Eval datasets are test-case data, not installed skill code.

Two identical findings in same file → one output.

Existing behaviour: when both sides carry the same concrete end_line, the finding is kept (no regression from the new fallback).

Finding whose context contains 'example:' gets downweighted.

Parameter description exceeding 500 chars → TP3 finding, confidence ~0.65.

Parameter description with exfiltration phrase → TP3 finding.

An extensionless file (inferred as 'other') in code-example context is downweighted, not dropped.

A transient failure costs only its own batch, not the whole fan-out.

Break response_schema, call _apply_patches, verify it raises and no methods are patched.

If SKILL.md has no name in frontmatter, use directory name.

When the optional nv_inference subpackage can't be imported, the default/``nv_inference`` selection degrades to ``NvBuildProvider``.

When anthropic_proxy is selected but not configured, falls back to OpenAI.

Finding whose context contains ``` gets confidence halved.

Same rule_id in different files should be independently filtered.

A finding in an executable file is NOT dropped when context contains a generic indicator. Validates that an attacker cannot suppress a genuin

Some LLMs return the findings array as a JSON string, not a list.

mcp_poisoned_tool fixture triggers TP1, TP2, and TP3 findings.

``from builtins import compile`` must still raise AST6.

``from builtins import __import__`` must still raise AST3.

``from builtins import eval`` must still raise AST2.

``from builtins import exec; exec(code)`` must still raise AST1.

``from builtins import exec as e; e(...)`` must still raise AST1.

``from builtins import exec`` with tainted input must raise TT5.

``from pathlib import Path`` should resolve p.read_text() correctly.

``from socket import socket`` should resolve s.recv() correctly.

Bare-imported ``import_module('os').system(...)`` must raise AST5.

Bare-imported ``import_module('os').system(input())`` must raise TT5.