Improve job scoring for QA vs generic SWE roles; document Ollama on 16GB GPU

- Cap suitability score when roleTypeMatch is low and target roles are set
- Harden scoring prompt (role-type gating, examples, headline in resume payload)
- QA-oriented mock scoring when LLM is unavailable
- .env.example: optional Ollama MODEL/LLM_BASE_URL block for self-hosted LLM

Made-with: Cursor

.env.example

@@ -5,6 +5,12 @@
 
 MODEL=google/gemini-3-flash-preview
 
+# Self-hosted Ollama (e.g. 16GB GPU): use a 22B-class model for scoring/tailoring; pull the tag on the server first.
+# MODEL=mistral-small:22b
+# LLM_PROVIDER=ollama
+# LLM_BASE_URL=http://127.0.0.1:11434
+# Heavier option (~may offload layers to CPU on 16GB): qwen2.5:32b
+
 # DEPRECATED (auto-copied to LLM_API_KEY for compatibility)
 # OPENROUTER_API_KEY=your_openrouter_api_key_here
 
@@ -64,11 +70,18 @@ UKVISAJOBS_HEADLESS=true
 # =============================================================================
 # Adzuna (multi-country API source) - optional
 # =============================================================================
-# App credentials from Adzuna developer account.
+# Register at https://developer.adzuna.com/admin/access_details
 ADZUNA_APP_ID=
 ADZUNA_APP_KEY=
-# Optional default per-term cap (can be overridden by UI run budget logic).
+# Default cap per search term (orchestrator run budget / settings can override).
 # ADZUNA_MAX_JOBS_PER_TERM=50
+# API page size (Adzuna max 50).
+# ADZUNA_RESULTS_PER_PAGE=50
+# Optional global `where` text for Adzuna. Pipeline runs usually use Settings → search cities
+# instead; leave unset unless you want a fixed location for standalone extractor use.
+# ADZUNA_LOCATION_QUERY=
+# Only for running the extractor CLI alone; the pipeline sets country from your run (us / ca / gb / …).
+# ADZUNA_COUNTRY=gb
 
 # =============================================================================
 # JobSpy - Job search configuration