Add immutable turn-level training export v1

[Zeng-Weijun]

Summary

• add immutable turn-training-export-v1 bundles for main-agent turns and session-memory compaction tasks
• preserve exact provider requests, raw responses, retry attempts, tool-result closure, termination metadata, and final patch references
• expose run_id and training_export_path through Python and OpenAI-compatible entrypoints
• reserve rollout targets before provider side effects and validate bundles with a standalone CLI
• preserve interrupted, failed-provider, failed-compaction, and missing-patch trajectories without publishing incomplete clean rollouts

Validation

• 39 turn export contract, API, and entrypoint tests passed
• existing OpenAI helper, Python API/tool, edge-case, and extra-tool offline checks passed
• compileall, YAML parse, git diff --check, and git show --check passed
• two independent adversarial reviews accepted commit 4e330c6 with no remaining P0/P1 findings

Follow-up smoke

This change was validated entirely with offline fake providers. A real relay task plus downstream DataFactory ingest smoke should be run after merge; no live API or external dataset mutation was performed in this PR.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.