feat: agent-workflows E2E implementation (#652)

* fix: agent-workflows E2E via terraphim-llm-proxy with Cerebras

Three bugs fixed to get all 5 agent workflow demos working end-to-end:

1. #[serde(flatten)] nesting bug: Role.extra field with flatten causes
   JSON "extra" key to nest as extra["extra"]["key"]. Added
   get_extra_str/get_role_extra_str helpers that check both flat and
   nested paths in agent.rs and multi_agent_handlers.rs.

2. rust-genai hardcoded Ollama endpoint: v0.4.4 hardcodes
   localhost:11434. Rewrote from_config_with_url to use
   ServiceTargetResolver to override endpoint at request time.

3. Model name adapter routing: rust-genai selects adapter from model
   name. Used openai:: namespace prefix (e.g. openai::cerebras:llama3.1-8b)
   to force OpenAI adapter for proxy-compatible endpoints.

Config switched from Ollama to terraphim-llm-proxy (bigbox:3456) with
Cerebras llama3.1-8b. 6-step prompt chain completes in ~10s vs minutes.

Co-Authored-By: Terraphim AI <noreply@anthropic.com>

* test(agent-workflows): fix Playwright browser tests for all 5 workflows

- Use HTTP URLs (localhost:3000) instead of file:// to avoid CORS blocking API calls
- Add correct button selectors per workflow (was using wrong IDs for routing, evaluator)
- Add per-workflow setup functions to fill required form inputs before triggering API calls
- Handle alert() dialogs in headless mode that were silently blocking execution
- Evaluator-Optimizer: generate mock content first, then trigger real /workflows/optimize API
- Skip fragile comprehensive test suite page, test individual workflows directly
- Add .gitignore for test artifacts (screenshots, reports, lockfile)

Results: 6 passed, 0 failed, 1 skipped in 57s via Cerebras through terraphim-llm-proxy

Co-Authored-By: Terraphim AI <noreply@anthropic.com>

* fix(agent-workflows): replace parallelization mock data with real LLM output

generatePerspectiveAnalysis() and generateAggregatedInsights() were returning
hardcoded mock data, ignoring actual API responses. Now parses LLM markdown
from parallel_tasks[].result into structured UI components (title, keyPoints,
insights, recommendations, confidence).

Co-Authored-By: Terraphim AI <noreply@anthropic.com>

---------

Co-authored-by: Terraphim CI <alex@terraphim.ai>
Co-authored-by: Terraphim AI <noreply@anthropic.com>

Author: 3 people