fix(anthropic): harden structured JSON parsing

[lzhs1995]

Summary

• accept one whole-response JSON fence across LF/CRLF, unlabelled, and case-insensitive json forms
• split empty structured output into kiro_structured_empty_output
• log privacy-preserving diagnostics without raw response content
• keep prose, unsupported labels, malformed fences, and trailing content fail-closed

Validation

• local linux/amd64 release build: pass
• focused structured_json_is_strict_and_canonical: 1 passed, 0 failed
• git diff --check: pass

Deployment boundary

Train A only. This PR does not enable Gate B flags, change RPM, modify credentials/proxies, or authorize shadow/canary/enforce. Any new kiro.rs digest invalidates the prior R2 evidence and requires a fresh R2 anchor.

[Copilot]

Pull request overview

This PR hardens Anthropic structured JSON output handling by making fenced JSON extraction more robust (LF/CRLF, unlabeled fences, case-insensitive json), introducing a dedicated error kind for empty structured output, and adding privacy-preserving diagnostics that avoid logging raw response content.

Changes:

• Broaden accepted whole-response JSON fences (unlabeled and json label, case-insensitive; LF/CRLF supported) while still failing closed on prose, unsupported labels, malformed fences, or trailing content.
• Split “empty structured output” into a distinct kiro_structured_empty_output error kind.
• Add warn-level structured diagnostics (lengths, fence classification, first non-whitespace classification, per-process salted digest, optional parse position) without logging raw content, plus tests for the new parsing rules.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.