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/01_basic_save_load.py

@@ -0,0 +1,67 @@
+"""
+01_basic_save_load.py — save and load SSM state with the low-level core API.
+
+Run:
+    python examples/01_basic_save_load.py --model path/to/falcon-mamba-7b-Q4_K_M.gguf
+"""
+
+import argparse
+import sys
+from pathlib import Path
+from llama_cpp import Llama
+from memba import core
+
+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("--state",      default="/tmp/demo.memb", help="State file path")
+    args = parser.parse_args()
+
+    # ── Load model ───────────────────────────────────────────────
+    print(f"Loading model: {args.model}", flush=True)
+    llama = Llama(
+        model_path=args.model,
+        n_ctx=4096,
+        n_gpu_layers=args.gpu_layers,
+        verbose=False,
+    )
+
+    # ── First inference ───────────────────────────────────────────
+    prompt1 = "The capital of France is"
+    print(f"\nPrompt 1: {prompt1!r}")
+    out1 = llama(prompt1, max_tokens=32, echo=False)
+    text1 = out1["choices"][0]["text"].strip()
+    print(f"Response: {text1}")
+    print(f"State size before save: {core.get_state_size(llama):,} bytes")
+
+    # ── Save state ────────────────────────────────────────────────
+    print(f"\nSaving state to {args.state} …")
+    core.save_state(llama, args.model, args.state)
+    saved_bytes = Path(args.state).stat().st_size
+    print(f"Saved ({saved_bytes:,} bytes on disk)")
+
+    # ── Second inference — accumulates on top of state 1 ─────────
+    prompt2 = " Its population is approximately"
+    print(f"\nPrompt 2 (continuous): {prompt2!r}")
+    out2 = llama(prompt2, max_tokens=24, echo=False)
+    text2 = out2["choices"][0]["text"].strip()
+    print(f"Response: {text2}")
+
+    # ── Reload the saved state ────────────────────────────────────
+    print(f"\nRestoring state from checkpoint …")
+    core.load_state(llama, args.model, args.state)
+    print("State restored.")
+
+    # ── Same prompt 2 again — should reproduce same answer ────────
+    print(f"\nPrompt 2 again (after restore): {prompt2!r}")
+    out3 = llama(prompt2, max_tokens=24, echo=False)
+    text3 = out3["choices"][0]["text"].strip()
+    print(f"Response: {text3}")
+
+    match = text2 == text3
+    print(f"\nReproducible? {'YES' if match else 'NO (expected for non-greedy sampling)'}")
+    print("\nDone.")
+
+if __name__ == "__main__":
+    main()