feat(n8n): add Ollama embeddings to duplicate detector

[mikemikimike]

Closes #5

Background

The duplicate issue detector currently requires OpenAI embeddings. That prevents fully local, private deployments even though the repository already uses Ollama for other workflow integrations.

Changes

• Add disabled Ollama HTTP Request nodes for both the live detector and the backfill flow.
• Keep OpenAI as the enabled default and connect both providers to the existing extraction and Qdrant paths, so users can switch providers without rebuilding the workflow.
• Accept both OpenAI (data[0].embedding) and Ollama (embeddings[0]) response shapes with a clear error for malformed responses.
• Document Ollama setup, provider switching, endpoint reachability, model dimensions, Qdrant recreation, and the no-double-processing requirement.
• Add a regression test covering the node configuration, connections, response handling, and required documentation.

Compatibility

OpenAI remains the default and existing OpenAI/Qdrant behavior is unchanged. Ollama is opt-in and uses nomic-embed-text by default. Because embedding dimensions are part of the Qdrant collection contract, the README requires a 768-dimensional collection for nomic-embed-text and a full backfill after switching.

Validation

• python -m pytest -q tests/test_duplicate_issue_detector_workflow.py — passed.
• python -m pytest -q tests --basetemp .pytest-tmp — passed (60 tests).
• python scripts/validate_contracts.py with PYTHONPATH=src — passed (5 contracts).
• python -m ruff check tests/test_duplicate_issue_detector_workflow.py — passed.
• python -m compileall -q src scripts tests — passed.
• JSON parsing of implementations/n8n/duplicate-issue-detector/workflow.json — passed.
• python -m build --wheel --no-isolation — passed.
• No live Ollama, Qdrant, GitHub, or OpenAI services were used; the regression test validates the exported workflow structure and response-shape handling offline.

AI disclosure

This pull request was prepared with AI assistance. The implementation was reviewed against the issue acceptance criteria, and all added behavior is covered by a deterministic local regression test.