as_ref());
    Ok(value)
}

/// [`insert_cached`], with a backend failure reported.
///
/// # Errors
///
/// [`CacheBackendError`] when the value cannot be serialized, or the backend
/// did not store it.
pub fn try_insert_cached<V>(
    cache: &dyn Cache,
    key: &str,
    value: V,
    ttl: Option<std::time::Duration>,
) -> Result<(), CacheBackendError>
where
    V: Clone + serde::Serialize + Send + Sync + 'static,
{
    let bytes = serde_json::to_vec(&value).map_err(|error| {
        CacheBackendError::new(format!("the value cannot be serialized: {error}"))
    })?;
    if record_or_replay_cache_insert(key, Some(&bytes), ttl) {
        return Ok(());
    }
    #[cfg(feature = "reporting")]
    let _helper = HelperCall::enter();
    cache.insert_value(key, Arc::new(value));
    let stored = cache.try_insert_raw_bytes(key, bytes, ttl);
    if stored.is_err() {
        // The write is already recorded, but it did not land.
        mark_failed_cache_op(FAILED_WRITE_NOTE);
    }
    stored
}

/// Why a capsule with a failed cache read is not replayable.
const FAILED_READ_NOTE: &str = "a cache read failed; the failure is not recorded, so replay \
     cannot take the same path";

/// Why a capsule with a failed cache write is not replayable.
const FAILED_WRITE_NOTE: &str = "a cache write failed; the capsule recorded it as stored, so \
     replay cannot take the same path";

/// Mark the capture scope incomplete after a failed cache read or write.
#[cfg(feature = "reporting")]
fn mark_failed_cache_op(note: &str) {
    if let Some(scope) = crate::capsule::current_scope() {
        scope.note(note);
        scope.mark_truncated();
    }
}

/// No capsule support compiled in: nothing records the failure.
#[cfg(not(feature = "reporting"))]
const fn mark_failed_cache_op(_note: &str) {}

/// Like [`insert_cached`], but store only if the shared epoch did not change.
///
/// `epoch` is the [`Cache::fill_epoch`] read after the miss and before the
/// compute. Returns `true` when the value was stored. `false` means the fill
/// was fenced out, or the epoch was [`FillEpoch::Unavailable`].
///
/// With [`FillEpoch::Unsupported`] this is [`insert_cached`].
pub fn insert_cached_fenced<V>(
    cache: &dyn Cache,
    key: &str,
    value: V,
    ttl: Option<std::time::Duration>,
    namespace: &str,
    epoch: FillEpoch,
) -> bool
where
    V: Clone + serde::Serialize + Send + Sync + 'static,
{
    // A replay has no fence: `insert_cached` sends the write to the tape, or
    // drops it during `autumn replay` with no tape. Only a sampled epoch
    // reaches the backend, so only it is bypassed, and only after the value
    // is encoded, as the fenced path does.
    let epoch = match epoch {
        FillEpoch::Sampled(_) if replaying() => {
            if serde_json::to_vec(&value).is_err() {
                return false;
            }
            FillEpoch::Unsupported
        }
        other => other,
    };
    let sampled = match epoch {
        FillEpoch::Unsupported => {
            insert_cached(cache, key, value, ttl);
            return true;
        }
        FillEpoch::Unavailable => return false,
        FillEpoch::Sampled(sampled) => sampled,
    };
    let Some(bytes) = serde_json::to_vec(&value).ok() else {
        return false;
    };
    // Raw bytes only: `insert_value` would write without the check.
    let stored = {
        #[cfg(feature = "reporting")]
        let _helper = HelperCall::enter();
        cache.insert_raw_bytes_if_epoch(key, bytes.clone(), ttl, namespace, sampled)
    };
    if stored {
        // Capture records the write after it happened. A replay never gets
        // here.
        let _ = record_or_replay_cache_insert(key, Some(&bytes), ttl);
    }
    stored
}

/// Read the shared fill epoch for a fill that just missed.
///
/// A replay serves cache effects from the tape (#1634), so it never reads the
/// backend. It gets [`FillEpoch::Unsupported`], and [`insert_cached_fenced`]
/// then writes to the tape. During `autumn replay`, code with no tape also
/// gets `Unsupported`, and the write is dropped.
#[must_use]
pub fn sample_fill_epoch(cache: &dyn Cache, namespace: &str) -> FillEpoch {
    if replaying() {
        return FillEpoch::Unsupported;
    }
    cache.fill_epoch(namespace)
}

/// Whether an effect tape serves the current task, or this process replays a
/// capsule.
///
/// Not `tape_active`: that marks the replay scope as entered. Only the
/// reporting layer may do that.
#[cfg(feature = "reporting")]
fn replaying() -> bool {
    crate::capsule::effects::current_tape().is_some() || replay_blocked()
}

/// No capsule support compiled in: there is no replay.
#[cfg(not(feature = "reporting"))]
const fn replaying() -> bool {
    false
}

// ── Failure-capsule seam (#1634) ─────────────────────────────────────────────
//
// The `capsule` module is behind the `reporting` feature, so each helper here
// has a no-op twin for builds without it — the seam stays one line at each
// call site whatever the feature set.

/// What a replay has to say about a cache read.
///
/// A named type rather than a nested `Option`: "no replay is in progress" and
/// "the replay says this key missed" are opposite instructions to the caller,
/// and `Option<Option<V>>` spells them the same way round as each other.
enum ReplayedRead<V> {
    /// No replay is in progress; read the real cache.
    NoTape,
    /// The tape has no value to serve — a recorded miss, an unrecorded key
    /// (already logged as a divergence), or a hit the recording could not
    /// serialize. The read replays as a miss.
    Miss,
    /// The recorded hit.
    Hit(V),
}

/// Serve a cache read from the capsule's effect tape, when one is active.
#[cfg(feature = "reporting")]
fn replayed_cache_get<V>(key: &str) -> ReplayedRead<V>
where
    V: Clone + serde::de::DeserializeOwned + Send + Sync + 'static,
{
    use crate::capsule::effects::CachedValue;
    let Some(tape) = crate::capsule::effects::current_tape() else {
        // Replay startup code has no tape. A raw backend it holds must not
        // be read, so the read is a miss.
        if replay_blocked() {
            return ReplayedRead::Miss;
        }
        return ReplayedRead::NoTape;
    };
    match tape.cache_get(key) {
        CachedValue::Hit(bytes) => {
            serde_json::from_slice::<V>(&bytes).map_or(ReplayedRead::Miss, ReplayedRead::Hit)
        }
        CachedValue::Miss | CachedValue::Unrecorded => ReplayedRead::Miss,
    }
}

/// No capsule support compiled in: never a replay.
#[cfg(not(feature = "reporting"))]
const fn replayed_cache_get<V>(_key: &str) -> ReplayedRead<V>
where
    V: Clone + serde::de::DeserializeOwned + Send + Sync + 'static,
{
    ReplayedRead::NoTape
}

/// Tee a cache read into the in-flight request's capsule.
///
/// A miss is recorded as a *keyed* entry with no value: replay then knows the
/// key was read (so it does not call it unrecorded) while still being honest
/// that it has nothing to serve, and the handler takes its miss branch.
///
/// A **hit** whose value will not serialize cannot be recorded that way. It
/// would be indistinguishable from a miss, so replay would take the miss branch
/// without a divergence and grade a run that never happened — the handler read
/// a value in production and reads nothing here. There is no honest recording
/// of a value that cannot be encoded, so the capsule declares itself
/// incomplete instead.
#[cfg(feature = "reporting")]
fn record_cache_get<V: serde::Serialize>(key: &str, value: Option<&V>) {
    let Some(scope) = crate::capsule::current_scope() else {
        return;
    };
    let encoded = value.map(|value| {
        serde_json::to_vec(value).map(|bytes| {
            use base64::Engine as _;
            base64::engine::general_purpose::STANDARD.encode(&bytes)
        })
    });
    let encoded = match encoded {
        None => None,
        Some(Ok(encoded)) => Some(encoded),
        Some(Err(_)) => {
            scope.note(UNRECORDABLE_CACHE_HIT_NOTE);
            scope.mark_truncated();
            None
        }
    };
    scope.record_cache(crate::capsule::CacheEffect::Get {
        key: key.to_owned(),
        value: encoded,
    });
}

/// Why a capsule that wrote an unserializable cache value is not replayable.
#[cfg(feature = "reporting")]
const UNRECORDABLE_CACHE_WRITE_NOTE: &str = "a cache write held a value that could not be serialized into the capsule, so replay would \
     suppress a mutation the recorded run really made";

/// Why a capsule that read an unserializable cache hit is not replayable.
#[cfg(feature = "reporting")]
const UNRECORDABLE_CACHE_HIT_NOTE: &str = "a cache read hit a value that could not be serialized into the capsule, and a hit with no \
     recorded value is indistinguishable from a miss; replay would take the miss branch the \
     recorded run never took";

