Add prototype log-to-training pipeline

[vandit98]

Summary

This PR adds an early Python-first prototype for the DMP 2026 log-to-training-data pipeline described in #1.

The implementation focuses on the first reviewable end-to-end slice:

• JSON/JSONL ingestion and normalization into a canonical event schema
• deterministic rule-based PII/secrets redaction with stable placeholders
• session segmentation into Q&A units and agent trajectories
• trajectory complexity tagging for training schedule buckets
• basic validation for tool-call/tool-result consistency
• LoRA-ready SFT JSONL export
• human-review DPO candidate JSONL export
• sample agent logs, schema docs, and implementation plan

Verification

• python -m pytest
• sample pipeline run produces 2 SFT rows, 1 DPO candidate row, PII counts, tag summary, and 0 validation issues

Notes

This is intentionally dependency-light and stdlib-only for the first pass, so mentors can review the data model and pipeline shape before we add heavier trainer integrations, NER models, Hugging Face datasets, PEFT/TRL dry runs, or organization-specific PII dictionaries.