'gh api --paginate' follows the Link header and merges results. For endpoints that return a JSON array at the top level, such as the issues or releases lists, the pages are concatenated into one array and the output is valid JSON. For endpoints that return an object wrapping an array, such as the search endpoints with their total_count and items keys, or the check-runs endpoint, each page is emitted as a separate top-level JSON object. The result is a stream of objects, not one document, and piping it into a strict JSON parser fails.
The fixes are to use '--jq' so that gh's built-in jq processes each page as it arrives, for example '--jq ".items[]"', which yields one record per line; or to pass '--slurp', which collects the pages into a single JSON array. Piping to an external jq with '--slurp' on the jq side also works.
Also remember that --paginate ignores a per_page you did not set; pass '-F per_page=100' to reduce the number of round trips and the rate limit cost.