Parses the model's reply into findings the diagnostics SDK can publish.
The model is asked for a strict JSON array of {"line", "concern"}, but
models sometimes wrap their output in prose or fenced JSON blocks. This parser is
deliberately tolerant (it extracts the outermost JSON array and drops
anything malformed) and never raises. A bad reply yields [], which is
treated as "no findings" (and clears stale ones).
Line numbers arrive 1-based and are converted to the 0-based lines the diagnostics pipeline uses.
Summary
Types
@type finding() :: %{line: non_neg_integer(), concern: String.t()}