Add intent extraction test suite and fix priority extraction

[anfredette]

• Add 39-scenario test suite for LLM intent extraction covering all 9 use cases, GPU/model extraction, priority enforcement, user counts, conversation history, and UI button strings. Run explicitly via make test-intent (requires Ollama); 2 smoke tests run as part of make test-integration.

• Rewrite the intent extraction prompt as a short, directive format with ordered pattern-matching rules, replacing the verbose prose prompt. Remove experience_class, complexity_priority, and additional_context from the prompt — fields that were either inferred deterministically or never consumed downstream. Consolidate the prompt as self-contained (remove INTENT_EXTRACTION_SCHEMA constant and schema_description parameter from extract_structured_data).

• Fix unreliable priority extraction: the LLM (qwen2.5:7b) was inferring priorities from use case type rather than from explicit user statements, causing incorrect scoring weights. The LLM now returns *_mentioned booleans alongside *_priority values, and post-processing resets priority to "medium" when the topic was not explicitly mentioned.

• Harden post-processing: case-insensitive normalization for domain_specialization, experience_class, use_case, and *_mentioned booleans; priority value aliases (e.g. "very_high" -> "high") applied before validation; logger.warning when an unrecognized use_case cannot be resolved by alias map or fuzzy match. Add 4 unit tests for case-insensitive normalization.

[anfredette]

@amito I suggest you test this with the use cases that were giving you trouble. And, better yet, we can add them to the test cases now or in a later pr.