Stop concurrent coroutines sharing a phpredis socket

[adhikjoshi]

The failure

Swoole binds a hooked phpredis socket to the coroutine that first reads from it. When a second coroutine issues a command on that same connection, the scheduler raises:

Swoole\Error: Socket#94 has already been bound to another coroutine#3,
reading of the same socket in coroutine#18 at the same time is not allowed

This is raised by the scheduler, not thrown through the calling frame — a try/catch (\Throwable) wrapped directly around the failing call does not catch it, and the worker dies. Worker deaths under load caused request pile-up, and in turn memory_limit fatals.

Root cause

Every occurrence observed in production came in through the same path:

RateLimiter->tooManyAttempts -> Repository->get -> RedisStore->get -> PhpRedisConnection->command

Illuminate\Cache\RateLimiter is registered by CacheServiceProvider as a boot-time singleton wrapping a cache Repository, so its RedisStore keeps pointing at the worker-level Redis manager. RequestScope already isolated redis and cache per coroutine, but an object constructed at boot never resolves them again — so every concurrent request throttled through one shared socket.

The fix — two independent defences

1. RequestScope scopes Illuminate\Cache\RateLimiter. The base limiter is cloned rather than rebuilt, so named RateLimiter::for() registrations made during boot survive; only the cache repository is swapped for the coroutine-local one.

2. CoroutineRedisManager makes the worker-level manager itself coroutine-aware. This covers boot-time singletons that this package never sees, because RedisStore re-resolves its connection on every operation. Notes:

• Strips persistent/persistent_id — persistent phpredis sockets are shared process-wide by id and would reintroduce the sharing.
• Installed via extend() rather than singleton(), because Illuminate's RedisServiceProvider is deferred and would otherwise overwrite the binding on first resolve.
• Resolves through a sibling manager rather than temporarily rewriting $this->config, since resolving performs I/O and a coroutine can park mid-connect.
• Leaves customized/subclassed managers alone rather than silently dropping app behaviour.
• Connections are released when the request coroutine finishes.

3. Sentry transport hardening. Sentry sends events with curl_exec, which some Swoole builds make coroutine-only. Laravel's shutdown handler reports fatals from outside any coroutine, so reporting an error could raise a second fatal (API must be called in the coroutine) that replaced the original in the logs. Captures now run with curl hooks temporarily cleared when outside a coroutine, and any transport failure is swallowed — losing one event beats losing the worker.

Verification

The end-to-end probe runs in a subprocess, because surviving this fatal is only observable as a clean exit. Run against the production runtime image (phpswoole/swoole:php8.3 — PHP 8.3.32, Swoole 6.2.1, phpredis 6.3.0):

manager	resolve	result
stock	base	FATAL — reproduces the production stack byte-for-byte
stock	scope	survived, distinct sockets
coroutine	base	survived — un-scoped boot singleton protected
coroutine	scope	survived — shipped configuration

The unfixed combination is asserted as a failing test, so if the harness ever stops reproducing the bug the passing cases cannot silently become vacuous.

136/136 tests pass (was 106 — 30 new), both locally and inside the production image.

Trade-off

Each coroutine now opens its own Redis connection per request rather than reusing a persistent one. This already applied to anything resolved during a request; this change extends it to the RateLimiter path, so the net increase is modest. If Redis maxclients or connect latency becomes a concern, a pooled free-list (mirroring the existing DatabasePool) is the follow-up — deliberately left out here, since returning a connection interrupted mid-reply to a pool carries real corruption risk and stability was the priority.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is enabled.}