Add memba MVP: C++ core, Python SDK, CLI, examples, experiments

C++ core (libmemba.so):
- include/memba/state.h — C API (state_new/free/save/load/get_size)
- src/state.cpp — MEMB file format: magic, version, SHA-256 model_id,
  CRC-32, opaque llama_state_*_data() blob
- src/cli.cpp — minimal demo binary with greedy sampler
- CMakeLists.txt + build.sh with llama.cpp submodule, CUDA auto-detect

Python SDK (memba):
- core.py — file I/O via llama-cpp-python's exposed C functions,
  unwraps _LlamaContext to access raw context pointer (≥0.3.x)
- session.py — high-level Session with auto-save/load, ChatML wrapper
  for instruct models, raw mode for base models
- cli.py — typer-based: chat (REPL), run (one-shot), list, rm, info

Examples:
- 01_basic_save_load.py, 02_chat_session.py

Experiments (throwaway POCs documenting product-direction findings):
- recall_poc.py — git log → state → cross-process query
- mood_poc.py — batch sentiment trajectory, Mamba vs Transformer
- mood_stream_poc.py, mood_batch_poc.py — variants
- diag_saveload.py — minimal save/load isolation test
- README.md documents the headline finding: save/load is byte-identical,
  but Falcon-Mamba-7B-Instruct does not retain facts across conversation
  turns even in-process — limits viable products to single-prompt analysis
  and persona priming.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

examples/02_chat_session.py

@@ -0,0 +1,58 @@
+"""
+02_chat_session.py — multi-turn chat that persists across Python processes.
+
+First run  : model has no prior context.
+Second run : model continues from the saved SSM state.
+
+Run twice:
+    python examples/02_chat_session.py --model path/to/falcon-mamba-7b-Q4_K_M.gguf
+    python examples/02_chat_session.py --model path/to/falcon-mamba-7b-Q4_K_M.gguf
+"""
+
+import argparse
+from memba import Session
+
+TURNS = [
+    "My name is Alex. I am a researcher studying ancient Roman aqueducts.",
+    "What is the most famous aqueduct I should know about?",
+    "How long did it take to build?",
+]
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model",      required=True,       help="Path to GGUF file")
+    parser.add_argument("--gpu-layers", type=int, default=0, help="GPU layers (0=CPU)")
+    parser.add_argument("--session",    default="aqueduct_research")
+    parser.add_argument("--state-dir",  default="~/.memba/states")
+    args = parser.parse_args()
+
+    print(f"Session: {args.session!r}")
+    print(f"Model  : {args.model}")
+    print("-" * 60)
+
+    sess = Session(
+        model_path=args.model,
+        session_id=args.session,
+        state_dir=args.state_dir,
+        n_gpu_layers=args.gpu_layers,
+        verbose=False,
+    )
+    print(f"State size on load: {sess.state_size:,} bytes\n")
+
+    # Only feed the turns that haven't been answered yet.
+    # A real app would track which turns were already fed; here we keep it simple
+    # and feed all TURNS every run — the SSM state update is idempotent in terms
+    # of demonstrating cross-process continuity.
+    for i, turn in enumerate(TURNS, 1):
+        print(f"[Turn {i}] User: {turn}")
+        reply = sess.chat(turn, max_tokens=200)
+        print(f"[Turn {i}] Model: {reply}")
+        print()
+
+    saved_path = sess.save()
+    print(f"State saved → {saved_path}")
+    print(f"State size  : {sess.state_size:,} bytes")
+    print("\nRun this script again — the model will continue from this checkpoint.")
+
+if __name__ == "__main__":
+    main()