Codex/persisted probes main

[mvyleteljr]

Summary

Adds the Xenon core support needed to replay and score persisted linear probes cleanly.

• Adds staged probe fine-tuning through ProbeSpec.train_stages and stage_epochs
• Adds optional persisted model export from ProbeSpec
• Adds persisted probe import and inference operation specs
• Wires persisted-probe operations into the API exports, operation registry, execution dispatch, and artifact-bound runtime specs
• Adds known-answer tests for staged fine-tuning and persisted probe export/import/inference replay

Why

Behavior Audit needs to score Tau2 traces across precomputed assistant, emotion, and high-stakes spaces. The high-stakes adapted probes must preserve staged fine-tuning metadata instead of being replayed as ordinary merged fixed train values.

Validation

uv run pytest tests/pipelines_v2/operations/test_execution_oracles.py tests/test_pipelines_v2_basics.py -q

Result: 203 passed, 10 warnings.

The warnings are existing numpy empty-slice warnings in probe selectivity/subset tests.