fix: bound the Redis read so a dead socket cannot park a worker forever

[adhikjoshi]

Companion to modelq#16. Same defect, different route.

Why this looked safe and wasn't

The worker already passed a timeout:

$taskData = $this->redis->blPop(['ml_tasks'], 1);

That 1 is the server's timeout. It tells Redis how long to hold the pop before answering "nothing". It says nothing about how long this process waits for that answer to arrive.

phpredis defaults its read timeout to 0 — wait forever — and connect() was called without one:

$this->redis->connect($host, $port);   // no connect timeout, no read timeout

So when the connection is silently dropped, the reply can never arrive, the read never returns, and the worker parks permanently on a socket the server has already discarded. Identical outcome to the Python side, where the same thing happened via an unbounded BLPOP.

On 2026-08-15 this took out seven server_types within fifteen minutes; sfx_server sat dead for two days before anyone noticed.

Changes

• connect() passes CONNECT_TIMEOUT and READ_TIMEOUT.
• OPT_TCP_KEEPALIVE is set so the kernel probes an idle peer, guarded by defined() since it needs phpredis 5+.
• The blocking pop uses BLPOP_TIMEOUT, deliberately held below READ_TIMEOUT. That ordering is the whole point — invert it and every idle poll aborts the read before Redis has replied, turning a correct empty result into an exception on every loop.

Blast radius

Only affects the path where ModelQ builds its own client. Callers injecting their own Redis are untouched — modelslab-frontend-v2 already sets a 5s connect and read timeout on the client it passes in, which is why the frontend never exhibited this.

Testing

3 unit tests, no Redis required. Mutation-tested:

mutation	result
replace self::BLPOP_TIMEOUT with a literal in blPop	1 test red
restore READ_TIMEOUT = 0 (phpredis default)	2 tests red

Suite goes 130 → 133. The same 13 pre-existing Integration failures remain (they need a live Redis); comm against the recorded baseline confirms none added. phpstan level 6: 36 errors before, 36 after, zero introduced.

Not a duplicate of #2

#2 removes the redundant prune_old_task_results SCAN storm — a throughput fix. This is a liveness fix. They touch different code and are independently mergeable in either order.

---

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