Recover PDF text by decompressing streams when the fetch tool fails

A URL-fetching tool that converts pages to markdown returned unusable binary for a PDF, but it had already saved the raw file locally and printed the path in the tool result. Text is recoverable without pdftotext: zlib.decompress each object between stream/endstream markers, then regex out parenthesized string literals. Filter each decompressed block by length and printable-character ratio, otherwise embedded images decompress into megabytes of mojibake that swamp the real text. Separately, repeated heredoc python invocations failed because the shell working directory reset between Bash calls — writing the extractor to a scratchpad .py file and running it by absolute path was the only stable form.