Summary

Closes #10. Adds rule P005 sqlalchemy-text-fstring, which catches sqlalchemy.text(f"...{var}") -- the same SQL-injection hazard P001 catches for cursor.execute(), but on the sqlalchemy.text() surface.

from sqlalchemy import text

# Flagged: P005 (error)
conn.execute(text(f"SELECT * FROM users WHERE id = {user_id}"))

# OK:
conn.execute(
    text("SELECT * FROM users WHERE id = :id"),
    {"id": user_id},
)

Why this matters

sqlalchemy.text() is the documented escape hatch for raw SQL in SQLAlchemy. Wrapping an f-string inside it defeats parameter binding and re-introduces the exact SQL-injection risk that P001-P004 catch on cursor.execute(). "text" was already registered in the scanner's EXECUTE_METHODS frozenset, so P001 silently fired on these cases with a generic .text() message; P005 replaces that with a sqlalchemy-specific message and suggestion so the warning lands correctly for the reader.

Changes

• sql_guard/rules/python_rules.py: new SqlalchemyTextFstring class (P005) checking hit.kind == "fstring" and hit.call_name == "text"; added to PYTHON_RULES. Severity error, matches P001 + what the issue asked for.
• sql_guard/rules/python_rules.py: P001's FStringInExecute.check now skips when call_name == "text". This mirrors the existing P004 pattern (BareVariableInExecute already has call_name != "text" in its guard) and prevents P001+P005 double-firing on the same line.
• tests/fixtures/python_hazards.py: new unsafe_sqlalchemy_text_fstring function as the fixture for P005.
• tests/test_python_scanner.py: new test_p005_sqlalchemy_text_fstring_flagged (severity + message check), test_p001_skips_text_to_avoid_double_firing_with_p005 (regression guard), and the end-to-end test_check_with_include_python_surfaces_p_rules now asserts P005 is in the surfaced rule ID set.
• Docstring header in python_rules.py bumped from "P001-P004" to "P001-P005".

Testing

• pytest tests/test_python_scanner.py -v: 23 passed, including the 3 new P005-related tests.
• pytest -q (full suite): 106 passed.
• ruff check sql_guard/ tests/: clean (caught an unused local during self-review, now removed).

Implementation notes

The issue said to "Model after P001 fstring-in-execute"; this PR does that structurally (same class shape, same check signature, same severity). The one extra change beyond a plain clone is the P001 skip-on-text guard, which keeps the findings output clean for sqlalchemy.text(f"...") callers -- one P005 finding per site, not a P001+P005 pair.

Estimated LOC in the issue: ~30 code + ~25 test. Actual: ~22 code + ~20 test (matches the estimate closely).

This contribution was developed with AI assistance (Claude Code).