fix: stabilize CPU training runtime

feat: enforce held-out benchmark workflow

fix: train_custom.py never writes to original yml — use temp file to prevent total_iter corruption

feat: add sequential model improvement roadmap (Phase 1 complete — yml config fixes)

feat: complete custom training pipeline bugfixes, dataset crawling, and validation

Fix training crash on CPU, reduce val_freq, add OpenCV env vars

perf: maximize CPU/RAM/disk for CPU training (LMDB, avx512, larger batch/queue)

feat: optimize CPU performance and add agents rules

feat: add custom stage 3 training runner and fix Windows/CPU resume path compatibility in BasicSR