docs: fix program naming in russian README

fix(demo): improve screencast demos and update scenario

feat(demo): add screencast scenario + inference/quantization demos for ЭВМ

docs: export V100 18-epoch metrics for paper plots

docs: add AGENTS.md for contributor and agent onboarding

fix(docker): use system pip so DataSphere test #6 passes

fix(docker): add jupyter user required by DataSphere image tests

docs: add 18-epoch experiment reports and per-epoch metrics JSON

perf(train): scale V100 params and fix bulk staging disk check

docs: add Russian experiment spec and README, remove EN MCE file

chore: add local Cursor folder to gitignore

perf(train): bulk stage frames on extended SSD before training

fix(train): lazy frame staging to avoid root disk exhaustion on Jobs

fix(jobs): write metrics and status to DataSphere output paths

perf(train): cap smoke runs and stage frames on local SSD

fix(train): force clean DataSphere exit and report job progress

feat(train): tag batch size in job launch and batch GPU forwards

fix(train): ensure clean DataSphere job exit after GPU training

fix(jobs): wire --params into DataSphere job input uploads

fix(extract): drop PyTorch from stage-1 CPU frame extraction

fix(jobs): make DataSphere requirements parseable for manual Python env

feat(train): two-stage pipeline with CUDA-safe GPU jobs

fix(jobs): fix train job cmd and submit asynchronously

docs(cloud): add DataSphere S3 connector setup guide

docs: adopt Conventional Commits standard

docs(cloud): document Object Storage upload and AWS signing pitfalls

feat: train at native 336×336 Kvasir resolution

docs: add ADR-001 and core unit tests

feat: add uv project and Datasphere Jobs launcher

Train at native 336×336 Kvasir resolution with real video loader.

Document training architecture in ADR-001 and add core unit tests.

Add uv project and Datasphere Jobs launcher for codec training.