fix: require .trained marker to load classifier, preventing use of ra…

[Pcmhacker-piro]

Summary
Closes #335
The IntentClassifier's _has_trained_weights method only checked for the existence of weight files (pytorch_model.bin/model.safetensors), but did not verify the model was actually fine-tuned. This allowed random stub models (e.g., from download_model.py) to be loaded for inference, producing unreliable prompt injection detection.
The fix introduces a .trained marker file that the training pipeline creates after fine-tuning completes. The classifier now requires this marker to exist alongside the weight files before loading the model as a fine-tuned classifier. Without it, the system falls back to the deterministic heuristic classifier.
Type of Change

• Bug fix
• New feature
• Documentation update
• Refactor
• Tests
• Infra / CI
  Checklist
• I have read CONTRIBUTING.md
• My code follows the project style
• I have added/updated tests where relevant
• Tests/lint pass locally (if available)
• I have not committed .env or any secrets
• I have updated documentation if needed
  CHANGED FILES
• backend/app/modules/guard/intent_classifier.py
• guard-sdk/src/aegisai_guard/intent_classifier.py
• backend/tests/test_guard_explain.py
  COMMITS
  67187e9 - fix: require .trained marker to load classifier, preventing use of random stub weights
  TESTING PERFORMED
  git diff --stat

3 files changed, 44 insertions(+), 6 deletions(-)

FINAL STATUS

• Branch Name: fix/guard-classifier-trained-marker
• Commit Hash: 67187e9
• PR Created: No (PAT lacks createPullRequest scope — open manually at the PR link above)
• Ready for Review: Yes

[Pcmhacker-piro]

hyy @SdSarthak
i fixed this issus so please check it and if you found any bug then informe