[chrisknvidia]

Complete recursive scans could report MEDIUM/SAFE even when their child reports were MEDIUM/CAUTION. Reuse the normal risk-band mapping for the aggregate while preserving execution-failure and incomplete-analysis overrides, scoring, output schemas, and the existing greater-than-50 exit threshold.

Fixes #835.

Validation:

• 517 focused CLI/report tests passed, including real subprocess regressions for JSON, SARIF, Markdown, and terminal.
• 31 fresh installed-wheel CLI cases passed across risk boundaries, strict flags, file output, and incomplete scans; real MCP stdio clean/medium controls passed.
• Full non-provider regression suite: 13,721 passed, 14 skipped, 4 expected failures (245 integration/provider tests deselected).
• Additional TG validation: 39 actual installed-wheel CLI cases, seven base/head comparisons, and 10 actual HTTP MCP checks passed. Covered failed children, real output/count bounds, baselines, write errors, and concurrent scans.
• Independent architecture, code, and manual security reviews completed without actionable findings. All six hosted checks passed, including the full unit workflow.
• Follow-up review: 36 focused boundary/output checks passed (one 60-second subprocess timeout under host load passed on an unchanged isolated rerun). Five additional installed-wheel CLI checks passed for direct/recursive JSON and SARIF, strict exits and file output.

Provider-backed validation was not rerun. Native review-bot passes were unavailable; the listed reviews were manual.

[chrisknvidia]

Ready for review at 2e5fd982. There are no outstanding review comments or known actionable findings.

All six hosted CI checks have passed.

[rng1995]

[SkillSpector Review]

@chrisknvidia, thank you for tracking down #835 and fixing it at the source. Sharing one score-to-verdict helper between single-skill and recursive scans removes the duplicated band table that let the two drift apart.

Value and readiness

Before this PR, a complete --recursive scan whose top score was 21-50 reported MEDIUM/SAFE in the aggregate JSON risk_recommendation and SARIF riskAssessment.recommendation, while every child report said MEDIUM/CAUTION. The PR moves the band mapping into risk_severity_and_recommendation() in nodes/report.py. _compute_risk_score and cli._multi_skill_risk_assessment now both call it. The two recursive overrides stay: a failed execution forces DO_NOT_INSTALL, and incomplete analysis raises SAFE to CAUTION.

The report.py change is a pure extraction with the same loop and fallback, so single-skill output does not change. I checked every score band against the failed and incomplete flags. The PR is never less strict than main. Scores 0-20 and 51+ are unchanged. Scores 21-50 move from SAFE to CAUTION only for complete, non-failed scans. Exit codes are unchanged because the > 50 threshold is still checked separately. The aggregate now matches the README mapping (LOW to SAFE, MEDIUM to CAUTION, HIGH/CRITICAL to DO_NOT_INSTALL).

Verdict: correct and ready to merge.

Material findings

None blocking. One optional test note:

1. Optional: Some of the new test cases do not guard the fix (tests/unit/test_cli_subprocess.py:149, tests/unit/test_cli.py:2629).

   • Markdown/terminal subprocess cases. A complete recursive scan in these formats prints only each child's own report. The "Recursive Risk Assessment" block is appended only if analysis_incomplete. So "MEDIUM" in result.stdout and "CAUTION" in result.stdout are satisfied by the marked child's own section on both main and this PR.
   • Score-25 rows in test_recursive_non_json_caps_preserve_aggregate_risk. These rows always hit a cap, which sets analysis_incomplete=True, and main already returned CAUTION for incomplete scans scoring 50 or less.
   • What I saw. Running the PR's tests against main's source gives 4 failed, 48 passed. The failures are medium_recommendation[json], medium_recommendation[sarif], and the bands cases for scores 21 and 50 (complete, not failed). The markdown/terminal cases and all 8 score-25 cap cases pass on main. Expected: a regression test fails on main.
   • Suggested change. Either drop markdown/terminal from that parametrization, or make those cases assert what they actually cover: no aggregate section in complete output, and each child section matching its direct scan. Keep the score-25 cap rows as non-regression coverage. The JSON/SARIF cases and the bands test are the real guards, and they work. Optionally, write the bands test's expected override values as a literal table instead of recomputing them with the same if/elif logic as the code under test.

   This finding also means the PR description overstates the Markdown/terminal coverage a little. Fix coverage is still adequate.

Verification and gaps

• PR head 2e5fd982, merged with current main (314 commits behind, merges cleanly). Main has not changed the recommendation logic since the branch base. The nearby SC9 score floors merge in unchanged.
• Focused suites on the merged tree: test_cli.py, test_cli_subprocess.py, nodes/test_report.py, test_multi_skill.py, test_batch_scan_reports.py gave 754 passed. test_sarif.py, test_mcp_server.py, nodes/test_analysis_completeness.py gave 104 passed. The PR's new tests pass on both the head and the merged tree (52/52).
• Real CLI runs with --recursive --no-llm and a custom YARA marker, merged tree vs main, in all four formats, at scores 0, 10, 20, 21, 25, 50 and 67. Also covered: mixed children, a symlinked child, an executable binary that triggers the 51 floor, a failed child (LLM requested with no credentials), and a 34-skill collection past the 32-skill budget. Flags tried: --fail-on-findings, --fail-on-incomplete, --min-coverage 100, --verbose, --output. The only difference was SAFE to CAUTION for complete 21-50 scans in JSON/SARIF. Exit codes matched in every case, and no input produced SAFE when any child was HIGH/CRITICAL, failed, or incomplete.
• A child with a degraded LLM stage is marked incomplete, so the aggregate is still raised to CAUTION.
• CI: all 6 checks passed on 2e5fd982.
• Not verified: provider-backed LLM runs. The "LLM partly failed but reports complete coverage" case could not be reached without a real provider. An unreachable endpoint marks the scan failed, which fails closed.