fix: expire every build key at write time

[prsimp]

Why

tools-redis-001, the ElastiCache node behind CI_QUEUE_URL, was pinned at 100% memory with 469,496 keys of which only 1,484 had a TTL. Under its volatile-lru policy the only evictable keys were ci-queue's live queue bookkeeping, so builds lost total / master-status mid-run and ruby spec lanes went red with no failing spec (Total tests in queue: 0, or CI::Queue::Redis::LostMaster).

The permanent keys were ci-queue's own. Measured on the node before cleanup:

Key	Count	Memory
processed	~3,900	4.62 GiB
worker:<n>:queue	~337,500	3.82 GiB
requeues-count	~112,300	0.99 GiB
owners / running / warnings	~11,100	0.09 GiB

That accounts for the whole 9.7 GiB.

What changed

TTLs are now set at write time. The EXPIRE for processed and worker:<id>:queue used to live at the tail of Worker#poll, inside a method whose bottom is rescue *CONNECTION_ERRORS. A worker that is SIGKILLed, cancelled, OOMs, or loses its Redis connection never reaches that line, so the key leaked forever. running, owners, requeues-count, warnings, test_failed_count and created-at never got an expiry at all.

The TTL is threaded through the Lua scripts as an argument, and through the spawned heartbeat monitor process. No key's lifetime now depends on a clean shutdown.

rspec-queue --report no longer certifies a vanished queue. An evicted queue reads as exhausted and its error-report hash reads as empty, so --report printed "No errors found" and exited 0 for a run whose results were gone. It now refuses to report a result when total is 0 or no progress was recorded, and says explicitly that this indicates eviction rather than a test failure.

Verification

Driving reserve/requeue/acknowledge by hand so the end-of-poll EXPIRE never runs — the leak path:

• Before: 9 of 13 keys written with ttl=-1 (created-at, owners, processed, requeues-count, running, test_failed_count, warnings, worker:1:queue, worker:2:queue).
• After: all 13 carry a TTL.

Covered by a new regression test, test_every_build_key_has_a_ttl, which fails against the previous Lua scripts naming exactly the leaked keys. Full redis_test.rb suite: 27 tests, 75 assertions, 0 failures.

For the report guard, against a real rspec-queue --report process with a completed queue:

Case	Before	After
healthy queue	exit 0	exit 0
total evicted	exit 0, "No errors found"	exit 1, eviction explained
master-status evicted	exit 1	exit 1

The middle row is the false green this closes.

Notes

Upstream Shopify/ci-queue has the same missing expires in reserve.lua, so this is an inherited bug rather than a fork regression. An upstream PR should follow separately.

Running the test suite locally on Ruby 3.3 also needs rexml in the gem's Gemfile; that is not included here to keep the diff scoped.

PE-3857