fix(llmops): extract a single valid JSON response

[hlubek]

Fixes #6. Skip non-JSON brace literals when extracting checker responses, reject multiple JSON objects, and require a non-null findings array. Adds 54 regression cases; full tests, vet, build, and lint passed for this final parser implementation.

Prompt wording changes were evaluated and reverted: observed scores were slightly worse, with no demonstrated benefit. Original prompts are retained.

Greptile Summary

Tightens parsing of JSON-shaped LLM checker responses.

• Extracts exactly one valid outermost JSON object while ignoring balanced non-JSON brace literals.
• Rejects ambiguous multiple objects, malformed or unbalanced responses, and null or missing findings arrays.
• Adds regression coverage for preflight and writing-guide response parsing.

Confidence Score: 5/5

The PR appears safe to merge, with no new actionable issues introduced since the previous review.

No blocking or non-blocking findings remain in the reviewed changes.

Important Files Changed

Filename	Overview
internal/llmops/preflight.go	Reworks shared JSON extraction and enforces a non-null preflight findings array without introducing a new issue since the previous review.
internal/llmops/writingguide.go	Enforces a non-null writing-guide findings array consistently with preflight parsing.
internal/llmops/json_response_test.go	Adds comprehensive regression cases covering valid prose, brace literals, malformed data, nested objects, and ambiguous responses.

_{Reviews (2): Last reviewed commit: "fix(llmops): restore prompts after incon..." | Re-trigger Greptile}