Establish Rust-first learned provider runtime

[epi13]

What changed

• records ADR 0001 accepting Rust as the default production language for learned micro-provider hosting, dispatch, SDKs, snapshot reuse, budget enforcement, and CPU-first implementations
• retains Python as the training, calibration, experimentation, export, and high-level MNEL control plane
• defines mnel-provider-c-abi/1 in matching Rust and C forms, with identity-bound snapshots, resource budgets, host-owned result storage, and diagnostic-only result semantics
• adds a safe Rust provider SDK and a persistent-host policy crate
• encodes admission tiers for Rust-native, quarantined WASM, and external experimental providers
• requires benchmark evidence and a threat-review identity before a non-Rust implementation can enter the trusted native tier
• adds the Python runtime manifest model, JSON schema, example manifest, and negative tests
• updates the README, roadmap, changelog, repository map, and CI

Why

Micro-providers are expected to be small and frequently invoked, so process startup, weight loading, serialization, copying, and snapshot construction can dominate inference. This PR fixes the implementation boundary before runtime work begins: persistent Rust hot path, Python research plane, compact identity-bound snapshot views, and a stable language-neutral ABI.

The contract deliberately contains no evaluator verdict, conformance, acceptance, causal-truth, or promotion field. Learned outputs remain diagnostic-only and can only inform the next bounded Forge question.

Developer impact

New provider work now has a concrete target:

1. implement CPU-first providers through the Rust SDK by default;
2. export Python-trained models into the versioned runtime boundary;
3. bind declarations, artifacts, calibration, feature extraction, queries, snapshots, and budgets in the runtime manifest;
4. use an evidence-backed exception for specialized C/C++/Zig or vendor-native kernels;
5. measure snapshot construction separately from warm inference.

Validation

Completed in the available execution environment:

• Python source compilation
• runtime-policy smoke test, including Rust-native admission and Python-native rejection
• JSON parsing for the manifest schema and checked-in example
• C11 header compilation with both GCC and Clang using -Wall -Wextra -Werror
• branch diff review against the current main head

Cargo is not installed in the execution environment used to prepare the PR, so a local Rust compile could not be run. The PR adds CI jobs for cargo fmt --check, Clippy with warnings denied, and the full Rust workspace test suite; the PR remains draft until those checks confirm the new crates.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.