Fix accidental local image-generation routing

[dovvnloading]

What changed

• Removed the keyword-based image-generation heuristic from send_message().
• Normal chat text now always uses the selected chat provider, including Ollama and other local runtimes.
• Image generation remains available through explicit node actions only.
• Added regression coverage and updated the architecture findings.

Root cause

send_message() silently routed messages containing image-related verb/noun pairs into api_provider.generate_image(). The provider correctly rejected that call outside API Endpoint mode, producing Image generation is only available in API Endpoint mode. even though the user was using local chat.

Validation

• python -m pytest graphlink_app\\tests
• 394 passed