/// No capsule support compiled in: nothing to record.
#[cfg(not(feature = "reporting"))]
const fn record_cache_get<V: serde::Serialize>(_key: &str, _value: Option<&V>) {}

/// Record a cache write, or divert it into the replay tape.
///
/// Returns `true` when a replay handled the write and the live backend must
/// not be touched.
#[cfg(feature = "reporting")]
fn record_or_replay_cache_insert(
    key: &str,
    bytes: Option<&[u8]>,
    ttl: Option<std::time::Duration>,
) -> bool {
    if let Some(tape) = crate::capsule::effects::current_tape() {
        if let Some(bytes) = bytes {
            tape.cache_insert(key, bytes, ttl.map(|ttl| ttl.as_secs()));
        }
        return true;
    }
    // Replay startup code has no tape. A raw backend it holds must not be
    // written.
    if replay_blocked() {
        return true;
    }
    if let Some(scope) = crate::capsule::current_scope() {
        if let Some(bytes) = bytes {
            use base64::Engine as _;
            scope.record_cache(crate::capsule::CacheEffect::Insert {
                key: key.to_owned(),
                value: base64::engine::general_purpose::STANDARD.encode(bytes),
                ttl_secs: ttl.map(|ttl| ttl.as_secs()),
            });
        } else {
            // The live backend accepted this write; the capsule cannot hold it.
            // Recording nothing would let replay suppress a real cache mutation
            // and still grade the run clean, which is the same falsification the
            // unserializable *read* refuses — caught here a function later.
            scope.note(UNRECORDABLE_CACHE_WRITE_NOTE);
            scope.mark_truncated();
        }
    }
    false
}

/// No capsule support compiled in: the live backend always takes the write.
#[cfg(not(feature = "reporting"))]
const fn record_or_replay_cache_insert(
    _key: &str,
    _bytes: Option<&[u8]>,
    _ttl: Option<std::time::Duration>,
) -> bool {
    false
}

/// Serve an untyped read during a replay. Returns `true` when a replay
/// handled it; the read is then a miss.
#[cfg(feature = "reporting")]
fn replayed_untyped_get(key: &str) -> bool {
    let Some(tape) = crate::capsule::effects::current_tape() else {
        // Replay startup code: a raw backend is not read.
        return replay_blocked();
    };
    tape.cache_untyped_get(key);
    true
}

/// No capsule support compiled in: never a replay.
#[cfg(not(feature = "reporting"))]
const fn replayed_untyped_get(_key: &str) -> bool {
    false
}

/// Record an untyped read. A miss records like a typed miss. A hit has no
/// value the capsule can hold, so the capsule is marked incomplete.
#[cfg(feature = "reporting")]
fn record_untyped_get(key: &str, hit: bool) {
    let Some(scope) = crate::capsule::current_scope() else {
        return;
    };
    if hit {
        scope.note(UNRECORDABLE_CACHE_HIT_NOTE);
        scope.mark_truncated();
    }
    scope.record_cache(crate::capsule::CacheEffect::Get {
        key: key.to_owned(),
        value: None,
    });
}

/// No capsule support compiled in: nothing to record.
#[cfg(not(feature = "reporting"))]
const fn record_untyped_get(_key: &str, _hit: bool) {}

/// Record an untyped write, or divert it during a replay. Returns `true` when
/// a replay handled it and the backend must not be touched.
#[cfg(feature = "reporting")]
fn record_or_replay_untyped_insert(key: &str) -> bool {
    if let Some(tape) = crate::capsule::effects::current_tape() {
        tape.cache_untyped_insert(key);
        return true;
    }
    // Replay startup code: a raw backend is not written.
    if replay_blocked() {
        return true;
    }
    if let Some(scope) = crate::capsule::current_scope() {
        scope.note(UNRECORDABLE_CACHE_WRITE_NOTE);
        scope.mark_truncated();
    }
    false
}

/// No capsule support compiled in: the backend always takes the write.
#[cfg(not(feature = "reporting"))]
const fn record_or_replay_untyped_insert(_key: &str) -> bool {
    false
}

// ── Capsule seam for the removal methods (#2351 item 2) ──────────────────

/// Wrap `cache` so a capsule records its removals and a replay never reaches
/// it. Idempotent: a wrapped backend is returned as it is, so `Arc` identity
/// (which single-flight keys on) stays stable.
#[cfg(feature = "reporting")]
#[must_use]
pub(crate) fn with_capsule_seam(cache: Arc<dyn Cache>) -> Arc<dyn Cache> {
    if cache.is_capsule_seam() {
        return cache;
    }
    Arc::new(CapsuleSeamCache(cache))
}

/// No capsule support compiled in: the backend is used as it is.
#[cfg(not(feature = "reporting"))]
#[must_use]
pub(crate) fn with_capsule_seam(cache: Arc<dyn Cache>) -> Arc<dyn Cache> {
    cache
}

/// The installed backend, with the removal methods on the capsule seam.
///
/// Reads and writes pass through: `get_cached` / `insert_cached` and the
/// untyped functions carry their own seam.
#[cfg(feature = "reporting")]
struct CapsuleSeamCache(Arc<dyn Cache>);

#[cfg(feature = "reporting")]
impl CapsuleSeamCache {
    /// Check a write that reaches the seam. A helper write with a tape never
    /// gets here, so with a tape this is a direct write: a divergence, as
    /// capture refuses such a capsule. Returns `true` when the backend must
    /// not be written.
    fn direct_write_blocked(key: &str) -> bool {
        let direct = !HELPER_CALL.with(std::cell::Cell::get);
        if let Some(tape) = crate::capsule::effects::current_tape() {
            if direct {
                tape.cache_untyped_insert(key);
            }
            return true;
        }
        if replay_blocked() {
            return true;
        }
        if direct && let Some(scope) = crate::capsule::current_scope() {
            scope.note(DIRECT_WRITE_NOTE);
            scope.mark_truncated();
        }
        false
    }

    /// Mark the capsule incomplete: the fill fence stopped a fill.
    fn fence_stopped_fill() {
        if let Some(scope) = crate::capsule::current_scope() {
            scope.note(FILL_FENCE_NOTE);
            scope.mark_truncated();
        }
    }

    /// Take a tape slot for a removal, in call order.
    fn reserve() -> Option<(Arc<crate::capsule::CaptureScope>, usize)> {
        crate::capsule::current_scope()
            .and_then(|scope| scope.reserve_cache().map(|index| (scope, index)))
    }

    /// Fill a slot taken with [`reserve`](Self::reserve).
    fn fill(
        slot: Option<(Arc<crate::capsule::CaptureScope>, usize)>,
        effect: crate::capsule::CacheEffect,
    ) {
        if let Some((scope, index)) = slot {
            scope.fill_cache(index, effect);
        }
    }

    /// Run a removal that returns no result. A backend counts a failed one
    /// with [`record_invalidation_failure`]. When the count goes up, replay
    /// cannot know the result, so the capsule is marked incomplete. A failure
    /// on another task can also do this; replay then only refuses.
    fn unchecked_removal(remove: impl FnOnce()) {
        let before = invalidation_failures_total();
        remove();
        if invalidation_failures_total() != before
            && let Some(scope) = crate::capsule::current_scope()
        {
            scope.note(UNCHECKED_REMOVAL_NOTE);
            scope.mark_truncated();
        }
    }
}

/// Set while this process replays a capsule. The capsule seam then never
/// reaches a backend, also on a task with no tape (a state initializer, a
/// detached task). It is never unset.
#[cfg(feature = "reporting")]
static REPLAY_BLOCKED: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);

/// Keep every cache backend offline from now on. `autumn replay` calls it
/// before it builds the app (#2351).
#[cfg(feature = "reporting")]
pub(crate) fn block_backends_for_replay() {
    REPLAY_BLOCKED.store(true, std::sync::atomic::Ordering::SeqCst);
}

#[cfg(feature = "reporting")]
fn replay_blocked() -> bool {
    #[cfg(test)]
    if TEST_REPLAY_BLOCKED.with(std::cell::Cell::get) {
        return true;
    }
    REPLAY_BLOCKED.load(std::sync::atomic::Ordering::SeqCst)
}

// The process-wide block cannot be set in a unit test: it would block every
// other test. This thread's block stands in for it.
#[cfg(all(test, feature = "reporting"))]
thread_local! {
    static TEST_REPLAY_BLOCKED: std::cell::Cell<bool> = const { std::cell::Cell::new(false) };
}

/// Whether the seam must not reach the backend: a tape is active, or this
/// process replays a capsule.
#[cfg(feature = "reporting")]
fn offline() -> bool {
    // Not `tape_active`: that marks the replay scope as entered. Only the
    // reporting layer may do that.
    crate::capsule::effects::current_tape().is_some() || replay_blocked()
}

/// Why a capsule with a removal of unknown result is not replayable.
#[cfg(feature = "reporting")]
const UNCHECKED_REMOVAL_NOTE: &str = "a cache removal with no result reported a failure; \
     replay cannot know whether the entry stayed";

