Add 2.4.0 changelog and docs for context thread affinity

Author: benoitc

CHANGELOG.md

@@ -1,5 +1,19 @@
 # Changelog
 
+## 2.4.0 (Unreleased)
+
+### Added
+
+- **Context thread affinity** - Contexts in MULTI_EXECUTOR mode are now assigned a
+  fixed executor thread at creation. All operations (call, eval, exec) from the same
+  context run on the same OS thread, preventing thread state corruption in libraries
+  like numpy and PyTorch that have thread-local state.
+
+### Changed
+
+- **Removed obsolete subinterp test references** - Test suites updated to reflect
+  the removal of subinterpreter mode. Tests now use `worker` or `owngil` modes.
+
 ## 2.3.1 (2026-04-01)
 
 ### Fixed

docs/context-affinity.md

@@ -357,6 +357,18 @@ Process-bound environments work by:
 
 For subinterpreters, environments are created inside the target interpreter to ensure memory safety - Python's subinterpreters have isolated memory allocators.
 
+### Thread Safety for C Extensions
+
+Contexts in MULTI_EXECUTOR mode have automatic thread affinity. Each context is
+assigned a dedicated executor thread at creation, ensuring that all Python operations
+run on the same OS thread. This is critical for C extensions with thread-local state:
+
+- **numpy** - Uses thread-local random state and BLAS threading
+- **torch** - Maintains thread-local state for CUDA and CPU operations
+- **tensorflow** - Thread-local session state
+
+No configuration is needed - thread affinity is enabled automatically.
+
 ## Best Practices
 
 1. **Use explicit contexts for stateful operations**: `Ctx = py:context(1)` ensures state persists

docs/scalability.md

@@ -93,6 +93,10 @@ Ctx = py:context(1),
 
 Runs N executor threads that share the GIL. Requests are distributed round-robin across executors. Good for I/O-bound workloads where Python releases the GIL during I/O operations.
 
+**Thread Affinity:** In MULTI_EXECUTOR mode, both workers and contexts are assigned
+a fixed executor thread. This ensures libraries with thread-local state (numpy, torch,
+tensorflow) always run on the same OS thread, preventing segfaults and state corruption.
+
 ## Choosing the Right Mode
 
 ### Mode Comparison