M3: Normalize structured AI review output

[dbrosio3]

Goal

Define and enforce one structured review output format across all local AI providers.

User value

Findings look consistent in local terminal output and provider-neutral future rendering regardless of provider.

Implementation notes

• Define JSON schema for findings: category, severity, confidence, file, line, message, suggestion, and source/provider metadata.
• Validate provider output and provide safe repair/fallback behavior where appropriate.
• Keep prompts explicit that source code/diffs are untrusted data.
• Preserve useful v1 prompt-injection framing.

Risks / tradeoffs

• Some providers may not reliably emit strict JSON without repair.
• Overly strict parsing may discard useful findings.

Suggested priority

P1

Milestone

M3: Local AI review

TBD considerations

• Exact severity/category taxonomy.
• Privacy policy for including full-file context in review payloads.

Acceptance criteria

• Structured output schema exists and is documented.
• Malformed provider output is handled predictably.
• Terminal rendering uses normalized findings, not provider-specific text.