/// Why a capsule that met a distributed fill lock is not replayable.
#[cfg(feature = "reporting")]
const FILL_LOCK_NOTE: &str = "a cache fill met a distributed fill lock; the lock outcome \
     is not recorded, so replay cannot take the same path";

/// Why a capsule whose fill the shared fence stopped is not replayable.
#[cfg(feature = "reporting")]
const FILL_FENCE_NOTE: &str = "the shared fill fence stopped a cache fill; replay \
     has no fence, so it cannot take the same path";

/// A recorded [`InvalidationError`].
#[cfg(feature = "reporting")]
fn recorded_error(
    result: &Result<(), InvalidationError>,
) -> Option<crate::capsule::CacheInvalidationError> {
    result
        .as_ref()
        .err()
        .map(|error| crate::capsule::CacheInvalidationError {
            attempts: error.attempts(),
            reason: error.reason().to_owned(),
        })
}

#[cfg(feature = "reporting")]
impl Cache for CapsuleSeamCache {
    // During a replay the backend is never reached. A typed or untyped read
    // or write is answered by its own seam before it gets here; anything
    // that still arrives is a miss, a dropped write, or no fill lock.
    fn get_value(&self, key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
        let direct = !HELPER_CALL.with(std::cell::Cell::get);
        // A helper read with a tape never gets here, so with a tape this is a
        // direct read: a divergence, as capture refuses such a capsule. It
        // does not consume a read a helper recorded.
        if let Some(tape) = crate::capsule::effects::current_tape() {
            if direct {
                tape.cache_direct_get(key);
            }
            return None;
        }
        if replay_blocked() {
            return None;
        }
        if direct && let Some(scope) = crate::capsule::current_scope() {
            scope.note(DIRECT_READ_NOTE);
            scope.mark_truncated();
        }
        self.0.get_value(key)
    }

    fn insert_value(&self, key: &str, value: Arc<dyn Any + Send + Sync>) {
        if Self::direct_write_blocked(key) {
            return;
        }
        self.0.insert_value(key, value);
    }

    fn try_get_value(
        &self,
        key: &str,
    ) -> Result<Option<Arc<dyn Any + Send + Sync>>, CacheBackendError> {
        let direct = !HELPER_CALL.with(std::cell::Cell::get);
        if let Some(tape) = crate::capsule::effects::current_tape() {
            if direct {
                tape.cache_direct_get(key);
            }
            return Ok(None);
        }
        if replay_blocked() {
            return Ok(None);
        }
        if direct && let Some(scope) = crate::capsule::current_scope() {
            scope.note(DIRECT_READ_NOTE);
            scope.mark_truncated();
        }
        self.0.try_get_value(key)
    }

    fn try_insert_raw_bytes(
        &self,
        key: &str,
        bytes: Vec<u8>,
        ttl: Option<Duration>,
    ) -> Result<(), CacheBackendError> {
        if Self::direct_write_blocked(key) {
            return Ok(());
        }
        self.0.try_insert_raw_bytes(key, bytes, ttl)
    }

    fn try_invalidate(&self, key: &str) -> Result<(), InvalidationError> {
        if let Some(tape) = crate::capsule::effects::current_tape() {
            return tape.cache_invalidate(key);
        }
        if replay_blocked() {
            return Ok(());
        }
        // `unchecked_removal` covers a backend that keeps the default
        // `try_invalidate` and counts its failure with
        // `record_invalidation_failure`. A returned error is recorded below.
        let slot = Self::reserve();
        let mut result = Ok(());
        Self::unchecked_removal(|| result = self.0.try_invalidate(key));
        Self::fill(
            slot,
            crate::capsule::CacheEffect::Invalidate {
                key: key.to_owned(),
                error: recorded_error(&result),
            },
        );
        result
    }

    fn invalidate(&self, key: &str) {
        if let Some(tape) = crate::capsule::effects::current_tape() {
            let _ = tape.cache_invalidate(key);
            return;
        }
        if replay_blocked() {
            return;
        }
        // The slot is taken now and filled after the call, so a backend that
        // panics leaves it unfilled and the capsule incomplete.
        let slot = Self::reserve();
        Self::unchecked_removal(|| self.0.invalidate(key));
        Self::fill(
            slot,
            crate::capsule::CacheEffect::Invalidate {
                key: key.to_owned(),
                error: None,
            },
        );
    }

    fn clear(&self) {
        if let Some(tape) = crate::capsule::effects::current_tape() {
            tape.cache_clear();
            return;
        }
        if replay_blocked() {
            return;
        }
        let slot = Self::reserve();
        Self::unchecked_removal(|| self.0.clear());
        Self::fill(slot, crate::capsule::CacheEffect::Clear);
    }

    fn invalidate_namespace(&self, namespace: &str) -> bool {
        if let Some(tape) = crate::capsule::effects::current_tape() {
            return tape.cache_invalidate_namespace(namespace).is_ok();
        }
        if replay_blocked() {
            return true;
        }
        // Coherence records the removal with its registered stores' part.
        if COHERENCE_REMOVAL.with(std::cell::Cell::get) {
            return self.0.invalidate_namespace(namespace);
        }
        let slot = Self::reserve();
        let done = self.0.invalidate_namespace(namespace);
        let result = if done {
            Ok(())
        } else {
            Err(InvalidationError::new(
                1,
                "the backend cannot drop a namespace, or its sweep failed",
            ))
        };
        Self::fill(
            slot,
            crate::capsule::CacheEffect::InvalidateNamespace {
                namespace: namespace.to_owned(),
                error: recorded_error(&result),
            },
        );
        done
    }

    fn insert_raw_bytes(&self, key: &str, bytes: Vec<u8>, ttl: Option<Duration>) {
        if Self::direct_write_blocked(key) {
            return;
        }
        self.0.insert_raw_bytes(key, bytes, ttl);
    }

    fn shares_fill_epoch(&self) -> bool {
        self.0.shares_fill_epoch()
    }

    // A replay takes the unfenced path. A capture that does not store its
    // fill cannot replay that path, so it is incomplete.
    fn fill_epoch(&self, namespace: &str) -> FillEpoch {
        if offline() {
            return FillEpoch::Unsupported;
        }
        let epoch = self.0.fill_epoch(namespace);
        if epoch == FillEpoch::Unavailable {
            Self::fence_stopped_fill();
        }
        epoch
    }

    fn insert_raw_bytes_if_epoch(
        &self,
        key: &str,
        bytes: Vec<u8>,
        ttl: Option<Duration>,
        namespace: &str,
        sampled: u64,
    ) -> bool {
        if Self::direct_write_blocked(key) {
            return false;
        }
        let stored = self
            .0
            .insert_raw_bytes_if_epoch(key, bytes, ttl, namespace, sampled);
        if !stored {
            Self::fence_stopped_fill();
        }
        stored
    }

    fn try_acquire_fill_lock(&self, key: &str, token: &str, ttl: Duration) -> FillLockStatus {
        if offline() {
            return FillLockStatus::Unsupported;
        }
        let status = self.0.try_acquire_fill_lock(key, token, ttl);
        // Replay answers `Unsupported`, so a lock path cannot replay.
        if status != FillLockStatus::Unsupported
            && let Some(scope) = crate::capsule::current_scope()
        {
            scope.note(FILL_LOCK_NOTE);
            scope.mark_truncated();
        }
        status
    }

    fn release_fill_lock(&self, key: &str, token: &str) {
        if offline() {
            return;
        }
        self.0.release_fill_lock(key, token);
    }

    fn invalidate_async<'a>(
        &'a self,
        key: &'a str,
    ) -> CacheFuture<'a, Result<(), InvalidationError>> {
        // The tape and the capture scope are task-locals, so they are read
        // here, on the calling task, and carried into the future.
        if let Some(tape) = crate::capsule::effects::current_tape() {
            return Box::pin(async move { tape.cache_invalidate(key) });
        }
        if replay_blocked() {
            return Box::pin(async { Ok(()) });
        }
        // The scope is read now; the tape position is taken when the future
        // first runs, which is when replay consumes its entry. It is filled
        // when the removal ends, so a cancelled removal leaves an unfilled
        // slot that marks the capsule incomplete.
        let scope = crate::capsule::current_scope();
        Box::pin(async move {
            let slot = scope.and_then(|scope| scope.reserve_cache().map(|index| (scope, index)));
            let result = self.0.invalidate_async(key).await;
            if let Some((scope, index)) = slot {
                scope.fill_cache(
                    index,
                    crate::capsule::CacheEffect::Invalidate {
                        key: key.to_owned(),
                        error: recorded_error(&result),
                    },
                );
            }
            result
        })
    }

