Fix: PrometheusMiddleware exports metrics from only one worker when --workers > 1

[sergey-chek]

Fixes #590

Problem

start_http_server() in startup() uses the default REGISTRY, which only contains metrics from the current process. With --workers 2, one worker binds port 9000 and serves only its own metrics, the other workers' mmap files are never read.

Root Cause

Each worker process writes its metrics to a separate mmap file (counter_{pid}.db) in PROMETHEUS_MULTIPROC_DIR - this part works correctly. However, start_http_server() without the registry= parameter serves from the default REGISTRY, which only calls Counter.collect() / Histogram.collect() on the current process's values (source).

Per prometheus_client multiprocess docs, a separate CollectorRegistry with MultiProcessCollector is required to aggregate mmap files from all processes.

Fix

• Pass CollectorRegistry + MultiProcessCollector to start_http_server() in startup()
• Remove duplicate os.environ assignment (line 42)

Testing

Tested with --workers 2 on taskiq 0.12.1:

• Before fix: 5 tasks executed (PID 10: 3, PID 11: 2) -> received_tasks_total = 3.0 (one worker only)
• After fix: 12 tasks executed (PID 10: 7, PID 11: 5) -> received_tasks_total = 12.0 (all workers aggregated)