Prevent incorrect coercion from digit-only strings to integers in fallback code if slots are not used (e.g. in PyPy).

Extend test to assure that normal conversion to C integers does not allow strings, but only (optimised) casting does.

Disable refnanny in PyPy since it otherwise risks to access partly initialised objects, e.g. tuples.

https://github.com/python/typeshed/pull/1554#discussion_r134561386

Remove accidental left-over name "gs" from Shadow.py that should not be part of the pure mode types.

Avoid some unnecessary strlen() calls in exception text formatting, while still keeping up the nice property of sharing const C message strings.

Revert "Do not compile Lexicon.py. It's really only used once and is much larger as a binary module than as a byte code file, but not slower to import."
Currently seems to be required when the scanner is compiled.
This reverts commit d66405c6.

Do not compile Lexicon.py. It's really only used once and is much larger as a binary module than as a byte code file, but not slower to import.

Include deprecation warning from CPython (3.3+) in __Pyx_PyNumber_IntOrLong() and avoid inlining error handling code.

Streamline __Pyx_PyNumber_IntOrLong() with direct slot calls instead of C-API function calls that only do what we already do on our side already.

Specialise coercion from bytes to char* (and friends) to avoid going through several type checks in the generic "__Pyx_PyObject_AsStringAndSize()" function.

Split uncommon case from unicode/bytes.endswith() implementation to suggest inlining the normal case.

Fix reference leak in the Py2.x __Pyx_ReleaseBuffer() implementation for types with an overridden __releasebuffer__ method.

Closes #1638.

Fix more tests: Correct the source character position of name nodes to improve error reporting (previously reported the character position *behind* the name).

Correct the source character position of name nodes to improve error reporting (previously reported the character position *behind* the name).

Closes #1824.