    fn invalidate_namespace_async<'a>(
        &'a self,
        namespace: &'a str,
    ) -> CacheFuture<'a, Result<(), InvalidationError>> {
        if let Some(tape) = crate::capsule::effects::current_tape() {
            return Box::pin(async move { tape.cache_invalidate_namespace(namespace) });
        }
        if replay_blocked() {
            return Box::pin(async { Ok(()) });
        }
        // Coherence records the removal with its registered stores' part.
        if COHERENCE_REMOVAL.with(std::cell::Cell::get) {
            return self.0.invalidate_namespace_async(namespace);
        }
        let scope = crate::capsule::current_scope();
        Box::pin(async move {
            let slot = scope.and_then(|scope| scope.reserve_cache().map(|index| (scope, index)));
            let result = self.0.invalidate_namespace_async(namespace).await;
            if let Some((scope, index)) = slot {
                scope.fill_cache(
                    index,
                    crate::capsule::CacheEffect::InvalidateNamespace {
                        namespace: namespace.to_owned(),
                        error: recorded_error(&result),
                    },
                );
            }
            result
        })
    }

    fn is_capsule_seam(&self) -> bool {
        true
    }
}

// ── CacheableResult trait ────────────────────────────────────────────

/// Helper trait used by `#[cached(result)]` to extract the `Ok` type
/// from a `Result<T, E>` return type at the type level.
///
/// This avoids the need for the proc macro to syntactically parse
/// generic arguments out of the return type.
pub trait CacheableResult {
    /// The success type to cache.
    type Ok: Clone;
    /// The error type (passed through uncached).
    type Err;

    /// Convert into a standard `Result` for pattern matching.
    ///
    /// # Errors
    ///
    /// Returns `Err` if the original result was an error.
    fn into_result(self) -> Result<Self::Ok, Self::Err>;
    /// Wrap a cached `Ok` value back into the original result type.
    fn from_ok(ok: Self::Ok) -> Self;
}

impl<T: Clone, E> CacheableResult for Result<T, E> {
    type Ok = T;
    type Err = E;

    fn into_result(self) -> Self {
        self
    }

    fn from_ok(ok: T) -> Self {
        Ok(ok)
    }
}

// ── Cache key helper ─────────────────────────────────────────────────

/// Build a cache key from a function name and its hashable arguments.
///
/// Used by `#[cached]` macro-generated code. The key is
/// `"{fn_name}:{hash_hex}"` where the hash is a 64-bit `DefaultHasher`
/// digest of the argument tuple.
#[must_use]
pub fn make_cache_key<K: Hash>(fn_name: &str, args: &K) -> String {
    let mut hasher = DefaultHasher::new();
    args.hash(&mut hasher);
    format!("{}:{:x}", fn_name, hasher.finish())
}

#[cfg(test)]
mod shared_fence_tests {
    //! A fake cross-replica backend: two handles share one store and one epoch.
    use super::*;
    use std::collections::HashMap;
    use std::sync::Mutex;

    #[derive(Default)]
    struct Shared {
        data: HashMap<String, Vec<u8>>,
        epochs: HashMap<String, u64>,
    }

    #[derive(Clone, Default)]
    struct Replica {
        shared: Arc<Mutex<Shared>>,
        epoch_down: bool,
        sweep_fails: bool,
    }

    impl Cache for Replica {
        fn get_value(&self, key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            let bytes = self.shared.lock().unwrap().data.get(key).cloned()?;
            Some(Arc::new(RawCacheBytes(bytes)))
        }
        fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {
            panic!("a shared-fence backend must not take the unfenced path");
        }
        fn invalidate(&self, key: &str) {
            self.shared.lock().unwrap().data.remove(key);
        }
        fn clear(&self) {
            self.shared.lock().unwrap().data.clear();
        }
        fn invalidate_namespace(&self, namespace: &str) -> bool {
            if self.sweep_fails {
                return false;
            }
            let mut shared = self.shared.lock().unwrap();
            *shared.epochs.entry(namespace.to_owned()).or_default() += 1;
            let prefix = format!("{namespace}:");
            shared.data.retain(|key, _| !key.starts_with(&prefix));
            true
        }
        fn insert_raw_bytes(&self, key: &str, bytes: Vec<u8>, _ttl: Option<Duration>) {
            self.shared
                .lock()
                .unwrap()
                .data
                .insert(key.to_owned(), bytes);
        }
        fn shares_fill_epoch(&self) -> bool {
            true
        }
        fn fill_epoch(&self, namespace: &str) -> FillEpoch {
            if self.epoch_down {
                return FillEpoch::Unavailable;
            }
            let shared = self.shared.lock().unwrap();
            FillEpoch::Sampled(shared.epochs.get(namespace).copied().unwrap_or(0))
        }
        fn insert_raw_bytes_if_epoch(
            &self,
            key: &str,
            bytes: Vec<u8>,
            _ttl: Option<Duration>,
            namespace: &str,
            sampled: u64,
        ) -> bool {
            let mut shared = self.shared.lock().unwrap();
            if shared.epochs.get(namespace).copied().unwrap_or(0) != sampled {
                return false;
            }
            shared.data.insert(key.to_owned(), bytes);
            true
        }
    }

    #[test]
    fn a_fill_on_another_replica_cannot_resurrect_an_invalidated_value() {
        let a = Replica::default();
        let b = Replica {
            shared: Arc::clone(&a.shared),
            epoch_down: false,
            sweep_fails: false,
        };
        // B misses and samples the shared epoch, then starts computing.
        let epoch = b.fill_epoch("ns");
        // A commits a write and invalidates the namespace.
        assert!(a.invalidate_namespace("ns"));
        // B finishes and tries to insert the pre-write value.
        let inserted = insert_cached_fenced(&b, "ns:k", "old".to_string(), None, "ns", epoch);
        assert!(
            !inserted,
            "the shared epoch moved, so the fill is fenced out"
        );
        assert!(get_cached::<String>(&a, "ns:k").is_none());
    }

    /// Codex review on #3222: during `autumn replay`, the fenced helpers do
    /// not reach a raw backend that startup code holds, with no tape.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_replay_block_keeps_the_fenced_helpers_offline() {
        let replica = Replica::default();
        let live = replica.fill_epoch("ns");
        super::TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(true));
        let sampled = sample_fill_epoch(&replica, "ns");
        let stored = insert_cached_fenced(&replica, "ns:k", "v".to_string(), None, "ns", live);
        super::TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(false));
        assert_eq!(sampled, FillEpoch::Unsupported, "no epoch read");
        assert!(stored, "the replay took the write");
        assert!(replica.shared.lock().unwrap().data.is_empty(), "no write");
    }

    /// Codex review on #3222: a replay keeps the fenced fill's result. A
    /// value that cannot be serialized, and an unreadable epoch, are not
    /// stored, as in production.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_replay_keeps_the_fenced_fill_result() {
        struct Unencodable;
        impl serde::Serialize for Unencodable {
            fn serialize<S: serde::Serializer>(&self, _: S) -> Result<S::Ok, S::Error> {
                Err(serde::ser::Error::custom("cannot encode"))
            }
        }
        impl Clone for Unencodable {
            fn clone(&self) -> Self {
                Self
            }
        }
        let replica = Replica::default();
        super::TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(true));
        let unencodable = insert_cached_fenced(
            &replica,
            "ns:k",
            Unencodable,
            None,
            "ns",
            FillEpoch::Sampled(0),
        );
        let unavailable =
            insert_cached_fenced(&replica, "ns:k", 1_u32, None, "ns", FillEpoch::Unavailable);
        super::TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(false));
        assert!(!unencodable, "the fenced path does not store it");
        assert!(!unavailable, "an unreadable epoch stores nothing");
        assert!(replica.shared.lock().unwrap().data.is_empty());
    }

    #[test]
    fn a_fill_inserts_when_no_invalidation_landed() {
        let a = Replica::default();
        let epoch = a.fill_epoch("ns");
        assert!(insert_cached_fenced(
            &a,
            "ns:k",
            "new".to_string(),
            None,
            "ns",
            epoch
        ));
        assert_eq!(get_cached::<String>(&a, "ns:k").as_deref(), Some("new"));
    }

    #[test]
    fn an_unreadable_epoch_skips_the_insert() {
        let a = Replica {
            epoch_down: true,
            ..Replica::default()
        };
        let epoch = a.fill_epoch("ns");
        assert_eq!(epoch, FillEpoch::Unavailable);
        assert!(!insert_cached_fenced(
            &a,
            "ns:k",
            "v".to_string(),
            None,
            "ns",
            epoch
        ));
        assert!(get_cached::<String>(&a, "ns:k").is_none());
    }

    #[test]
    fn a_fragment_render_on_another_replica_cannot_resurrect_an_invalidated_value() {
        let a = Replica::default();
        let b = Replica {
            shared: Arc::clone(&a.shared),
            epoch_down: false,
            sweep_fails: false,
        };
        let html = fragment::cache_fragment_in(Some(&b), "ns", "id", "v1", None, || {
            // A writes and invalidates while B is still rendering.
            assert!(a.invalidate_namespace("ns"));
            maud::html! { "old" }
        });
        assert_eq!(html.0, "old", "the fenced-out caller still gets its markup");
        let again = fragment::cache_fragment_in(Some(&b), "ns", "id", "v1", None, || {
            maud::html! { "new" }
        });
        assert_eq!(again.0, "new", "the stale markup must not be cached");
    }

