feat(loader, daemon): Implement strict ABI contamination tracking and worker isolation

This commit introduces a comprehensive overhaul of the worker daemon and loader to prevent ABI (Application Binary Interface) conflicts and ensure process isolation, dramatically improving stability when switching between versions of C-extension-heavy packages like TensorFlow, NumPy, and PyTorch.

Previously, reusing a worker process that had already loaded an ABI package (e.g., `numpy==1.24.3`) for a task requiring an incompatible version (e.g., `numpy==2.0.0`) could lead to segmentation faults or silent data corruption, as the OS linker would not reload the underlying `.so` file. This change implements a robust system to detect, manage, and prevent this "ABI contamination."

**Key Changes & Improvements:**

*   **ABI Contamination Tracking & Eviction:**
    *   Workers now track whether they have loaded an ABI-sensitive package (like TensorFlow, JAX, NumPy, etc.) and report this "contaminated" state back to the daemon.
    *   The daemon will now automatically evict contaminated workers instead of returning them to the idle pool, ensuring a fresh, clean process is always used for tasks that could have conflicting ABI requirements.

*   **Robust JAX/TensorFlow Dependency Hot-Patching:**
    *   The worker now intelligently detects broken or partially-initialized imports (specifically the `tensorflow -> jax -> ml_dtypes` conflict).
    *   Instead of failing, it now purges the broken module and injects a functional stub in its place, allowing TensorFlow 2.13+ to load successfully even with an older `ml_dtypes` present in the environment.

*   **Centralized & Atomic Filesystem Locking:**
    *   Replaced all manual `shutil.move` and disparate `filelock` logic with a new centralized, re-entrant locking mechanism (`fs_lock_queue`).
    *   All package "cloaking" and "uncloaking" operations are now atomic and process-safe, eliminating filesystem race conditions between concurrent loaders.

*   **Proactive ABI Conflict Delegation:**
    *   The `omnipkgLoader` can now detect a potential ABI conflict *before* activating a package (e.g., sees `numpy.core._multiarray_umath` is already mapped).
    *   It raises a `ProcessCorruptedException`, which is caught to automatically delegate the task to a clean, ephemeral daemon worker (`run_once` mode), seamlessly avoiding the in-process crash.

*   **Massive Daemon Performance Boost (Overlay Fast Path):**
    *   Nested `omnipkgLoader` calls inside a daemon worker now use a highly optimized "fast path."
    *   It bypasses all filesystem cloaking and locking, simply swapping `sys.path` entries in-memory. This reduces nested activation overhead from 100+ms to <1ms.

*   **Enhanced Cleanup and Diagnostics:**
    *   The loader's exit (`__exit__`) logic is now far more aggressive, performing deep scans for "zombie" modules and stale bubble paths left in `sys.path` to ensure a clean state.
    *   Added a new `_prof` helper and enhanced test output for better timing and debugging of worker initialization.

Modified:
  • src/omnipkg/core.py (+31/-12 lines)
  • src/omnipkg/isolation/patchers.py (+1/-1 lines)
  • src/omnipkg/isolation/worker_daemon.py (+164/-23 lines)
  • src/omnipkg/loader.py (+1856/-455 lines)
  • src/tests/test_tensorflow_switching.py (+18/-4 lines)

[gitship-generated]

Author: 1minds3t

src/omnipkg/core.py

@@ -88,9 +88,19 @@
 _DBG = os.environ.get("OMNIPKG_DEBUG", "0") == "1"
 
 def _dbg(msg: str):
-    """Lightweight debug printer; no-op unless OMNIPKG_DEBUG=1."""
-    if _DBG:
-        print(f"[DEBUG-CORE] {msg}", file=sys.stderr, flush=True)
+       if _DBG:
+               print(f"[DEBUG-CORE] {msg}", file=sys.stderr, flush=True)
+
+# ── PROFILE HELPER ───────────────────────────────────────────────────────────
+# Set UV_FFI_PROFILE=1 in environment to see timing info.
+# ─────────────────────────────────────────────────────────────────────────────
+_PROF = os.environ.get("UV_FFI_PROFILE", "0") == "1"
+
+def _prof(msg: str):
+       """Lightweight timing printer; no-op unless UV_FFI_PROFILE=1."""
+       if _PROF:
+               # We use stderr so it doesn't pollute actual program output
+               print(f"[TIMING] {msg}", file=sys.stderr, flush=True)
 
 def _get_dynamic_omnipkg_version():
     """
@@ -5825,13 +5835,13 @@ def _create_bubble_manifest(
         manifest_path = bubble_path / ".omnipkg_manifest.json"
 
