feat: add AI evaluation for SA-01.01, SA-02.01, SA-03.01, SA-03.02

[vinayada1]

Summary

Add optional AI evaluation paths for four OSPS security-assessment requirements:

Requirement added to the AI path	Evaluation focus
OSPS-SA-01.01: design documentation	System actions, actors, and interactions
OSPS-SA-02.01: external interfaces	Documentation of external software interfaces and their contracts
OSPS-SA-03.01: security assessment	Identification and prioritization of likely and impactful security problems
OSPS-SA-03.02: threat modeling	Attack surface, critical paths, threats, and mitigations

Scope

This PR's new AI paths only evaluate evidence explicitly declared in Security Insights; they do not look for additional evidence on their own. Design review uses the declared detailed guide; interface review uses the detailed and quickstart guides; security-assessment and threat-model review use the declared self and third-party assessment evidence.

The AI evaluates the contents of those artifacts rather than treating a filename or declaration as proof of sufficient documentation. This PR does not add repository-wide evidence discovery, source-code scanning, arbitrary external fetching, or downstream-link traversal. Existing deterministic checks remain separate from these new AI paths.

Evidence and result handling

• Fetch only same-repository GitHub blob/raw text URLs through the configured Contents API client, preserving the declared ref. Validate paths, formats, regular-file content, and UTF-8; do not follow redirects or download URLs.
• Cache successful and failed reads across checks. Bound retrieval to 64 KiB per document, 16 distinct URLs, and a 64 KiB assembled evidence packet.
• Defer unsupported, unavailable, comment-only, or incomplete declared evidence to NeedsReview without asking the model to grade a partial packet.
• Keep AI optional. Without AI configuration or relevant declarations, use the deterministic path. Configured-provider, configuration, and response-validation failures return NeedsReview/Low; they do not preserve an earlier Pass or Fail.
• Require affirmative documentary support for AI failures; absent or unfetched evidence alone should defer. Tighten the risk rubric to distinguish explicit likelihood-and-impact prioritization from generic security posture.
• Return Low confidence for AI NeedsReview results and cap design passes at Medium.

Important SA-02.01 limitation

Every SA-02.01 AI pass recommendation is returned as NeedsReview/Low, requiring human confirmation. Live grading sometimes accepted insufficient interface documentation despite prompt tightening, so this is enforced in code, not merely requested by the prompt.

The original AI verdict, explanation, citations, and supplied material remain in the evidence for human review. AI Failed and NeedsReview responses retain their normal result handling. This intentionally also defers otherwise valid AI pass recommendations.

Non-AI behavior changes

SA-01.01 now applies the published-release applicability gate already used by the other three checks and guards optional SI documentation fields. Its deterministic missing-design result now explicitly uses Medium confidence. Release detection still uses GitHub Releases, not tags alone or releases distributed elsewhere.

Validation

• Passed focused security-assessment and reusable-step tests, including prompt goldens, optionality/fallbacks, and SA-02 pass deferral with Low/Medium/High model confidence while preserving original evidence.
• Passed go test -race ./data ./evaluation_plans/osps/sec_assessment ./evaluation_plans/reusable_steps -count=1.
• Passed targeted golangci-lint, make -B build, and git diff --check after the final policy change.
• Passed the native .github/scripts/ci.sh integration with all four supported catalog IDs.
• Full go test ./... -count=1 and golangci-lint run ./... passed on the updated upstream before the final SA-02 policy change; the related checks above were rerun afterward.
• Replayed genuine SI-declared Meshery and FDC3-Sail evidence through gpt-5.4-mini three times each: all nine final scanner results matched the expected Meshery SA-03.01 NeedsReview, FDC3-Sail SA-01.01 Passed, and FDC3-Sail SA-02.01 NeedsReview outcomes.

The committed opt-in live test exercises production grading with caller-supplied captured fixtures; it does not replace full scanner retrieval tests or run in the normal unit suite. These finite examples do not establish general model accuracy. Existing SI reader rate limiting and legacy-schema incompatibilities remain outside this PR.