test: regression coverage for providers, search, portal, and workflows

[cursor]

Risky behavior now covered

• Litigation provider factory (get_provider): ollama/lm_studio/openrouter selection, missing API key guard, OpenRouter attribution headers and extra_body config
• Web search routing (search_web): Tavily vs DuckDuckGo backend selection and domain attribution parsing
• Portal generate post-processing: CSS injection, personality/vote styling, transcript index generation
• Transcript export: markdown-to-HTML conversion, personality styling, exports dir resolution, timestamped filename titles
• Litigation runner: slugify edge cases, case allocation without registry, courtroom save location, Scribe certification footer
• Prompt assembly: matter/feasibility inclusion, hearing-type variants, spectator section toggle
• Research report workflow: report ID format, HTML wrapping for email, save path/header

Test files added/updated

• tests/test_litigation_providers.py (new, 8 tests)
• tests/test_workflows_search.py (new, 6 tests)
• tests/test_portal_generate.py (new, 3 tests)
• tests/test_research_report.py (new, 5 tests)
• tests/test_litigation_run.py (+7 tests)
• tests/test_export_transcript.py (+4 tests)
• tests/test_prompt_assembler.py (+4 tests)

Why these tests materially reduce regression risk

These paths sit on core courtroom flows with large blast radius: LLM provider selection feeds both litigation deliberation and the research-report pipeline; portal post-processing and export determine whether transcripts render correctly; and transcript save/case-number logic prevents duplicate Case Nos and broken paths. All 98 tests pass (python3 -m pytest tests/).

Repos skipped

• Agent_RPG_Simulator: no new production code since last coverage run (docs-only commit 7cb8be5)
• Agent_Officespace: README/docs only