feat(inference): operator-tunable sampling knobs (top_k/min_p/repeat_penalty) for stage requests

GenerationConfig only carried temperature/top_p/max_tokens/stop/seed. Added nullable
topK/minP/repeatPenalty, serialized to the llama.cpp and OpenAI-compat request bodies
via @EncodeDefault(NEVER) so an unset knob is omitted (the model keeps its own default)
and behavior is unchanged unless the operator opts in.

Surfaced as a new [sampling] config section feeding the default stage GenerationConfig
(the main agentic loop) through TomlWorkflowLoader + ExecutionPlanCompiler; the former
hardcoded temperature=0.7/topP=1.0 stage defaults now come from config. Talkie
chat/narration keep their own generation settings.

Vikunja #46 (task 76) — sampling half.

core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt

@@ -662,6 +662,15 @@ object ConfigLoader {
             )
         }
 
+        val samplingSection = sections["sampling"] ?: emptyMap()
+        val sampling = SamplingConfig(
+            temperature = asDouble(samplingSection["temperature"], 0.7),
+            topP = asDouble(samplingSection["top_p"], 1.0),
+            topK = samplingSection["top_k"]?.let { asInt(it) },
+            minP = samplingSection["min_p"]?.let { asDouble(it) },
+            repeatPenalty = samplingSection["repeat_penalty"]?.let { asDouble(it) },
+        )
+
         return CorrexConfig(
             server = server,
             tui = tui,
@@ -675,6 +684,7 @@ object ConfigLoader {
             project = project,
             personalization = personalization,
             orchestration = orchestration,
+            sampling = sampling,
         )
     }
 

core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt

@@ -16,9 +16,26 @@ data class CorrexConfig(
     val project: ProjectConfig = ProjectConfig(),
     val personalization: PersonalizationConfig = PersonalizationConfig(),
     val orchestration: OrchestrationKnobs = OrchestrationKnobs(),
+    val sampling: SamplingConfig = SamplingConfig(),
     val health: HealthConfig = HealthConfig(),
 )
 
+/**
+ * Sampling knobs sent to the llama-server on every *stage* inference request (the main agentic
+ * loop). [temperature] and [topP] default to the former hardcoded stage values. [topK], [minP] and
+ * [repeatPenalty] are null by default, meaning the request omits them and the model keeps its own
+ * default — set them to tighten a rambling local model. maxTokens is not here: it is pinned per-stage
+ * to the token budget. Talkie chat/narration have their own [GenerationSettings]/[NarrationSettings].
+ */
+@Serializable
+data class SamplingConfig(
+    val temperature: Double = 0.7,
+    val topP: Double = 1.0,
+    val topK: Int? = null,
+    val minP: Double? = null,
+    val repeatPenalty: Double? = null,
+)
+
 /**
  * Continuous health watch (observability-spec §4). When [enabled], a background monitor polls
  * the on-disk footprint every [intervalMs] and records a degraded/restored event on the edge.

core/config/src/main/kotlin/com/correx/core/config/CorrexConfigWriter.kt

@@ -102,6 +102,13 @@ object CorrexConfigWriter {
         b.kv("recovery_route_budget", cfg.orchestration.recoveryRouteBudget)
         b.kv("intent_route_budget", cfg.orchestration.intentRouteBudget)
 
+        b.section("sampling")
+        b.kv("temperature", cfg.sampling.temperature)
+        b.kv("top_p", cfg.sampling.topP)
+        cfg.sampling.topK?.let { b.kv("top_k", it) }
+        cfg.sampling.minP?.let { b.kv("min_p", it) }
+        cfg.sampling.repeatPenalty?.let { b.kv("repeat_penalty", it) }
+
         b.section("personalization")
         b.kv("enabled", cfg.personalization.enabled)
         b.kv("learn", cfg.personalization.learn)

core/inference/src/main/kotlin/com/correx/core/inference/GenerationConfig.kt

@@ -13,4 +13,9 @@ data class GenerationConfig(
     val maxTokens: Int,
     val stopSequences: List<String> = emptyList(),
     val seed: Long? = null,          // null = non-deterministic; set for replay
+    // Sampling knobs. null = omit from the request so the provider/model keeps its own default,
+    // preserving prior behavior. Serialized only when set (top_k / min_p / repeat_penalty).
+    val topK: Int? = null,
+    val minP: Double? = null,
+    val repeatPenalty: Double? = null,
 )