fix(loader): detect linker-locked TF .so via /proc/self/maps; guard zombie purge

Problem: after loading TF 2.13.0 and exiting the context, a second
with-block requesting TF 2.12.0 would still activate in-process,
corrupting interpreter state and crashing.

Two root causes fixed in loader.py:

1. /proc/self/maps fallback for cext detection
   The existing _tf_loaded_pids check requires omnipkg.isolation.patchers
   to be imported first. In bare subprocesses (test runner, user scripts)
   that module is never loaded, so _abi_so_mapped stayed False and the
   loader fell through to in-process activation of the wrong version.
   Fix: read /proc/self/maps directly — OS-level, survives sys.modules
   purging, no import dependencies.

2. Zombie purge was wiping linker-locked TF modules
   _aggressive_module_cleanup correctly preserved pywrap_tensorflow and
   friends, but the POST-EXIT zombie scan (which checks __file__ for
   .omnipkg_versions paths) then purged those same modules, leaving Load 2
   with no Python layer for a still-mapped C backend — causing reinit
   failures. Fix: zombie purge now skips all tensorflow.* modules when
   pywrap_tensorflow is present in sys.modules, matching the same guard
   already used by _aggressive_module_cleanup.

3. Graceful no-daemon path
   When _abi_so_mapped=True but no daemon is available, execution
   previously fell through to Strategy 1/2/3 (in-process activation).
   Now returns self immediately with _abi_conflict_detected=True —
   caller gets the already-loaded version, no crash, no corrupt state.

Also in this commit:
- refactor(tests): move src/tests/ → src/omnipkg/tests/ so the test
  package lives inside the omnipkg namespace and does not pollute
  site-packages for end users
- fix(pyproject.toml): update packages.find include to omnipkg.tests*
- chore(.gitignore): exclude __pycache__/ and *.pyc

Author: 1minds3t

.gitignore

@@ -209,3 +209,8 @@ src/omnipkg/_vendor/wheels/
 .omnipkg_baseline
 ci-wheel-cache/
 build/
+
+# Python cache
+__pycache__/
+*.pyc
+*.pyo

pyproject.toml

@@ -152,7 +152,7 @@ omnipkg = "omnipkg.dispatcher:main"
 
 [tool.setuptools.packages.find]
 where = ["src"]
-include = ["omnipkg*", "omnipkg.*", "tests*"]
+include = ["omnipkg*", "omnipkg.*", "omnipkg.tests*"]
 exclude = [
     "build*",
     "dist*",

src/omnipkg/cli.py

@@ -32,7 +32,7 @@
 from .core import omnipkg as OmnipkgCore
 
 project_root = Path(__file__).resolve().parent.parent
-TESTS_DIR = Path(__file__).parent.parent / "tests"
+TESTS_DIR = Path(__file__).parent / "tests"
 DEMO_DIR = Path(__file__).parent
 try:
     FILE_PATH = Path(__file__).resolve()

src/omnipkg/isolation/patchers.py

@@ -428,7 +428,6 @@ class DummyBackend:
 
             # Only check for reload if THIS WORKER previously loaded TF successfully
             if current_pid in _tf_loaded_pids and "tensorflow" not in sys.modules:
-                safe_print("☢️  [OMNIPKG] FATAL TENSORFLOW RELOAD DETECTED!")
                 raise ProcessCorruptedException(
                     "Attempted to reload TensorFlow in a process where its C++ libraries were already initialized."
                 )

src/omnipkg/loader.py

@@ -4085,8 +4085,13 @@ def __exit__(self, exc_type, exc_val, exc_tb):
             # but they may leave modules from a different version in sys.modules.
             # At depth=1 we own the final cleanup, so we do a full pass over
             # the known C-extension switcher families.
+            _patchers_mod = sys.modules.get("omnipkg.isolation.patchers")
+            _tf_pids = getattr(_patchers_mod, "_tf_loaded_pids", set()) if _patchers_mod else set()
+            _tf_linker_locked = os.getpid() in _tf_pids
             _CE_SWITCHERS = ("numpy", "torch", "scipy", "pandas", "tensorflow")
             for _sw in _CE_SWITCHERS:
+                if _sw == "tensorflow" and _tf_linker_locked:
+                    continue
                 _stale = [
                     k for k in list(sys.modules)
                     if k == _sw or k.startswith(_sw + ".")
@@ -5719,6 +5724,13 @@ def __enter__(self):
             _pkg_lower in omnipkgLoader.ABI_PACKAGES
             and _abi_indicator_map.get(_pkg_lower, f"{_pkg_lower}._") in sys.modules
         )
+        # Also check _tf_loaded_pids — TF purges its modules on __exit__ but
+        # the .so stays mapped. sys.modules check alone misses this case.
+        if not _abi_so_mapped and _pkg_lower == "tensorflow":
+            _patchers = sys.modules.get("omnipkg.isolation.patchers")
+            _tf_pids = getattr(_patchers, "_tf_loaded_pids", set()) if _patchers else set()
+            if os.getpid() in _tf_pids:
+                _abi_so_mapped = True
         if _abi_so_mapped and self._worker_fallback_enabled and DAEMON_AVAILABLE:
             if not self.quiet:
                 safe_print(