fix(sdk): parse structured AI assessments

[TimeToBuildBob]

Summary

The SDKs AssessCode() and BatchAssessCode() return CodeAssessment with always-empty Issues and Recommendations fields, despite the model producing structured output. The SDK docs and example worker branch on these empty fields, encouraging dead control flow.

Changes

1. internal/discovery/sdk/ai.go: Both AssessCode() and BatchAssessCode() now request raw JSON output and parse it using the existing internal/ai JSON parsing utilities.
2. internal/discovery/sdk/ai_internal_test.go: New package-local tests covering direct JSON parsing, fenced JSON, batch snippet ID preservation, and malformed/missing/duplicate batch results.
3. Richer CodeAssessment fields: Issues, Recommendations, and Confidence now come from real model output instead of empty placeholders.
4. Batch mode now preserves per-snippet results instead of cloning one response text into every result.

Verification

go test ./internal/discovery/sdk

Closes #... (no issue opened for this, it was a scout finding)