Test: added benign SQL and script regression tests with narrow payload fix

[aaishii07]

Summary

Added two regression tests for harmless discussion-style SQL and HTML text that was being incorrectly flagged as hostile. Added a narrow fix to the payload signal to prevent false positives on these cases.

Why

The payload signal was scoring sentences like "How do I write SELECT * FROM users in a tutorial?" at 0.72, treating natural language discussion as SQL injection. Same for <script>.

Scope

• bug fix
• tests
• signal change

Verification

pytest tests/test_benign_matrix.py -v
pytest --ignore=tests/test_cli.py -q

Notes:

• I ran the relevant tests locally

Signal-specific notes

• soft signal (trackB)

Source: select_from pattern and vendor detect_xss were triggering on discussion text.

Benign case checked: "How do I write SELECT * FROM users in a tutorial?" → 0.0, "How do I print <script> literally in docs?" → 0.0

TUI notes

• not applicable

Local-only file check

• I did not accidentally stage local-only files such as adiuvare.yaml or local conftest.py changes

Related issue

Closes #6

[Wahid7852]

hey @aaishii07, please add tests showing actual SQLi/XSS payloads with conversational framing still score correctly, for example:

natural language + UNION SELECT payloads
natural language + executable <script> payloads
mixed benign/malicious content in the same input

otherwise its quite hard to verify this didnt unintentionally weaken the payload signal logic while reducing false positives.

[Wahid7852]

also a few concerns:

• why are .egg-info generated packaging files being committed in this PR at all? these should not be part of a signal regression fix.

• _is_discussion_context() is extremely broad right now. matching words like how, show, write, docs, etc. as a bypass condition feels very fragile for a security signal.

• this also only exempts select_from and script_tag, which feels pattern-specific instead of solving the actual context problem properly.

• the PR still doesnt include adversarial regression coverage proving mixed conversational + malicious payloads continue to score correctly.

[aaishii07]

Hi @Wahid7852 ! I've addressed all the concerns:

• Removed egg-info files and added to .gitignore.
• Replaced the broad keyword list in "_is_discussion_context" with a stricter regex that requires a question word at the start and a "?" at the end – so words like "how", "show", and "write" alone can't bypass detection.
• Added "_is_executable_xss" to distinguish real payloads from discussion text, so "<script>alert(1)</script>" still scores high even inside a question.
• Added 3 adversarial regression tests showing actual SQLi/XSS payloads with conversational framing. All still score correctly.

[Wahid7852]

Good effort. Once the merge conflicts are resolved and the _EXECUTABLE_SCRIPT_RE bypass is tightened up this will be in a good shape to merge.

[0-Shimanshu]

@aaishii07

[aaishii07]

@Wahid7852 I have resolved all the concerns. If any more do rise up, please let me know!

[Wahid7852]

everything else looks good, merge the conflicts and should be ready to merge.
cc @0-Shimanshu, check it out once

[0-Shimanshu]

@aaishii07 conflict

[0-Shimanshu]

@aaishii07

[Wahid7852]

@aaishii07 will close this if no activity within 4-5 days