feat: add split leakage and trajectory consistency validators

[VarshiniGunti]

Summary

This PR introduces focused data-integrity guardrails for the logs-to-training pipeline by adding two validation layers:

1. Cross-split leakage detection (train/val/test)
2. Agent trajectory consistency checks (tool-call/tool-result linkage)

The scope is intentionally narrow and infrastructure-focused so it complements existing broad pipeline PRs without duplicating end-to-end ingestion/export implementations.

What’s Added

1. Split Leakage Validator

• Added training_setup_logs/integrity.py
• Normalizes text and computes stable fingerprints to detect near-duplicate overlaps across splits
• Flags cross-split collisions with split names and sample indices for debugging
• Useful for preventing train/validation/test contamination and accidental overlap between derived subsets

2. Trajectory Consistency Validator

• Added training_setup_logs/trajectory_validator.py
• Validates tool-use sequence integrity:
  • tool_call must include tool_call_id
  • tool_result must include tool_call_id
  • each tool_result must map to a prior tool_call
• Produces structured issues that can be used to exclude or quarantine invalid trajectories

3. Validator CLI

• Added training_setup_logs/validate_cli.py
• New commands:
  • validate-splits
  • validate-trajectories
• Writes machine-readable JSON reports for CI/review workflows
• Returns non-zero exit code when violations are found, enabling straightforward gating

4. Tests and Fixtures

• Added tests:
  • tests/test_integrity.py
  • tests/test_trajectory_validator.py
• Added example fixtures:
  • examples/train_sample.jsonl
  • examples/val_sample.jsonl
  • examples/test_sample.jsonl
  • examples/trajectory_valid.jsonl
  • examples/trajectory_invalid.jsonl

5. Documentation

• Updated README.md with usage commands and validator workflow for this PR prototype

Why This Matters

This PR addresses two critical failure modes in training data preparation:

• Leakage risk: near-identical prompts across splits can inflate offline metrics and harm generalization.
• Broken trajectories: inconsistent tool traces degrade behavior-cloning quality and can introduce invalid supervision.

By adding these checks early, downstream SFT/DPO preparation can rely on cleaner, policy-aligned data.

Validation Performed

• python -m py_compile for new modules/tests
• python -m pytest -q (all tests passed)
• validate-splits CLI run on sample split files (leakage intentionally detected)
• validate-trajectories CLI run on valid and invalid trajectory fixtures (expected pass/fail behavior)

Files Included

• training_setup_logs/integrity.py
• training_setup_logs/trajectory_validator.py
• training_setup_logs/validate_cli.py
• tests/test_integrity.py
• tests/test_trajectory_validator.py
• examples/train_sample.jsonl
• examples/val_sample.jsonl
• examples/test_sample.jsonl
• examples/trajectory_valid.jsonl
• examples/trajectory_invalid.jsonl
• README.md