[snchimata]

Summary

• Prepare 0.5.2 across six existing Cargo registry packages, the Python binding, npm root/five platforms and optional Select Python runtime.
• Default the approved experimental Select worker to pinned FP32 ONNX Select v2 CPU; preserve explicit approval, original text/order, protected groups, bounded admission/IPC/deadline and deterministic fallback. Core remains model-free. Legacy PyTorch backend remains explicit.
• Add Select wheel/sdist prepublication checks and registry-version consistency checks.
• Incrementally update README and the offline quickstart notebook, with disabled-by-default ONNX scoring pinned to the published full model commit, Safetensors guidance and separate legacy v1 evidence. Add notebook regression tests.

Validation

• Rust fmt/clippy/workspace: 692 local passes, 1 existing ignored subprocess helper; six Cargo publication packages packaged/verified. Audit/deny passed.
• npm ci and 12 TypeScript/platform tests passed.
• Windows abi3 Python wheel built; 22 Python tests passed, including full offline notebook execution without model imports/downloads and mocked opt-in model path.
• Select wheel/sdist built/twine-checked; 17 installed CPU/mock contracts passed. Final wheel imports without ML dependencies.
• Deterministic fidelity smoke passed (not model qualification).
• Actual ONNX reference/native worker/CLI integration passed three synthetic forwards, exact worker/reference scores, required/source-order text and 79/80 token target.
• Independent Safetensors CPU validation passed all 76 tensors bit-for-bit and two synthetic score/rank/packing cases, no reconversion. Owned worker/guardian exited/waited, prior failed scopes preserved.

Model publication (authorized separately)

Public Apache-2.0 ONNX + Safetensors preview: https://huggingface.co/snchimata/tokenfold-select-v2/tree/78736d9afea819322f212d6d53eedcf68ac7a39b
All 45 remote files verified against the exact approved payload. Explicit maintainer approval covered final model/tokenizer terms and recorded source applicability. V1 remains unchanged. No weights are in Git or registry packages.

Release boundaries

Wait for all exact-head protected-main checks; no bypass or direct push. Tag v0.5.2 only on merged main, then existing release/publishing workflows and registry verification. New tokenfold-select PyPI publishing uses the existing credential; report any scope failure rather than silently omitting a package.

V2 is an opt-in experimental preview, NOT representative quality/latency/full-8k/GA qualified. Historical PyTorch FP16 results do not benchmark the ONNX or Safetensors CPU references. Keep original full-v2 gates for subsequent iterations.

[snchimata]

Safetensors has been added to the separate local model release bundle alongside ONNX; ONNX remains the approved CLI worker default. No weights are included in this PR or registry packages.

The alternative contains 76 FP32 tensors / 98,864,641 values, model.safetensors, a strict custom joint-group reference loader/architecture, corrected distribution config, tokenizer, environment and computational manifest. It is not a stock AutoModelForSequenceClassification artifact. ONNX computational files and manifest remain byte-identical.

Current validation state: both saved synthetic Safetensors parity cases have exact source score/rank agreement, but the run exited with an unsupported terminal-report status label after those cases. This failed scope is preserved, not relabeled passed. The reporting fix was tested without model execution; one fresh CPU-only validation of the existing weights is awaiting explicit approval. No reconversion is planned. Combined format readiness remains false until that validation succeeds.

No Hub upload, tag or registry publication performed. The outstanding model target/visibility/terms and new-project PyPI credential checks still apply.

[snchimata]

The explicitly approved fresh CPU-only validation of the existing Safetensors files has now passed (no reconversion):

• All 76 FP32 tensors / 98,864,641 values matched the original source checkpoint bit-for-bit; strict complete state loading passed.
• Both synthetic requests (1/8 groups, 20/125 tokens) matched the source scores exactly, with identical ranks and packing; two source plus two Safetensors forwards.
• CUDA stayed unavailable/uninitialized. The native worker and independent original-handle guardian both exited zero and were stopped/waited; all 61 frozen input pins were verified before documentation updates.
• The prior failed scopes remain failed and unchanged. The accepted terminal report and ready markers were independently audited.

The local combined model release bundle is resealed: 44 allowlisted files, ONNX default plus Safetensors alternative. ONNX and Safetensors computational files/manifests did not change during finalization; README, provenance, status and successful public validation evidence were updated. Static model-card/link/example/integrity/ONNX graph/Safetensors header checks passed without further model execution.

This clears format/loader mechanics only, not full-8k, representative quality, latency, redistribution or release gates. No upload, tag or registry publication occurred. Final model target/visibility/terms, maintainer merge and publisher scope checks remain outstanding. PR code/head is unchanged at 5199f73; all required CI checks are green.