Release packed RecurQuant cache with held-out MBPP confirmation

[Labeeb2339]

What changed

• adds physically packed INT4/INT8 recurrent-state storage for Qwen3.5 Gated DeltaNet caches;
• exposes the frozen mixed-v0.2 policy through the Python API, CLI, and Colab quickstart;
• preserves packed payloads exactly during beam reorder;
• publishes the frozen MBPP development and 500-task held-out confirmation records;
• adds independent artifact/checkpoint verification with hash-anchored CI;
• documents compatibility, prior art, claim boundaries, and reproducible commands;
• prepares the 0.2.0a1 package and generated evidence charts.

Held-out result

On the pinned Qwen/Qwen3.5-0.8B-Base revision, the frozen layer-0 INT8 plus 17-layer INT4 layout reduced task-macro excess NLL from 2.949743 to 0.803713 versus uniform INT4, a 72.75% reduction, over 500 MBPP test tasks and 30,244 teacher-forced reference-code tokens.

The paired mixed-vs-uniform improvement was 2.1460 nats/token with 95% CI [2.0922, 2.1999]. The layout used exactly 2,564,096 resident recurrent-state bytes. Every preregistered quality gate passed.

This is a scoped teacher-forced fidelity and resident-byte result. It is not a novelty, breakthrough, generated-code correctness, speed, peak-memory, whole-model-memory, or cross-model claim.

Verification

• full local test suite passed;
• Ruff lint and deterministic asset checks passed;
• artifact-only and raw-checkpoint confirmation verification passed;
• pinned full-checkpoint CUDA quickstart passed with exact byte accounting;
• final wheel/sdist passed strict Twine checks;
• isolated final-wheel import, demo, and confirmation CLI passed;
• Gitleaks found no leaks across all commits;
• dependency audit found no known vulnerabilities in installed dependencies.

The raw 34.4 MB checkpoint is kept out of Git and will be attached to the v0.2.0a1 prerelease as a 9.1 MB archive so the full raw-array verification remains available.