[Bug]: parse_json_markdown loses literal JSON fence markers inside valid payloads

Finding: In an external report against llama-index-core (reporter-reproduced at run-llama/llama_index main commit 962940d; not reproduced by KFA or independently; the reporter says the helper is unchanged at 81f0e06), parse_json_markdown reportedly returns None for a complete, valid fenced JSON object or array when a JSON key or string value contains the literal json fence marker (https://github.com/run-llama/llama_index/issues/23391). Reported symptom: no exception is raised. The minimal example wraps the object {"text": "```json"} in a json fence and calls the helper. The reporter says the actual result is None, where the original object was expected (https://github.com/run-llama/llama_index/issues/23391). The reporter says the same failure occurs with marker-containing keys, nested objects, array roots, and strings holding Markdown snippets. Reporter's diagnosis (not maintainer-confirmed): the extraction step splits the text on the json fence marker and takes the second piece, so a marker inside the payload is treated as another delimiter and truncates the selection. The empty extraction then reaches the YAML fallback, which returns None (https://github.com/run-llama/llama_index/issues/23391). This is a hypothesis from the reporter's reading of the code, not a documented or maintainer-confirmed cause. Scope stated by the reporter: complete valid object/array JSON followed by its closing fence. Raw/unfenced JSON, missing closing fences, malformed JSON, and YAML-tolerated payloads are outside the proposed fix (https://github.com/run-llama/llama_index/issues/23391). Evidence status: the reporter says a proposed regression file gives 6 failed / 11 passed on the baseline and 17 passed with their prepared fix. These were isolated-helper checks using the actual utils.py and base.py, PyYAML 6.0.3, Python 3.12.14, pytest 8.4.2, with networking blocked. Package initialization, repository conftest, full-package tests and downstream callers were not run (https://github.com/run-llama/llama_index/issues/23391). The report discloses AI assistance (Codex) and makes no claim that the human contributor ran the tests or reviewed the code line by line. Fix status: the supplied data shows no merged commit, maintainer response, or released version. The prepared patch is an unmerged proposal and is not a confirmed or shipped fix. Limitation: this is an external report only; KFA did not reproduce it, and the affected released versions are unknown. Anyone parsing LLM output that may contain a literal json fence marker inside JSON content should treat a None result from this helper as possibly wrong until a fix is verified.