Fix _strip_code_fences truncating JSON when content contains inner backticks

[poog26]

The current implementation uses content.split("```json")[1].split("```")[0] to strip markdown code fences. This finds the first ``` after the opening fence — so when the extracted JSON itself contains literal triple-backtick characters, the split matches those inner backticks and truncates the JSON mid-string.

When this bites: Hindsight's fact-extraction LLM produces output wrapped in ```json ... ``` fences. If one of the extracted facts describes code fence behavior (e.g. "the model wraps output in json fences"), the JSON now contains ` json as a literal string value, andsplit("```")finds that inner occurrence. Result: truncated JSON →json.loads` fails → retry loop (up to 4 retries per chunk).

The fix replaces string-splitting with line-based fence detection. Per the markdown spec, code fences must appear at the start of a line. Inner triple backticks inside JSON string values are preserved because they aren't at line boundaries.

Before (broken):

_strip_code_fences('```json\n{"facts": [{"what": "uses ```json fences"}]}\n```')
# → '{"facts": [{"what": "uses '  (truncated!)

After (fixed):

_strip_code_fences('```json\n{"facts": [{"what": "uses ```json fences"}]}\n```')
# → '{"facts": [{"what": "uses ```json fences"}]}'  (intact)

Edge cases: normal fences ✓, inner backticks preserved ✓, no fences no-op ✓, opening-only (truncated output) returned unchanged ✓, text around fence block handled correctly ✓.

[benfrank241]

thanks for this! just had to add a test. merging