feat: add adversarial prompt-injection document corpus

[yuehua1213]

What changed

• add an MIT-licensed manifest of six synthetic indirect prompt-injection samples
• cover TXT, Markdown comments, metadata, hidden DOCX table text, hidden XLSX rows, and policy-looking PDF text
• add a deterministic offline generator for all declared document formats
• add regression tests for manifest quality, expected behavior, generated artifacts, and hidden carrier preservation

Why

This implements the initial corpus requested in #20. DocSentinel already tests several injection strings directly, but it did not have a reusable, documented, multi-format corpus that guardrail changes could evaluate offline.

Each sample declares a trusted goal, attack carrier, synthetic payload, risk classification, and the same expected safety boundary: document content remains untrusted data and cannot change the deterministic rule-engine decision.

Validation

• Before implementation: test collection failed because the corpus generator did not exist
• After implementation: 4 passed
• Ruff: all changed Python files pass
• No external LLM or network call is used

python -m pytest --noconftest tests/test_prompt_injection_corpus.py -q
ruff check scripts/generate_prompt_injection_corpus.py tests/test_prompt_injection_corpus.py

The contribution was prepared with AI assistance and manually reviewed and tested, following CONTRIBUTING_WITH_AI.md.

Closes #20.