[Golden007-prog]

What does this PR do?

Follow up on the requested history limit and adversarial-input coverage in the review of #239, which has now merged.

HTML benchmark reports accept at most 20 historical evaluations, preserving their explicit order. Excess history raises ValueError before any brief is rendered. The writer renders before opening its destination, so a rejected history neither truncates an existing report nor creates an empty one. The limit is documented as a count of historical briefs, not a byte limit on an individual evaluation.

Tests cover the 20/21 boundary, preserving existing or absent output files on rejection, and inert handling of script/style breakouts, SVG event handlers, iframe/image payloads, JavaScript links, and encoded HTML across history labels and metrics. Existing numeric-meter restrictions, Markdown parity, and input immutability are also checked.

How was this tested?

• The two excess-history regression cases failed before the guard.
• python -m pytest tests/test_authority_canonical.py tests/test_authority_benchmark.py -q: 66 passed.
• python -m pytest tests/ -q: 1,328 passed, 3 skipped.
• ruff check safeai/ tests/: passed.
• git diff --check: passed.

The repository-wide ruff check . additionally reports three existing findings in examples/crewai_example.py and two fuzz harnesses; those files are unchanged. The merged main tree was verified identical to the tested parent tree before this follow-up was based on it.

Checklist

• Tests pass locally (python -m pytest tests/ -q)
• Lint passes (ruff check safeai/ tests/)

AI-assistance disclosure: implementation and verification assisted by OpenAI Codex.

[IkarusAILabs]

This is a solid follow-up to PR #239 that addresses the DoS concern