Efficiency proof

Each claim below has the measurement or test that backs it.

Claim	Proof
Picking k of Q jobs costs O(Q + k log Q), not O(Q log Q)	engine/pipeline.py:select (lazy heapq); B1: 5.5 ms vs 31.6 ms for Q=50k, identical picks
Applying scales with containers, not one machine's browser	--scale applier=N; 5 appliers × 4 tabs ran from the bank on 2026-10-10 (pipeline log: dispatch: ... -> batches/inbox (container worker) × 5)
No job is applied twice across workers	B3: 200/200 exactly-once over 3 runs; test_worker_claims_each_batch_once
Tabs overlap waiting (page loads, uploads)	round-robin scheduler, S1: ×2.05 vs sequential
Email codes need no human	relay log: code -> torcrobotics_... within one 20 s tick; the first tick on the bank delivered 3 codes
The queue holds only fillable work	prune(): on 2026-10-10 only 17 of 229 queued jobs were fillable; after the 30-day scan, 488 usable kept and 179 dropped
Freshest jobs first	strict bands (hour, day, week, cold), test_select_prefers_fresh_band
Disk stays bounded	engine clean every pass; first run freed 459 MB; pending resumes protected
Zero model cost	no API key anywhere in requirements.txt or .env.example; AI only through your own MCP client