JobQueue now skips bands swept within MinResweepHours (config, default 6h) instead of re-scraping the whole catalogue continuously — the dominant cost on the metered residential proxy. Roughly linear savings with no data loss (full pagination retained); 0 disables it. Worker logs the real compressed transferSize per job (what the proxy bills) rather than the ~6.5x-larger decompressed length, so spend is visible.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
_relay waited for both pipe directions (gather), leaking a task holding two sockets on every half-closed tunnel — visible as a flood of pending-task lines under load. Tear the tunnel down when either side closes (FIRST_COMPLETED + close both writers), matching the .NET LocalForwardingProxy's WhenAny. Also move the worker's noVNC to an ephemeral host port so replicas don't collide under 'docker compose up --scale worker=N'.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Brings up the pull-model scraper: the .NET C2 hands skin+wear jobs to Python nodriver workers that scrape cs.money and post results back, plus the supporting Core/EFCore data model, migrations, and docker-compose orchestration.
IPRoyal proxying lets workers scale horizontally with a distinct residential exit IP each: every worker process mints its own sticky session at startup, and an in-process forwarding proxy injects the gateway auth so Chromium talks only to an auth-free localhost endpoint (zero CDP). On a Cloudflare challenge a worker rotates to a fresh session/IP and re-warms. Verified end-to-end against live IPRoyal: distinct US residential exits per worker and IP rotation on demand.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
