feat(security): integrate llm-sanitizer for prompt-injection detection

[warnes]

Summary

Replaces the bespoke 11-pattern substring list in cv_orchestrator.py with a centralised prompt_safety module backed by the llm-sanitizer library.

What changed

scripts/utils/prompt_safety.py (new)

Centralised security module exposing three public functions:

Function	Purpose
scan_text_for_injection(text, min_risk)	Boolean check; union of llm-sanitizer rule scan + supplementary substring check for cv-builder-specific phrases in DOM text fragments
sanitize_instruction_text(text)	Strips injection content via redact() then word-boundary regex pass; returns (cleaned_text, findings) in legacy cv-builder dict format
scan_for_safety_alert(text, sensitivity)	Warn-only scanner for job-description ingestion; returns alert dict or None

Detection coverage added over the old implementation:

• ✅ Zero-width characters (U+200B etc.)
• ✅ Base64-encoded payloads
• ✅ Homoglyph substitution attacks
• ✅ Data exfiltration patterns
• ✅ Role-play / jailbreak phrases
• ✅ XML/chat-delimiter system prompt markers

scripts/utils/cv_orchestrator.py

• Removes _LAYOUT_AGENT_INSTRUCTION_PATTERNS constant
• _sanitize_layout_instruction_text: delegates to sanitize_instruction_text() — return shape unchanged
• _sanitize_layout_context_html, _sanitize_layout_instruction_html: replaces any(pattern in text.lower() for pattern in ...) with scan_text_for_injection(); all BeautifulSoup DOM manipulation is unchanged

scripts/routes/job_routes.py

Three job-description ingestion routes (submit_job, fetch_job_url, load_job_file) now scan the received text via scan_for_safety_alert():

• Indicators are logged as warnings (not blocking — job descriptions are user-selected content)
• safety_alert dict is included in the JSON response when findings are present

tests/test_prompt_safety.py (new)

21 unit tests covering all three public functions, including:

• Detection of classic injection phrases
• Zero-width character detection (new vector)
• Clean text correctly returns no findings
• Fragment field truncation to ≤ 500 chars

Test results

343 existing Python tests: all pass
21 new tests in test_prompt_safety.py: all pass
22 layout instruction tests: all pass

Notes

llm-sanitizer must be installed in the cvgen conda environment before running the app or tests:

/usr/local/Caskroom/miniconda/base/envs/cvgen/bin/pip install -e /path/to/llm-sanitizer

The library is at Warnes-Innovations/llm-sanitizer (local sibling repo; AGPL-3.0).