Build the v0.4 consumer batched runtime - #30

Summary

• add typed, mask-isolated padded update batches for SFT, offline KD and strict OPD
• keep optimizer-group semantics while making physical batch size independently configurable
• add a local typed role graph without Ray, DataProto, FSDP or distributed API emulation
• add one-base student/teacher/optional-reference adapter ownership with failure-safe role restoration and standard student PEFT export
• publish a preregistered RTX 4080 runtime matrix, profiler summary, data-bound Pareto SVG and systems-only frozen adapter

Measured result

The final matrix used Qwen3-0.6B NF4 weights with FP32 compute, eight fixed SQLite trajectories, one strict-OPD update, one warmup and three measured repetitions per cell.

• dual batch-4: 3.866 trajectories/s, 3.035 GiB peak reserved, 1.63x sequential throughput
• shared batch-4: 3.475 trajectories/s, 2.227 GiB peak reserved, 1.54x sequential throughput
• sharing saved 26.6% reserved memory at batch-4 but was 10.1% slower than dual ownership
• auto was not optimal because padding all eight trajectories added 1,636 tokens
• trajectory and teacher-target digests were identical across all eight cells
• all 12 preregistered loss/full-gradient/post-update-logit comparisons passed

Preregistration revisions 1.1 and 1.2 were public before the sole headline run. They record the non-headline diagnostics that replaced nondeterministic SDPA with eager attention and BF16 compute with FP32 compute so the original strict numerical gates remained meaningful.

Scientific scope

This is a systems release. It does not claim a task-quality gain, universal speedup across GPUs or models, a batched rollout server, or verified/distributed verl parity. The shared benchmark teacher is explicitly not a newly qualified teacher. No compatible preregistered 4B/7B adapter was available, so those cells remain not_run with reasons.

Validation

• Ruff check and format check: pass
• mypy: 84 source files, pass
• full non-GPU/non-network suite: 1,448 passed, 6 deselected, 86.26% branch coverage
• available GPU suite: 5 passed
• available network suite: 3 passed
• focused HF/config/runtime suites: pass, including one-base loading, optional frozen reference, role restoration and shared/dual update equivalence
• both shipped Qwen recipes validate without downloads
• package build and twine check: pass for wheel and sdist
• Markdown/link, generated PyPI README and JSON Schema checks: pass
• generated SVG byte comparison: pass; native and ~820 px visual inspection: pass
• immutable calculator JSON SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

Frozen artifacts

• result: a302da31af99f1d29f1efd4e6b3dbeb6ea4ac956bba102ca8a1bee8dff0319eb
• profiler: 66111cd7fc876cf1befea3297a1a51bcd99252c0bf8989c029381e1dc155a98b
• Pareto SVG: 98645a668a7832423d28b621262292619615917f037adf7219ff1bf071fb2fea