feat(contracts): add native loss and metric semantics

[GBeurier]

Summary

• transfer the absent score-provider scope into the native DAG-ML roadmap and record the integrated training/runtime/TCV1 authority from PR feat(runtime): integrate native training handoff #20
• add backend-neutral LossSpec, MetricSpec, distinct role references, and one generic implementation descriptor with opaque local registry keys
• add versioned built-in catalogs, compile-time loss resolution, strict schemas/fingerprints, executable-payload rejection, fixtures, and an independent Python oracle
• add typed metric evaluation task/result contracts, exact provider registry dispatch, decomposition/reduction validation, and a built-in adapter that reuses the existing numerical kernels
• route native score reports through validated provider results while preserving existing aggregate and per-target values

Independent review

• ownership amendment 8206fcc: approved with three documentation followups
• ownership followup 7c83262: approved, no findings
• semantic contracts fcd1afc: approved with four schema/Rust parity followups
• parity followup 7e9a363: approved, all prior findings resolved, no findings
• metric provider f062bb7: approved; one low-severity followup records that the legacy regression scoring API lacks a real classification task-kind input

Validation

• cargo test --workspace: 787 passed, 2 ignored
• cargo clippy --workspace --all-targets -- -D warnings: clean
• python3 -m pytest parity/criteria/tests/test_criteria_contracts.py parity/training/tests/test_training_replay_contracts.py -q: 22 passed
• python3 scripts/validate_contracts.py: passed
• python3 scripts/validate_training_replay_contracts.py: passed

Stack

Depends on draft PR #20 (agent/loss-runtime-integration).

Followups

• controller FIT_CV/REFIT loss transport and execution attestation will be a separate stacked PR
• binding-local executable registries will follow the controller protocol instead of expanding this contract PR
• replace the legacy generic regression/classification scoring entry point with a typed task-kind surface before exposing provider injection there