    /// A caller-owned shared store, registered with `register_namespace_store`,
    /// must get its epoch raised by the namespace invalidation. A plain
    /// `clear` would sweep it and leave the epoch alone.
    #[test]
    fn invalidating_a_registered_shared_store_raises_its_epoch() {
        let _guard = GLOBAL_CACHE_TEST_LOCK
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        clear_global_cache();
        let store = Replica::default();
        let handle = Arc::new(store.clone());
        let _ = coherence::register_namespace_store("tests::shared_fence_ns", handle);
        let before = store.fill_epoch("tests::shared_fence_ns");
        assert!(coherence::invalidate_namespace("tests::shared_fence_ns"));
        assert_ne!(
            store.fill_epoch("tests::shared_fence_ns"),
            before,
            "an in-flight fill on another replica must be fenced out"
        );
    }

    /// A shared store that fails to bump its epoch makes the result `false`.
    /// A store with no shared epoch keeps the old `clear` contract.
    #[test]
    fn a_failed_registered_shared_store_fails_the_invalidation() {
        let _guard = GLOBAL_CACHE_TEST_LOCK
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        clear_global_cache();
        let failing = Replica {
            sweep_fails: true,
            ..Replica::default()
        };
        let _ = coherence::register_namespace_store("tests::failing_shared_ns", Arc::new(failing));
        assert!(!coherence::invalidate_namespace("tests::failing_shared_ns"));
    }

    #[cfg(feature = "cache-moka")]
    #[test]
    fn a_registered_store_without_a_shared_epoch_still_counts_as_cleared() {
        let _guard = GLOBAL_CACHE_TEST_LOCK
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        clear_global_cache();
        let _ = coherence::register_namespace_store(
            "tests::plain_ns",
            Arc::new(MokaCache::new(4, None)),
        );
        assert!(coherence::invalidate_namespace("tests::plain_ns"));
    }

    /// A replay serves cache effects from the tape. It must not read the live
    /// epoch, and the fenced write must land on the tape (#1634).
    #[cfg(feature = "reporting")]
    #[tokio::test]
    async fn a_replay_never_touches_the_live_epoch_and_still_writes_the_tape() {
        use crate::capsule::{
            CapsuleEffects,
            effects::{CachedValue, ReplayEffects},
        };
        let live = Replica {
            epoch_down: true,
            ..Replica::default()
        };
        let tape = std::sync::Arc::new(ReplayEffects::new(CapsuleEffects::default()));
        let seen = std::sync::Arc::clone(&tape);
        crate::capsule::effects::with_effect_tape(tape, async {
            let epoch = sample_fill_epoch(&live, "ns");
            assert_eq!(
                epoch,
                FillEpoch::Unsupported,
                "replay must skip the backend"
            );
            assert!(insert_cached_fenced(
                &live, "ns:k", 7_u32, None, "ns", epoch
            ));
        })
        .await;
        assert!(
            live.shared.lock().unwrap().data.is_empty(),
            "a replay must not write the live backend"
        );
        assert!(matches!(seen.cache_get("ns:k"), CachedValue::Hit(_)));
    }

    #[cfg(feature = "reporting")]
    fn capture_scope() -> Arc<crate::capsule::CaptureScope> {
        Arc::new(crate::capsule::CaptureScope::new(
            "fence-test".to_owned(),
            Arc::new(crate::capsule::CaptureSettings::default()),
            Arc::new(crate::log::filter::ParameterFilter::default()),
        ))
    }

    /// The capsule seam wraps every installed backend. It must keep the
    /// shared fence of the backend.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_capsule_seam_keeps_the_shared_fence() {
        let a = Replica::default();
        let b = with_capsule_seam(Arc::new(Replica {
            shared: Arc::clone(&a.shared),
            ..Replica::default()
        }));
        assert!(b.shares_fill_epoch());
        let epoch = sample_fill_epoch(b.as_ref(), "ns");
        assert_eq!(epoch, FillEpoch::Sampled(0));
        assert!(a.invalidate_namespace("ns"));
        assert!(!insert_cached_fenced(
            b.as_ref(),
            "ns:k",
            "old".to_string(),
            None,
            "ns",
            epoch
        ));
        assert!(a.shared.lock().unwrap().data.is_empty());
    }

    /// A fenced fill under capture is recorded. A fill the fence stops, or
    /// an epoch that cannot be read, is not recorded and replay takes the
    /// unfenced path, so the capsule is incomplete.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_fill_the_fence_stops_marks_the_capsule_incomplete() {
        use futures::executor::block_on;
        let fill = |replica: Replica, invalidate: bool| {
            let shared = Arc::clone(&replica.shared);
            let cache = with_capsule_seam(Arc::new(replica));
            let scope = capture_scope();
            block_on(crate::capsule::capture::with_capture_scope(
                Arc::clone(&scope),
                async {
                    let epoch = sample_fill_epoch(cache.as_ref(), "ns");
                    if invalidate {
                        shared.lock().unwrap().epochs.insert("ns".to_owned(), 9);
                    }
                    let _ = insert_cached_fenced(cache.as_ref(), "ns:k", 1_u32, None, "ns", epoch);
                },
            ));
            scope
        };
        let stored = fill(Replica::default(), false);
        assert!(!stored.is_truncated(), "a stored fill is replayable");
        assert_eq!(stored.effects_snapshot().cache.len(), 1);
        assert!(fill(Replica::default(), true).is_truncated());
        let down = Replica {
            epoch_down: true,
            ..Replica::default()
        };
        assert!(fill(down, false).is_truncated());
    }

    /// A direct fenced write is like a direct `insert_raw_bytes`.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_direct_fenced_write_marks_the_capsule_incomplete() {
        use futures::executor::block_on;
        let cache = with_capsule_seam(Arc::new(Replica::default()));
        let scope = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                let _ = cache.insert_raw_bytes_if_epoch("ns:k", b"1".to_vec(), None, "ns", 0);
            },
        ));
        assert!(scope.is_truncated());
    }

    #[cfg(feature = "cache-moka")]
    #[test]
    fn a_backend_without_a_shared_epoch_inserts_as_before() {
        let moka = MokaCache::new(10, None);
        let epoch = moka.fill_epoch("ns");
        assert_eq!(epoch, FillEpoch::Unsupported);
        assert!(insert_cached_fenced(
            &moka,
            "ns:k",
            "v".to_string(),
            None,
            "ns",
            epoch
        ));
        assert_eq!(get_cached::<String>(&moka, "ns:k").as_deref(), Some("v"));
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn cache_key_deterministic() {
        let k1 = make_cache_key("get_user", &(42_i64,));
        let k2 = make_cache_key("get_user", &(42_i64,));
        assert_eq!(k1, k2);
    }

    #[test]
    fn cache_key_differs_by_fn_name() {
        let k1 = make_cache_key("get_user", &(42_i64,));
        let k2 = make_cache_key("find_user", &(42_i64,));
        assert_ne!(k1, k2);
    }

    #[test]
    fn cache_key_differs_by_args() {
        let k1 = make_cache_key("get_user", &(1_i64,));
        let k2 = make_cache_key("get_user", &(2_i64,));
        assert_ne!(k1, k2);
    }

    #[test]
    fn cache_key_no_args() {
        let k = make_cache_key("get_config", &());
        assert!(k.starts_with("get_config:"));
    }

    #[cfg(feature = "cache-moka")]
    #[test]
    fn insert_cached_and_get_cached_round_trip() {
        let cache = MokaCache::new(10, None);
        insert_cached(&cache, "key", "hello".to_string(), None);
        let val: Option<String> = get_cached(&cache, "key");
        assert_eq!(val.as_deref(), Some("hello"));
    }

    #[cfg(feature = "cache-moka")]
    #[test]
    fn get_cached_raw_bytes_slow_path() {
        // Simulate a cross-replica backend: store RawCacheBytes directly, then
        // verify get_cached deserializes it back to the concrete type.
        let cache = MokaCache::new(10, None);
        let bytes = serde_json::to_vec(&42_i32).unwrap();
        cache.insert_value("k", Arc::new(RawCacheBytes(bytes)));
        let val: Option<i32> = get_cached(&cache, "k");
        assert_eq!(val, Some(42));
    }

    #[cfg(feature = "cache-moka")]
    #[test]
    fn get_cached_miss_returns_none() {
        let cache = MokaCache::new(10, None);
        let val: Option<String> = get_cached(&cache, "missing");
        assert!(val.is_none());
    }

