fix: preserve work intent when normalizing permissions

[giscebot]

Summary

• Treat work_intent as canonical when normalizing intent-classifier output.
• Normalize inconsistent write_permission values to match work_intent instead of downgrading work_allowed to review_only.
• Expose normalization_warnings in intent-classifier metadata when a field is corrected.

Why

A trusted PR comment asking @giscebot to add an e2e test produced raw LLM output with work_intent=work_allowed and write_permission=none. The previous normalizer silently converted that to review_only, making the dashboard look like the LLM had classified the request as read-only.

Tests

• pytest tests/test_intent_classifier.py tests/test_parser.py