feat(python): execute typed local criteria

[GBeurier]

Summary

• resolve and execute required Python-local losses from native NodeTask contracts during FIT_CV and REFIT, with attestations emitted only after successful callback execution
• execute process-local Python metrics from typed MetricEvaluationTask values and build validated native results
• expose direct register_local_loss / register_local_metric helpers that build signed host-local references while keeping callable code process-local
• record typed early-stopping outcomes independently from selection and reporting scores
• reject unknown in-process selection metrics instead of silently substituting RMSE
• document the current L0-L7 implementation boundary and add positive stable-identity reconstruction coverage
• consume the exact paired dag-ml-data branch consistently across the stacked CI jobs

Validation

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace: 802 passed, 2 ignored
• cargo run -p dag-ml-cli -- validate-graph examples/minimal_graph.json
• PyO3 Clippy with warnings denied
• PyO3 Rust tests: 25 passed
• Python binding tests: 22 passed
• Python extension source freshness check passed
• D4 replay oracle and migration suite: 163 passed
• conformal/robustness oracle suite: 55 passed
• primary and D4 training/replay contract validators passed
• shared contract validation against dag-ml-data#15 passed
• Sphinx documentation build passed with warnings denied

Review

Each implementation batch received an independent read-only review. Findings were addressed before publication:

• early-stopping schema review requested clearer cross-field semantics; the schema now documents the relationship explicitly
• lifecycle review detected a stale checked-in extension; the extension was rebuilt and the freshness gate passes
• direct registration and unknown-selection-metric reviews found no actionable defects
• final status/test review found an R/MATLAB ownership overstatement; the roadmap now attributes those facades to nirs4all-core over the C ABI
• replay-pack review confirmed all 91 base artifacts and all 136 transitive replay artifacts are byte-current; no drift is masked by the refreshed pins
• conformal-oracle review confirmed the restricted import guard permits only the pinned Unicode 17 backport beyond the standard library and still excludes production dependencies
• peer-aware CI review found and removed an unused MSRV peer checkout; Rust/Python/WASM now consume the single dynamically resolved peer ref

Residual boundaries are explicit in the roadmap: full Python nirs4all controller migration awaits approval, and real nirs4all controller integration remains required in every official language before the ecosystem custom-loss objective is complete.