    /// A backend that keeps only the required methods, to test the defaults.
    struct MinimalBackend {
        invalidated: std::sync::Mutex<Vec<String>>,
    }

    impl Cache for MinimalBackend {
        fn get_value(&self, _key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            None
        }
        fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {}
        fn invalidate(&self, key: &str) {
            self.invalidated
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner)
                .push(key.to_owned());
        }
        fn clear(&self) {}
    }

    #[test]
    fn default_try_methods_delegate_and_succeed() {
        let backend = MinimalBackend {
            invalidated: std::sync::Mutex::new(Vec::new()),
        };
        assert!(matches!(backend.try_get_value("k"), Ok(None)));
        assert!(backend.try_insert_raw_bytes("k", vec![1], None).is_ok());
        backend
            .try_invalidate("k")
            .expect("the default cannot fail");
        assert_eq!(
            *backend
                .invalidated
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner),
            ["k"],
            "the default try_invalidate must reach invalidate"
        );
    }

    /// Fails every `try_*` read and write.
    struct DownBackend;

    impl Cache for DownBackend {
        fn get_value(&self, _key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            None
        }
        fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {}
        fn invalidate(&self, _key: &str) {}
        fn clear(&self) {}
        fn try_get_value(
            &self,
            _key: &str,
        ) -> Result<Option<Arc<dyn Any + Send + Sync>>, CacheBackendError> {
            Err(CacheBackendError::new("down"))
        }
        fn try_insert_raw_bytes(
            &self,
            _key: &str,
            _bytes: Vec<u8>,
            _ttl: Option<Duration>,
        ) -> Result<(), CacheBackendError> {
            Err(CacheBackendError::new("down"))
        }
        fn try_invalidate(&self, _key: &str) -> Result<(), InvalidationError> {
            Err(InvalidationError::new(1, "down"))
        }
    }

    /// The capsule seam must pass a backend failure through, and must not mark
    /// a helper call as a direct one.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_capsule_seam_passes_try_failures_through() {
        let cache = with_capsule_seam(Arc::new(DownBackend));
        for op in [
            (|c: &dyn Cache| assert!(try_insert_cached(c, "k", 1_i64, None).is_err()))
                as fn(&dyn Cache),
            |c| assert!(try_get_cached::<i64>(c, "k").is_err()),
        ] {
            // A failed read or write cannot replay, so it marks the capsule.
            let scope = capture_scope();
            block_on(crate::capsule::capture::with_capture_scope(
                Arc::clone(&scope),
                async { op(cache.as_ref()) },
            ));
            assert!(scope.is_truncated(), "a failed cache op is not replayable");
        }
        // A failed removal is recorded with its error, so replay can match it.
        let scope = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async { assert!(cache.try_invalidate("k").is_err()) },
        ));
        let _ = scope.effects_snapshot();
        assert!(!scope.is_truncated(), "a recorded removal error replays");

        // Control: a direct call is not a helper call, so it marks the capsule.
        let direct = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&direct),
            async {
                let _ = cache.try_get_value("k");
            },
        ));
        assert!(direct.is_truncated(), "a direct read is not recorded");
    }

    /// A value whose `Serialize` always fails.
    #[derive(Clone)]
    struct Unserializable;

    impl serde::Serialize for Unserializable {
        fn serialize<S: serde::Serializer>(&self, _: S) -> Result<S::Ok, S::Error> {
            Err(serde::ser::Error::custom("no"))
        }
    }

    #[test]
    fn try_insert_cached_reports_a_serialization_failure() {
        let cache = MokaCache::new(8, None);
        let err = try_insert_cached(&cache, "k", Unserializable, None).expect_err("no bytes");
        assert!(err.reason().contains("cannot be serialized"));
    }

    #[test]
    fn try_cached_helpers_report_backend_failure() {
        let err = try_insert_cached(&DownBackend, "k", 1_i64, None).expect_err("write failed");
        assert_eq!(err.reason(), "down");
        let err = try_get_cached::<i64>(&DownBackend, "k").expect_err("read failed");
        assert_eq!(err.reason(), "down");
    }

    /// Holds bytes that are not JSON.
    struct CorruptBackend;

    impl Cache for CorruptBackend {
        fn get_value(&self, _key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            None
        }
        fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {}
        fn invalidate(&self, _key: &str) {}
        fn clear(&self) {}
        fn try_get_value(
            &self,
            _key: &str,
        ) -> Result<Option<Arc<dyn Any + Send + Sync>>, CacheBackendError> {
            Ok(Some(Arc::new(RawCacheBytes(b"not json".to_vec()))))
        }
    }

    #[test]
    fn try_get_cached_reports_bytes_it_cannot_decode() {
        let err = try_get_cached::<i64>(&CorruptBackend, "k").expect_err("corrupt bytes");
        assert!(err.reason().contains("cannot be decoded"), "{err}");
    }

    /// A decode fault marks the capsule, as a backend fault does.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_decode_fault_marks_the_capsule_incomplete() {
        let scope = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                assert!(try_get_cached::<i64>(&CorruptBackend, "k").is_err());
            },
        ));
        assert!(scope.is_truncated());
    }

    #[test]
    fn try_cached_helpers_round_trip_in_process() {
        let cache = MokaCache::new(8, None);
        try_insert_cached(&cache, "k", 7_i64, None).expect("stored");
        assert_eq!(try_get_cached::<i64>(&cache, "k").expect("read"), Some(7));
        assert_eq!(try_get_cached::<i64>(&cache, "gone").expect("read"), None);
    }

    fn block_on<F: std::future::Future>(future: F) -> F::Output {
        tokio::runtime::Builder::new_current_thread()
            .build()
            .expect("runtime")
            .block_on(future)
    }

    #[test]
    fn default_invalidate_async_calls_the_sync_method() {
        let backend = MinimalBackend {
            invalidated: std::sync::Mutex::new(Vec::new()),
        };
        block_on(backend.invalidate_async("k")).expect("the default cannot fail");
        assert_eq!(
            *backend
                .invalidated
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner),
            vec!["k".to_owned()]
        );
    }

    #[test]
    fn default_invalidate_namespace_async_reports_an_unsupported_backend() {
        let backend = MinimalBackend {
            invalidated: std::sync::Mutex::new(Vec::new()),
        };
        let err = block_on(backend.invalidate_namespace_async("ns"))
            .expect_err("a backend that cannot sweep must say so");
        assert_eq!(err.attempts(), 1);
    }

    /// Logs every call that reaches the backend.
    #[cfg(feature = "reporting")]
    #[derive(Default)]
    struct SpyBackend {
        calls: std::sync::Mutex<Vec<String>>,
    }

    #[cfg(feature = "reporting")]
    impl SpyBackend {
        fn log(&self, call: impl Into<String>) {
            self.calls
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner)
                .push(call.into());
        }
        fn calls(&self) -> Vec<String> {
            self.calls
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner)
                .clone()
        }
    }

    #[cfg(feature = "reporting")]
    impl Cache for SpyBackend {
        fn get_value(&self, key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            self.log(format!("get {key}"));
            Some(Arc::new(7_u32))
        }
        fn insert_value(&self, key: &str, _value: Arc<dyn Any + Send + Sync>) {
            self.log(format!("insert {key}"));
        }
        fn invalidate(&self, key: &str) {
            self.log(format!("invalidate {key}"));
        }
        fn clear(&self) {
            self.log("clear");
        }
        fn invalidate_namespace(&self, namespace: &str) -> bool {
            self.log(format!("invalidate_namespace {namespace}"));
            true
        }
    }

    /// #2351 item 2: a replayed run never reaches the installed backend
    /// through the untyped functions or the removal methods.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_replayed_run_never_reaches_the_installed_backend() {
        let spy = Arc::new(SpyBackend::default());
        let state =
            crate::state::AppState::for_test().with_cache(Arc::clone(&spy) as Arc<dyn Cache>);
        let tape = Arc::new(crate::capsule::ReplayEffects::new(
            crate::capsule::CapsuleEffects::default(),
        ));
        block_on(crate::capsule::with_effect_tape(Arc::clone(&tape), async {
            let cache = state.cache().expect("installed");
            cache.invalidate("widgets");
            cache.clear();
            let _ = cache.invalidate_namespace("widgets");
            let _ = cache.invalidate_async("widgets").await;
            insert(cache.as_ref(), "widgets", 41_u32);
            assert_eq!(get::<u32>(cache.as_ref(), "widgets"), None);
        }));
        assert!(spy.calls().is_empty(), "{:?}", spy.calls());
        assert_eq!(tape.divergences().len(), 6, "{:?}", tape.divergences());
    }

    /// #2351 item 2: capture records a removal, and the backend still gets it.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_captured_removal_is_recorded_and_still_applied() {
        let spy = Arc::new(SpyBackend::default());
        let state =
            crate::state::AppState::for_test().with_cache(Arc::clone(&spy) as Arc<dyn Cache>);
        let scope = Arc::new(crate::capsule::CaptureScope::new(
            "cache-test".to_owned(),
            Arc::new(crate::capsule::CaptureSettings::default()),
            Arc::new(crate::log::filter::ParameterFilter::default()),
        ));
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                let cache = state.cache().expect("installed");
                cache.invalidate("widgets");
                cache.clear();
            },
        ));
        assert_eq!(spy.calls(), vec!["invalidate widgets", "clear"]);
        assert_eq!(
            scope.effects_snapshot().cache,
            vec![
                crate::capsule::CacheEffect::Invalidate {
                    key: "widgets".to_owned(),
                    error: None,
                },
                crate::capsule::CacheEffect::Clear,
            ]
        );
    }

