fix(agent-eval): bump LLM-judge default model off EOL claude-sonnet-4 (#965)

The LLM-judge default was pinned to claude-sonnet-4-20250514, which is
deprecated and now returns a 404 not_found_error. CI doesn't set
EVAL_SCORE_MODEL, so it fell back to the dead default and every scenario
failed. Bump the default (and the .env.example / help-text references)
to claude-sonnet-4-6.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: alexluong

docs/agent-evaluation/.env.example

@@ -36,4 +36,4 @@ EVAL_TEST_DESTINATION_URL=
 # Scoring is ON by default after each scenario (heuristic + LLM). Opt out:
 # EVAL_NO_SCORE_HEURISTIC=1
 # EVAL_NO_SCORE_LLM=1
-# EVAL_SCORE_MODEL=claude-sonnet-4-20250514
+# EVAL_SCORE_MODEL=claude-sonnet-4-6

docs/agent-evaluation/src/llm-judge.ts

@@ -8,7 +8,7 @@ import { basename, dirname, join } from "node:path";
 import { extractTranscriptScoringText } from "./score-transcript.js";
 
 const ANTHROPIC_MESSAGES_URL = "https://api.anthropic.com/v1/messages";
-const DEFAULT_SCORE_MODEL = "claude-sonnet-4-20250514";
+const DEFAULT_SCORE_MODEL = "claude-sonnet-4-6";
 const MAX_TRANSCRIPT_CHARS = 180_000;
 
 export interface LlmCriterionJudgment {

docs/agent-evaluation/src/score-eval.ts

@@ -60,7 +60,7 @@ Score an eval transcript.
   npm run score -- --llm [--write]      # Anthropic judge (needs ANTHROPIC_API_KEY)
   npm run score -- --llm --no-heuristic # LLM only (no regex heuristic)
 
-Heuristic: src/score-transcript.ts. LLM: reads scenarios/*.md Success criteria + assistant text; model from EVAL_SCORE_MODEL (default claude-sonnet-4-20250514).
+Heuristic: src/score-transcript.ts. LLM: reads scenarios/*.md Success criteria + assistant text; model from EVAL_SCORE_MODEL (default claude-sonnet-4-6).
 
 Options:
   --run <path>      transcript.json, a run directory, or legacy flat *-scenario-NN.json