fix: fall back when Redis has no BLMOVE instead of crash-looping

[adhikjoshi]

The failure

BLMOVE needs Redis ≥ 6.2. Every GPU worker container in production is still on Redis 6.0.16, and 1.0.17 — which introduced the BLMOVE-based claim — calls it unconditionally.

Reproduced on a live flux-klein box while upgrading it:

ERROR - Worker 0 crashed with error: unknown command `BLMOVE`, with args
        beginning with: `ml_tasks`, `inflight:18187a468ab3:0`, `LEFT`, `RIGHT`, `15`,

…repeating roughly thirty times a second, forever. The worker never registered, ml_tasks never drained, and nothing looked wrong from outside — ps showed modelq running and burning CPU, while every request came back "status": "queued" with an ETA climbing 30s at a time.

Nothing in production has hit this yet only because every deployed container still runs an older ModelQ (0.1.21 / 1.0.9 / 1.0.10 / 1.0.14 across the boxes I checked). Worker requirements.txt files already pin modelq==1.0.17, so the next image rebuild would have.

The fix

Probe BLMOVE once, cache the answer, and on unknown command claim tasks with an atomic Lua LPOP+RPUSH polled on a short interval.

Both properties the BLMOVE claim exists for are preserved:

• Atomic — the task is never absent from both the queue and the in-flight list.
• FIFO — claims come off the head, matching the rpush producer.

Only the blocking is emulated.

BRPOPLPUSH is deliberately not the fallback. It is the obvious Redis 6.0 substitute and it is atomic — but it pops the tail, which against an rpush producer silently turns the queue LIFO and starves the oldest request.

Only unknown command disables BLMOVE. Any other ResponseError (LOADING, etc.) leaves a capable server on the fast path.

Tests

tests/test_blmove_fallback.py — claim, atomicity, FIFO order, empty-queue timeout, probe-once (the crash loop was 30 exceptions/second), the modern-Redis control case, and the unrelated-error case.

Mutation-tested: forcing _blmove_supported() to return True turns five of them red with the exact production ResponseError.

Test extra bumped to fakeredis[lua] so the Lua path is actually exercised.

Full suite: 114 passed.

---

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