analyzer_statuses reports completed while counting out-of-scope binary members as failed, so SkillEvaluator rejects the report #800

Summary

When a skill contains binary content that SkillSpector itself classifies as out_of_scope
(reason_code: binary_content), the static-pattern analyzers count those units as failed
but still report status: "completed". The report is therefore internally inconsistent:

• analysis_completeness.status is "complete", is_complete: true, coverage_percent: 100.0
• 15 static_patterns_* analyzers report status: "completed" with completed: 13, failed: 8 out of planned_work: 21
• the same 8 units are listed in scope_exclusions as outcome: "out_of_scope", fatal: false

SkillEvaluator's report-contract check rejects this combination
(analysis_completeness.analyzer_statuses has a status that contradicts its work accounting)
and marks the whole Tier 1 security scan INCOMPLETE. Any downstream tool that relies on
SkillEvaluator therefore receives no security opinion for such skills.

Versions

• SkillSpector v2.11.2 (uv tool install "skillspector @ git+https://github.com/NVIDIA/SkillSpector.git@v2.11.2") — reproduced on Windows 11, Python 3.13
• SkillSpector v2.12.0 — same result in a python:3.13-slim Linux container
• SkillEvaluator v0.3.0 and v0.5.0 — both reject the report the same way
• Semgrep 1.177.0 present; --no-llm

Reproduce

Public sample: trailofbits/overtly-malicious-skills at commit 4ffbf9461ef0505f9ce76a0d3694a18ec33ea531,
skill skills/context-loader (it ships a .instructions.docx.txt, a DOCX archive containing embedded .ttf fonts).

git clone https://github.com/trailofbits/overtly-malicious-skills
git -C overtly-malicious-skills checkout 4ffbf9461ef0505f9ce76a0d3694a18ec33ea531

skillspector scan overtly-malicious-skills/skills/context-loader --no-llm --format json --output ss.json

skillevaluator validate overtly-malicious-skills/skills/context-loader \
  --tiers 1 --checks security --no-dedup -r json -o out

Actual

ss.json (excerpt):

"analysis_completeness": { "status": "complete", "is_complete": true, "coverage_percent": 100.0, "fully_inspected_files": 12, ... }

{"analyzer_id": "static_patterns_prompt_injection", "status": "completed",
 "planned_work": 21, "completed": 13, "partial": 0, "skipped": 0, "failed": 8, "unaccounted": 0}

"scope_exclusions": [
  {"outcome": "out_of_scope", "phase": "static", "reason_code": "binary_content",
   "message": "Binary content is unsupported by this analyzer.",
   "path": ".instructions.docx.txt!/word/fonts/AlegreyaSans-bold.ttf", "fatal": false, ...},
  ... 8 entries, all `word/fonts/*.ttf` members of the DOCX
]

The same completed + failed: 8 pattern appears on all 15 static_patterns_* analyzers
(static_patterns_tool_misuse reports degraded with partial: 1, failed: 8).

SkillEvaluator output:

Security Scan: INCOMPLETE
skillspector JSON field 'analysis_completeness.analyzer_statuses' has a status that contradicts its work accounting; security scan did not complete

Expected

Units excluded as out_of_scope should be accounted consistently, e.g. as skipped
(or excluded from planned_work), not as failed, so that an analyzer whose only
non-completed units are non-fatal scope exclusions can report completed without
contradicting its own counts. Alternatively, if failed is intended, the analyzer status
should not be completed.

Impact

In a 105-skill local library, the SkillEvaluator security scan came back INCOMPLETE for 102
skills with this same error message. The root cause was confirmed only for the sample above;
the other 101 were not traced individually.

Related but distinct: #557 (static_yara reporting completed after dropping rule files), and
SkillEvaluator NVIDIA/SkillEvaluator#137 (documentation-only skills reported incomplete).