feat(config): add RSAConfig + named profiles (default-16gb, paper-16k, paper-headline-40k)

src/markovian_rsa_mlx/config.py

@@ -0,0 +1,71 @@
+"""RSAConfig dataclass + named profiles. Zero MLX dependencies — purely declarative."""
+from __future__ import annotations
+from dataclasses import dataclass, replace
+from typing import Literal
+
+
+@dataclass(frozen=True)
+class RSAConfig:
+    """Markovian RSA hyperparameters.
+
+    Defaults target the 16 GB practical profile (N=4, T=2, chunk=16K),
+    not the paper headline (N=16, chunk=40K). See named constructors for
+    paper-aligned profiles.
+    """
+    rounds: int = 2
+    parallel: int = 4
+    aggregation_subsample: int = 4
+    chunk_tokens: int = 16384
+    tail_tokens: int = 4096
+    final_tokens: int | None = None
+    temperature: float = 1.0
+    top_p: float = 0.95
+    top_k: int = -1
+    seed: int | None = None
+    serial: bool = False
+    auto_serial: bool = True
+    memory_fraction: float = 0.80
+    max_context_tokens: int | None = None
+    aggregation_template: Literal["zaya_v1"] = "zaya_v1"
+    answer_selection: Literal["first_final_candidate"] = "first_final_candidate"
+
+    def __post_init__(self) -> None:
+        if self.chunk_tokens <= 0:
+            raise ValueError("chunk_tokens must be positive")
+        if self.tail_tokens <= 0:
+            raise ValueError("tail_tokens must be positive")
+        if self.parallel < 1:
+            raise ValueError("parallel must be >= 1")
+        if self.rounds < 1:
+            raise ValueError("rounds must be >= 1")
+        if self.aggregation_subsample > self.parallel:
+            raise ValueError(
+                f"aggregation_subsample ({self.aggregation_subsample}) "
+                f"must be <= parallel ({self.parallel})"
+            )
+        if not 0.0 < self.memory_fraction <= 1.0:
+            raise ValueError("memory_fraction must be in (0, 1]")
+
+    def effective_final_tokens(self) -> int:
+        return self.final_tokens if self.final_tokens is not None else self.chunk_tokens
+
+    @classmethod
+    def default_16gb(cls) -> "RSAConfig":
+        return cls(parallel=2, chunk_tokens=16384, aggregation_subsample=2)
+
+    @classmethod
+    def paper_16k(cls) -> "RSAConfig":
+        return cls(parallel=4, chunk_tokens=16384, top_p=1.0)
+
+    @classmethod
+    def paper_headline_40k(cls) -> "RSAConfig":
+        return cls(
+            parallel=16,
+            aggregation_subsample=4,
+            chunk_tokens=40960,
+            final_tokens=40960,
+            top_p=1.0,
+        )
+
+    def replace(self, **kwargs) -> "RSAConfig":
+        return replace(self, **kwargs)