train: add replayable memory extraction debugging

[chenjw]

Summary

• add a generic trace-backed replay framework with entry and mock points, Pydantic serialization, and debug_trace integration
• refine trajectory and experience extraction around exact-case evidence, candidate experiences, and post-validation gates
• modularize the retained experience gates and trace each gate result
• reduce duplicated VLM trace payloads and add training run metadata and git-note support

Validation

• ruff check on changed Python files: passed
• ruff format --check on changed Python files: passed (71 files)
• relevant changed tests: 386 passed, 3 failed
  • test_patch_merge_uses_original_messages_for_output_language: expected zh-CN, received 中文
  • test_embedding_handler_binds_registered_operation_telemetry
  • test_resource_service_add_resource_reports_queue_summary: test double lacks the current poll_interval argument
• full pytest -q: collection blocked by the repository test environment (pytest_html_report_title hook plugin unavailable; 12 collection errors)

Draft status

This remains a draft while tau2 extraction quality is being tuned. The latest airline task 22 epoch-1 post-write evaluation was 0/8 even though the generated experiences were retrieved and read, so the experience content and trajectory evidence still need follow-up.