    /// #2351 item 2: an untyped hit cannot be serialized, so the capsule says
    /// it is incomplete.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_captured_untyped_hit_or_write_marks_the_capsule_incomplete() {
        let spy = Arc::new(SpyBackend::default());
        let scope = Arc::new(crate::capsule::CaptureScope::new(
            "cache-test".to_owned(),
            Arc::new(crate::capsule::CaptureSettings::default()),
            Arc::new(crate::log::filter::ParameterFilter::default()),
        ));
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                assert_eq!(get::<u32>(spy.as_ref(), "widgets"), Some(7));
            },
        ));
        assert!(scope.is_truncated());
        let scope = Arc::new(crate::capsule::CaptureScope::new(
            "cache-test".to_owned(),
            Arc::new(crate::capsule::CaptureSettings::default()),
            Arc::new(crate::log::filter::ParameterFilter::default()),
        ));
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                insert(spy.as_ref(), "widgets", 1_u32);
            },
        ));
        assert!(scope.is_truncated());
    }

    /// Codex review on #3222: a direct `get_value` during a replay is an
    /// unrecorded read. It does not consume a read a helper recorded:
    /// capture refuses a capsule with a direct read.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_direct_read_on_replay_is_a_divergence_not_a_helper_read() {
        let cache = with_capsule_seam(Arc::new(SpyBackend::default()) as Arc<dyn Cache>);
        let recorded = crate::capsule::CapsuleEffects {
            cache: vec![crate::capsule::CacheEffect::Get {
                key: "user:7".to_owned(),
                value: None,
            }],
            ..Default::default()
        };
        let tape = Arc::new(crate::capsule::ReplayEffects::new(recorded));
        block_on(crate::capsule::with_effect_tape(Arc::clone(&tape), async {
            assert!(cache.get_value("user:7").is_none());
        }));
        let finished = tape.finish();
        assert!(
            finished
                .iter()
                .any(|divergence| divergence.actual.contains("direct cache read")),
            "{finished:?}"
        );
        assert!(
            !finished
                .iter()
                .any(|divergence| divergence.actual.contains("user:7")),
            "the key is withheld: {finished:?}"
        );
        assert!(
            finished
                .iter()
                .any(|divergence| divergence.kind
                    == crate::capsule::EffectDivergenceKind::Unconsumed),
            "the helper's recorded read is still owed: {finished:?}"
        );
    }

    /// Review fix: during a replay the wrapper reaches the backend for no
    /// read, write or fill lock.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_wrapper_reaches_no_backend_during_a_replay() {
        let spy = Arc::new(SpyBackend::default());
        let cache = with_capsule_seam(Arc::clone(&spy) as Arc<dyn Cache>);
        let tape = Arc::new(crate::capsule::ReplayEffects::new(
            crate::capsule::CapsuleEffects::default(),
        ));
        block_on(crate::capsule::with_effect_tape(Arc::clone(&tape), async {
            assert!(cache.get_value("k").is_none());
            cache.insert_value("k", Arc::new(1_u32));
            cache.insert_raw_bytes("k", b"1".to_vec(), None);
            assert_eq!(
                cache.try_acquire_fill_lock("k", "t", Duration::from_secs(1)),
                FillLockStatus::Unsupported
            );
            cache.release_fill_lock("k", "t");
            assert_eq!(cache.fill_epoch("ns"), FillEpoch::Unsupported);
        }));
        assert!(spy.calls().is_empty(), "{:?}", spy.calls());
        // Codex review on #3222: only the reporting layer enters the replay
        // scope. A cache call must not, or the missing-layer warning is lost.
        assert!(!tape.scope_entered());
    }

    /// An async removal takes its tape position when it first runs, the
    /// point where replay consumes its entry (Codex review on #3222).
    #[cfg(feature = "reporting")]
    #[test]
    fn an_async_removal_is_recorded_when_it_first_runs() {
        let spy = Arc::new(SpyBackend::default());
        let cache = with_capsule_seam(Arc::clone(&spy) as Arc<dyn Cache>);
        let scope = Arc::new(crate::capsule::CaptureScope::new(
            "cache-test".to_owned(),
            Arc::new(crate::capsule::CaptureSettings::default()),
            Arc::new(crate::log::filter::ParameterFilter::default()),
        ));
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                let first = cache.invalidate_async("a");
                cache.clear();
                first.await.expect("removed");
            },
        ));
        assert_eq!(
            scope.effects_snapshot().cache,
            vec![
                crate::capsule::CacheEffect::Clear,
                crate::capsule::CacheEffect::Invalidate {
                    key: "a".to_owned(),
                    error: None,
                },
            ]
        );
    }

    /// Codex review on #3222: when production had a global cache, a replay
    /// answers a repository invalidation from the tape, so it is consumed.
    #[cfg(feature = "reporting")]
    #[test]
    fn replay_installs_a_seam_cache_when_the_builder_had_one() {
        let _guard = GLOBAL_CACHE_TEST_LOCK
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        let recorded = crate::capsule::CapsuleEffects {
            cache: vec![crate::capsule::CacheEffect::InvalidateNamespace {
                namespace: "posts".to_owned(),
                error: None,
            }],
            state_cache: true,
            ..Default::default()
        };
        let installed = install_replay_cache(true).expect("installed");
        assert!(installed.is_capsule_seam());
        assert!(Arc::ptr_eq(
            &installed,
            &global_cache().expect("the global cache")
        ));
        let tape = Arc::new(crate::capsule::ReplayEffects::new(recorded));
        let complete = block_on(crate::capsule::with_effect_tape(
            Arc::clone(&tape),
            coherence::invalidate_namespace_async("posts"),
        ));
        clear_global_cache();
        assert!(complete);
        assert!(tape.finish().is_empty(), "{:?}", tape.finish());

        assert!(install_replay_cache(false).is_none());
        assert!(global_cache().is_none());
    }

    /// Codex review on #3222: a cache only a startup hook installed is put
    /// in place after the state initializers, as production's initializers
    /// did not see it. A cache an initializer installed is kept.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_startup_hook_cache_is_installed_after_the_initializers() {
        let _guard = GLOBAL_CACHE_TEST_LOCK
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        let recorded = crate::capsule::CapsuleEffects {
            state_cache: true,
            ..Default::default()
        };
        let state = crate::state::AppState::for_test();
        install_late_replay_cache(&state, &recorded);
        let late = state.cache().expect("installed after the initializers");
        assert!(late.is_capsule_seam());

        let initializer = crate::state::AppState::for_test();
        initializer.set_cache(Arc::new(SpyBackend::default()));
        let kept = initializer.cache().expect("the initializer's cache");
        install_late_replay_cache(&initializer, &recorded);
        assert!(Arc::ptr_eq(
            &kept,
            &initializer.cache().expect("still there")
        ));

        let none = crate::state::AppState::for_test();
        install_late_replay_cache(&none, &crate::capsule::CapsuleEffects::default());
        clear_global_cache();
        assert!(none.cache().is_none());
    }

    /// A backend whose sync removals fail, and whose fill lock is held.
    #[cfg(feature = "reporting")]
    struct FailingSyncBackend;

    #[cfg(feature = "reporting")]
    impl Cache for FailingSyncBackend {
        fn get_value(&self, _key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            None
        }
        fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {}
        fn invalidate(&self, _key: &str) {
            record_invalidation_failure();
        }
        fn clear(&self) {
            record_invalidation_failure();
        }
        fn try_acquire_fill_lock(
            &self,
            _key: &str,
            _token: &str,
            _ttl: Duration,
        ) -> FillLockStatus {
            FillLockStatus::Held
        }
    }

    #[cfg(feature = "reporting")]
    fn capture_scope() -> Arc<crate::capsule::CaptureScope> {
        Arc::new(crate::capsule::CaptureScope::new(
            "cache-test".to_owned(),
            Arc::new(crate::capsule::CaptureSettings::default()),
            Arc::new(crate::log::filter::ParameterFilter::default()),
        ))
    }

    /// Codex review on #3222: a sync removal that reports a failure marks
    /// the capsule incomplete.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_failed_sync_removal_marks_the_capsule_incomplete() {
        for remove in [
            (|cache: &dyn Cache| cache.invalidate("k")) as fn(&dyn Cache),
            |cache: &dyn Cache| cache.clear(),
        ] {
            let cache = with_capsule_seam(Arc::new(FailingSyncBackend));
            let scope = capture_scope();
            block_on(crate::capsule::capture::with_capture_scope(
                Arc::clone(&scope),
                async { remove(cache.as_ref()) },
            ));
            assert!(scope.is_truncated());
        }
    }

