Define MNCS-native training pipeline and lineage direction

[epi13]

Summary

Establishes the architectural direction for making MNCS the semantic owner of MNEL training rather than treating an external Python/ML stack as the source of truth.

This PR adds docs/MNCS_NATIVE_TRAINING_PIPELINE.md and links the work into the project roadmap.

Goals

• Treat model training as an evidence-producing MNCS computation.
• Close the lineage gap between source observations, training, model artifacts, and deployment.
• Define the intended integration contract with mncs-lineage.
• Keep external numerical/accelerator runtimes available as execution backends without surrendering semantic ownership.
• Use the training pipeline as a deliberate stress test of mncs-language, its compiler, and standard library.
• Promote reusable primitives—numeric, graph, stream, parallel compute, checkpoint, artifact, observability, deterministic randomness—into shared MNCS layers when they are useful outside ML.
• Prioritize small/task-specific MNEL models as early native-training candidates.

Implementation strategy

The document defines a phased migration:

1. MNCS-owned training manifests and lineage over existing trainers.
2. MNCS-native dataset construction and evaluation semantics.
3. MNCS training primitives.
4. Backend lowering to CPU/accelerator/external numerical runtimes.
5. Fully native training for selected micro-models.
6. Full lineage closure through training, checkpoints, transformations, admission, and deployment.

Non-goals

This does not propose rewriting CUDA, ROCm, BLAS, PyTorch, or every existing trainer immediately, nor does it change MNEL's diagnostic-only learned-provider authority or hard-gate evaluation rules.

Why now

MNEL already has native provider-runtime work, distributed sufficient-statistic training, and Python-trained artifact export work in progress. Defining the ownership boundary now prevents those transitional mechanisms from becoming the permanent architecture and turns future training work into useful pressure on the MNCS language and standard library.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3fd37cb0b7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Reconcile native training with the accepted Python policy

When Phase 5 is implemented, this instruction directly contradicts the still-accepted policy in docs/decisions/0001-rust-provider-runtime.md:25-34, which says Python remains MNEL's training language. Because this document is labeled only as architectural direction and does not supersede or amend that ADR, contributors now have two conflicting implementation policies for micro-model training; update the ADR's status/decision or explicitly define native training as an exception.

Useful? React with 👍 / 👎.