Initial commit: Future-Entropy Sampler prototype

llama.cpp-based implementation of the countbayesie Future-Entropy Sampler,
built via a direct ctypes binding to libllama.so. Includes a one-shot CLI
(entropy_cli.py), an OpenAI-compatible server (entropy_server.py), and the
tuning/benchmark scripts used to derive the alpha/confidence-threshold
defaults documented in the README.

Model path and llama.cpp library path are now read from environment
variables (ENTROPY_SAMPLER_MODEL, LLAMA_CPP_LIB) instead of being hardcoded,
and gguf-py is pulled from PyPI instead of a local llama.cpp checkout, so
this runs on any machine with a compatible llama.cpp build and model.

tune_offset_sine.py

@@ -0,0 +1,27 @@
+"""Offset-sine sweep: dip the wave deeper into safe/coherent territory than
+it peaks into entropy-chasing territory, to see if it keeps the occasional
+surprise from tune_alpha.py's best run (period=16 amp=0.6) without ever
+grazing the +0.3-and-above coherence cliff found there.
+"""
+from entropy_sampler import EntropySampler, sine_alpha_schedule
+from tune_alpha import PROMPT, MAX_NEW_TOKENS, run_labeled
+
+CONFIGS = [
+    {"period_tokens": 16, "amplitude": 0.6, "offset": -0.2},  # peak +0.4 / trough -0.8
+    {"period_tokens": 16, "amplitude": 0.7, "offset": -0.3},  # peak +0.4 / trough -1.0
+    {"period_tokens": 12, "amplitude": 0.8, "offset": -0.2},  # shorter period, same bias
+    {"period_tokens": 16, "amplitude": 0.6, "offset": 0.0},   # unbiased baseline for comparison
+]
+
+if __name__ == "__main__":
+    sampler = EntropySampler(top_k=12, top_n_future=20, n_ctx=4096)
+    try:
+        for cfg in CONFIGS:
+            schedule = sine_alpha_schedule(**cfg)
+            label = (
+                f"sine period={cfg['period_tokens']} amp={cfg['amplitude']} "
+                f"offset={cfg['offset']}"
+            )
+            run_labeled(sampler, label, schedule)
+    finally:
+        sampler.close()