Refactor module name resolution

codetracer-python-recorder/CHANGELOG.md

@@ -9,7 +9,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/)
 - Balanced call-stack handling for generators, coroutines, and unwinding frames by subscribing to `PY_YIELD`, `PY_UNWIND`, `PY_RESUME`, and `PY_THROW`, mapping resume/throw events to `TraceWriter::register_call`, yield/unwind to `register_return`, and capturing `PY_THROW` arguments as `exception` using the existing value encoder. Added Python + Rust integration tests that drive `.send()`/`.throw()` on coroutines and generators to guarantee the trace stays balanced and that exception payloads are recorded.
 
 ### Changed
-- Module-level call events now use the actual dotted module name (e.g., `<my_pkg.mod>` or `<boto3.session>`) instead of the generic `<module>` label. `RuntimeTracer` derives the name via the shared module-identity helper, caches the result per code object, and falls back to `<module>` only for synthetic or nameless frames. Added Rust + Python tests plus README documentation covering the new semantics.
+- Module-level call events now prefer the frame's `__name__`, fall back to filter hints, `sys.path`, and package markers, and no longer depend on the legacy resolver/cache. The globals-derived naming flag now defaults to enabled so direct scripts record `<__main__>` while package imports emit `<pkg.mod>`, with CLI and environment overrides available for the legacy resolver.
 
 ## [0.2.0] - 2025-10-17
 ### Added

codetracer-python-recorder/README.md

@@ -84,7 +84,8 @@ action = "drop"
 
 ## Trace naming semantics
 
-- Module-level activations no longer appear as the ambiguous `<module>` label. When the recorder sees `co_qualname == "<module>"`, it derives the actual dotted package name (e.g., `<my_pkg.mod>` or `<boto3.session>`) using project roots, `sys.modules`, and frame metadata.
+- Module-level activations no longer appear as the ambiguous `<module>` label. When the recorder sees `co_qualname == "<module>"`, it first reuses the frame's `__name__`, then falls back to trace-filter hints, `sys.path` roots, and package markers so scripts report `<__main__>` while real modules keep their dotted names (e.g., `<my_pkg.mod>` or `<boto3.session>`).
+- The globals-derived naming flow ships enabled by default; disable it temporarily with `--no-module-name-from-globals`, `codetracer.configure_policy(module_name_from_globals=False)`, or `CODETRACER_MODULE_NAME_FROM_GLOBALS=0` if you need to compare against the legacy resolver.
 - The angle-bracket convention remains for module entries so downstream tooling can distinguish top-level activations at a glance.
 - Traces will still emit `<module>` for synthetic filenames (`<stdin>`, `<string>`), frozen/importlib bootstrap frames, or exotic loaders that omit filenames entirely. This preserves previous behaviour when no reliable name exists.
 

codetracer-python-recorder/benches/trace_filter.rs

@@ -52,7 +52,7 @@ fn run_workload(engine: &TraceFilterEngine, dataset: &WorkloadDataset) {
         for &index in &dataset.event_indices {
             let code = dataset.codes[index].as_ref();
             let resolution = engine
-                .resolve(py, code)
+                .resolve(py, code, None)
                 .expect("trace filter resolution should succeed during benchmarking");
             let policy = resolution.value_policy();
             for name in dataset.locals.iter() {
@@ -66,7 +66,7 @@ fn prewarm_engine(engine: &TraceFilterEngine, dataset: &WorkloadDataset) {
     Python::with_gil(|py| {
         for code in &dataset.codes {
             let _ = engine
-                .resolve(py, code.as_ref())
+                .resolve(py, code.as_ref(), None)
                 .expect("prewarm resolution failed");
         }
     });

codetracer-python-recorder/codetracer_python_recorder/cli.py

@@ -120,6 +120,15 @@ def _parse_args(argv: Sequence[str]) -> RecorderCLIConfig:
             "'proxies+fd' also mirrors raw file-descriptor writes."
         ),
     )
+    parser.add_argument(
+        "--module-name-from-globals",
+        action=argparse.BooleanOptionalAction,
+        default=None,
+        help=(
+            "Derive module names from the Python frame's __name__ attribute (default: enabled). "
+            "Use '--no-module-name-from-globals' to fall back to the legacy resolver."
+        ),
+    )
 
     known, remainder = parser.parse_known_args(argv)
     pending: list[str] = list(remainder)
@@ -181,6 +190,8 @@ def _parse_args(argv: Sequence[str]) -> RecorderCLIConfig:
                 policy["io_capture_fd_fallback"] = True
             case other:  # pragma: no cover - argparse choices block this
                 parser.error(f"unsupported io-capture mode '{other}'")
+    if known.module_name_from_globals is not None:
+        policy["module_name_from_globals"] = known.module_name_from_globals
 
     return RecorderCLIConfig(
         trace_dir=trace_dir,

codetracer-python-recorder/resources/trace_filters/builtin_default.toml

@@ -34,7 +34,7 @@ exec = "skip"
 reason = "Skip builtins module instrumentation"
 
 [[scope.rules]]
-selector = 'pkg:glob:*_distutils_hack*'
+selector = 'pkg:literal:_distutils_hack'
 exec = "skip"
 reason = "Skip setuptools shim module"