fix(docs-eval): reduce LLM judge verdict variance

[alexluong]

The docs-agent-eval job on run 29590304174 failed because the LLM judge scored scenario 02 as fail on a transcript that objectively succeeded (heuristic 9/9, live event id printed and verified). The saved llm-score.json shows the verdict was incoherent: two criteria have pass: false while their own evidence text concludes pass ("Pass for TS script", "The final TS script DID run successfully"), and the judge read the scenario's scripted turn 1 ("Option 1. Let's do it in TypeScript") as the user pushing back on a wrong deliverable, calling the SDK version "an afterthought".

Two changes in llm-judge.ts: pin temperature: 0 on the judge call (it previously used the API default of 1.0 — maximum variance for what is a classification task), and add a system-prompt paragraph stating that user turns are scripted by the harness, so an option-choice follow-up is scenario design rather than pushback, and criteria satisfied by the final artifact pass regardless of which turn produced it.

Parse/prompt tests pass (npm run test). Not re-judged against the saved failing transcript locally (no Anthropic key on this machine); to verify: download the run's artifact and npm run score -- --run <scenario-02 dir> --llm. Neither change makes an LLM fully deterministic — if flakes persist, majority-of-3 judging is the next step.

🤖 Generated with Claude Code