Task comments and fixed headers/scaffolds; deterministic prefix assembly and first-function scope extraction without semantic repair. Same 350 tasks/tests and 96–256 token budgets. All 350 reference bodies pass static roundtrip checks. Full runtime validation and model score improvements remain unverified.
Bound browser candidate hangs with a phase-aware watchdog
Batch 64 with OOM backoff, length/budget buckets, cached decoding and no per-token CPU streamer. Versioned 96–256 output budgets. Preserve full2000 as an archived config. Full sandbox reference execution and GPU throughput remain unverified.
Allow cached model resolution during offline evaluation
2000 generated examples; exact language/difficulty allocation, isolated runners and Hugging Face evaluation. Static audit passes. Full reference and negative-control execution remains unverified because a Docker/Podman sandbox is unavailable.