         # --- ADDED EXPLICIT PATH PRINTING ---
-        print(f"   [MANIFEST] Writing schema 2.0 manifest to: {manifest_path}")
-
+        _dbg(f"Writing schema 2.0 manifest to: {manifest_path}")
         try:
             with open(manifest_path, "w") as f:
                 json.dump(manifest_data, f, indent=2)
-            print(f"   [MANIFEST] ✅ Successfully wrote {len(json.dumps(manifest_data))} bytes.")
+            _dbg(f"Successfully wrote {len(json.dumps(manifest_data))} bytes to manifest.")
         except Exception as e:
+            # THIS IS THE CORRECTED PART - We actually print the critical error
             print(f"   [MANIFEST] ❌ CRITICAL ERROR writing manifest: {e}")
 
         try:
@@ -17126,11 +17136,11 @@ def _get_latest_version_from_pypi(
         safe_print(
             f" -> Finding latest COMPATIBLE version for '{package_name}' using background caching..."
         )
-        print(f"[TIMING] after safe_print: {(_t.perf_counter()-_t0)*1000:.1f}ms", flush=True)
+        _prof(f"after safe_print: {(_t.perf_counter()-_t0)*1000:.1f}ms")
         import requests as http_requests
-        print(f"[TIMING] after import requests: {(_t.perf_counter()-_t0)*1000:.1f}ms", flush=True)
+        _prof(f"after import requests: {(_t.perf_counter()-_t0)*1000:.1f}ms")
         py_context = python_context_version or self.current_python_context
-        print(f"[TIMING] after py_context: {(_t.perf_counter()-_t0)*1000:.1f}ms", flush=True)
+        _prof(f"after py_context: {(_t.perf_counter()-_t0)*1000:.1f}ms")
         py_context = python_context_version or self.current_python_context
 
         if not hasattr(self, "pypi_cache"):
@@ -17139,10 +17149,19 @@ def _get_latest_version_from_pypi(
         cached_version = self.pypi_cache.get_cached_version(package_name, py_context)
         if cached_version:
             import time as _t
-            _t0 = _t.perf_counter(); stale = self.pypi_cache.is_cache_entry_stale(package_name, py_context, max_age_seconds=3600); print(f"[TIMING] is_cache_entry_stale: {(_t.perf_counter()-_t0)*1000:.1f}ms stale={stale}", flush=True)
+            _t0 = _t.perf_counter()
+            stale = self.pypi_cache.is_cache_entry_stale(
+                    package_name,
+                    py_context,
+                    max_age_seconds=3600
+            )
+            _prof(f"is_cache_entry_stale: {(_t.perf_counter()-_t0)*1000:.1f}ms stale={stale}")
+
             if stale:
-                _t0 = _t.perf_counter(); self._start_background_cache_refresh(package_name, py_context); print(f"[TIMING] _start_background_cache_refresh: {(_t.perf_counter()-_t0)*1000:.1f}ms", flush=True)
-            return cached_version
+                    # 2. Background refresh with hidden timing
+                    _t0 = _t.perf_counter()
+                    self._start_background_cache_refresh(package_name, py_context)
+                    _prof(f"_start_background_cache_refresh: {(_t.perf_counter()-_t0)*1000:.1f}ms")
 
         # USE THE ROBUST TEST INSTALLATION APPROACH FIRST
         safe_print(f"    🧪 Using pip to find latest compatible version for Python {py_context}...")

src/omnipkg/isolation/patchers.py

@@ -282,7 +282,7 @@ class DummyBackend:
         # ═══════════════════════════════════════════════════════════
         module = _original_import_func(name, globals, locals, fromlist, level)
 
-        if is_tf_import and module:
+        if is_tf_import and module and name == "tensorflow":
             import os
             _tf_loaded_pids.add(os.getpid())  # Mark THIS worker as having loaded TF
             _patch_numpy_for_tf_recursion()
@@ -611,4 +611,4 @@ def _force_complete_initialization(module_name):
                 try:
                     importlib.import_module(submod)
                 except ImportError:
-                    continue
+                    continue