fix(eval): distinguish pending and completed Judge results

[BigDataDZ]

When an L3 Judge returned malformed text, reflection was degraded but the eval report showed judgePending=0 and the dashboard counted the attempted call as a finished judgment. Keep every degraded outcome pending and add judgeAttempted/judgeCompleted fields while retaining legacy judgeAvailable/judgeRan semantics. Display completed judgments separately from fixture pass rates, and clarify that correctly rejecting a negative fixture also counts as a passing fixture.

Extract the L3 case runner to respect the existing size ratchet. Production reflection and its parser acceptance rules remain unchanged. Completion means an existing-parser-accepted, nondegraded verdict; this does not add strict schema validation.

Validation:

• Regression first reproduced missing pending status for malformed/empty/unrelated responses; focused eval/reflection tests 26/26 passed.
• Full npm test after integrating current main: 503 tests, 491 passed, 12 optional skips, 0 failures.
• Real Edge browser: pending-only, mixed and completed-only states passed with no page errors; original UI failed the regression. Parent independently reran browser plus eval tests: 11/11 passed, no skips.
• npm run lint: 0 errors, 7 existing warnings; npm run eval: 8/8 fixtures, 2 L3 judgments pending.
• Independent task and broad final reviews approved. Added the new browser regression to the existing Chromium CI job. The initial PR checks passed, including the new Chromium test; after integrating current main, exact-head checks rerun real PostgreSQL, browser acceptance and both Docker builds.