Stabilize live browser eval runtime

[justsml]

What changed

• Stabilized the live browser-eval lifecycle: shared Juice Shop runs serialize safely, wait for target readiness, and browser /api/chat requests now cancel before the row deadline.
• Made browser-side aborts report as timeout rather than a generic harness error.
• Added strict Mastra inference-span polling so valid telemetry is not rejected before persistence completes.
• Preserved the fresh smoke, repeat, and timeout artifacts, including LFM tuning evidence and the repaired DeepSeek/GLM/Qwen/GPT runs.

Why

Long-running live evaluations could leave a browser fetch pending after the Node-side timeout, while shared resettable Docker targets could race between rows. Those issues produced harness-only failures and prevented reliable repeat runs.

Validation

• pnpm exec tsc --noEmit --pretty false
• pnpm exec vitest run tests/security-chat/browser-e2e-failures.test.ts
• Production /api/health check
• Fresh strict-evidence GPT smoke
• DeepSeek Flash repeats 2 and 3, plus GLM repeat 2, with persisted streams, usage, model-inference spans, and target cleanup.