[b-re-w]

Closes #203

Defect fixed: a function holding a list copy-in that also ran user code compiled, and diverged from CPython. On develop, f(xs, cb) (cb() sets xs[0] = 100.0, then return xs[0]) returned 1.0 compiled against 100.0 interpreted. A store followed by a callback was also overwritten by write-back.

Fix: a new verifier rule, verify/array-open. A function with an ArrayParam must be closed: no GetAttr, CallObject, Truth, CompareObj, ObjToFloat or OBJ BinOp. It also may not use FieldGet or New, whose slow paths run user code after a class change. The rule applies over the whole call tree (fixpoint). Such a function stays interpreted with a reason.

Tests added:

• verifier: CallObject, GetAttr, Truth, an open callee, and the closed case still accepted;
• end to end: test_array_function_with_a_callback_matches_cpython.

Red before the fix: 5 failed, 1 passed; the passing one is the accepted closed case.

Tests corrected: these encoded the defect and now use closed array functions or non-array functions:

• test_len_bound_is_zero_to_two_to_the_62;
• test_the_same_open_node_without_entry_globals_is_accepted;
• the scale and ip base functions;
• one fuzz mutation that assumed scale had an ArrayParam;
• fill in test_cgen_class_changes.py. It was an array function whose slot read ran a property installed after init, which is exactly this defect. It now returns the entry global instead, so the scenario still checks the snapshot against the property.

Documentation corrected: the ArrayParam contract in ir.py (the verifier, not the front end, enforces it), and the SPEC N-9 table row.

Verification (local):

• Full TypedPython suite: 962 passed, 3 skipped.
• Mutation: disabling the rule makes the end-to-end test fail.
• The five benchmarks compile exactly the same functions as on develop.