Remove CompositeTracer wrapper layer (item 2 of #115)

[islamicfinanceaaoifi-ops]

This PR addresses item (2) of #115 — "CompositeTracer is now largely just a wrapper around CTracer. Let's try to remove the CompositeTracer layer entirely."

Item (3) (porting TracingModule to C) is left for a follow-up PR so the diff stays reviewable; happy to take that on next if this approach looks right.

What changed

The old CompositeTracer class owned a self.ctracer = CTracer() instance and forwarded most methods through. The session lifecycle — push/pop the patching_module, wire sys.monitoring (3.12+) or sys.settrace (older), start/stop the tracer — lived in its __enter__/__exit__. This PR folds those responsibilities into CTracer itself and deletes the wrapper.

C side (_tracers.c, _tracers.h)

• New patching_module slot on CTracer, exposed via tp_getset. Defaults to None.
• New __enter__ / __exit__ C methods that push the patching_module, configure sys.monitoring on 3.12+ (or stash sys.gettrace() and enable opcode tracing on the current frame for older Python), call start()/stop(), then unwind on exit.
• New trace_caller C method (no-op on 3.12+; on older Python enables opcode tracing on the calling frame — same as the previous sys._getframe(2)-based Python implementation).
• CTracer.push_module now invokes sys.monitoring.restart_events() itself on 3.12+ (only when the tracer is enabled — restarting has no effect otherwise). That restart_events() call used to live in the now-deleted CompositeTracer.push_module.
• Helpers _ch_load_sys_monitoring and _ch_call_monitoring_method centralize the sys.monitoring lookup so the lifecycle methods stay readable.

Python side (tracers.py)

• CompositeTracer class is gone. The module singleton is now just COMPOSITE_TRACER = CTracer() with patching_module = PatchingModule() set explicitly.
• is_tracing, NoTracing, ResumedTracing, and PushedModule now go through the CTracer instance directly instead of COMPOSITE_TRACER.ctracer.

Caller renames

The wrapper provided two name-only shims; callers move to the unified CTracer API:

• COMPOSITE_TRACER.pop_config(m) -> COMPOSITE_TRACER.pop_module(m)
• COMPOSITE_TRACER.set_postop_callback(callback, frame) -> COMPOSITE_TRACER.push_postop_callback(frame, callback) (argument order swap)

Touches opcode_intercept.py (7 callsites), enforce.py, core.py, tracers_test.py.

Verified locally (Python 3.14.3 / Windows 11 / MSVC 14.50)

pytest crosshair/tracers_test.py crosshair/_tracers_test.py     # 12 / 12 pass (3 skipped baseline)
pytest crosshair/condition_parser_test.py                       # 27 / 27 pass
pytest crosshair/enforce_test.py crosshair/statespace_test.py   # 16 / 16 pass
pytest crosshair/util_test.py crosshair/opcode_intercept_test.py
pytest crosshair/fuzz_core_test.py                              # 1294 / 1294 pass
pytest crosshair/core_test.py crosshair/diff_behavior_test.py crosshair/path_cover_test.py   # 106 / 106 pass

libimpl/ also runs through; the only failure I saw (datetimelib_ch_test::test_builtin[check_datetime_astimezone]) reproduces on unmodified main and isn't related to this refactor — current main CI is red for the same family of path-exploration failures.

Design notes / questions for reviewer

• I kept the public name COMPOSITE_TRACER so this PR doesn't ripple a rename through downstream consumers. If you'd rather call it TRACER or similar now that it's just a CTracer, happy to follow up.
• sys.monitoring.restart_events() inside push_module is now guarded by self->enabled. The previous Python wrapper called it unconditionally. Let me know if you'd prefer the unconditional behavior — I went conservative on the perf side since it's only meaningful when the tracer is actively running.
• _PyFrame_GetBackBorrow in trace_caller walks back one frame to mirror the old sys._getframe(2) semantics (top Python frame on the eval stack is the caller of trace_caller, one back is the function we want to trace).
• prev_traceobj is stored on the struct for the older-Python branch so __exit__ can restore sys.settrace to whatever was set before the session.

Ready for item (3) (porting TracingModule.trace_op + _CALL_HANDLERS and the handle_call_* family to C) as a follow-up — let me know if you'd like me to open that PR next.

---

IssueHunt Summary

Referenced issues

This pull request has been submitted to:

• #115: Improve trace callback performance

---

[pschanely]

@islamicfinanceaaoifi-ops thanks for helping! Please check contributing.rst for getting the formatting/linting etc right.

sys.monitoring.restart_events() inside push_module is now guarded by self->enabled

I expect that we'll still need this to happen because the tracer can be re-enabled later. If the tests pass anyway, I'd want to know whether we need test coverage there, or why my understanding is flawed! :)