[b-re-w]

Closes #150. The wordfreq speedup it aimed at is blocked on #232; see below.

Feature added (each new operation calls the C API the interpreter's bytecode calls, so results, exceptions and user __hash__/__eq__/__index__/__format__ calls are identical):

• str constants and f-strings (PyObject_Format, PyUnicode_Join);
• o[i], o[i] = v, in (PyObject_GetItem, PyObject_SetItem, PySequence_Contains);
• empty {}/[], and for x in o / for a, b in o with CPython's unpacking errors;
• in-place augmented assignment on objects (PyNumber_InPlace*). This also fixes a latent bug: lst += x was lowered as a plain add.
• Exact-type fast paths, each calling the same C function the generic path ends in: exact-dict get and set, and exact str + str.

Cost rule (typedpython/cost.py): an impure function that has loops, none of which does native work, stays interpreted, with policy/cost and a reason. Straight-line impure functions keep compiling, as before.

• Without the rule, wordfreq ran at 0.8x CPython (1.335 against develop): word_of was 2.6x slower compiled.
• Cause: object-mode ints are generic PyNumber_* calls, while CPython 3.15 specialises the same code.
• Details in benchmarks/typedpython/results/2026-10-17-wordfreq-object-mode.md. The real fix is TypedPython: native ints in impure functions, with in-place promotion on overflow #232.

Tests added:

• test_str_dict.py, 29 tests: differential against the interpreter for results, exception repr and user-hook call logs, plus mutation tests on the fast-path guards. Red before: 26 of 27 failed.
• test_cost_rule.py, 7 tests: the object-only loop stays interpreted, a native loop compiles, straight-line code compiles, callers of a rejected function are rejected, and a mutation that disables the rule fails. Red: 3 of 6 failed with the rule disabled.

Tests changed: where a fact changed.

• test_frontend: inplace=True.
• test_runtime: the invalid-op code moved from 9 to 12.
• test_verify_classes: a str constant is now valid.
• test_module_report and the test_pipeline* files: their "unsupported" example became s[1:].
• test_free_threaded: its fixture uses cost_rule=False.

Measured: wordfreq with the rule against develop is 1.003 (0.974 to 1.027), still 1.1x against CPython. No regression.

Verification (local, 3.15.0rc1t, numpy present):

• Full suite: 1174 passed, 3 skipped.
• The five benchmarks compile the same sets as develop. The Kotlin bridge fixture (drive, checked) still compiles.