    /// Codex review on #3222: a fill lock outcome other than `Unsupported`
    /// marks the capsule incomplete.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_distributed_fill_lock_marks_the_capsule_incomplete() {
        let cache = with_capsule_seam(Arc::new(FailingSyncBackend));
        let scope = capture_scope();
        let status = block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async { cache.try_acquire_fill_lock("k", "t", Duration::from_secs(1)) },
        ));
        assert_eq!(status, FillLockStatus::Held);
        assert!(scope.is_truncated());

        let moka = with_capsule_seam(Arc::new(SpyBackend::default()));
        let scope = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                let _ = moka.try_acquire_fill_lock("k", "t", Duration::from_secs(1));
            },
        ));
        assert!(
            !scope.is_truncated(),
            "a backend with no lock is replayable"
        );
    }

    /// Codex review on #3222: a cache installed by any route is recorded,
    /// a startup hook (`RedisCachePlugin`) included. Replay runs no startup
    /// hook, so it must install a cache in its place.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_cache_from_any_install_route_is_recorded() {
        let _guard = GLOBAL_CACHE_TEST_LOCK
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        let builder = crate::state::AppState::for_test()
            .with_cache(Arc::new(SpyBackend::default()) as Arc<dyn Cache>);
        assert!(builder.has_cache());
        let startup = crate::state::AppState::for_test();
        assert!(!startup.has_cache());
        startup.set_cache(Arc::new(SpyBackend::default()));
        clear_global_cache();
        assert!(startup.has_cache());
    }

    /// Codex review on #3222: during `autumn replay`, a cache call with no
    /// tape (a state initializer) does not reach the backend.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_replay_block_keeps_the_backend_offline_with_no_tape() {
        let spy = Arc::new(SpyBackend::default());
        let cache = with_capsule_seam(Arc::clone(&spy) as Arc<dyn Cache>);
        TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(true));
        block_on(async {
            assert!(cache.get_value("k").is_none());
            cache.insert_value("k", Arc::new(1_u32));
            cache.insert_raw_bytes("k", b"1".to_vec(), None);
            cache.invalidate("k");
            cache.clear();
            assert!(cache.invalidate_namespace("ns"));
            assert!(cache.invalidate_async("k").await.is_ok());
            assert!(cache.invalidate_namespace_async("ns").await.is_ok());
            assert_eq!(
                cache.try_acquire_fill_lock("k", "t", Duration::from_secs(1)),
                FillLockStatus::Unsupported
            );
        });
        TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(false));
        assert!(spy.calls().is_empty(), "{:?}", spy.calls());
    }

    /// Codex review on #3222: during `autumn replay`, the cache helpers do
    /// not reach a raw backend that startup code holds, with no tape.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_replay_block_keeps_a_raw_backend_offline_in_the_helpers() {
        let spy = SpyBackend::default();
        TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(true));
        assert_eq!(get::<u32>(&spy, "k"), None);
        insert(&spy, "k", 1_u32);
        assert_eq!(get_cached::<u32>(&spy, "k"), None);
        insert_cached(&spy, "k", 1_u32, None);
        TEST_REPLAY_BLOCKED.with(|blocked| blocked.set(false));
        assert!(spy.calls().is_empty(), "{:?}", spy.calls());
    }

    /// A backend whose sync removals panic.
    #[cfg(feature = "reporting")]
    struct PanickingBackend;

    #[cfg(feature = "reporting")]
    impl Cache for PanickingBackend {
        fn get_value(&self, _key: &str) -> Option<Arc<dyn Any + Send + Sync>> {
            None
        }
        fn insert_value(&self, _key: &str, _value: Arc<dyn Any + Send + Sync>) {}
        fn invalidate(&self, _key: &str) {
            panic!("backend failed");
        }
        fn clear(&self) {
            panic!("backend failed");
        }
        fn invalidate_namespace(&self, _namespace: &str) -> bool {
            panic!("backend failed");
        }
    }

    /// Codex review on #3222: a sync removal whose backend panics leaves
    /// its slot unfilled, so the capsule is incomplete and records no false
    /// success.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_panicking_sync_removal_leaves_the_capsule_incomplete() {
        for remove in [
            (|cache: &dyn Cache| cache.invalidate("k")) as fn(&dyn Cache),
            |cache: &dyn Cache| cache.clear(),
            |cache: &dyn Cache| {
                let _ = cache.invalidate_namespace("ns");
            },
        ] {
            let cache = with_capsule_seam(Arc::new(PanickingBackend));
            let scope = capture_scope();
            let caught = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                block_on(crate::capsule::capture::with_capture_scope(
                    Arc::clone(&scope),
                    async { remove(cache.as_ref()) },
                ));
            }));
            assert!(caught.is_err(), "the backend panics");
            let _ = scope.effects_snapshot();
            assert!(
                scope.is_truncated(),
                "an unfilled slot marks the capsule incomplete"
            );
        }
    }

    /// Codex review on #3222: a direct `get_value` call through the seam
    /// marks the capsule incomplete during capture, and is a divergence
    /// during a replay. A helper read stays recorded.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_direct_get_value_is_not_silently_substituted() {
        let cache = with_capsule_seam(Arc::new(SpyBackend::default()));
        let scope = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                let _: Option<u32> = get_cached(cache.as_ref(), "k");
            },
        ));
        assert!(!scope.is_truncated(), "a helper read is recorded");
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async {
                let _ = cache.get_value("k");
            },
        ));
        assert!(scope.is_truncated(), "a direct read cannot be recorded");

        let tape = Arc::new(crate::capsule::ReplayEffects::new(
            crate::capsule::CapsuleEffects::default(),
        ));
        let read = block_on(crate::capsule::with_effect_tape(Arc::clone(&tape), async {
            cache.get_value("k")
        }));
        assert!(read.is_none());
        assert_eq!(tape.divergences().len(), 1, "{:?}", tape.divergences());
    }

    /// Codex review on #3222: a direct `insert_value` or `insert_raw_bytes`
    /// call through the seam marks the capsule incomplete during capture, and
    /// is a divergence during a replay. A helper write stays recorded.
    #[cfg(feature = "reporting")]
    #[test]
    fn a_direct_write_is_not_silently_dropped() {
        let cache = with_capsule_seam(Arc::new(SpyBackend::default()));
        let scope = capture_scope();
        block_on(crate::capsule::capture::with_capture_scope(
            Arc::clone(&scope),
            async { insert_cached(cache.as_ref(), "k", 1_u32, None) },
        ));
        assert!(!scope.is_truncated(), "a helper write is recorded");
        for write in [
            (|cache: &dyn Cache| cache.insert_value("k", Arc::new(1_u32))) as fn(&dyn Cache),
            |cache: &dyn Cache| cache.insert_raw_bytes("k", b"1".to_vec(), None),
        ] {
            let scope = capture_scope();
            block_on(crate::capsule::capture::with_capture_scope(
                Arc::clone(&scope),
                async { write(cache.as_ref()) },
            ));
            assert!(scope.is_truncated(), "a direct write cannot be recorded");

            let tape = Arc::new(crate::capsule::ReplayEffects::new(
                crate::capsule::CapsuleEffects::default(),
            ));
            block_on(crate::capsule::with_effect_tape(Arc::clone(&tape), async {
                write(cache.as_ref());
            }));
            assert_eq!(tape.divergences().len(), 1, "{:?}", tape.divergences());
        }
    }

    /// The seam wraps a backend once, so `Arc` identity stays stable.
    #[cfg(feature = "reporting")]
    #[test]
    fn the_capsule_seam_wraps_a_backend_once() {
        let spy: Arc<dyn Cache> = Arc::new(SpyBackend::default());
        let once = with_capsule_seam(spy);
        let twice = with_capsule_seam(Arc::clone(&once));
        assert!(Arc::ptr_eq(&once, &twice));
    }

    #[test]
    fn invalidation_error_reports_attempts_and_reason() {
        let err = InvalidationError::new(3, "READONLY");
        assert_eq!(err.attempts(), 3);
        assert_eq!(err.reason(), "READONLY");
        assert_eq!(
            err.to_string(),
            "cache invalidation failed after 3 attempt(s): READONLY"
        );
    }

    #[test]
    fn invalidation_failures_counter_is_monotonic() {
        let before = invalidation_failures_total();
        record_invalidation_failure();
        assert!(invalidation_failures_total() > before);
    }

    #[test]
    fn cacheable_result_ok_round_trips() {
        let r: Result<i32, &str> = Result::from_ok(42);
        assert_eq!(r, Ok(42));
        assert_eq!(r.into_result(), Ok(42));
    }

    #[test]
    fn cacheable_result_err_passes_through() {
        let r: Result<i32, &str> = Err("oops");
        assert_eq!(r.into_result(), Err("oops"));
    }
}