[SizzleUnrlsd]

Summary

Implements the approach agreed in #239.

What proves a value:

• A membership test proves the value when every member is a constant that cannot change before the test:
  • a tuple, list or set display, or a dict display by its keys, tested directly or through .keys();
  • frozenset, set, tuple or list of one of these;
  • a module-level tuple, or frozenset of a display, of constants, bound once and never rebound.
• The module-level case is a new ModuleCollectionsAnalysis. It shares the "bound once, never rebound" rule with ModuleStringsAnalysis through module_bindings.
• A name the semantic tables know as an instance (ALLOWED = frozenset(...)) is loaded as the Symbol of frozenset. The refutation therefore recognises it by the span of the name in the function's HIR.
• A lookup in a dict display ({...}[key], .get(key[, default])) carries the taint of its values, of what was stored in it and of the default, never the taint of the key.

What does not prove a value:

• A list, set or dict display of the function counts only while the function only reads it. This closes a false negative on main: allowed = ["ls"]; allowed.append(x); if cmd in allowed was refuted and is now a hotspot. The same holds for a display passed to a call, or with an unpacked element (("ls", *extra)).
• Module-level lists, sets and dicts keep today's verdict: a hotspot for a membership test, a vulnerability for a lookup.
• As for module-level strings, rebinding a module-level name from another module is not seen.

Verification

• tests/test_allowlist_shapes.py, new: 34 tests through the bundled command-injection rule. 13 of them fail on main. No existing test changed.
• Full suite green, apart from test_packaging.py::test_the_version_has_one_source_of_truth, which also fails on main (locally installed metadata). ruff check and mypy clean.
• Regression suite: 18 of 19 repositories match their snapshots. full-stack-fastapi-template fails on coverage (43 files and 198 functions observed, 42 and 194 expected) on main too; its findings are identical with and without this change.
• healthchecks and mitmproxy: identical findings, and no measurable time difference.