Issue #273: Security | Hidden PDF text poisons resume extraction and inflates hiring scores

It only edits the work prompt rules and appends the text again as a user message, which adds no trust boundary or filtering.

It only adds an optional job_description parameter to evaluation, with no change to the poisoned-text handling.

It only splits the text into sections for token savings; extraction still takes all spans, the hidden text is still passed to the LLMs, and evaluator.py is untouched.

It extracts per-section snippets by regex from the same unfiltered text, so hidden spans still pass through and evaluation is unchanged.

It only replaces logger calls with print statements and does not touch extraction, templates or evaluation.

It adds certificate and volunteer extractors that use the same unfiltered text and raw template insertion.

It only adds None handling for malformed JSON and does nothing about hidden text or output validation.

It validates the file extension, MIME type and size, but the content of the PDF is still extracted unfiltered.

It moves settings into a config file, and its MAX_* constants in evaluator.py are only defined there and never applied to clamp the output.

It closes the document and adds validation, with no change to what text is extracted or how scores are checked.

It only adds a None check on the extracted JSON and a log-level change.

It switches to page.get_text('text'), which still extracts every span including invisible ones, and adds no visibility filtering, grounding or clamping.

It reworks header detection and section splitting but still extracts all text with no visibility check, and evaluation is unchanged.

It switches to structured output parsing, which enforces the schema but not the truth of the content, and has no visibility filtering, grounding or clamping.

Its evaluator.py and pdf.py changes only switch to structured output parsing, and the contributors-API change in github.py concerns fork detection, not grounding of the resume claims.

It adds job description analysis and a job_match_score and does not touch hidden text or grounding.

It adds JD-augmented scoring and batch processing, and still builds the evaluation from raw LLM JSON with no grounding or clamping.

It adds retries, JSON repair, caching and parallelism, and the hidden text still flows through with the LLM JSON accepted unchecked.

It only makes the role and organization configurable in the evaluation template.

It only changes a log level and logging stream routing.

It only adds PoC PDFs, scripts and write-ups under security/ and explicitly states it implements no fix.

It passes role-based score weights into the prompt, which are only prompt text and not enforced by clamping or grounding in code.

It adds a provider and retry logic for invalid JSON in evaluator.py, with no sanitization, grounding or score clamp.

It expands the extraction prompts, but the text is still inserted raw with no untrusted-data boundary, and the extraction and evaluation code are unchanged.

It adds ignore_graphics=True to to_markdown so more text is kept, which makes hidden-text extraction no better and arguably worse.