[SizzleUnrlsd]

Summary

Opens Phase 7 of docs/architecture.md: framework model plugins (§15, §25, §36). Detectors stay generic; frameworks only enrich the model.

• Acceptance tests first (test: define framework model behavior), implementation follows.
• Three engine mechanisms, all framework-agnostic:
  • a module-level name = Symbol(...) binds name to that symbol in the symbol table, so app.route resolves to python.flask.Flask.route when app = Flask(__name__);
  • inside functions, call results and with contexts carry their callee's symbol (call-chain rule), so conn.cursor().execute resolves to python.sqlite3.connect.cursor.execute; the call graph exposes these derived symbols and the taint engine reads sources through them;
  • a new EntryPoint model taints every parameter of a function decorated by its symbol, which is how route handler arguments become HTTP input.
• A source covers the symbols below it (flask.request.args also covers request.args.get), and a method called on a tainted object returns tainted data.
• Shipped model plugins: models/flask (request sources, Flask.route / Blueprint.route entry points, HTML and path sinks, escape sanitizers), models/fastapi (FastAPI and APIRouter route entry points, response sinks), models/sqlalchemy (raw execution sinks on engines, connections and sessions), and sqlite3 execution sinks in the stdlib models.

End to end with --check: a Flask route building a sqlite3 query from request.args is a SQL injection, render_template_string with request data is XSS unless escaped, and a <host> route parameter reaching os.system is a command injection. Same for FastAPI route parameters.

Not in this PR: Django (views are undecorated, so their request parameter needs parameter typing), Requests/httpx models, PossibleTargets in the call graph.

Part of #30