fix(e13-s5): accept markdown-fenced JSON from real LLMs in --ai parse

[imneov]

What

Live-testing the --ai path against a real Anthropic model (claude-sonnet-4-5, via a relay) surfaced a bug that mocked tests could not: the model returns perfectly correct content but wraps the JSON in a ```json … ``` markdown fence, despite the system prompt asking for no markdown. ParseCoreJSON's strict decode then failed on the leading fence, so --ai refused every real response — the feature was effectively unusable against live models.

Fix

internal/aiparse now strips a single outer markdown code fence before the strict JSON decode. Strictness is not weakened: after stripping it still DisallowUnknownFields, rejects nested objects / arrays / nulls / trailing data / empty objects, and only peels one well-formed outer fence (multiple/incomplete/odd-language fences fall through to refusal).

Evidence

• Unit tests: fenced live-response fixture parses; bare ``` fence parses; fenced_unknown_field still refuses (strictness not bypassed); all prior non-conforming cases still refuse.
• Live end-to-end (real relay, synthetic secret-free blob) after the fix:
  base_url=https://api.anthropic.com model=claude-3-5-sonnet-20241022 api_key={{CLAUDECM_SECRET_1}} small_fast_model=claude-3-5-haiku-20241022 — parsed successfully; placeholder preserved for local re-injection.
• Fresh-context adversarial review: APPROVE (strictness preserved, stripping robust, request/secret/credential path untouched).

Response-parse side only; no change to the outbound redaction, secret handling, or credential path.

🤖 Generated with Claude Code