Fix eval explorer comparisons and complete Kimi low coverage

[justsml]

What changed

• require meaningful row and task coverage before model profiles enter headline charts
• order Anthropic, OpenAI, and Google first, then other companies alphabetically
• classify root exploit-hunter/network-lab results as Network Labs
• omit radar axes without scored coverage instead of drawing misleading blank spokes
• complete the missing Kimi K3 Native low-reasoning model-tool tail with an 8,192-token ceiling
• retain the zero-cost native-temperature failure diagnostic and its external failure research
• regenerate the checked-in explorer dataset with the new results

Why

Sparse canaries and mixed-suite outliers were distorting headline token, time, and cost charts. Network browser rows were also misclassified because the category matcher required a trailing slash. The historical Kimi report referenced a clean completion run that had never been checked in.

Kimi completion

The corrected native profile uses temperature=0.6, maxTokens=8192, and low reasoning. The nine-row tail cost $1.200962, used 34/408 tool calls, scored 198/218 deterministically, and completed all 9 judges. Seven rows passed; two valid quality failures remain visible.

Validation

• pnpm -s run run:vitest run tests/security-chat/eval-explorer-build-data.test.ts
• node --check evals/explorer/app.js
• pnpm eval:explorer