Refactor llm-judge.ts from raw API-only to dual-mode: CLI first
(spawns claude with --output-format text, unsets CLAUDECODE), API
fallback. Backend selectable via LLM_JUDGE_BACKEND env var.
Fix pre-existing race condition in config.rs tests where parallel
test execution caused env var mutations to interfere. Added static
Mutex to serialize env-mutating tests.
Adds 6 new E2E scenarios in phase-b.test.ts covering multi-project grid
rendering, independent tab switching, status bar fleet state, and
LLM-judged agent response quality evaluation via Claude API.
Includes llm-judge.ts helper (raw Anthropic API fetch, haiku-4-5,
structured verdicts with confidence thresholds).
7 human-authored test scenarios (22 tests) using data-testid
selectors. Test fixture generator for isolated environments.
JSON results store (no native deps). WebDriverIO config updated
with TCP readiness probe and multi-spec support.
Tauri creates one app session per spec file; multiple files caused
invalid session id on subsequent specs. WebDriver clicks on Svelte 5
components inside scrollable panels dont trigger onclick handlers
via WebKit2GTK/tauri-driver - use browser.execute() JS clicks.
Also removed tauri-plugin-log (redundant with telemetry::init()).
