fix(service): hash-tag job cancellation keys for Redis Cluster

[TomasPalsson]

What breaks

On the sandbox (Redis in cluster mode), every programmatic run (run_tools_with_bash, and any PTC job) fails. The worker errors right away with:

CROSSSLOT Keys in request don't hash to the same slot

The API never learns that the job failed, so it waits the full JOB_TIMEOUT (about 5 min). Then it returns a bare Code execution failed. Plain /exec runs like echo hi still work.

Cause

The claim/commit/cancel Lua scripts in job-cancellation.ts touch three keys for one job: <key>, <key>:result and <key>:execution. The key had no hash tag, so on a cluster those keys hash to different slots. reconcile() also sent one MGET across many jobs, which crosses slots as well.

Fix

-  return `${JOB_CANCELLATION_PREFIX}:${encodeURIComponent(queueName)}:${encodeURIComponent(jobId)}`;
+  return `${JOB_CANCELLATION_PREFIX}:${hashTag(`${encodeURIComponent(queueName)}:${encodeURIComponent(jobId)}`)}`;

reconcile() now does one GET per job instead of a cross-slot MGET.

Verification

• A new unit test asserts that a job's keys share one slot. It fails before the fix and passes after; bun test src/job-cancellation.test.ts passes 24/24.
• Run against a real single-node cluster-mode Redis (--cluster-enabled yes), the old code fails with the exact production CROSSSLOT error. The new code completes claim, commit and read.
• tsc --noEmit reports the same 7 existing errors before and after; none are in the changed files.
• Not run: job-cancellation-commit.test.ts, because it needs a local redis-server.

Rollout

Deploy the API and the worker together, because the key names change.