flyingsquirrel0419/layercache

Breaking

• Public CacheStack, namespace, wrap(), getOrSet(), and mget() reads now return undefined for misses and negative-cache entries instead of null, resolving issue #90.
• Read-through fetchers now cache null as a regular value by default, and getOrThrow() returns stored nulls while throwing only for undefined. Set cacheNullValues: false when null represents absence.
• Automatically derived structured wrap() keys use the j2: schema, so old j: entries become cold misses and expire naturally.
• Write coordination and generation cleanup now apply finite default limits and can reject saturated work with CacheWriteSaturationError.

Security

• Fixed implicit Express/Hono private URL caching so requests containing sensitive query parameters, including OAuth client_secret and client assertions, bypass implicit URL-only caching unless a custom keyResolver is supplied.
• Fixed namespace clearing so CacheNamespace.clear() invalidates the delimiter-bound namespace prefix (for example tenant:), preventing sibling namespace eviction such as a clearing ab.
• Fixed DiskLayer protected reads so plaintext legacy payloads are rejected when encryptionKey or signingKey is configured. Added allowLegacyPlaintext as an explicit migration-only opt-in.
• Fixed generation cleanup to stream old-generation keys into bounded batches and cap its unique-key discovery set at 10,000 entries by default.
• Changed OpenTelemetry instrumentation to export layercache.key_hash by default. Raw layercache.key attributes now require includeRawKeyAttributes: true.
• Hardened snapshot commits by rejecting symlinked target parents immediately before the final rename.
• Hardened CLI invalidation so every wildcard-only pattern such as *, **, and ?* requires --force, matching the default full-cache invalidation guard.
• Hardened CLI Redis connection errors by scrubbing Redis URLs from nested error messages before printing them.
• Added a Redis invalidation warning when RedisInvalidationBus is constructed without signingSecret and a logger is provided.
• Hardened the docs playground by keeping the token-bearing worker message sender outside user-code global reach, in addition to sandboxed iframe, Blob Worker, and token checks.
• Prevented plain objects from forging native structured cache-key tags and rotated structured argument keys from j: to j2: so previously ambiguous entries are not reused.
• Prevented stale single and bulk writes from repopulating invalidated keys or deleting newer mset() values by sharing one per-key ordering boundary and checking fences after backend writes.
• Bounded retained write-ordering state with configurable global, active-key, and per-key limits that reject saturation with CacheWriteSaturationError.
• Replaced reusable per-key epoch values with monotonic tokens so maintenance pruning cannot create an ABA window for stale writes.
• Fixed per-key fetch rate limiting so a ready key can preempt a later shared drain timer chosen by a throttled key.

Changed

• Updated Vite and related development-only transitive dependencies to resolve the release audit finding.
• Updated the README and benchmarking guide with a fresh reproducible 4.0.0 baseline covering direct, HTTP, payload, slow-Redis, queue-amplification, memory-pressure, invalidation, and distributed single-flight scenarios.
• Corrected benchmark TTLs to use the documented millisecond units and updated Autocannon plus its Hyperid override so the HTTP benchmark runs on current Node.js with zero npm audit findings.
• Updated API, migration, serialization, resilience, integration, architecture-decision, and security documentation for v4 miss/null semantics, finite cleanup/write limits, structured-key rotation, HTTP credential handling, CLI guard, and playground isolation boundary.

Tests

• Added regression coverage for all security fixes and hardening changes, including shared drain-timer preemption, disposal of queued rate-limited work, epoch rollover, empty serialized writes, symlinked snapshot bases, and serializer undefined behavior. The unit suite reports 672 passing tests and 24 skipped tests; the real Redis integration suite adds 25 passing tests.
• Added docs-web content assertions for the updated security-sensitive documentation.